Datasets:

timaeus
/

pile-arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 100k rows

text stringlengths 4 2.78M	meta dict
--- abstract: 'I present 1.5- and 8.4-GHz observations with all configurations of the NRAO VLA of the wide-angle tail source . The source has a pair of relatively symmetrical, well-collimated inner jets, one of which terminates in a compact hot spot. Archival [ROSAT]{} PSPC data confirm that 3C130’s environment is a luminous cluster with little sign of sub-structure in the X-ray-emitting plasma. I compare the source to other wide-angle tail objects and discuss the properties of the class as a whole. None of the currently popular models is entirely satisfactory in accounting for the disruption of the jets in 3C130.' author: - \| M.J. Hardcastle$^1$[^1]$^{,2}$\ $^1$ H.H. Wills Physics Laboratory, University of Bristol, Royal Fort, Tyndall Avenue, Bristol BS8 1TL\ $^2$ Mullard Radio Astronomy Observatory, Cavendish Laboratory, Madingley Road, Cambridge, CB3 0HE title: 'Jets, plumes and hot spots in the wide-angle tail source ' --- \#1C\#2 [\#1C\#2]{} \#1[\#1 ]{} \#1[\#1 ]{} \#1 radio continuum: galaxies – galaxies: jets – galaxies: active – galaxies: individual: Introduction ============ is a FRI radio source at redshift 0.109 (Spinrad [[et al]{}.]{}1985). Its 178-MHz luminosity is $7.6 \times 10^{25}$ W Hz$^{-1}$ sr$^{-1}$, slightly above the nominal FRI-FRII boundary of $\sim 2 \times 10^{25}$ W Hz$^{-1}$ sr$^{-1}$ (Fanaroff & Riley 1974, hereafter FR). Leahy (1985, 1993) and Jägers and de Grijp (1985) present intermediate-resolution VLA maps of the central regions of the source, while Jägers (1983) has a lower-resolution WSRT image which shows the whole source and its field; the source extends for $\sim 1.5$ Mpc. Saripalli [[et al]{}.]{} (1996) present high-frequency maps made with the Effelsberg 100-m telescope. The host galaxy is classed as a DE2 by Wyndham (1966) and appears to lie in a cluster, although strong galactic reddening makes optical identification of the cluster members difficult. The [Einstein]{} detection of extended X-ray emission (Miley [[et al]{}.]{} 1983), the nearby aligned sources (Jägers 1983) and the many mJy radio sources in the field at 1.5 GHz make it plausible that the object is the dominant member of a large cluster. Leahy (1985) also attempts to constrain the RM distribution of the source, but notes that it depolarizes rapidly (particularly in the S lobe) so that few good measurements are available; this could be taken as evidence for a dense magneto-ionic environment for the source (cf. Hydra A, Taylor [[et al]{}.]{} 1990). \[definition\] 3C130 is a wide-angle tail (WAT) radio source. The term WAT has been used to describe many different types of object. Here I shall use it to refer to those FRI sources which are associated with central cluster galaxies (e.g. Owen & Rudnick 1976) and have luminosities comparable to or exceeding the Fanaroff-Riley break between FRI and FRII. I shall follow Leahy (1993) in using the behaviour of the jets at the base as another defining feature. At high resolution one or two well-collimated jets \[‘strong-flavour’ jets, by the classification of Leahy (1993)\] are seen (e.g. O’Donoghue, Owen & Eilek 1990), extending for some tens of kpc before broadening, often at a bright flare point, into the characteristic plumes or tails. These jets are very similar to the jets seen in FRII radio galaxies, and quite different from the behaviour of jets in more typical FRIs, where a collimated inner jet, if visible at all, decollimates rapidly (on scales of a few kpc at most) and comparatively smoothly into a bright ‘weak-flavour’ jet with a large opening angle.[^2] WATs, according to this definition, never have a weak-flavour jet, but make the transition between strong-flavour jet and diffuse, bent tail in a single step. The requirement that WATs be central cluster galaxies excludes objects (e.g. , Blundell 1996, Hardcastle [[et al]{}.]{}1997a; , Leahy 1997) where the ‘tails’ are likely to be simply ordinary FRII lobes which have been disrupted by unusual host-galactic dynamics. The condition on jet behaviour allows us to exclude objects such as the twin sources in (Owen [[et al]{}.]{}1985; Hardcastle 1996) which are associated with a dominant cluster galaxy and sometimes classed as WATs but whose inner jets are similar to those of typical powerful FRIs. Because of the requirements of this definition, wide-angle tail sources make up a small minority of the radio source population. For this reason, the detailed properties of their jets and tails have not been well studied, although a number have been imaged for studies of source dynamics (O’Donoghue [[et al]{}.]{} 1989). The only objects which have been the subject of detailed study in the radio are (Leahy 1984; Eilek [[et al]{}.]{} 1984) and , Hydra A (Taylor [[et al]{}.]{} 1990), although M87, Virgo A (e.g. Biretta & Meisenheimer 1993) exhibits some of the properties of a WAT. In this paper I present multi-configuration, multi-frequency VLA observations of a further powerful WAT. Throughout this paper I use a cosmology in which $H_0 = 50{\rm\ km\,s^{-1}\,Mpc^{-1}}$ and $q_0 = 0$. At the distance of , one arcsecond is equivalent to a projected length of 2.72 kpc. B1950.0 co-ordinates are used throughout. Observations ============ was observed with the VLA as part of a programme of detailed observations of FRI radio galaxies. Dates and integration times are shown in table \[obstab\]. and were used as primary flux calibrators; the nearby point sources 0537+531 and 0435+487 were used as phase calibrators, and (where was not observed) was used as a polarization angle reference. A bandwidth of 50 MHz was used, except at A configuration, where 25 MHz was used to reduce bandwidth smearing. The data were reduced within [aips]{} in the standard way. The datasets from each configuration were initially reduced separately, each undergoing several iterations of CLEANing and phase self-calibration. The B, C and D-configuration datasets were then phase-calibrated, using the appropriate baselines, with images made from the higher-resolution datasets, with which they were then merged without reweighting. Thus the B-configuration data were phase calibrated with an image made from the A-configuration data and merged with it to form an AB dataset; images made with this at low resolution were used to phase calibrate the C-configuration data and the two were merged to form an ABC dataset, and so on. This process ensures phase consistency in the data while removing the need for a self-calibration of the final merged dataset. Maps were made using the [aips]{} task IMAGR, with tapering of the [[uv]{}]{} plane where low-resolution maps were required. The robustness parameter in IMAGR was used to temper the uniform weighting of the [[uv]{}]{} plane, to improve the signal-to-noise ratio. In all cases the restoring beam was a circularly symmetrical Gaussian, well matched to the Gaussian fit to the dirty beam, and the resolution quoted is its FWHM. The total-intensity map at the highest resolution was made with a combination of IMAGR and the maximum-entropy imaging task VTESS; IMAGR was used to clean off the bright point-like components, the residual image was deconvolved with VTESS, and the point-like components subsequently restored. Results ======= Overall source structure ------------------------ Fig. \[3C130\_low\] shows the large-scale structure of the source. There are several pronounced bends, in spite of the overall straightness of the source. The sudden change in direction at the end of the south tail is particularly noticeable; this feature is similar to several seen in the small sample of O’Donoghue, Owen and Eilek (1990). The source disappears into the noise on these images and is longer than the $\sim 1$ Mpc seen here. The core -------- The radio core of did not vary over the timescales of the observations either at 8.4 or 1.5 GHz, within the errors imposed by the uncertainty of absolute flux calibration at the VLA. Its flux at 8.4 GHz was 29.0 mJy and at 1.5 GHz 12.4 mJy. The best position for the core is RA 04 48 57.34, DEC +51 59 49.7. The jets and hot spots ---------------------- The high-resolution images in Figs \[3C130.060g\] and \[3C130.060c\] show two very well-collimated jets emerging from the core. The jets are reasonably symmetrical. The northern jet in is brighter, noticeably so at bends; over the inner section where both jets are straight (approximately 9 arcsec) the difference in brightness is roughly a factor 1.4. \[This symmetry in the brightness of jets is reasonably common among WAT sources, compared to FRII radio galaxies or quasars (e.g. O’Donoghue [[et al]{}.]{} 1993). The ‘archetype’ of the class, , appears to be unusual in having a very one-sided jet.\] The bends in the jets, particularly the northern one, are very striking. The beams may be ballistic, implying some short-timescale wobble of the collimator (‘garden-hose’ behaviour), but if this is the case it is surprising that the jets are brighter at bends and that there is no antisymmetry between the jet and counterjet. If they are not ballistic it is equally remarkable that they remain collimated while undergoing oscillations of such large amplitude in so short a distance. The northern jet terminates in a hot spot, but there is a long filament which leaves the hot spot to the north, possibly suggesting some continued collimated outflow. At this resolution there is little compact structure at the end of the southern jet; the ‘hot spot’ seen in the maps of Leahy (1985, 1993) is resolved, with a size of around a second of arc. By contrast, the northern hot spot is only just resolved at the full resolution of the dataset (0.24 arcsec; maps not shown) and its brightest component has a minor axis of $\sim 0.3$ arcsec. This use of the term ‘hot spot’ is stronger than that of O’Donoghue [[et al]{}.]{}(1993), who only used it to indicate a brighter, broader region; the hot spot seen here is comparable in compactness with those in nearby FRIIs (e.g. Black [[et al]{}.]{} 1992; Leahy [[et al]{}.]{} 1997; Hardcastle [[et al]{}.]{}1997a) and is superposed on a brighter region which corresponds to the ‘hot spot’ of O’Donoghue et al. The northern jet is resolved at the bends at full resolution, and has a cross-sectional width of up to 0.8 arcsec. The polarization map (Fig. \[3C130.060c\]) includes a correction for Ricean bias and shows all points with polarized and total intensity greater than three times the respective off-source r.m.s.noise values. The position-angle vectors are perpendicular to the observed $E$-field, and so show the direction of the apparent magnetic field if Faraday rotation is negligible. Although we expect a non-negligible rotation measure (discussed further below, section \[rotm\]), these angles remain the best guess of the magnetic field direction. On this basis, the jets have apparent magnetic field parallel to their length where polarization is detected; the field follows the bends in the northern jet. This is as expected for a strong-flavour jet (e.g. Saikia & Salter 1988). The field in the hot spot is transverse to the jet direction and parallel to the hot spot’s direction of extension; this is similar to the field configuration in many FRII hot spots (Hardcastle [[et al]{}.]{} 1997a) but also to that in the termination knots of M87’s jet (Owen, Hardee & Cornwell 1989). Further out, the magnetic field is parallel to the plumes, and the degree of polarization is high. This appears to be the behaviour in the best-studied WATs (e.g. Taylor [[et al]{}.]{} 1990, O’Donoghue [[et al]{}.]{}1990, Patnaik [[et al]{}.]{} 1984; Saikia & Salter 1988, and references therein) but is quite different from the behaviour observed in the weak-flavour jets of normal FRIs, in which the field is transverse to the jet axis, sometimes with a longitudinal sheath (e.g. Hardcastle [[et al]{}.]{} 1996; Laing 1996; Hardcastle [[et al]{}.]{} 1997b). Depolarization, rotation measure and spectral index {#rotm} --------------------------------------------------- Using matched-baseline maps, I confirm earlier findings that the source is rapidly depolarized at low frequencies. The mean depolarization between 1.5 and 8.4 GHz (averaged over the areas with good signal-to-noise in both maps) of the northern plume is 0.2, and that of the southern plume 0.1. It may be noteworthy that the southern lobe, with a weaker jet and no bright compact hot spot, is the more depolarized: this may be an example of a Laing-Garrington effect (Laing 1988; Garrington [[et al]{}.]{} 1988) in WATs, although Saripalli [[et al]{}.]{} (1996) suggest that there are substantial variations in the degree of polarization with radio frequency. There is weak evidence that the inner 50 arcsec of both lobes is more depolarized than the outer parts, which would be consistent with depolarization by a medium associated with the galaxy or cluster. There are no systematic observations of depolarization in this class of source. The rotation measure (RM) distribution is not constrained by the rotation of polarization angle between 8.4 and 1.5 GHz. Rotations through all possible angles take place over the source, so there are variations in RM of more than 36 rad m$^{-2}$ on arcsecond scales. This is consistent with the RM measurements of Leahy (1985). Good maps at a higher frequency are needed to constrain the RM distribution adequately. Saripalli [[et al]{}.]{} (1996) report measurements suggesting an integrated galactic RM of $\sim 300$ rad m$^{-2}$ in the region of . From the fact that the polarization vectors are well aligned with one another (and consistent with those in the lower-resolution maps of Saripalli [[et al]{}.]{}) in the 8.4-GHz maps, and seem to follow bends in the source where these are present, we may guess that the rotation measure towards any point in the source is not much greater than this value, which would produce a $20\dgr$ rotation in polarization position angle at 8.4 GHz. The spectral index of the source steepens rapidly with distance from the core. Fig. \[3C130-spix\] shows a map of spectral index; the matched baselines of the maps ensure that the steepening is not an effect of undersampling. This spectral behaviour is expected in the standard model in which the plumes flow slowly away from the source \[compare the spectral index maps of Hydra A by Taylor [[et al]{}.]{}(1990)\]. Note the comparatively flat ($\alpha \approx 0.5$) spectral index of the jets and of the material they flow into. The northern hot spot has a spectral index flatter than the material that surrounds it. I determined spectral ages for regions along the (straighter) southern tail, using a minimum energy for the relativistic electron distribution corresponding to $\gamma = 100$, an initial electron-energy power-law index of 2 to reflect the hot spot spectral indices of 0.5, no energy contribution from relativistic protons, filling factor unity, and equipartition magnetic fields; I took flux measurements of regions of the plume between 30 and 110 arcsec, measured along the plume, from the radio core. The ageing field used was 0.46 nT, which was the mean of the equipartition fields fitted at various points along the tail; there was little variation in equipartition field strength with distance, so that this field is a good approximation to the correct self-consistent value. The model included the effects of inverse-Compton scattering from the CBR, which at this redshift produces energy loss equivalent to that due to a magnetic field of 0.40 nT. Using a model with effective pitch-angle scattering of electrons (Jaffe & Perola 1973), the plot of age against projected distance along the source was well fitted by a straight line with gradient $\sim 1.2 \times 10^4$ km s$^{-1}$, inferred ages being of the order of $10^7$ years (as found by Jägers & De Grijp 1985). The intercept was non-zero, reflecting the presence of steeper-spectrum material surrounding the jet termination; derived velocities were similar if the intercept was made zero by choosing a steeper initial power-law index (2.53). These inferred ages and velocities in the tails are comparable to those found by spectral age methods in other WAT sources (e.g. Taylor [[et al]{}.]{} 1990; O’Donoghue [[et al]{}.]{} 1993) and would imply outflow which is considerably faster than the sound speed in the external medium, given the temperature of the gas around 3C130 (discussed below); this is perhaps surprising in view of the absence of any evidence for post-hot spot shock structures in the tails and of their generally relaxed appearance. The usual caveats apply to velocities determined by spectral-ageing methods, but it should be noted that most factors that can affect the velocity (including a contribution to the energy density from relativistic protons, a particle filling factor less than unity, and significant projection of the radio source) will produce velocities higher than the value given above. Only if the assumptions involved in the spectral ageing analysis are seriously wrong – for example, if there is significant [in situ]{} particle acceleration in the tails or significant magnetic field inhomogeneity – can the tail velocity be much lower than this value. Evidence for such processes is discussed in Eilek (1996) and references therein. X-ray observations ------------------ Miley [[et al]{}.]{} (1983) report on [Einstein]{} IPC observations of . Serendipitously, the source is also included in the field of a 39.4 ks [ROSAT]{} PSPC pointed observation, taken from the public archives, of the X-ray emitting supernova remnant RX04591+5147 (Pfeffermann, Aschenbach & Predehl 1991; Reich [[et al]{}.]{}1992). Although the X-ray source associated with is 32 arcmin away from the pointing centre of the PSPC, and is thus badly vignetted, the observations have superior signal-to-noise to the [Einstein]{} data and show details of the X-ray structure of the source. The cluster is detected at $2200 \pm 100$ PSPC counts between 0.1–2.4 keV (derived from a circle of 11 arcmin radius about the centre of the X-ray source, using a background annulus between 11 and 17.5 arcmin), in spite of the reduced sensitivity of the PSPC at this off-axis distance. Because of the difficulty of measuring the background in the presence of extended emission from the SNR, and because the shadows of the ring and one of the radial struts pass close to the source, the derived count rate of $\sim 6 \times 10^{-2}$ s$^{-1}$ is uncertain. A rough correction for vignetting would imply an on-axis count rate of $9 \times 10^{-2}$ s$^{-1}$. Using the Post-Reduction Offline Software (PROS) within IRAF, I made spectral fits to the data, correcting for the off-axis location of the source. A single Raymond-Smith model provided a good fit ($\chi^2 = 15.8$ with 26 degrees of freedom), giving a best-fit temperature $kT = 2.9^{+9}_{-2}$ keV; the fitted galactic $N_H$ was $0.9^{+0.5}_{-0.2} \times 10^{22}$ cm$^{-2}$ \[cf. the value of $0.4 \times 10^{22}$ cm $^{-2}$ predicted by interpolation from Stark [[et al]{}.]{} (1992)\]. Errors quoted for $N_H$ and $kT$ are $1\sigma$ for two interesting parameters. Abundances were poorly constrained; 70 per cent solar abundance gave marginally the best fit. With this best-fit model, the 0.1-2.4 keV luminosity of the cluster is $5 \times 10^{37}$ W, consistent with the luminosity derived, on crude spectral assumptions, from the [Einstein]{} data by Miley [[et al]{}.]{} (1983); the cluster is thus comparable in X-ray luminosity to rich Abell clusters, and the temperature consistent with the temperature-luminosity relation (e.g. David [[et al]{}.]{} 1993). There is no evidence for a lower temperature in the central regions of the source, and so no evidence that a cooling flow is present; this appears to be normal for the host clusters of WATs (Norman, Burns & Sulkanen 1988; Gómez [[et al]{}.]{} 1997) although Schindler & Prieto (1997) suggest that a weak cooling flow is present in Abell 2634, the host cluster of 3C465, and Hydra A inhabits a cooling flow with high mass deposition rates (David [[et al]{}.]{}1990). The best-fit Gaussian to the off-axis point-spread function (PSF) of ROSAT at this distance from the pointing centre has $\sigma \approx 70$ arcsec (Hasinger [[et al]{}.]{} 1995). For a radio-X-ray comparison I have smoothed the broad-band (0.1-2.4 keV) X-ray image with a Gaussian of this size; this should allow the coma-induced asymmetry of the PSF to be neglected. The X-ray images (Fig. \[3C130-xray\]) show an extended structure on scales comparable to the length of the radio source (i.e. $\sim 1$ Mpc). The cluster gas seems reasonably symmetrical about the radio source, in contrast to the clumpy structures, with offset radio sources, seen in some lower-luminosity WAT hosts even at lower spatial resolution (e.g. Burns [[et al]{}.]{} 1994; Gómez [[et al]{}.]{} 1997). The distortion of the X-ray isophotes to the northeast coincides with, and may be related to, the kink ($\sim 150$–300 kpc from the nucleus) in the northern tail; there is no structure in the X-ray emission which can be related to the sudden change in direction at the end of the southern radio tail, however. Given the large and asymmetrical PSF, I have not attempted to fit radial profiles to the X-ray data. Discussion ========== Approaches to the source dynamics of WATs in the literature (e.g.Burns 1981; Eilek [[et al]{}.]{} 1984; O’Donoghue [[et al]{}.]{} 1993) have concentrated on the large-scale bends seen in the tails. It is instructive to ask a rather different question: why are these sources, with well-collimated strong-flavour jets, compact hot spots, and high radio luminosities, not classical double radio galaxies? In FRII objects of this radio power, radio-emitting plasma is thought to flow back from the hot spots into the ‘cocoon’ left behind as the hot spot and associated shocks propagate into the external medium, forming the radio lobes (e.g. Scheuer 1974; Williams 1991). In WATs, the jet appears to terminate in a shock in the same way. Norman [[et al]{}.]{} (1988) argue that strong shocks are necessary to explain the single-step transition between jets and plumes, and observations of compact hot spots in these objects, such as that seen in 3C130, support this model. However, the situation after the shock is different in the two classes of object. In WATs lobes are not formed. Instead, the hot spot is at the base of the tail; by analogy with the standard model for FRIIs, we may assume that the emitting material in the tail has passed through and been excited in the hot spot, and this is borne out by the spectral index results in 3C130. The tails may immediately deviate from the axis defined by the jets (e.g. ) or appear to continue in a straight line (e.g. ) but in no case does there appear to be lobe emission [significantly]{} closer to the core than the hot spot.[^3] The fact that there is no cocoon may explain the brightness and two-sidedness of the strong-flavour jets in WATs compared to those in FRIIs; a direct interaction with the (comparatively dense) external medium might be expected to make the beam more dissipative and perhaps to slow the regions of the beam responsible for the emission to only weakly relativistic velocities. This would explain the low values ($0.2c$) of ‘jet velocity’ estimated from sidedness by O’Donoghue [[et al]{}.]{} (1993) compared to the much higher values (0.6–$0.7c$) estimated from the sidedness of jets in FRII quasars (Bridle [[et al]{}.]{} 1994; Wardle & Aaron 1997) and their prominence and sidedness in FRII radio galaxies (Hardcastle [[et al]{}.]{} in prep.). Hardcastle [[et al]{}.]{} (1997a) proposed a similar explanation for the prominence and two-sidedness of the jets in the peculiar FRII 3C438. However, in the absence of classical double lobes and the associated discontinuity between radio-emitting plasma and shocked external medium, why are there jet termination shocks in WATs? It is well known that the difference between WATs and classical doubles is the local environment; whereas WATs always lie at the centres of clusters, FRII radio galaxies of comparable powers tend to avoid them (e.g. Prestage & Peacock 1988). An explanation for the peculiar properties of WATs compared to their classical double counterparts must turn on this environmental difference. A suggestion along these lines by Leahy (1984), applied to , invoked motion of the galaxy through the cluster, causing it to leave behind a passive wake of radio-emitting material; in this type of model the post-hot-spot material is left behind by the motion of the galaxy and so never forms a lobe. However, the motions of central cluster galaxies are not expected to be large (Eilek 1984; Pinkney [[et al]{}.]{} 1993 and references therein) and in any case such a model cannot account, without invoking projection effects implausibly often, for the large population of WATs in which one or both tails are more or less aligned with the inner jets (as in 3C130). Burns [[et al]{}.]{} (1994) suggest a model in which WATs have an origin in the merger of a cluster with a group or subcluster. This is motivated by the X-ray substructure which they find in many WAT host clusters. Large-scale, high-velocity residual motions of gas could then be responsible for the bending of the radio tails, while the merger would provide tidally stripped gas to fuel the AGN. In an extension of this work Gómez [[et al]{}.]{} (1997) show that the majority of WAT hosts in a larger sample show some X-ray substructure, with an alignment between the direction of the X-ray elongation and the angle that bisects the tails, consistent with such a model. 3C130, however, is clearly a WAT despite the location of its host at the centre of a smooth, approximately symmetrical distribution of X-ray emitting gas and its (apparently) straight tails. It appears that strong cluster inhomogeneity, though it may be necessary for bent tail formation, is not necessary for the existence of a WAT; in particular it does not, on its own, explain the jet shock/hot spot behaviour discussed above. Loken [[et al]{}.]{} (1995) discuss the physics of a jet propagating across the boundary between the interstellar medium of the host galaxy and the hotter, less dense intracluster medium, and suggest that this may be the reason for the disruption of the inner, well-collimated jet at a hot spot. They then postulate velocity shear across the boundary, as described above, to account for jet bending. The structures seen in numerical simulation when the jet simply crosses a contact discontinuity with crosswind do not resemble WATs strongly, however. If the jets are taken to cross a shock front instead (cf.Norman [[et al]{}.]{} 1988), then the simulations of Loken [[et al]{}.]{} are more convincing in their resemblance to WATs, but we again face the problem of the smoothness of the large-scale X-ray emission in 3C130; there is little evidence in this source for the recent cluster merger that Loken [[et al]{}.]{} invoke to produce such a shock. Neither the sonic radius of a possible cooling flow nor the shock front associated with a putative nuclear or galactic wind are in the appropriate place to produce the internal shocks in WAT jets (Soker & Sarazin 1988; Smith, Kennel & Coroniti 1993). Because of the low resolution of the X-ray data, cluster-merger models cannot be ruled out for 3C130. Producing such straight plumes in such a model while still causing both jets to disrupt requires a rather special geometry for the merger and/or convenient projection effects, however. If hot spots in WATs represent jet termination shocks, it is perhaps surprising that only a single hot spot is seen in 3C130 and that there are no clear hot spot candidates in several of the sources of O’Donoghue [[et al]{}.]{} (1993). It is possible that there are intrinsically similar hot spots but that relativistic beaming effects are affecting their visibility. In 3C130 the hot spot in the N lobe is approximately ten times brighter than the most comparable feature in the S lobe, which using standard results requires flow or advance velocities greater than $0.3c$. More high-resolution observations of these objects are needed to test such a model. Conclusions =========== A compact hot spot is detected at the base of one plume of the WAT , and the jets are shown to have longitudinal magnetic field. The source is thus very like a classical double in some respects. The data support the model in which WATs are objects whose jets make the transition from super- to sub-sonic velocities in one step, rather than decelerating gradually, by showing a bright sub-kpc structure (comparable to those seen in classical double radio sources) associated with the termination of a jet. Archival [ROSAT]{} PSPC observations of 3C130 show it to lie in a luminous cluster with $kT \sim 2.9$ keV. There is little sign of substructure in the X-ray, in contrast to many other WATs; this may be related to the nearly straight tails of 3C130. The lack of strong substructure seems to be inconsistent with recent models for jet disruption in WATs. ACKNOWLEDGEMENTS {#acknowledgements .unnumbered} ================ I am grateful to Julia Riley and Guy Pooley for suggesting the original radio observations of this source, and thank Mark Birkinshaw, Julia Riley and Diana Worrall for useful comments. I acknowledge a research studentship from the UK Particle Physics and Astronomy Research Council (PPARC) and support from PPARC grant GR/K98582. The National Radio Astronomy Observatory is operated by Associated Universities Inc., under co-operative agreement with the National Science Foundation. This project made use of Starlink facilities. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This research has made use of data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC), provided by NASA’s Goddard Space Flight Center. Blundell K.M., 1996, MNRAS, 283, 538 Bridle A.H., Hough D.H., Lonsdale C.J., Burns J.O., Laing R.A., 1994, AJ, 108, 766 Burns J.O., 1981, MNRAS, 195, 523 Burns J.O., Rhee G., Owen F.N., Pinkney J., 1994, ApJ, 423, 94 David L.P., Arnaud K.A., Forman W., Jones C., 1990, ApJ, 356, 32 David L.P., Slyz A., Jones C., Forman W., Vrtilek S.D., 1993, ApJ, 412, 479 Eilek J.A., 1996, in Hardee P.E., Bridle A.H., Zensus J.A., eds, Energy Transport in Radio Galaxies and Quasars, ASP Conference Series vol. 100, San Francisco, p. 281 Eilek J.A., Burns J.O., O’Dea C.P., Owen F.N., 1984, ApJ, 278, 37 Fanaroff B.L., Riley J.M., 1974, MNRAS, 167, 31P Garrington S., Leahy J.P., Conway R.G., Laing R.A., 1988, Nat, 331, 147 Gómez P.L., Pinkney J., Burns J.O., Wang Q., Owen F.N., Voges W., 1997, ApJ, 474, 580 Hasinger G., Boese G., Predehl P., Turner T.J., Yusaf R., George I.M., Rohrbach G., 1995, MPE/OGIP Calibration Memo CAL/ROS/93-015, version 1995 May 08 Hardcastle M.J., 1996, PhD thesis, University of Cambridge Hardcastle M.J., Alexander P., Pooley G.G., Riley J.M., 1996, MNRAS, 278, 273 Hardcastle M.J., Alexander P., Pooley G.G., Riley J.M., 1997a, MNRAS, 288, 859 Hardcastle M.J., Alexander P., Pooley G.G., Riley J.M., 1997b, MNRAS, 288, L1 Jägers W.J., 1983, A&A, 125, 172 Jägers W.J., De Grijp M.H.K., 1985, A&A, 143, 176 Jaffe W.J., Perola G.C., 1973, A&A, 26, 423 Laing R.A., 1988, Nat, 331, 149 Laing R.A., 1996, in Hardee P.E., Bridle A.H., Zensus J.A., eds, Energy Transport in Radio Galaxies and Quasars, ASP Conference Series vol. 100, San Francisco, p. 241 Leahy J.P., 1984, MNRAS, 208, 323 Leahy J.P., 1985, PhD thesis, University of Cambridge Leahy J.P., 1993, in Röser H.-J., Meisenheimer K., eds, Jets in Extragalactic Radio Sources, Springer-Verlag, Heidelberg, p. 1 Leahy J.P., 1997, in preparation Loken C., Roettiger K., Burns J.O., 1995, ApJ, 445, 80 Miley G.K., Norman C., Silk J., Fabbiano G., 1983, A&A, 122, 330 Norman M.L., Burns J.O., Sulkanen M.E., 1988, Nat, 335, 146 O’Donoghue A.A., Owen F.N., Eilek J.A., 1990, ApJS, 72, 75 O’Donoghue A.A., Eilek J., Owen F., 1993, ApJ, 408, 428 Owen F.N., Hardee P.E., Cornwell T.J., 1989, ApJ, 340, 698 Owen F.N., O’Dea C.P., Inoue M., Eilek J.A., 1985, ApJ, 294, L85 Patnaik A.R., Banhatti D.G., Subrahmanya C.R., 1984, MNRAS, 211, 775 Pfeffermann E., Aschenbach B., Predehl P., 1991, A&A, 246, L28 Pinkney J., Rhee G., Burns J.O., Hill J.M., Oegerle W., Batuski D., Hintzen P., 1993, ApJ, 416, 36 Prestage R.M., Peacock J.A., 1988, MNRAS, 230, 131 Reich W., Furst E., Arnal E.M., 1992, A&A, 256, 214 Saikia D.J., Salter C.J., 1988, ARA&A, 26, 93 Saripalli L., Mack K.-.H., Klein U., Strom R., Singal A.K., 1996, A&A, 306, 708 Scheuer P.A.G., 1974, MNRAS, 166, 513 Scheuer P.A.G., 1982, in Heeschen, D.S., Wade C.M., eds, Extragalactic Radio Sources, IAU Symposium 97, Reidel, Dordrecht, p. 163 Schindler S., Prieto M.A., 1997, A&A, 327, 37 Smith S.J., Kennel C.F., Coroniti F.V., 1993, ApJ, 412, 82 Soker N., Sarazin C.L., 1988, ApJ, 327, 66 Spinrad H., Djorgovski S., Marr J., Aguilar L., 1985, PASP, 97, 932 Stark A.A., Gammie C.F., Wilson R.W., Bally J., Linke R.A., Heiles C., Hurwitz M., 1992, ApJS, 79, 77 Taylor G.B., Perley R.A., Inoue M., Kato T., Tabara H., Aizu K., 1990, ApJ, 360, 41 Wardle J.F.C., Aaron S.E., 1996, in Hardee P.E., Bridle A.H., Zensus J.A., eds, Energy Transport in Radio Galaxies and Quasars, ASP Conference Series vol. 100, San Francisco, p. 123 Williams A.G., 1991, in Hughes P.A., ed., Beams and Jets in Astrophysics, Cambridge University Press, Cambridge, p. 342 Wyndham J.D., 1966, ApJ, 144, 459 ------- ---------------- ----------- ---------------- ----------- Conf. Date $t_{int}$ Date $t_{int}$ (mins) (mins) A 1995/08/06$^a$ 120 1995/07/23$^a$ 45 B 1995/11/28 120 1995/11/28 30 C 1994/11/10 55 1994/11/10 50 D 1995/03/06 20 1995/03/06 15 ------- ---------------- ----------- ---------------- ----------- : VLA observations of []{data-label="obstab"} [^1]: E-mail: [email protected] [^2]: There are a few exceptions to this behaviour; (Hardcastle [[et al]{}.]{} 1996) does appear to show an inner ‘strong-flavour’ jet and a bright knot at the base of the ‘weak-flavour’ jet. But even here the transition from strong to weak flavours occurs on scales of $\sim 1$ kpc. [^3]: Whether the bends in the jet seen in 3C130 are due to ballistic motion or to buffeting by the IGM, it is clear that the position of the jet termination point, however it is formed, must change with time. The hot spot will therefore move about in the base of the plume in a manner similar to that described in the ‘dentist’s drill’ model of Scheuer (1982) for the end points in FRIIs. We do not therefore expect to see the hot spot at a particular place in the tail in all cases.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Symplectic N-body integrators are widely used to study problems in celestial mechanics. The most popular algorithms are of 2nd and 4th order, requiring 2 and 6 substeps per timestep, respectively. The number of substeps increases rapidly with order in timestep, rendering higher-order methods impractical. However, symplectic integrators are often applied to systems in which perturbations between bodies are a small factor $\epsilon$ of the force due to a dominant central mass. In this case, it is possible to create optimized symplectic algorithms that require fewer substeps per timestep. This is achieved by only considering error terms of order $\epsilon$, and neglecting those of order $\epsilon^2$, $\epsilon^3$ etc. Here we devise symplectic algorithms with 4 and 6 substeps per step which effectively behave as 4th and 6th-order integrators when $\epsilon$ is small. These algorithms are more efficient than the usual 2nd and 4th-order methods when applied to planetary systems.' author: - 'J. E. Chambers' - 'M. A. Murison' title: 'Pseudo-High-Order Symplectic Integrators' --- Introduction ============ Symplectic integrators are widely used to study problems in celestial mechanics. These integrators have two advantages over most other algorithms. First, they exhibit no long-term build up in energy error. Second, the motion of each object about the central body can be “built in”, so that the choice of step size, $\tau$, is determined by the perturbations between bodies, whose magnitude is a factor $\epsilon$ smaller than the forces due to the central body ([@wis91]). The most popular algorithm is the second-order symplectic integrator. The error at each step is proportional to $\epsilon \tau^3$, so that the likely error for an integration as a whole is $\sim\epsilon \tau^2$. The second-order method is easy to implement, consisting of only two substeps, including one force evaluation, per time step. It is also very fast for integrations requiring moderate accuracy. For more accurate integrations, it is better to use the fourth-order method ([@for90]). Here, the error at each step is proportional to $\epsilon \tau^5$, although each step is computationally more expensive since it consists of 6 substeps. Yoshida (1990) has developed 6th and 8th-order symplectic integrators. However, these do not appear to be competitive in most situations, due to the large number of substeps required. Here we show how to construct what are effectively high-order (4th, 6th etc.) symplectic integrators that require fewer substeps per time step than those in current use. The trick is to take into account the dependence of each error term on $\epsilon$ when choosing the coefficients for each substep. The algorithms are designed by eliminating error terms proportional to $\epsilon$ up to the desired order of the timestep. Error terms proportional to $\epsilon^2, \epsilon^3$ etc., in low orders of the timestep, still exist. However, in many situations these terms are negligible, and the integrators behave as if they are of higher order than the leading error term in $\tau$ suggests. Section 2 gives a quick review of how symplectic integrators are traditionally constructed using Lie algebra. In Section 3, we show how to build more efficient symplectic algorithms using fewer substeps. Section 4 contains results of test integrations that compare the new algorithms with traditional symplectic integrators. Symplectic Integrators ====================== Symplectic integrators for the N-body problem can be constructed starting from Hamilton’s equations of motion: $$\begin{aligned} \frac{dx_i}{dt}&=&\frac{\partial H}{\partial p_i} \nonumber \\ \frac{dp_i}{dt}&=&-\frac{\partial H}{\partial x_i}\end{aligned}$$ where $x_i$ and $p_i$ are the coordinates and momenta of each body respectively, and $H$ is the Hamiltonian of the system. Using these equations, the rate of change of any dynamical quantity $q({\bf x},{\bf p},t)$ can be expressed as $$\frac{dq}{dt}=\sum_{i=1}^{3N} \left(\frac{\partial q}{\partial x_i} \frac{\partial H}{\partial p_i}-\frac{\partial q}{\partial p_i} \frac{\partial H}{\partial x_i}\right)\equiv\{q,H\}\equiv Fq \label{eq20}$$ where $\{,\}$ are Poisson brackets, and $F$ is a differential operator. The formal solution of equation (\[eq20\]) is $$q(t)=e^{\tau F}q(t-\tau) =\left(1+\tau F+\frac{\tau ^2F^2}{2}+\ldots\right)q(t-\tau)$$ where $F^2q = F(Fq)$ etc. Now suppose that we are able to split the Hamiltonian into two pieces, $H_A$ and $H_B$, so that each part of the problem can be solved relatively easily in the absence of the other. The solution for $q$ becomes $$q(t)=e^{\tau(A+B)} q(t-\tau) \label{eq30}$$ where $A$ and $B$ are differential operators related to $H_A$ and $H_B$ respectively, in the same way that $F$ is related to $H$. The Baker-Campbell-Hausdorff (BCH) formula states that, for any noncommutative operators $A$ and $B$, $$\exp(A)\cdot\exp(B)=\exp(C)$$ where $C$ is a series consisting of nested commutators, $$C=A+B+\frac{1}{2}[A,B]+\frac{1}{12}[A,A,B]+\frac{1}{12}[B,B,A]+\cdots$$ where the commutator $[A,B]=AB-BA\neq 0$ in general (see, for example, Yoshida 1990 or Forest and Ruth 1990). Here, we have used the nested commutator notation $[A,B,C]=[A,[B,C]]$, etc.. Hence, if we evolve $q$ under the two parts of the Hamiltonian separately, one after the other, we have $$\exp(\tau A)\cdot\exp(\tau B)q(t-\tau)= \exp\left[\tau F+\frac{\tau^2}{2}[A,B]+\cdots\right]q(t-\tau) \label{eq40}$$ This is identical to the righthand side of equation (\[eq30\]) to $O(\tau)$, and so equation (\[eq40\]) represents a first-order integrator. Each step of the integrator consists of 2 substeps, with the whole step giving an error of $O(\tau^2)$. Alternatively, we can say that the integrator exactly solves a problem whose Hamiltonian is given by $$H_{integ}=H+\frac{\tau}{2}\{H_B,H_A\}+O(\tau^2)$$ (see, for example, [@sah92]). Provided that $\tau$ is small, and $\{H_B,H_A\}$ remains bounded, the energy of the integrated system will always be near to that of the real system. Other integrators can be found by combining exponential operators in such a way that they are equivalent to equation (\[eq30\]) up to a given order in $\tau$. For example, we have the second-order symplectic integrator $$\begin{aligned} {\rm S2A}&=& \exp\left(\frac{\tau}{2}A\right)\cdot\exp(\tau B)\cdot \exp\left(\frac{\tau}{2}A\right) \\ &=&\exp\left[\tau F+\frac{\tau^3}{12}[B,B,A]-\frac{\tau^3}{24}[A,A,B] +\ldots\right]\end{aligned}$$ When many integration steps are performed one after another, the $\exp(\tau A/2)$ terms at the end of one step and the start of another can be combined. Hence, the second-order integrator also consists of only 2 substeps, except at the beginning and the end of an integration. Another second-order integrator is $$\begin{aligned} {\rm S2B}&=& \exp\left(\frac{\tau}{2}B\right)\cdot\exp(\tau A)\cdot \exp\left(\frac{\tau}{2}B\right) \nonumber \\ &=&\exp\left[\tau F+\frac{\tau^3}{12}[A,A,B]-\frac{\tau^3}{24}[B,B,A] +\ldots\right] \label{eq70}\end{aligned}$$ The distinction between S2A and S2B (which at first sight appear to be the same) will become apparent in the next section, when we consider situations in which $A$ and $B$ are of different magnitude. Forest and Ruth (1990) give a fourth-order symplectic integrator with 6 substeps per step: $$\begin{aligned} {\rm S4B}&=& \exp\left(\frac{\tau B}{2c}\right)\cdot \exp\left(\frac{\tau A}{c}\right)\cdot \exp\left[\frac{\tau B(1-k)}{2c}\right]\cdot \exp\left(\frac{-\tau kA}{c}\right)\cdot \exp\left[\frac{\tau B(1-k)}{2c}\right] \\ &\cdot& \exp\left(\frac{\tau A}{c}\right)\cdot \exp\left(\frac{\tau B}{2c}\right) \\ &=&\exp[\tau F+O(\tau^5)]\end{aligned}$$ where $k=2^{1/3}$ and $c=2-k$. Note that the middle 3 substeps move in the opposite direction to the integration as a whole. Higher-order integrators require progressively more substeps. Yoshida (1990) gives examples of 6th and 8th-order integrators using 14 and 30 substeps respectively. In the next section, we will show how to create what are [effectively]{} 4th and 6th order integrators (and in principle, 8th-order etc.) using fewer substeps than are required conventionally. Constructing Pseudo-Order Integrators ===================================== Up to this point we have not considered the details of how $H$ is split. Suppose that one part of the Hamiltonian is much smaller than the other, [[i.e.]{}]{} $H=H_A+\epsilon H_B$, where $\epsilon\ll 1$. Now consider the error terms in the second-order integrator of equation (\[eq70\]): $${\rm S2B}=\exp\left[\tau F+\frac{\epsilon\tau^3}{12}[A,A,B] -\frac{\epsilon^2\tau^3}{24}[B,B,A]+\cdots\right]$$ One of the $O(\tau^3)$ error terms is smaller than the other by a factor of $\epsilon$. Similarly, for the fourth order integrator: $${\rm S4B}=\exp[\tau F+O(\epsilon\tau^5)+O(\epsilon^2\tau^5) +O(\epsilon^3\tau^5)+O(\epsilon^4\tau^5)]$$ Some of these error terms are insignificant compared to others, and yet this was not taken into account when constructing the integrator. The only design criterion was that S4B should contain no error terms below the fifth power in the timestep. If we take into account the dependence of the error terms on both $\tau$ and $\epsilon$, we can design more efficient symplectic integrators. To construct the new integrators, we again employ the BCH formula. Adapting the expression for the BCH formula given by Yoshida (1990), we have: $$\begin{aligned} &&\exp(a_1\tau A)\cdot\exp(b_1\epsilon\tau B) \\ &=&\exp\left[(a_1A+\epsilon b_1B)\tau +\epsilon\tau^2\left(\frac{a_1b_1}{2}\right) [A,B] +\epsilon\tau^3\left(\frac{a_1^2b_1}{12}\right) [A,A,B] +\epsilon^2\tau^3\left(\frac{a_1b_1^2}{12}\right) [B,B,A]\right. \\ &+&\left.\epsilon^2\tau^4\left(\frac{a_1^2b_1^2}{24}\right) [A,B,B,A] -\epsilon\tau^5\left(\frac{a_1^4b_1}{720}\right) [A,A,A,A,B] -\epsilon^4\tau^5\left(\frac{a_1b_1^4}{720}\right) [B,B,B,B,A] +\cdots\right]\end{aligned}$$ where $a_1$ and $b_1$ are constants. Additional fifth-order commutators are present; however, we will only require terms that contain either $A$ or $B$ once, since these are the type of error term we are seeking to eliminate. Applying the BCH formula twice, Yoshida (1990) gives an expression for a symmetric product of three exponential operators: $$\begin{aligned} &&\exp(b_1\epsilon\tau B)\cdot\exp(a_1\tau A)\cdot \exp(b_1\epsilon\tau B) \nonumber \\ &=&\exp\left[(a_1A+2\epsilon b_1B)\tau +\epsilon\tau^3\left(\frac{a_1^2b_1}{6}\right) [A,A,B] -\epsilon^2\tau^3\left(\frac{a_1b_1^2}{6}\right) [B,B,A]\right. \nonumber \\ &-&\left.\epsilon\tau^5\left(\frac{a_1^4b_1}{360}\right) [A,A,A,A,B] +\epsilon^4\tau^5\left(\frac{7a_1b_1^4}{360}\right) [B,B,B,B,A] +\cdots\right] \label{eq100}\end{aligned}$$ Again we have neglected fifth-order terms that contain both $A$ and $B$ more than once. Note that there are no terms containing even powers of the timestep: Yoshida shows that this is a general property of any symmetric arrangement of exponential operators. From now on we will consider only symmetrical integrators because of this property. We need to extend equation (\[eq100\]) once more to get a pseudo-fourth order integrator, and twice more for a pseudo-sixth order one. By substituting $a_2A$ for $b_1B$ in equation (\[eq100\]), and substituting the righthand side of equation (\[eq100\]) for $a_1A$, we get: $$\begin{aligned} &&\exp(a_2\tau A)\cdot\exp(b_1\epsilon\tau B)\cdot\exp(a_1\tau A) \cdot\exp(b_1\epsilon\tau B)\cdot\exp(a_2\tau A) \nonumber\\ &=&\exp\left\{(a_1+2a_2)\tau A+2b_1\epsilon\tau B +\epsilon\tau^3\left(\frac{b_1}{6}\right)[(a_1+2a_2)^2-6a_2(a_1+a_2)] \,[A,A,B]\right. \nonumber \\ &+&\epsilon^2\tau^3\left(\frac{b_1^2}{6}\right)(4a_2-a_1)\,[B,B,A] -\epsilon\tau^5\left(\frac{b_1}{360}\right) [(a_1+2a_2)^4-30a_2^2(a_1+a_2)^2]\,[A,A,A,A,B] \nonumber \\ &-&\left.\epsilon^4\tau^5\left(\frac{b_1^4}{360}\right) (16a_2-7a_1)\,[B,B,B,B,A]+\cdots\right\} \label{eq110}\end{aligned}$$ Finally, substituting the righthand side of equation (\[eq110\]) for $a_1A$ in equation (\[eq100\]), and replacing $b_1B$ with $b_2B$, we arrive at $$\begin{aligned} &&\exp(b_2\epsilon\tau B)\cdot\exp(a_2\tau A)\cdot \exp(b_1\epsilon\tau B)\cdot\exp(a_1\tau A)\cdot \exp(b_1\epsilon\tau B)\cdot\exp(a_2\tau A)\cdot \exp(b_2\epsilon\tau B) \nonumber\\ &=&\exp\left\{(a_1+2a_2)\tau A+2(b_1+b_2)\epsilon\tau B +\epsilon\tau^3\left[\frac{(b_1+b_2)(a_1+2a_2)^2-6a_2b_1(a_1+a_2)}{6} \right][A,A,B]\right. \nonumber \\ &-&\epsilon^2\tau^3\left[\frac{(a_1+2a_2)(b_1+b_2)^2-6a_2b_1^2}{6}\right] [B,B,A] \nonumber \\ &-&\epsilon\tau^5\left[\frac{(b_1+b_2)(a_1+2a_2)^4-30a_2^2b_1(a_1+a_2)^2} {360}\right] [A,A,A,A,B] \nonumber \\ &+&\left.\epsilon^4\tau^5\left[\frac{7(a_1+2a_2)(b_1+b_2)^4 -60a_2b_1^2(b_1+b_2)^2+30a_2b_1^4}{360}\right] [B,B,B,B,A]+\cdots\right\} \label{eq120}\end{aligned}$$ The first stage in converting these general expressions into specific integrators is to make the coefficients of the linear $A$ and $B$ terms equal to 1. This places two constraints on the values of the constants. We can then get what is effectively a 4th-order integrator by simply eliminating the $[A,A,B]$ term from equation (\[eq110\]). The leading error terms will now be $O(\epsilon^2\tau^3)$ and $O(\epsilon\tau^5)$. Provided that $\epsilon$ is small enough, the largest error term will be $O(\epsilon\tau^5)$, and the integrator effectively will be of fourth order in the timestep. Applying these conditions, we require $$\begin{aligned} a_1+2a_2&=&1 \nonumber \\ 2b_1&=&1 \nonumber \\ 1-6a_2(1-a_2)&=&0 \label{eq150}\end{aligned}$$ where we have used the first two of equations (\[eq150\]) in deriving the third. Alternatively, we may construct an integrator in which each step begins by advancing $H_B$ instead of $H_A$. Unlike conventional symplectic integrators, such as $S2A$ and $S2B$, we cannot use the same set of coefficients when exchanging $A$ and $B$. Instead, we must derive a new set of coefficients by interchanging $A$ and $\epsilon B$ in equation (\[eq110\]) and then eliminating the new $[A,A,B]$ term. When we do this, the first two of equations (\[eq150\]) remain as before, but the third expression becomes $$6a_2-1=0 \label{eq160}$$ To get a pseudo-6th-order integrator, we eliminate terms containing $[A,A,B]$ and $[A,A,A,A,B]$. This will produce an extra constraining equation, so we need an extra constant. We get this by using an integrator with the form of equation (\[eq120\]) instead of equation (\[eq110\]). The corresponding equations for the constants are $$\begin{aligned} a_1+2a_2&=&1 \nonumber \\ 2(b_1+b_2)&=&1 \nonumber \\ 1/2-6a_2b_1(1-a_2)&=&0 \nonumber \\ 1/2-30a_2^2b_1(1-a_2)^2&=&0 \label{eq170}\end{aligned}$$ If we prefer an integration step that begins by advancing $H_A$, we can interchange $A$ and $\epsilon B$ in equation (\[eq120\]), and eliminate the new $[A,A,B]$ and $[A,A,A,A,B]$ terms. In this case, the last two of equations (\[eq170\]) become $$\begin{aligned} 1/4-6a_2b_1^2&=&0 \nonumber \\ 7/16-15a_2b_1^2+30a_2b_1^4&=&0 \label{eq180}\end{aligned}$$ The leading error terms for each of these integrators are $O(\epsilon^2\tau^3)$ and $O(\epsilon\tau^7)$. The latter will be dominant if $\epsilon$ is small enough, so that the algorithms behave as 6th-order integrators. Pseudo-4th and 6th-Order Examples --------------------------------- Solving equations (\[eq150\]) and (\[eq160\]), we obtain two pseudo-4th-order integrators: $$\begin{aligned} {\rm S4A}&=&\exp\left[\frac{\tau A}{2}\left(1-\frac{1}{\surd 3}\right)\right] \cdot\exp\left(\frac{\epsilon\tau B}{2}\right) \cdot\exp\left(\frac{\tau A}{\surd 3}\right) \cdot\exp\left(\frac{\epsilon\tau B}{2}\right) \cdot\exp\left[\frac{\tau A}{2}\left(1-\frac{1}{\surd 3}\right)\right] \\ &=&\exp\left[\tau F +\epsilon^2\tau^3\left(\frac{2-\surd 3}{24}\right)[B,B,A] -\frac{\epsilon\tau^5}{4320}[A,A,A,A,B]+\cdots\right] \\ && \\ {\rm S4B}&=&\exp\left(\frac{\epsilon\tau B}{6}\right) \cdot\exp\left(\frac{\tau A}{2} \right) \cdot\exp\left(\frac{2\epsilon\tau B}{3} \right) \cdot\exp\left(\frac{\tau A}{2} \right) \cdot\exp\left(\frac{\epsilon\tau B}{6} \right) \\ &=&\exp\left[\tau F +\frac{\epsilon^2\tau^3}{72}[B,B,A] +\frac{\epsilon\tau^5}{2880}[A,A,A,A,B]+\cdots\right]\end{aligned}$$ where the asterisk in S4A\* indicates that it only behaves as a 4th order integrator for certain values of $\tau$. Equations \[eq170\] and \[eq180\] give two pseudo-6th-order integrators: $$\begin{aligned} {\rm S6A}&=&\exp\left[\frac{\tau A}{2}\left(1-\frac{3}{\surd 15}\right)\right] \cdot\exp\left(\frac{5\epsilon\tau B}{18}\right) \cdot\exp\left(\frac{3\tau A}{2\surd 15}\right) \cdot\exp\left(\frac{4\epsilon\tau B}{9}\right) \cdot\exp\left(\frac{3\tau A}{2\surd 15}\right) \\ &\cdot&\exp\left(\frac{5\epsilon\tau B}{18}\right) \cdot\exp\left[\frac{\tau A}{2}\left(1-\frac{3}{\surd 15}\right)\right] \\ &=&\exp\left[\tau F+\epsilon^2\tau^3\left(\frac{54-13\surd 15}{648}\right) [B,B,A] +O(\epsilon\tau^7)\right] \\ && \\ {\rm S6B}&=&\exp\left(\frac{\epsilon\tau B}{12}\right) \cdot\exp\left[\frac{\tau A}{2}\left(1-\frac{1}{\surd 5}\right)\right] \cdot\exp\left(\frac{5\epsilon\tau B}{12}\right) \cdot\exp\left(\frac{\tau A}{\surd 5}\right) \cdot\exp\left(\frac{5\epsilon\tau B}{12}\right) \\ &\cdot&\exp\left[\frac{\tau A}{2}\left(1-\frac{1}{\surd 5}\right)\right] \cdot\exp\left(\frac{\epsilon\tau B}{12}\right) \\ &=&\exp\left[\tau F+\epsilon^2\tau^3\left(\frac{13-5\surd 5}{288}\right) [B,B,A] +O(\epsilon\tau^7)\right]\end{aligned}$$ Unlike the 4th-order algorithm of Forest and Ruth (1990) and the 6th-order integrators of Yoshida (1990), the algorithms above contain no substeps that move in the opposite direction to the main integration. An additional solution exists for each of equations (\[eq150\]), (\[eq170\]) and (\[eq180\]), however these have error terms with larger numerical coefficients than the integrators we show here. The same method can be used to generate a pseudo-8th-order integrator and so on. Each higher order will require only 2 more substeps than the previous one, since only one more commutator needs to be eliminated in each case. For example, to create a pseudo-8th-order integrator requires the elimination of the $[A,A,A,A,A,A,B]$ term in addition to those that are absent from the pseudo-6th-order case. However, depending on the system to be integrated, there will come a point at which the $\epsilon^2\tau^3$ error term becomes the most important. In principle, one could devise another set of integrators that eliminates terms in $\epsilon^2\tau^m$ for small $m$, in addition to terms in $\epsilon\tau^m$. However, achieving each new order will generally require the elimination of more than one commutator term, so that these integrators increase in complexity much more rapidly than those described here. Murison and Chambers (1999) have independently derived the two 4th-order integrators above, among others, using a symbolic algebra package. Further results from that approach will follow in another paper. We note that the pseudo-order algorithms can be adapted to use independent timesteps for each planet ([[c.f.]{}]{} [@sah94]), or to include close encounters ([@dun98; @cha99]). Numerical Comparisons ===================== In this section, we test the pseudo-4th and 6th order integrators derived in Section 3 against the well-known 2nd and 4th-order symplectic algorithms. We use the “mixed-variable” method of Wisdom and Holman (1991), in which the Hamiltonian is divided into a Keplerian part, $H_K$, and an interaction part, $H_I$. Under $H_K$, each object moves on an unperturbed Keplerian orbit about the central body. Under $H_I$, each object remains fixed while receiving an impulse due to the gravitational perturbations of all the other objects except the central body. As suggested by Wisdom and Holman, we use Jacobi coordinates rather than barycentric coordinates. The integrations themselves were carried out using a modified version of the [Mercury]{} integrator package ([@cha97]). The pseudo-order integrators require that the ratio $\epsilon=H_I/H_K\ll 1$. In our first test, we integrate the orbits of the 4 inner planets of the solar system in the absence of the outer planets. In this case $\epsilon\sim 10^{-5}$. Figure 1 shows the results of a 10000-year integration using the conventional 2nd and 4th-order symplectic integrators, S2B and S4B, and the pseudo-order integrators S4B\* and S6B\. For each integration, the maximum relative energy error is shown as a function of the step size. For the 2nd and 4th-order integrators, the maximum energy error is roughly proportional to $\tau^2$ and $\tau^4$ respectively, where $\tau$ is the timestep. This is what we would expect to find. For the pseudo-4th and 6th-order integrators, the maximum energy error varies as $\tau^4$ and $\tau^6$. That is, they behave as 4th and 6th-order integrators, as we anticipated, despite the fact that they contain error terms of lower order in the timestep. Using the mean relative energy error per integration instead of the maximum error gives results similar to Figure 1. The corresponding slopes are $2.10\pm 0.05$ for S2B, $3.9\pm 0.3$ for S4B, $4.6\pm 0.3$ for S4B\ and $6.4\pm 0.4$ for S6B\. Figure 2 shows the amount of CPU time required for the integrations shown in Figure 1. For energy errors of 1 part in $10^6$ or $10^8$ there is not much to choose between the four algorithms. For higher levels of accuracy, S4B outperforms S2B. However, the pseudo integrators S4B\ and S6B\* do even better. At an accuracy of 1 part in $10^{10}$, they are roughly an order of magnitude faster than the conventional second-order integrator, and 3 times faster than the 4th-order integrator. For accuracies of better than $10^{-11}$, S6B\* shows greater performance than S4B\. The pseudo-4th order integrator is more efficient than the real 4th-order integrator for two reasons. It requires fewer substeps per time step, and it has a slightly smaller leading error term. As a more interesting test, we integrated the whole planetary system (Mercury to Pluto) for 10000 years. Figure 3 shows the energy-error results of these integrations. The behaviour of S2B, S4B and S4B\ is similar to that for the integrations of the terrestrial planets. However, the energy error for S6B\* varies roughly as $\tau^5$ rather than $\tau^6$. It is not obvious why this should be, although the difference from the terrestrial-planet integration (Figure 1) is presumably due to the fact that $\epsilon$ is two orders of magnitude larger in this case. Figure 4 shows the CPU time required for the integrations of the 9 planets. The results are similar to the integration of the inner planets, except that S6B\* has only a marginal advantage over S4B\* at the highest levels of accuracy. Since writing the original draft of this manuscript, we have become aware of the symplectic corrector method of Wisdom [[et al.]{}]{} (1996), which substantially improves the efficiency of the second-order symplectic integrator. We present the pseudo-order integrators as an alternative strategy for designing accurate algorithms. It is possible to devise other symplectic correctors using the same approach we use in Section 3 to design the integrator kernel: that is, by considering the dependence of the resulting error terms on $\epsilon$ as well as $\tau$ ([@mik97; @rau99]). Finally we suggest that it may be possible to design symplectic correctors to improve the performance of pseudo-order algorithms, since the pseudo-order methods exhibit similar high-frequency oscillations in energy error to the second and 4th-order symplectic integrators (see Figure 5). In summary, we conclude that the new pseudo-order integrators outperform the widely-used 2nd and 4th-order algorithms at all reasonable values of the energy error, for problems involving a dominant central mass. Research at Armagh Observatory is grant-aided by the Dept. of Education, Northern Ireland. The test integrations described in this paper were carried out using computers purchased on a PPARC research grant. Chambers, J.E. 1999, , 304, 793 Chambers, J.E. and Migliorini, F. 1997, BAAS, 29, 1024. Duncan, M.J., Levison, H.F., and Lee, M.H. 1998, , 116, 2067 Forest, E. and Ruth, R.D. 1990, Physica D, 43, 105 Mikkola, S. 1997, Cel. Mech. Dyn. Astron., 67, 145 Murison, M.A. and Chambers, J.E. 1999, abstract at 1999 meeting of Division on Dynamical Astronomy of the AAS Rauch, K. and Holman, M. 1999, , 117, 108 Saha, P., and Tremaine, S. 1992, , 104, 1633 Saha, P., and Tremaine, S. 1994, , 108, 1962 Wisdom, J. and Holman, M. 1991, , 102, 1528 Wisdom, J., Holman, M. and Touma, J. 1996, Fields Institute Comm., 10, 217 Yoshida, H. 1990, Phys. Lett. A, 150, 262 Figure Captions {#figure-captions .unnumbered} =============== Figure 1: : Maximum relative energy error versus step size for a 10000-year integration of the 4 terrestrial planets using various symplectic integrators. Figure 2: : Maximum relative energy error versus CPU time for a 10000-year integration of the 4 terrestrial planets using various symplectic integrators. Figure 3: : Maximum relative energy error versus step size for a 10000-year integration of the 9 planets using various symplectic integrators. Figure 4: : Maximum relative energy error versus CPU time for a 10000-year integration of the 9 planets using various symplectic integrators. Figure 5: : Relative energy error versus time for 10000-year integrations of the 9 planets using various symplectic integrators.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Inert Doublet Model(IDM) is one of the simplest extensions beyond Standard Model(SM) with an extended scalar sector, which provide a scalar dark matter particle candidate. In this paper, we investigate the double charged Higgs production at $\gamma\gamma$ collider. By scanning the whole parameter space, we obtain the parameter points corresponding to the correct relic abundance of dark matter. After applying all theoretical and experimental constraints, the parameter space for the existence of dark matter is extremely restricted. We perform the analysis for the signal of $H^+H^-$ production in the IDM and the SM backgrounds, and the optimized selection conditions are chosen in kinematic variables to maximize signal significance. Comparing signal with backgrounds, we obtain the parameter points which can be detected at future $\gamma\gamma$ collider experiments.' author: - 'Yang Guo-He$^a$' - Song Mao$^a$ - Li Gang$^a$ - 'Zhang Yu$^{b,a}$' - 'Guo Jian-You$^a$' title: 'Searches for dark matter via charged Higgs pair production in the Inert Doublet Model at $\gamma\gamma$ collider' --- Introduction ============ After the discovery of the Higgs boson [@higgs1; @higgs2], the Standard Model (SM) of particle physics has achieved great success in describing the particles up to energies of about 1 TeV [@Baak:2012kk]. However, there are still many questions as well as many unexplained phenomena remain, such as, the symmetry of matter and antimatter, the sources of CP violation, the nature of dark matter (DM) particle. All these mean that the Standard Model is perhaps just a low energy approximation of a more fundamental theory. At the same time, the Standard Model of Big Bang Cosmology, known as “$\Lambda$CDM”, is successful in describing the Universe large scale structure formation and evolution, the state of the early Universe and the abundance of the different form of matter and energy [@Spergel:2003cb; @DelPopolo:2008mr; @DelPopolo:2013qba], whose predictions are supported by new observation ($e.g.$, lensing of the CMB [@Smith:2007rg; @das], B-mode polarisation [@Hanson:2013hsb], the kinetic Sunyaev Zeldovich(SZ) effect). The astrophysical and cosmological observational evidences have confirmed the existence of DM and provided the DM density in the universe [@Bertone04]. However, the Standard Model of particle physics could not provide enough dark matter. Until now, we have little information about the properties of dark matter particles. Among all the DM candidates, Weakly Interacting Massive Particles (WIMPs) is a promising option. Since it offers the DM candidates to interpret the relic abundance naturally in rebuilding the thermal history of the universe [@Feng2008a]. Among various extended scenarios beyond SM, the Inert Doublet Model (IDM) is one of the simplest models to explain the WIMP dark matter. In this model, an isospin doublet scalar field is added to the SM Higgs sector, which is assumed to be odd under a discrete $Z_2$ symmetry. After electroweak symmetry breaking, four $Z_2$ odd scalar particles are generated, $i.e.$, one CP-even $H$, one CP-odd $A$ and two charged $H^\pm$ scalar bosons. Among them, the lightest scalar boson may serve as a dark matter candidate. The $Z_2$ symmetry ensures that these new scalar particles can not decay into final states only including the SM particles. In addition, the additional isospin doublet scalar does not directly interact with the SM fermions at tree level. Their interactions with the Standard Model particles are achieved via gauge coupling and the quartic term with the SM Higgs in the scalar potential. The lightest scalar Higgs in the IDM, as a dark matter candidate, needs to be able to reconstruct the correct DM relic abundance. In Ref [@chile_DM; @Dolle:2009fn], they found that three allowed mass regimes for the lightest Higgs satisfy the requirement of relic abundance. The scalar dark matter particle has also been explored by various direct, indirect experiments and high energy colliders. In current direct detection experiments, the dark matter mass has been constrained to be around one half of the SM-like Higgs boson mass (125 GeV) or above about 500 GeV [@chile_DM; @IDM_DM_th3; @Compressing_IDM]. In Ref [@Eiteneuer:2017hoh], the authors investigate the constraint for IDM parameter space from dark matter annihilation induced gamma-rays in dwarf spheroidal galaxies. The phenomenology for the IDM at hadron colliders has been studied in the literature, such as $H^+H^-$, $HH^{\pm}$, $HA$ pair production, followed by the subsequent decay chains $A \to ZH$, $H^{\pm} \to W^{\pm}H$ [@IDM_DM_th8; @IDM_DM_th9; @Compressing_IDM; @LHC_direct_Cao; @LHC_direct_Su; @LHC_run-I_bound1; @LHC_run-I_bound2; @song_ctp; @Ahriche:2018ger]. The prospects for discovery of scalar dark matter particle in the IDM at future lepton colliders has been discussed [@IDM_ILC1; @IDM_ILC2; @Kalinowski:2018ylg]. Moreover, the constraint for IDM using vector boson fusion is also investigated in Ref. [@Dercks:2018wch]. With the option of an $e^+e^-$ collider, it also can be run in $\gamma\gamma$ mode (at an energy scale similar to that of the primary electron-positron design). The charged Higgs pair can product directly in the IDM at $\gamma\gamma$ collider mode. Compared with $e^+e^-$, $\gamma\gamma$ collider can provide higher cross section in the high energy region because the charged Higgs pair is mainly dominated by s-channel diagrams at the tree-level in $e^+e^-$ collider. In this paper, we will investigate the production of charged Higgs pair in $\gamma\gamma$ collider. The paper is organized as follows. In Section II we briefly describe the framework of inert doublet model. In Section III we calculate the relic abundance of dark matter in the IDM. In Section IV, we summarize all the theoretical and phenomenological constraints on the scalar potential of the IDM. In Section V, we present the numerical results of the total and differential cross sections for the charged Higgs pair production. In Sec VI, we analyse the charged Higgs pair signatures at $\gamma\gamma$ collider with its subsequent decay $H^{\pm} \to W^{\pm}H$. Finally, a short summary is given in Section VII. The Inert Higgs Doublet Model ============================= The inert doublet model is one of the simplest extension of Standard Model (SM), which contain two SU(2) complex scalar fields $\Phi_{1}$ and $\Phi_{2}$, which are invariant in discrete $Z_2$ symmetry. The scalar field $\Phi_{1}$ is almost the same as the SM Higgs field, which is $Z_2$ even with hypercharge $Y=1$. Under $Z_2$ symmetry, $\Phi_{1}$ satisfy the transformation $\Phi_{1} \rightarrow \Phi_{1}$. The field $\Phi_2^{}$ is odd under the $Z_2^{}$ symmetry with hypercharge $Y=1/2$, which satisfy the transformation $\Phi_{2} \rightarrow - \Phi_{2}$ under $Z_2$. Under the electroweak symmetry $SU(2)_{L}\times U(1)_{Y}$ and the discrete $Z_2$ symmetry, the Higgs sector potential of the IDM is $$\begin{aligned} V(\Phi_1,\Phi_2) =&\mu^2_1\|\Phi_1\|^2+\mu^2_2\|\Phi_2\|^2 +\frac{1}{2}\lambda_1\|\Phi_1\|^4+\frac{1}{2}\lambda_2\|\Phi_2\|^4 \nonumber\\ &+\lambda_3\|\Phi_1\|^2\|\Phi_2\|^2 +\lambda_4\|\Phi_1^\dagger\Phi_2\|^2 +\frac{1}{2}\{ \lambda_5(\Phi_1^\dagger\Phi_2)^2+h.c.\}, \label{eq:potential}\end{aligned}$$ In the case of CP-conservation, all the parameters are real. The theoretical constraints for these coupling parameters from perturbative unitarity are given in Refs. [@Kanemura:1993hm; @Akeroyd:2000wc]. In $Z_2$ symmetry, $\Phi_2$ has zero vacuum expectation value (VEV), and the SM like field, $\Phi_1$ takes part in the electroweak symmetry breaking (EWSB). After the EWSB, the doublet scalar fields are expanded around physical vacuum. $$\Phi_1^{}= \begin{pmatrix} G^+\\ \frac{1}{\sqrt{2}}(h+v+i G^0) \end{pmatrix} ,\ \Phi_2^{}= \begin{pmatrix} H^+\\ \frac{1}{\sqrt{2}}(H+iA) \end{pmatrix},$$ where $G^+$ and $G^0$ are the charged and neutral Goldstone bosons that are manifested as the longitudinal components of the gauge bosons, and $h$ is the SM-like Higgs boson with mass $m_h = 125$ GeV. The vacuum expectation value (VEV) of $\Phi_1$ is $v=246$ GeV. The second doublet field $\Phi_2$ contain four $Z_2^{}$ odd scalar bosons, a CP even neutral scalar boson $H$, a CP odd neutral scalar boson $A$, and two charged Higgs bosons $H^\pm$. After EWSB, the masses of these scalar bosons are given as $$\begin{aligned} m_h^2&=\lambda_1 v^2, \label{eq:mhsq}\\ m^2_{H^+}&=\mu_2^2+\frac{1}{2}\lambda_3 v^2, \label{eq:mHpsq}\\ m_H^2&=\mu_2^2+\frac{1}{2}(\lambda_3+\lambda_4+\lambda_5)v^2, \label{eq:mHsq}\\ m_A^2&=\mu_2^2+\frac{1}{2}(\lambda_3+\lambda_4-\lambda_5)v^2. \label{eq:mAsq} \end{aligned}$$ Assuming $\lambda_5<0$, the lightest CP even neutral scalar boson $H$ is stable, and could be a candidate of dark matter. The IDM scalar sector can be specified by a total of six free parameters: $$\begin{aligned} \{ \lambda_1, ~\lambda_2, ~\lambda_3, ~\lambda_4, ~\lambda_5, ~\mu_2 \}. \end{aligned}$$ We introduce the useful abbreviations $\lambda_L = \frac{1}{2}(\lambda_3+\lambda_4+\lambda_5)$ and $\lambda_s = \frac{1}{2}(\lambda_3+\lambda_4-\lambda_5)$. Through the above equations, the six parameters can be changed into a set of more meaningful parameters, $$\begin{aligned} \{ m_{H^{\pm}}, m_A, m_H, m_h, \lambda_L, \lambda_2 \},\end{aligned}$$ where $m_{H^{\pm}}, m_A, m_H$ are the four $Z_2^{}$ odd scalar boson masses. $\lambda_L$ correspond to the coupling of the dark matter and SM-like Higgs boson, which is relevant for dark matter annihilation. The quartic coupling $\lambda_2$ correspond to self-interaction in the dark sector. Thermal Relic abundance of dark matter ====================================== The dark matter relic abundance is obtained by solving the non-equilibrium Boltzmann equation, $$\frac { dn_{\chi} } {dt} + 3 \rm{H} n_{\chi} = - \left\langle \sigma \textit{v} \right\rangle \left( n_{ \chi }^{2} - \left(n_{\chi}^{ eq } \right)^{2}, \right)$$ where $n_{\chi}$ is the number density of dark matter particles, and $\rm{H}$ is the expansion rate of the universe, $\left\langle \sigma v \right\rangle$ is the thermally averaged annihilation cross section. By convention, we introduce the comoving number density $Y = n/s $, as well as substituting temperature for $x = \frac{m_{\chi} }{T}$, then the derivative of $Y$ with respect to $x$ is $$\frac { d Y } { d x } = - \frac {s} { \rm{H} x^{ 2 } } \left\langle \sigma v \right\rangle \left( Y^{ 2 } - \left( Y^{eq} \right)^{2} \right).$$ Solving this equation, we can get the number density as the temperature. By integrating the function from $x = x_f$ to $x \to \infty$, we get the number density $Y_{\infty }$. Using the result $Y_{\infty }$, we get the final relic density $ \Omega h^{2} $, $$\Omega h^{2} \equiv \frac {\rho_{\chi} } {\rho_{c} } h^{2} = \frac {m_{\chi} Y_{\infty} s_{\infty} } { 1.05 \times 10^{-5} \mathrm {GeV} \mathrm {cm}^{-3} },$$ where ${\rho_{c} } $ is the critical energy density of the universe, ${\rho_{\chi} } $ is the energy density of dark matter and $s_{\infty}$ is the entropy density in the present universe. We use the software Micromegas [@Belanger:2018ccd] to calculate the relic abundance of dark matter. The SM input relevant parameters are chosen as: $$\begin{aligned} &&m_{W}=80.379 {{\rm\ GeV}}, m_{Z}=91.1876 {{\rm\ GeV}}, m_{H}=125.18 {{\rm\ GeV}}, \nonumber \\ &&G_F=1.1663787\times 10^{-5}~{\rm GeV}^{-2}, m_{b}=4.18 {{\rm\ GeV}}, m_{t}=173.0 {{\rm\ GeV}},\nonumber \\ &&\alpha_s(m_Z)=0.1181, \alpha=7.297352\times 10^{-3} , m_{c}=1.275 {{\rm\ GeV}}.\end{aligned}$$ We choose ${m_{H}, m_{A}, m_{H^{\pm}}, \lambda_{L}, \lambda_{2}}$ as five independent input parameters of the IDM. The annihilation cross section is only calculated to the Leading Order (LO), thus the relic density is not sensitive to the parameter $\lambda_{2}$. If assuming the mass hierarchy $m_{H^{\pm} } \geq m_{A} > m_{H}$ for the inert scalar bosons, the lightest scalar $H$ will be a dark matter particle candidate. In general, the roles of $m_{H}$ and $m_{A}$ are interchangeable. Usually, its mass can be divided into three regions: [1]{}.low mass (1-80[[ GeV]{}]{}), [2]{}.intermediate mass (80-500[[ GeV]{}]{}), [3]{}.high mass (500-1000[[ GeV]{}]{}). Since $\lambda_2$ is only related to the self-coupling of inert particles, its variation has little effect on the relic abundance of dark matter particle, thus we fix it as $\lambda_2=0.01$. Then, we scan the other three mass parameters $m_{H}, m_{A}, m_{H^{\pm}}$ from 1[[ GeV]{}]{} to 1000[[ GeV]{}]{} and $\lambda_L$ from $-0.75$ to 6.28. When the DM relic density is in agreement with the Planck’s measurements: $0.119 < \Omega_{DM}h^2 < 0.121$, these data points are saved. ![ \[fig1\] (color online) The projection of reserved points in $m_H-m_{A}$ (left), $m_H-m_{H^\pm}$ (middle), $m_H-\lambda_L $(right) plane. ](relic.eps){width="95.00000%"} We put these reserved points project onto two dimension plane in Fig.\[fig1\]. The mass range of $ m_H$ is mainly concentrated on low mass region in the three figures, which below 150 GeV. The mass range of $m_{H^\pm}$ is mainly above 80 GeV. When $\lambda_L$ is close to zero, we can easily find a appropriate point in large $m_H$ area corresponding to correct relic abundance. When $\lambda_L$ is larger than 1, $m_H$ can only reserved in 1-3 GeV. We select six groups of parameters as benchmark points, which are listed in the following: [BP1]{}: $\lambda_2$ = 0.01, $\lambda_L$ =-0.067137, $m_H$ = 48.47607, $m_{A} $=159.37519, $m_{H^\pm} $=171.83007\ [BP2]{}: $\lambda_2$ = 0.01, $\lambda_L$ =-0.061830, $m_H$ =49.07796, $m_{A} $=104.58929, $m_{H^\pm} $=177.79335\ [BP3]{}: $\lambda_2$ = 0.01, $\lambda_L$ = 4.37294, $m_H$ =1.57134, $m_{A} $=183.29679, $m_{H^\pm} $= 192.24964\ [BP4]{}: $\lambda_2$ = 0.01, $\lambda_L$ = 0.13490, $m_H$ =32.23363, $m_{A} $=158.16124, $m_{H^\pm} $=194.83879\ [BP5]{}: $\lambda_2$ = 0.01, $\lambda_L$ = -0.15581, $m_H$ =3.1979, $m_{A} $=196.59793, $m_{H^\pm} $=201.52571\ [BP6]{}: $\lambda_2$ = 0.01, $\lambda_L$ = 5.83958, $m_H$ =1.87152, $m_{A} $=173.96496, $m_{H^\pm} $=186.41257\ ![ \[fig2\] (color online) Relic density parameter $\Omega h^{ 2 }$ as a function of $ m_H , m_{A}, m_{H^\pm}$ when other parameters are fixed (left). Relic density parameter $\Omega h^{2}$ as a function of $\lambda_L$ when other parameters are fixed (right). ](density1.eps "fig:"){width="49.00000%"} ![ \[fig2\] (color online) Relic density parameter $\Omega h^{ 2 }$ as a function of $ m_H , m_{A}, m_{H^\pm}$ when other parameters are fixed (left). Relic density parameter $\Omega h^{2}$ as a function of $\lambda_L$ when other parameters are fixed (right). ](density2.eps "fig:"){width="49.00000%"} In Fig.\[fig2\], we present the function of the relic abundance of dark matter with the parameters $m_H$, $m_A$ $m_{H^\pm}$ and $\lambda_L$. Since all above six set of parameters have similar characteristics by drawing, so we take BP1 as an example. From Fig.\[fig2\](left), we can see that, when the parameter $m_H$ or $m_A$ close to a half of higgs mass, the relic density is greatly reduced. This is due to that these two particle are easily merged into a on-shell Higgs and then decay into SM particles, thus the dark matter relic density has been dramatically reduced. When $m_A$ or $m_{H^\pm}$ get into high mass region, they don’t affect the relic density. In Fig.\[fig2\](right), we provide the relic density $\Omega h^{2}$ as a function of $\lambda_L$ with the fixed other parameters. The relic density parameter $\Omega h^{2}$ first increases and then decreases rapidly at the whole range of $\lambda_L$, and reaches its maximum value near zero. Even at some benchmark points, the point that corresponding to the correct relic abundance is not near zero, but the upward and downward trend always exists and reaches the maximum value near zero region. Constraints on the Model Parameters =================================== In this section, we summarize all the theoretical and experimental limitations for the extended scalar sector potential of the IDM. First, the perturbation of the theory requires all the scalar coupling constants cannot exceed $4\pi$ [@GarciaCely:2014jha]. $$\begin{aligned} \| \lambda_{1,2,3,4,5}\| \leq 4\pi \,,\hspace{10pt} \|\lambda_3+\lambda_4\pm\lambda_5\| < 4\pi \,,\hspace{10pt} \|\lambda_4\pm\lambda_5\| < 8\pi \,,\hspace{10pt} \|\lambda_3+\lambda_4\| < 4\pi.\end{aligned}$$ In order to obtain a stable vacuum, the following parameters must be positive [@Gunion:2002zf; @Gustafsson:2010zz; @Khan:2015ipa], $$\begin{aligned} \lambda_1^{} > 0,\ \lambda_2^{} > 0,\ \sqrt{\lambda_1^{} + \lambda_2^{}} + \lambda_3^{} > 0,\ \lambda_3^{}+\lambda_4^{} \pm \|\lambda_5^{}\| > 0. \label{eq:vs}\end{aligned}$$ The unitary of the S-matrix for processes $2 \to 2$ scattering at the perturbative level requires all the couplings [@Ginzburg:2004vp; @Branco:2011iw], $$\begin{aligned} \|\lambda_3 \pm \lambda_4 \| \leq 8 \pi, \quad \|\lambda_3 \pm \lambda_5 \| \leq 8 \pi, \quad \|\lambda_3+ 2 \lambda_4 \pm 3\lambda_5 \| &\leq 8 \pi{\nonumber}\\ \|-\lambda_1 - \lambda_2 \pm \sqrt{(\lambda_1 - \lambda_2)^2 + \lambda_4^2} \| & \leq 8 \pi{\nonumber}\\ \|-3\lambda_1 - 3\lambda_2 \pm \sqrt{9(\lambda_1 - \lambda_2)^2 + (2\lambda_3 + \lambda_4)^2} \| & \leq 8 \pi{\nonumber}\\ \|-\lambda_1 - \lambda_2 \pm \sqrt{(\lambda_1 - \lambda_2)^2 + \lambda_5^2} \| &\leq 8 \pi.{\nonumber}\label{eq:unitary}\end{aligned}$$ The Peskin-Takeuchi $S,\;T,\;U$ parameters are strictly limited by the electroweak precision observables. The deviation from the SM prediction $\Delta S$ and $\Delta T$ are experimentally given $\Delta S = 0.03\times0.09$, $\Delta T = 0.07\times0.08$. The contribution from the IDM can be calculated as in Ref.[@Arhrib:2012ia]. This typically prohibit large mass splittings among inert states, but for DM masses with $M_{H^0}\gtrsim 500$ GeV relatively small splittings are already required, especially when combined with the relic density constraint [@IDM_DM_th2]. The experimental constraints for the inert scalars are mainly come from the large electron-positron collider (LEP) and large hadron collider(LHC) at CERN. First, the constraints on the new scalar particles at LEP come from the measurements of the $Z\rightarrow A H$, $Z\rightarrow H^+ H^-$, $W^\pm\rightarrow A H^\pm$ and $W^\pm\rightarrow H H^\pm$ decay, which imply that $M_{A^0} + M_{H^0} \geq M_{Z}$, $2M_{H^\pm}\geq M_{Z}$, $M_{H^\pm} + M_{H,A} \geq M_{W}$. Secondly, SUSY searches at at LEPII leads to constraints on the charged Higgs mass: the charged Higgs mass is constrained by $M_{H^\pm}\gtrsim 70$ GeV [@Pierce:2007ut], the bound on $M_{H}$ is also involved: if $M_{H}<80$ GeV, then $\|M_{A}-M_{H}\| \le 8$ GeV, or else, $M_{A} \ge 110$ GeV[@LEP-II-H]. The constraints on IDM at the LHC come mainly from the SM Higgs boson decay width. The new couplings from IDM can either increase the invisible branching ratio and/or alter the strength of the Higgs boson and diphoton coupling [@Arhrib:2012ia; @Krawczyk:2013jta; @Swiezewska:2012eh; @Goudelis:2013uca], which strictly limited the mass of the inert lightest scalar particle less than $M_h/2$, and has little restriction for the masses above $M_h/2$. Direct di-leptons plus missing energy searches have also been performed to restrict the inert scalar masses in the region of $M_{H}\lesssim 60$ GeV and $M_A \lesssim 150$ GeV [@Belanger:2015kga]. From these constraints we find that the IDM is strongly restricted if the mass of inert scalar particles are less than 100 GeV and have little constraints for masses above 500GeV. charged Higgs pair production at $\gamma\gamma$ collider ======================================================== In order to maintain the symmetry of $Z_2$, the scalar particles in the IDM always produce in pairs at the collider. The lightest scalar particle $H$ in the IDM is stable and can be a dark matter candidate, other scalar particles will eventually decay into $H$ associated SM particles, such as $A \to H Z$, $H^{\pm} \to H W^{\pm}$. These scalar particles only couple to the Higgs boson and electroweak gauge bosons of the Standard Model, thus the production cross section for double scalar particles is usually small. However, the charged Higgs boson $H^\pm$ can couple to photons through electromagnetic interactions. Predictably, the cross section of the double charged Higgs production in $\gamma\gamma$ collider is considerable. In this section, we consider the following process at $\gamma\gamma$ collider, $$\begin{aligned} \label{process} \gamma \gamma \to H^+ H^-.\end{aligned}$$ All the tree-level Feynman diagrams are presented in Fig.\[born\]. The cross section of this process is only related to the mass of $H^{\pm}$ and is independent of the other four parameters $\lambda_{L}, \lambda_{2}, m_{H}, m_{A}$ in the IDM. The hard photon beam of the $\gamma\gamma$ collider can be obtained by using the laser backscattering technique at $e^+e^-$ linear collider [@Com1; @Com2; @Com3]. We denote $\hat{s}$ and s as the center-of-mass energies of the $\gamma\gamma$ and $e^+e^-$ systems, respectively. After calculating the cross section $\hat{\sigma}(\hat{s})$ for the subprocess $\gamma\gamma \to H^+ H^-$ in photon collision mode, the total cross section at an $e^+e^-$ linear collider can be obtained by folding $\hat{\sigma}(\hat{s})$ with the photon distribution function that is given in Ref.[@function; @Ginzburg:1999wz]. The cross section for the $e^+e^- \to \gamma\gamma \to H^+ H^-$ process is expressed as $$\sigma_{tot}(e^+e^- \to \gamma\gamma \to H^+ H^- ,~s)=\int^{x_{max}}_{2m_H^{\pm}/\sqrt{s}} dz\frac{d{\cal L}_{\gamma\gamma}}{dz} \hat{\sigma}(\hat{s}=z^2 s ).$$ The distribution function of photon luminosity is expressed as $$\begin{aligned} \frac{d{\cal L}_{\gamma\gamma}}{dz}=2z\int_{z^2/x_{max}}^{x_{max}} \frac{dx}{x} f_{\gamma/e}(x)f_{\gamma/e}(z^2/x),\end{aligned}$$ where $f_{\gamma/e}$ is the photon structure function, the fraction of the energy of the incident electron carried by the back-scattered photon $x$, which are interfaced by the CompAZ code [@Zarnecki:2002qr]. At low $x$ part ($x\leq0.1$), the photon spectrum is not properly described and underestimated, and it is qualitatively better for larger values of fraction $x$ of the longitudinal momentum of the electron beam. However, for $x > 2(1+\sqrt{2})\simeq 4.8$, the high energy photons can disappear through $e^+e^-$ pair creation in its collision with a following laser photon. The Feynman Rules are extracted by the program FeynRules [@FeynRules2.0] from the Lagrangian of IDM, then outputs to Universal FeynRules Output(UFO) files [@UFO]. For the cross-section calculation and simulation for signal and backgrounds, we make use of the Monte Carlo event generator MadGraph@NLO(MG5)[@MadGraph5]. PYTHIA6 [@Pythia] is utilized for parton shower and hadronization with the options of ISR and RSR included. Delphes [@Delphes] is then employed to account for the detector simulations and MadAnalysis5 for analysis, where the (mis-)tagging efficiencies and fake rates are assumed to be their default values in Delphes. The IDM mediator width is automatically computed by using the MadWidth module for each parameter point. In Fig.\[squre\] (left), we present the cross sections as functions of the colliding energy $\sqrt{s}$ for process $e^+e^- \to \gamma \gamma \to H^+ H^-$ by taking $m_H^{\pm} = 50,100,150,200,250~ {\rm GeV}$, separately. From this figure, we can see that, with the increment of the colliding energy $\sqrt{s}$, the total cross section for the process $e^+e^- \to \gamma \gamma \to H^+ H^-$ increases rapidly at first. When the colliding energy $\sqrt{s}$ reaches about 1 TeV, the total cross section increases slightly. Consequently, we can obtain larger cross section for process $e^+e^- \to \gamma \gamma \to H^+ H^-$ by raising the colliding energy $\sqrt{s}$. In Fig.\[squre\] (right), the total cross section is plotted for different mass of $m_H^{\pm}$ at $e^+ e^-$ collider by taking $\sqrt{s}$ = 500, 1000, 1500 and 3000 GeV. With the increment of charge Higgs mass $m_H^{\pm}$, the total cross section is decreasing. When its mass is close to a half of the centre of mass energy, the cross section quickly is approaching zero. Signal and Background ===================== Since the lightest scalar boson $H$ is stable, the charge Higgs $H^{\pm}$ particles will eventually decay into the $H$ and SM particles. In this section, we perform the Monte Carlo simulation and explore the sensitivity in photon-photon collider through the following channel, $$\begin{aligned} \label{process2} \gamma \gamma \to H^{+} H ^{-} \to W^{+} W^{-} H H,\end{aligned}$$ $H$ assumed as the lightest scalar particle in the IDM leave missing energy in detector and make it almost impossible to reconstruct events. $W$ boson decay to a electron or muon and its antineutrino. The Feynman diagrams for the process $\gamma \gamma \to H^{+} H ^{-} \to W^{+} W^{-} H H$ are presented in Fig.\[fig5\]. The dominant signal for pure leptonic channel is $\ell^{+} \ell^{-}+\slashed{E}_{T}$ in the IDM, where $\ell$ = [e, $\mu$]{}, which can obtain from either $H^{\pm} \to W^{\pm} H$, with $W^{\pm} \to \ell^{\pm} \nu $ or $H^{\pm} \to W^{\pm} A$, with $W^{\pm} \to \ell^{\pm} \nu, A \to H Z, Z \to \nu \overline {\nu}$, depend on the choice of parameters. The contribution from second decay chain can be neglect comparing the first case. Thus, we will focus on the process $\gamma \gamma \to H^{+} H ^{-} \to W^{+} W^{-} H H $, with the decay $W^{\pm} \to \ell^{\pm} \nu $. The cross sections of the production $\gamma \gamma \to H^{+} H ^{-}\to W^{+} W^{-} H H$ in the IDM with $\sqrt{s}=500$ GeV for the benchmark points are given in table \[cross\]. BP1 BP2 BP3 BP4 BP5 BP6 -------------- ------- ------- ------- ------- ------- ------- -- $\sigma(fb)$ 1.362 1.014 4.318 3.296 3.683 5.168 : \[BP\] The cross sections for the process $e^+e^- \to \gamma \gamma \to H^{+} H^{-}\to W^{+} W^{-} H H$ in the IDM with $\sqrt{s}=500$ GeV. \[cross\] In the pure leptonic channel, the signal of this process is two leptons $l^+l^-$ plus missing $E_{T}$, and the main backgrounds of the Standard Model are mainly $W^{+} W^{-}$ , Drell-Yan process, top-quark pair production ($t\bar{t}$), $WZ$, and $ZZ$ processes. For the Drell-Yan process, the two leptons are always back-to-back, and the missing $E_{T}$ is very small, which can be easily distinguished from the large missing $E_{T}$ signal. The final state of top-quark pair production contains a large number of hadrons, which can also be well eliminated in the photon photon collider. $WZ$ and $ZZ$ processes can also suppressed seriously by the two leptons invariant mass cut of $Z$ boson. These backgrounds can be neglected after suitable cuts. Therefore, we will not list the these backgrounds in the following analysis. We will analysis the main irreducible background $W^{+} W^{-}$ production. In our simulation, we first employ some basic cuts for the selection of events: $$\begin{aligned} p_T^\ell > 20 ~GeV, ~\|\eta_\ell\| < 2.0, ~\Delta R_{\ell\ell} > 0.4,\end{aligned}$$ where $p_T^{\ell}$ and $\eta_{\ell}$ are the transverse momentum and the pseudorapidity of the leptons. $\Delta R = \sqrt{\Delta \phi^2 + \Delta\eta^2}$ is the particle separation among the leptons in the final state with $\Delta \phi$ and $\Delta\eta$ being the separation in the azimuth angle and rapidity. The $\eta_{\ell}$ acceptance region avoids the gap between barrel and endcap, where the misidentification probability is the highest. According to the differential distribution between the signal and background, we can improve the ratio of signal to background by making suitable kinematical cuts. In Fig.\[fig7\], we show the distributions of some kinematical variables for the signal and background at 500 GeV. We first select $N(\ell) = 2$ , the signal almost concentrate low $p_{T}^{\ell}$ region, so we reject $p_{T}^{\ell} > $ 70 GeV. Then, because of signal decrease faster than background in high invariant mass region, $M(\ell^+,\ell^-) < 125 ~\rm {GeV}$ is required. Finally, we require the transverse missing energy $\slashed E_{T} > 95 $ GeV to improve the discovery significance. For a short summary, we list all the cut-based selections here: - Basic cut: $p_T^{\ell}>20$ GeV, $\|\eta_{\ell}\|<2.0$ and $\Delta R_{\ell\ell}> 0.4$; - Cut 1 means the basic cuts plus requiring select $ N(\ell) =2 $ ; - Cut 2 means Cut 1 plus requiring $p_T^{\ell} < 70~\rm {GeV}$; - Cut 3 means Cut 2 plus requiring the invariant mass of two leptons $M(\ell^+, \ell^-) < 125~\rm {GeV}$ ; - Cut 4 means Cut 3 plus requiring the transverse missing energy $\slashed E_{T} > 95~\rm{GeV}$. Cuts Signal Background $S/\sqrt{B+S}$ ------------ --------------------- --------------------- ---------------- -- -- -- Basic cuts $3.039\times10^{3}$ $2.616\times10^{6}$ 1.880 Cut 1 $1.956\times10^{3}$ $1.623\times10^{6}$ 1.534 Cut 2 $1.592\times10^{3}$ $1.036\times10^{6}$ 1.563 Cut 3 $1.528\times10^{3}$ $7.294\times10^{5}$ 1.787 Cut 4 $3.438\times10^{2}$ $2.904\times10^{2}$ 13.653 : The number of events for the signal ( $W^{+} W^{-} H H$ ) in BP2 and main backgrounds ( $W^{+} W^{-}$) after the cut flows are listed in the brackets at the 500 GeV with integrated luminosity $L = 3000 fb^{-1}$. The values of discovery significance $S/\sqrt{B+S}$ at each step of cut are also shown. \[cutflow\] The results of the number for the signal in BP2 and backgrounds (with luminosity $= 3000 fb^{-1}$) are shown in Table \[cutflow\] at each step of the cuts. The values of the discovery significance $S/\sqrt{B+S}$ are also shown, where $S$ and $B$ are the numbers of signal and total background events, respectively. After applying several cuts, the background can be reduced greatly, the discovery significance $S/\sqrt{B+S}$ can reach $13.653\sigma$. Thus, we have potential for observing the IDM effect though the charged Higgs $H^{\pm} $ pair in some parameter space with large luminosity at the $ \gamma\gamma $ collider. ![ \[fig8\] (color online) The signal background ratio at different reference points in $m_H-m_{H^\pm} $ plane.](SBRATIO.eps){width="80.00000%"} In Fig.\[fig8\], we present the distribution of the parameter point for the discovery significance $S/\sqrt{B+S}$ in the plane $m_{H^{+}}-m_H$ with the integrated luminosity of $3000fb^{-1}$ at $\sqrt{s}$ = 500 GeV. The parameter points with different colour represent the value of the significance. We investigate the effects of coupling parameter $\lambda_2$, $\lambda_L$ and the scalar even particle mass $m_A$ and find that the cross section has little change when varying these input parameters. From Fig.\[fig8\], we find that the parameter points with high significance are mainly concentrated in the range of $m_H$ from 10 to 50 GeV and $m_{H^{+}}$ from 110 to 180 GeV. If the CEPC or ILC can be built, these parameter points in the IDM model has potential to be detected or excluded. Summary ======= The Inert Doublet Model is one of the most simple extension of the Standard Model, which provide a scalar DM particle candidate. In this paper, we have investigated the double charged Higgs $ H^{\pm} $ pair production in IDM at the $ \gamma\gamma $ collider. Assuming that the lightest scalar Higgs is the dark matter particle, we have calculated the corresponding relic abundance, scanned the IDM parameter space, and obtained the parameter points satisfying the relic abundance of dark matter in our universe. We analyzed the pure lepton decay process of the double charged Higgs $H^{\pm} $ and the backgrounds of the Standard Model, and optimised the selection criteria employing suitable cuts on the kinematic variables to maximise the signal significance. We found that with high luminosity option of the $ \gamma\gamma $ collider, this channel has the potential to probe the IDM in the mass range of 1-250 GeV. In a scenario with light dark matter of mass about 10-50 GeV, charged Higgs in the mass range of around 110-180 GeV provides the best possibility with a signal significance of about $10 \sigma $ at an integrated luminosity of about 3000 $fb^{-1}$. Acknowledgments =============== This work was supported by the National Natural Science Foundation of China (No.11205003, No.11305001, No.11575002, No.11935001). [99]{} ATLAS Collaboration (G. Aad [et al.]{}), Phys. Lett. [B]{} 716, 1 (2012). CMS Collaboration (S. Chatrchyan [et al.]{}), Phys. Lett. [B]{} 716, 30 (2012). M. Baak [et al.]{}, Eur. Phys. J. C [72]{} (2012) 2205 \[arXiv:1209.2716 \[hep-ph\]\]. D. N. Spergel [et al.]{} \[WMAP Collaboration\], Astrophys. J. Suppl. [148]{} (2003) 175 \[astro-ph/0302209\]. A. Del Popolo, Astron. Rep. [51]{} (2007) 169 \[arXiv:0801.1091 \[astro-ph\]\]. A. Del Popolo, Int. J. Mod. Phys. D [23]{} (2014) 1430005 doi:10.1142/S0218271814300055 \[arXiv:1305.0456 \[astro-ph.CO\]\]. K. M. Smith, O. Zahn and O. Dore, Phys. Rev. D [76]{} (2007) 043510 \[arXiv:0705.3980 \[astro-ph\]\]. S. Das, B. D. Sherwin, P. Aguirre, et al., Phys. Rev. Lett. 107, 021301 (2011), arXiv:1103.2124. D. Hanson [et al.]{} \[SPTpol Collaboration\], Phys. Rev. Lett. [111]{} (2013) no.14, 141301 doi:10.1103/PhysRevLett.111.141301 \[arXiv:1307.5830 \[astro-ph.CO\]\]. G. Bertone, D. Hooper and J. Silk, Phys. Rept. [405]{} (2005) 279, \[hep-ph/0404175\]. J. L. Feng and J. Kumar, Phys. Rev. Lett. [101]{} (2008) 231301, \[arXiv:0803.4196\]. M. A. Diaz, B. Koch and S. Urrutia-Quiroga, arXiv:1511.04429 \[hep-ph\]. E. M. Dolle and S. Su, Phys. Rev. D [80]{} (2009) 055012 \[arXiv:0906.1609 \[hep-ph\]\]. A. Goudelis, B. Herrmann and O. Stål, JHEP [1309]{}, 106 (2013). N. Blinov, J. Kozaczuk, D. E. Morrissey and A. de la Puente, Phys. Rev. D [93]{}, no. 3, 035020 (2016). B. Eiteneuer, A. Goudelis and J. Heisig, Eur. Phys. J. C [77]{} (2017) no.9, 624 \[arXiv:1705.01458 \[hep-ph\]\]. A. Arhrib, Y. L. S. Tsai, Q. Yuan and T. C. Yuan, JCAP [1406]{}, 030 (2014). A. Ilnicka, M. Krawczyk and T. Robens, Phys. Rev. D [93]{}, no. 5, 055026 (2016). Q. H. Cao, E. Ma and G. Rajasekaran, Phys. Rev. D [76]{}, 095011 (2007). E. Dolle, X. Miao, S. Su and B. Thomas, Phys. Rev. D [81]{}, 035003 (2010). G. Belanger, B. Dumont, A. Goudelis, B. Herrmann, S. Kraml and D. Sengupta, Phys. Rev. D [91]{}, no. 11, 115011 (2015). P. Poulose, S. Sahoo and K. Sridhar, arXiv:1604.03045 \[hep-ph\]. N. Wan, N. Li, B. Zhang, H. Yang, M.-F. Zhao, M. Song, G. Li, and J.-Y. Guo, Searches for Dark Matter via Mono-W Production in Inert Doublet Model at the LHC, Commun. Theor. Phys. [69]{} (2018), no. 5, 617. A. Ahriche, A. Arhrib, A. Jueid, S. Nasri and A. de La Puente, arXiv:1811.00490 \[hep-ph\]. M. Aoki, S. Kanemura and H. Yokoya, Phys. Lett. B [725]{}, 302 (2013). M. Hashemi, M. Krawczyk, S. Najjari and A. F. Żarnecki, JHEP [1602]{}, 187 (2016). J. Kalinowski, W. Kotlarski, T. Robens, D. Sokolowska and A. F. Zarnecki, JHEP [1812]{}, 081 (2018) doi:10.1007/JHEP12(2018)081 \[arXiv:1809.07712 \[hep-ph\]\]. D. Dercks and T. Robens, arXiv:1812.07913 \[hep-ph\]. S. Kanemura, T. Kubota and E. Takasugi, Phys. Lett. B [313]{} (1993) 155. A. G. Akeroyd, A. Arhrib and E. M. Naimi, Phys. Lett. B [490]{} (2000) 119; G. B¨¦langer, F. Boudjema, A. Goudelis, A. Pukhov and B. Zaldivar, Comput. Phys. Commun. 231 (2018), 173-186 \[arXiv:1801.03509 \[hep-ph\]\]. C. A. Garcia Cely, PhD thesis, Technische Universität München (TUM), 2014. J. F. Gunion and H. E. Haber, [Phys.Rev.]{} [D67]{} (2003) 075019, \[[[ hep-ph/0207010]{}](http://xxx.lanl.gov/abs/hep-ph/0207010)\]. M. Gustafsson, [PoS]{} [CHARGED2010]{} (2010) 030, \[[[arXiv:1106.1719]{}](http://xxx.lanl.gov/abs/1106.1719)\]. N. Khan and S. Rakshit, [[arXiv:1503.0308]{}](http://xxx.lanl.gov/abs/1503.0308). I. F. Ginzburg and M. Krawczyk, [Phys.Rev.]{} [D72]{} (2005) 115013, \[[[hep-ph/0408011]{}](http://xxx.lanl.gov/abs/hep-ph/0408011)\]. G. Branco, P. Ferreira, L. Lavoura, M. Rebelo, M. Sher, et al., [Phys.Rept.]{} [516]{} (2012) 1–102, \[[[ arXiv:1106.0034]{}](http://xxx.lanl.gov/abs/1106.0034)\]. A. Arhrib, R. Benbrik, and N. Gaur, [Phys.Rev.]{} [D85]{} (2012) 095021, \[[[arXiv:1201.2644]{}](http://xxx.lanl.gov/abs/1201.2644)\]. T. Hambye, F.-S. Ling, L. Lopez Honorez and J. Rocher, JHEP [0907]{}, 090 (2009) Erratum: \[JHEP [1005]{}, 066 (2010)\] A. Pierce and J. Thaler, [JHEP]{} [0708]{} (2007) 026, \[[[ hep-ph/0703056]{}](http://xxx.lanl.gov/abs/hep-ph/0703056)\]. E. Lundstrom, M. Gustafsson and J. Edsjo, Phys. Rev. D [79]{}, 035013 (2009). B. Swiezewska and M. Krawczyk, [Phys. Rev.]{} [D88]{} (2013), no. 3 035019, \[[[arXiv:1212.4100]{}](http://xxx.lanl.gov/abs/1212.4100)\]. M. Krawczyk, D. Sokolowska, P. Swaczyna, and B. Swiezewska, [JHEP]{} [1309]{} (2013) 055, \[[[ arXiv:1305.6266]{}](http://xxx.lanl.gov/abs/1305.6266)\]. A. Goudelis, B. Herrmann, and O. Stal, [JHEP]{} [09]{} (2013) 106, \[[[ arXiv:1303.3010]{}](http://xxx.lanl.gov/abs/1303.3010)\]. G. Belanger, B. Dumont, A. Goudelis, B. Herrmann, S. Kraml, et al., [[arXiv:1503.0736]{}](http://xxx.lanl.gov/abs/1503.0736). I. Ginzburg, G. Kotkin, V. Serbo and V. Telnov, Pizma ZhETF, [34]{} (1981) 514; JETP Lett. [34]{} (1982) 491. Preprint INP 81-50, 1981, Novosibirsk. I. Ginzburg, G. Kotkin, V. Serbo and V. Telnov, Nucl. Instr. and Meth. [205]{} (1983) 47, Preprint INP 81-102, 1991, Novosibirsk. I. Ginzburg, G. Kotkin, S. Panfil, V. Serbo and V. Telnov, Nucl. Instr. and Meth. [219]{} (1984) 5. G. Jikia. [Nucl. Phys.]{}, 1992, B[374]{}: 83; O. J. P. Eboli $et$ $al$., Phys. Rev. D [47]{}, 1889(1993). I. F. Ginzburg and G. L. Kotkin, Eur. Phys. J. C [13]{} (2000) 295 \[hep-ph/9905462\]. A. F. Zarnecki, Acta Phys. Polon. B [34]{} (2003) 2741 \[hep-ex/0207021\]. A. Alloul, N. D. Christensen, C. Degrande, C. Duhr, and B.Fuks, Comput. Phys. Commun. 185, 2250-2300 (2014), \[arXiv:1310.1921\]. C. Degrande, C. Duhr, B. Fuks, D. Grellscheid, O.Mattelaer, and T. Reiter, Comput. Phys. Commun. 183, 1201-1214 (2012), \[arXiv:1108.2040\] J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S. Shao, T. Stelzer, P. Torrielli, and M. Zaro, JHEP 1407, 079 (2014), \[arXiv:1405.0301\] T. Sjostrand, S. Mrenna, P.Z. Skands, JHEP 0605, 026 (2006). J.de Favereau,et al,DELPHES 3 Collaboration, JHEP 1402,057(2014). M.Cacciari,G.P.Salam,G.Soyez, JHEP 0804,063(2008). A. Ilnicka, M. Krawczyk and T. Robens, arXiv:1505.04734 \[hep-ph\].	{ "pile_set_name": "ArXiv" }
--- abstract: \| Let $H$ be a graph on $h$ vertices. The number of induced copies of $H$ in a graph $G$ is denoted by $i_H(G)$. Let $i_H(n)$ denote the maximum of $i_H(G)$ taken over all graphs $G$ with $n$ vertices. Let $f(n,h) = \Pi_{i}^h a_i$ where $\sum_{i=1}^h a_i = n$ and the $a_i$ are as equal as possible. Let $g(n,h) = f(n,h) + \sum_{i=1}^h g(a_i,h)$. It is proved that for almost all graphs $H$ on $h$ vertices it holds that $i_H(n)=g(n,h)$ for all $n \le 2^{\sqrt{h}}$. More precisely, we define an explicit graph property ${\cal P}_h$ which, when satisfied by $H$, guarantees that $i_H(n)=g(n,h)$ for all $n \le 2^{\sqrt{h}}$. It is proved, in particular, that a random graph on $h$ vertices satisfies ${\cal P}_h$ with probability $1-o_h(1)$. Furthermore, all extremal $n$-vertex graphs yielding $i_H(n)$ in the aforementioned range are determined. We also prove a stability result. For $H \in {\cal P}_h$ and a graph $G$ with $n \le 2^{\sqrt{h}}$ vertices satisfying $i_H(G) \ge f(n,h)$, it must be that $G$ is obtained from a balanced blowup of $H$ by adding some edges inside the blowup parts. The [inducibility]{} of $H$ is $i_H = \lim_{n \rightarrow \infty} i_H(n)/\binom{n}{h}$. It is known that $i_H \ge h!/(h^h-h)$ for all graphs $H$ and that a random graph $H$ satisfies almost surely that $i_H \le h^{3\log h}h!/(h^h-h)$. We improve upon this upper bound almost matching the lower bound. It is shown that a graph $H$ which satisfies ${\cal P}_h$ has $i_H =(1+O(h^{-h^{1/3}}))h!/(h^h-h)$. author: - 'Raphael Yuster [^1]' bibliography: - 'references.bib' title: On the exact maximum induced density of almost all graphs and their inducibility --- Introduction ============ We consider the basic problem of maximizing the number of induced copies of a given graph $H$ in a graph with $n$ vertices. This problem has been studied already for four decades by various researchers such as in [@BHLP-2016; @BEHJ-1995; @BNT-1986; @BS-1994; @EL-2015; @exoo-1986; @FV-2012; @HHN-2014; @ht-2017; @hirst-2014; @huang-2014; @PG-1975] and its answer (even asymptotically) seems difficult already for some very small graphs. Our main result gives an exact answer to this problem for almost all graphs $H$, and for all $n$ that are not too huge. Let $i_H(G)$ denote the number of induced copies of $H$ in a graph $G$. The related extremal graph-theoretic parameter of interest is therefore the maximum of $i_H(G)$ taken over all graphs $G$ with $n$ vertices, denoted by $i_H(n)$. It is a simple exercise to prove that the maximum induced density of $H$ in graphs with $n$ vertices, namely $i_H(n)/\binom{n}{h}$ tends to a limit, which is denoted by $i_H = \lim_{n \rightarrow \infty}i_H(n)/\binom{n}{h}$ and is called the [inducibility]{} of $H$. The first to define and study these natural parameters were Pippenger and Golumbic [@PG-1975]. Let $f(n,h) = \Pi_{i}^h a_i$ where $\sum_{i=1}^h a_i = n$ and the $a_i$ are as equal as possible. Let $g(n,h)$ be the function recursively defined as $g(n,h)=0$ if $n < h$ and otherwise by $g(n,h) = f(n,h) + \sum_{i=1}^h g(a_i,h)$. Having noticed that the so-called [nested blowups]{} of $H$ on $n$ vertices (see definition in the next section) contain at least $g(n,h)$ induced copies of $H$, namely that $i_H(n) \ge g(n,h)$, they used this fact to prove that $i_H \ge h!/(h^h-h)$ for every graph $H$. However, as it turns out, determining $i_H$ and, moreover, determining $i_H(n)$ seems to be a very difficult task in general. In fact, there are only very few families of graphs and sporadic cases for which these are known. It is easy to see that $i_H(n)=i_{H^C}(n)$ where $H^C$ is the complement of $H$. Clearly, $i_{K_h}(n)=i_{I_h}(n)=\binom{n}{h}$ and that cliques and their complements are the only graphs with unit inducibility. Applying a result of Goodman [@goodman-1959] it is also a simple exercise to determine $i_{K_{1,2}}(n)$ and deduce that $i_{K_{1,2}}=3/4$ thereby obtaining a complete solution for all graphs on three vertices. The inducibility of most graphs on four vertices is known, see [@EL-2015; @exoo-1986]. However, the inducibility of $P_4$, the path on four vertices, is still not completely determined. See the website of the flag algebra software of Vaughan [@flagmatic-site] for the best upper bound obtained using the flag algebra method of Razborov [@razborov-2007] and Even-Zohar and Linial [@EL-2015] for the best lower bound. More generally, the inducibility and maximum induced density of complete bipartite graphs is well understood [@BNT-1986; @BS-1994; @PG-1975]. For example, $i_{K_{h,h}}=\binom{2h}{h}/4^h$. Bollob[á]{}s, Egawa, Harris, and Jin [@BEHJ-1995] proved that if $H$ is a sufficiently large balanced blowup of some complete graph $K_r$, then $i_H(n)$ is obtained by blowups of $K_r$. Recently, Hatami, Hirst, and Norine [@HHN-2014] proved that if $H$ is a sufficiently large balanced blowup of some graph $K$, then any graph which attains $i_H(n)$ for $n$ sufficiently large must itself essentially be a blow-up of $K$. An intriguing conjecture raised in the paper of Pippenger and Golumbic [@PG-1975], which is yet unsolved, is that the inducibility of the cycle $C_h$ for $h \ge 5$ is equal to the aforementioned lower bound $h!/(h^h-h)$. The case $h=5$ was only very recently solved by Balogh, Hu, Lidick[y]{}, and Pfender [@BHLP-2016] with sophisticated application of flag algebra. In fact, they have proved that $i_{C_5}(5^k)=g(5^k,5)$ and that the unique extremal graph is the corresponding nested blowup of $C_5$. They also proved that for $n$ sufficiently large, $i_{C_5}(n) = g(n,5)$. So, in this case nested blowups (as opposed to balanced blowups) are extremal graphs. Nested blowups appear as extremal constructions also in many other extremal problems, as shown by Pikhurko [@pikhurko-2014]. Which other graphs have the property that their inducibility is close (or equal) to the generic lower bound $h!/(h^h-h)$? Clearly, if we can prove that the nested blowups are extremal for $c_H(n)$ for all $n$, then a by-product would be that $i_H=h!/(h^h-h)$. Likewise, if we can prove that the nested blowups are extremal for $c_H(n)$ for a very large $n$ as a function of $H$, then we can get very close to $h!/(h^h-h)$. The main result of this paper does the latter, and does it for almost all graphs $H$. Recall a random graph on $h$ vertices is the distribution $G(h,1/2)$ on $h$-vertex graphs where each pair of vertices is an edge with probability $1/2$. Recall also that a graph property ${\cal P}_h$ is a family of $h$-vertex graphs closed under isomorphism. We say that a graph $H \sim G(h,1/2)$ satisfies ${\cal P}_h$ asymptotically almost surely (alternatively, that ${\cal P}_h$ is satisfied by almost all graphs), if $\Pr [H \in {\cal P}_h] = 1-o_h(1)$. In Section 3, we describe an explicit graph property ${\cal P}_h$ that we call [strongly asymmetric graphs]{}. The following two theorems together form the crux of our result. \[t:1\] Let $H \sim G(h,1/2)$. $$\Pr [H \in {\cal P}_h] = 1-o_h(1)\;.$$ \[t:main\] $ $ 1. For all $n \le 2^{\sqrt{h}}$, if $H \in {\cal P}_h$, then $i_H(n)=g(n,h)$. Furthermore, the extremal graphs yielding $i_H(n)$ are precisely all the nested blowups of $H$ on $n$ vertices. 2. For all $n \le 2^{\sqrt{h}}$, if $H \in {\cal P}_h$ and $G$ is an $n$-vertex graph with $i_H(G) \ge f(n,h)$, then $G$ must be a balanced blowup of $H$ with some edges added inside the blowup parts. Notice that Theorem \[t:main\] is a completely deterministic statement. We emphasize this since there are graphs that satisfy ${\cal P}_h$ and that are quite far from a typical element of $G(h,1/2)$ (for example, there are graphs in ${\cal P}_h$ that contain linear sized cliques). Another point to mention is that one can extend the range for which Theorem \[t:main\] holds to all $n \le 2^{h^{1-\epsilon}}$ for a given $\epsilon > 0$, at the price of redefining ${\cal P}_h = \emptyset$ when $h$ is sufficiently small, depending on $\epsilon$. It is known that random graphs $H$ on $h$ vertices satisfy $i_H \le h^{O(\log h)}h!/(h^h-h)$ asymptotically almost surely, as has been observed by Even-Zohar and Linial. Using our result we can improve this upper bound by essentially removing the $h^{O(\log h)}$ factor, thereby coming extremely close to the generic lower bound. \[t:inducibility\] If $H \in {\cal P}_h$, then $$i_H \le \frac{h!}{h^h-h} \cdot \left( 1+\frac{4}{h^{h^{1/3}}} \right)\;.$$ The rest of this paper has the following structure. The next section consists of definitions and description of the basic objects required for the rest of the paper. Strongly asymmetric graphs are defined in Section 3, where we also prove that almost all graphs are strongly asymmetric, namely Theorem \[t:1\]. The proof of the main result, Theorem \[t:main\], appears in Section 4. Its main ingredient uses a concept of [consistent large sets]{} defined there. The proofs of the properties of consistent large sets and the main lemma showing that they can be grouped to a single consistent large set are given in Sections 5 and 6 respectively. Section 7 shows how to apply our main result to inducibility, proving Theorem \[t:inducibility\]. The final section contains a conjecture and some concluding remarks. Preliminaries ============= Throughout this paper we generally do not omit floors and ceilings as these are important for the proof to work for the case where $n$ is close to $h$, as we do not want to treat this (seemingly easier) case separately. For positive integers $h$ and $n$, a sequence of nonnegative integers $a_1,\ldots,a_h$ whose sum is $n$ is an [$(h,n)$-partition]{}. We say that the sequence is [equitable]{} if any two elements in the sequence differ by at most $1$. Otherwise, we say that the sequence is non-equitable. Clearly, if $n \pmod h = t$ where $0 \le t \le h-1$, then an equitable $(h,n)$-partition has $t$ elements of order $\lceil n/h \rceil$ and $h-t$ elements of order $\lfloor n/h \rfloor$. Let $f(n,h)$ be the product of the elements of an equitable $(h,n)$-partition. Then, $$f(n,h) = \lceil n/h \rceil^t \lfloor n/h \rfloor^{h-t}\;.$$ We note that $f(n,h)=0$ for $n < h$ and that $f(h,h)=1$. More generally, $f(kh,h)=k^h$ for a positive integer $k$. The following lemma summarizes a few properties of $f(n,h)$ and non-equitable partitions. \[l:simple\] $$f(n-1,h) = \frac{f(n,h)}{\lceil n/h \rceil}\left( \lceil n/h \rceil -1 \right)\;.$$ Consequently, $$f(n-k,h) \le f(n,h)\left(1-\frac{1}{\lceil n/h \rceil} \right)^k\;.$$ If $a_1,\ldots,a_h$ is a non-equitable $(h,n)$-partition, then $$\Pi_{i=1}^h a_i \le f(n,h)\left(1-\frac{1}{\lceil n/h \rceil^2}\right)\;.$$ ------------------------------------------------------------------------ [[Proof.]{} ]{}The first part, as well as its obvious consequence are straightforward hence we prove only the third part. Consider placing $n$ elements in $h$ bins where the $i$th bin contains $a_i$ elements. One can obtain such a placement by starting with an equitable $(h,n)$-partition of the elements into the bins and repeatedly moving elements from some bin whose current size is $a$ to a bin whose current size is $b$ where $a \le b$. After each move the product of the bin sizes reduces by a factor of $(a-1)(b+1)/ab \le 1-1/(ab)$. Before the first move we have $a \le b \le \lceil n/h \rceil$. So already after the first move the product reduces by a factor of at most $1-\frac{1}{\lceil n/h \rceil^2}$. ------------------------------------------------------------------------ We next define a recursive variant of $f(n,h)$. Let $h$ and $n$ be positive integers. Let $g(n,h)=0$ for $n < h$. For $n \ge h$, define $g(n,h) = f(n,h) + \sum_{i=1}^h g(a_i,h)$ where $a_1,\ldots,a_h$ is an equitable $(h,n)$-partition. It is immediately observed that $g(n,h)=f(n,h)$ for all $n \le h(h-1)$ while $g(n,h) > f(n,h)$ for $n > h(h-1)$. When $n=h^k$ for a positive integer $k$ we obtain by induction that $$\label{e:ghk} g(h^k,h)= \sum_{i=1}^k h^{(k-1)h-(i-1)(h-1)} = \frac{h^{h(k-1)}(1- h^{k(1-h)})}{1-h^{1-h}}\;.$$ Let $H$ be a graph on $h$ vertices and assume that $V(H)=[h]=\{1,\ldots,h\}$. Let $a_1,\ldots,a_h$ be an $(h,n)$-partition, where the $a_i$ are positive integers. The graph $H(a_1,\ldots,a_h)$ is defined as follows. Its vertex set is the disjoint union of independent sets $A_1,\ldots,A_h$ with $\|A_i\|=a_i$, hence it has $n$ vertices. Its edges are defined as follows. For each pair $i,j$ of distinct vertices of $H$, if $ij \in E(H)$ then the bipartite graph in $H(a_1,\ldots,a_h)$ between $A_i$ and $A_j$ is complete, whereas if $ij \notin E(H)$ then the bipartite graph in $H(a_1,\ldots,a_h)$ between $A_i$ and $A_j$ is empty. We say that $H(a_1,\ldots,a_h)$ is a [blowup]{} of $H$. We call $A_i$ the [part]{} of the blowup corresponding to vertex $i \in V(H)$. We call a pair of distinct vertices of the blowup a [part pair]{} if both are in the same part and call them a [blowup pair]{} if they are in distinct parts. If $a_1,\ldots,a_h$ is an equitable $(h,n)$-partition, then $H(a_1,\ldots,a_h)$ is called an $n$-vertex [balanced blowup]{} of $H$ and is denoted by $H(n)$. A few simple facts regrading blowups follow. An induced copy of $H$ in $H(a_1,\ldots,a_h)$ can be obtained by selecting one vertex from each part. This immediately shows that $i_{H}(H(a_1,\ldots,a_h)) \ge \Pi_{i=1}^h a_i$. In particular, an $n$-vertex balanced blowup of $H$ gives that $i_{H}(H(n)) \ge f(n,h)$ which implies that $i_{H}(n) \ge f(n,h)$. Notice, however that the last inequality is not sharp for all $n > h(h-1)$. Indeed, consider $H(n)$. One of the parts has order at least $h$. Now, instead of letting this part be an independent set, add an induced copy of $H$ to it. This adds at least one additional induced copy of $H$ to the graph, while keeping all the induced copies having one vertex in each part. Hence, $i_{H}(n) > f(n,h)$ for $n > h(h-1)$. The last example immediately triggers the following construction, so called the family of $n$-vertex [nested balanced blowups]{} of $H$, denoted by $H^(n)$. For $n < h$ we define $H^(n)$ to be the set of all graphs on $n$ vertices. For $n \ge h$ we define $H^(n)$ as follows. Take a balanced blowup $H(n)$. Now, replace each independent set $A_i$ with an element of $H^(a_i)$ (note: if $a_i=a_j$ we are allowed to replace $A_i$ with an element of $H^(a_i)$ that is different from the element of $H^(a_i)$ that replaced $a_j$). An element of $H^(n)$ is called a nested balanced blowup of $H$. A few simple facts regrading nested balanced blowups follow from the definition. First notice that for all $h \le n \le h(h-1)$ we have that the elements of $H^(n)$ are just the usual balanced blowups of $H$. Another obvious but interesting point to observe is that for a positive integer $k$, if $n=h^k$, then $H^(n)$ has only one element. Also, by immediate induction, $i_H(X) \ge g(n,h)$ for any $X \in H^(n)$ which implies that $i_{H}(n) \ge g(n,h)$. Notice, however, that this inequality is not always tight. In fact, in some cases, $H^(n)$ can itself contain an element having more than $g(n,h)$ induced copies of $H$. Consider for example, $H=K_{1,2}$. It is immediate to check that the unique element of $K_{1,2}^(9)$ contains more than $3^3+3=g(9,3)$ induced copies of $K_{1,2}$. We end this section by recalling the definition of isomorphism between graphs. Let $H_1$ and $H_2$ be two graphs of the same order. A bijective function $f:V(H_1) \rightarrow V(H_2)$ is called an [isomorphism between $H_1$ and $H_2$]{} if $(u,v) \in E(H_1)$ if and only if $(f(u),f(v)) \in E(H_2)$. We say that $H_1$ and $H_2$ are [isomorphic]{} if there exists an isomorphism between them. An isomorphism between a graph and itself is called an [automorphism]{}. If $f$ is a bijection between $H_1$ and $H_2$, then a stationary point of $f$ is a vertex $v$ such that $f(v)=v$. Notice that the number of stationary points is at most $\|V(H_1) \cap V(H_2)\|$. Strongly asymmetric graphs ========================== The purpose of this section is to define the family of strongly asymmetric graphs and prove that almost all graphs reside in this family. However, before we do that, we need a couple of definitions. In what follows, [changing]{} an edge in a graph means either removing it, or adding it. Let $H_1$ and $H_2$ be two graphs of the same order, where $H_1$ is a blowup of some other graph $H$. We say that $H_1$ and $H_2$ are [blowup $m$-close]{} if one can change at most $m$ blowup edges of $H_1$ and arbitrarily change part edges of $H_1$ such that after the change, the obtained graph is isomorphic to $H_2$. We say that they are [blowup $m$-far]{} if they are not blowup $m$-close. Notice that if $H_1$ is a trivial blowup (i.e. $H_1$ is just an arbitrary graph), then the latter definition just means that we can change at most $m$ edges of $H_1$ such that after the change, the obtained graph is isomorphic to $H_2$. In this case we say that the graphs are $m$-close or $m$-far. More specifically, we need the following definition: Let $H_1$ and $H_2$ be two graphs of the same order, and let $f$ be a bijection from $V(H_1)$ to $V(H_2)$. We say that $f$ is [$m$-close to an isomorphism between $H_1$ and $H_2$]{} if one can change at most $m$ edges of $H_1$ such that $f$ is an isomorphism between the obtained graph and $H_2$. Otherwise, we say that $f$ is [$m$-far from an isomorphism between $H_1$ and $H_2$]{}. Let $H$ be a graph and let $S \subset V(H)$. Let $A(S) \subseteq V(H) \setminus S$ be the set of vertices with the following property. Each $w \in A(S)$ is either adjacent to all vertices of $S$ or non-adjacent to all vertices of $S$. We call $A(S)$ the [agreement set]{} of $S$. We say that $Q \subset V(H)$ is a [distinguishing set]{} if $Q$ is not contained in the agreement set of any two vertices outside of $Q$. Let $H$ be a graph on vertex set $[h]$. Recall from the previous section that a permutation $\pi : [h] \rightarrow [h]$ is an automorphism of $H$ if $(\pi(i), \pi(j))$ is an edge of $H$ if and only if $(i,j)$ is an edge of $H$. Clearly, the set of all automorphisms of $H$ is a group with respect to the composition operator, denoted by $Aut(H)$. Following Erdös and Rényi [@ER-1963], we say that $H$ is [asymmetric]{} if $Aut(H)$ consists only of the identity permutation and [symmetric]{} otherwise. Is is easy to verify that all graphs on at most $5$ vertices (and more than one vertex) are symmetric. However, there are already asymmetric graphs on $6$ vertices. The smallest one (with respect to the number of edges) is obtained from the path on vertices $1,2,3,4,5$ (in this order) by adding vertex $6$ and connecting it to vertices $3$ and $4$. Erdös and Rényi [@ER-1963] proved that almost all graphs are asymmetric. We will need a stronger notion of asymmetry in graphs. For a positive integer $h$ we say that a graph $H$ on $h$ vertices is [strongly asymmetric]{}, or, synonymously, that it belongs to the family ${\cal P}_{h}$, if all the following conditions hold for $H$: 1. The degree of every vertex of $H$ is larger than $0.4h$ and smaller than $0.6h$. 2. For every pair of vertices, the order of their agreement set is not larger than $0.55h$ and for any triple of vertices, the order of their agreement set is not larger than $0.3h$. 3. $H$ has a distinguishing set of order at most $3\log h$ [^2]. 4. Let $S \subset [h]$ with $\|S\| \le 0.7h$. Let $B$ be any blowup of $H[S]$ with at least $0.8h$ vertices and at most $h$ vertices, and let $K \subseteq [h]$. Furthermore, assume that each blowup part of $B$ contains no more than $h/100$ vertices. Then $B$ is blowup $10^{-5}h^2$-far from $H[K]$. 5. Let $J \subset [h]$ and $K \subset [h]$ where $\|J\|=\|K\|=\lceil 0.7h \rceil$ and let $\pi$ be a bijection from $J$ to $K$ having at least $0.1h$ non-stationary points. Then $\pi$ is $10^{-5}h^2$-far from an isomorphism between $H[J]$ and $H[K]$. It should be noted that although not stated explicitly, strong asymmetry implies asymmetry. Indeed, suppose $H$ is strongly asymmetric. Then, Condition 5 implies that any automorphism of $H$ has at most $0.1h$-non stationary points. But if an automorphism is non-trivial this means that there is a pair of distinct vertices whose agreement set is at least $0.9h-2$, which is impossible by Condition 2. We next show that almost all graphs on $h$ vertices do in fact belong to ${\cal P}_h$. Recall that ${\cal G}(h,\frac{1}{2})$ is the probability space of all graphs on vertex set $[h]$ where each pair of vertices are connected with an edge with probability $\frac{1}{2}$, and the $\binom{h}{2}$ choices are independent. \[t:phe\] For all $h$ sufficiently large, if $H \sim G(h,1/2)$, then $$\Pr [H \in {\cal P}_h] \ge 1-\frac{2}{h}\;.$$ Hence Theorem \[t:1\] is just a more explicit restatement of Theorem \[t:phe\]. The proof of Theorem \[t:phe\] follows from the five following lemmas, each considering one item in the definition of ${\cal P}_{h}$. While the first three are very elementary and their proofs are only presented for completeness, the latter two are more technical. \[l:cond-1\] If $h$ is sufficiently large and $H \sim G(h,1/2)$, then with probability at least $1-1/(5h)$ all degrees of $H$ are larger than $0.4h$ and smaller than $0.6h$. [[Proof.]{} ]{}The degree of a vertex of $H$ is a random variable with distribution $Bin(h-1,1/2)$ so its expectation is $(h-1)/2$. The probability that it deviates from its expectation by more than a constant factor is exponentially small in $h$, and there are only $h$ vertices to consider. ------------------------------------------------------------------------ \[l:cond-2\] If $h$ is sufficiently large and $H \sim G(h,1/2)$, then with probability at least $1-1/(5h)$, $H$ has the property that for every pair of vertices, the order of their agreement set is not larger than $0.55h$ and for any triple of vertices, the order of their agreement set is not larger than $0.3h$. [[Proof.]{} ]{}Consider a pair of vertices. Their agreement set is a random variable with distribution $Bin(h-2,1/2)$ so its expectation is $(h-2)/2$. The probability that it deviates from its expectation by more than a constant factor is exponentially small in $h$, and there are less than $h^2$ pairs to consider. Consider a triple of vertices. Then their agreement set is a random variable with distribution $Bin(h-3,1/4)$ so its expectation is $(h-3)/4$. The probability that it deviates from its expectation by more than a constant factor is exponentially small in $h$, and there are less than $h^3$ triples to consider. ------------------------------------------------------------------------ \[l:cond-3\] If $h$ is sufficiently large and $H \sim G(h,1/2)$, then with probability at least $1-1/h$, $H$ has a distinguishing set of at most $\lfloor 3 \log h \rfloor$ vertices. [[Proof.]{} ]{}Fix some $Q \subset [h]$ with $\|Q\|=\lfloor 3 \log h \rfloor$. We prove that with probability at least $1-1/h$, $Q$ is a distinguishing set. For two given vertices outside of $Q$, the probability that $Q$ is contained in their agreement set is precisely $2^{\|-Q\|}$. As there are less than $h^2/2$ pairs of vertices outside of $Q$, the probability that $\|Q\|$ is not distinguishing is (by the union bound) at most $(h^2/2) 2^{\|-Q\|}$ which satisfies $$\frac{h^2}{2}2^{\|-Q\|} \le \frac{h^2}{2}2^{1-3 \log h} = \frac{1}{h}\;.$$ ------------------------------------------------------------------------ \[l:cond-4\] If $h$ is sufficiently large and $H \sim G(h,1/2)$, then the following holds. Let $S \subset [h]$ with $\|S\|=s \le 0.7h$. Let $(a_1,\ldots,a_s)$ be an $(s,k)$-partition with $h \ge k \ge 0.8h$ and with $a_i \le h/100$ for $i=1,\ldots,s$. Let $K \subseteq [h]$ with $\|K\|=k$. Then, the probability that $B=H[S](a_1,\ldots,a_s)$ is blowup $10^{-5}h^2$-close to $H[K]$ is at most $e^{-0.001h^2}$. [[Proof.]{} ]{}Denote the parts of $B$ by $A_1,\ldots,A_s$ where $\|A_i\|=a_i \le h/100$. Let us fix a set $M$ of $m \le \lfloor 10^{-5}h^2 \rfloor$ blowup pairs of $B$ and fix a bijection $\pi\,:\, V[B] \rightarrow K$. Let $B_M$ be the graph obtained from $B$ after changing $M$ (making pairs in $M$ that are edges into non-edges and making pairs in $M$ that are non-edges into edges). Call a vertex of $B$ [problematic]{} if it appears in at least $h/30$ elements of $M$. Notice that being problematic has nothing to do with the structure of $H$, since it only depends on $S$ and $M$ ($S$ is just a subset of $[h]$ and $M$ is just a subset of blowup pairs, while blowup pairs only depend on the partition). As each pair in $M$ contributes two to the count towards being problematic, the number of problematic vertices is at most $$\label{e:num-prob} \frac{2m}{h/30} \le 0.0006h\;.$$ Consider now the following maximum set of pairs $\{x_1,y_1\},\ldots,\{x_\ell,y_\ell\}$ that are pairwise disjoint (i.e altogether they contain $2\ell$ vertices) and have the following property. For all $j=1,\ldots,\ell$, $\{x_j,y_j\}$ both belong to the same part (no matter which) and both are non problematic. We lower bound $\ell$ as follows. Since $\ell$ is maximum, from each part $A_i$ for $i=1,\ldots,s$, there is at most one uncovered non-problematic vertex in our set of pairs. Thus, $$\label{e:l} \ell \ge \frac{k-0.0006h-s}{2} \ge 0.04h$$ where we have used that $k \ge 0.8h$ and $s \le 0.7h$. Consider some $\{x_j,y_j\}$ pair and suppose they belong to the part $A_i$. Although both $x_j$ and $y_j$ are non-problematic, they can still occur each in at most $h/30$ elements of $M$, so after ignoring the at most $2\cdot h/30=h/15$ vertices in pairs in $M$ to which at least one of $x_j$ or $y_j$ belongs and after ignoring all the problematic vertices we still have a set $Z$ of vertices of $B$ of size at least $$k-0.0006h-\frac{h}{15}-a_i \ge 0.8h-0.006h-\frac{h}{15}-\frac{h}{100} \ge 0.7h$$ that has the following property: For each $z \in Z$, $(x_j,z)$ and $(y_j,z)$ are blowup pairs that are not in $M$. Again, we stress that $Z$ has nothing to do with the structure of $H$ since it only depends on $S$, on $M$, and on problematic vertices. In other words, regardless of the structure of $H$, in the graph $B_M$, the set $Z$ will be in the agreement set of $x_j$ and $y_j$. Now consider $H[K]$. In order for $\pi$ to be an isomorphism between $B_M$ (regardless of any additional changes to part pairs of $B$) and $H[K]$ we need for the agreement set of each pair $\{\pi(x_1),\pi(y_1)\},\ldots,\{\pi(x_\ell),\pi(y_\ell)\}$ to be of size at least $0.7h$. What is the probability of this occurring? Consider some pair $\{\pi(x_j),\pi(y_j)\}$. Since $H[K]$ is a random graph on $k$ vertices, their expected agreement set in $H[K]$ is a random variable $X$ with distribution $Bin(k-2,1/2)$. The probability that it is larger than $0.7h$ is bounded by the standard Chernoff bound (see appendix in [@AS-2004]): $$\Pr[X > 0.7h] \le \Pr[X > 0.7(k-2)] < e^{-2 \cdot 0.2^2(k-2)} < e^{-0.06h}$$ where we have used our assumption that $h \ge k \ge 0.8h$. Since agreement sets of distinct pairs among $\{\pi(x_1),\pi(y_1)\},\ldots,\{\pi(x_\ell),\pi(y_\ell)\}$ are pairwise independent, we have from the last inequality and from (\[e:l\]) that the probability that $\pi$ is an isomorphism between $B_M$ and $H[K]$ is at most $$e^{-0.06h\ell} < e^{-0.002h^2}\;.$$ Now, to complete the proof of the lemma, we need to show that [no matter]{} which set of at most $\lfloor 10^{-5}h^2 \rfloor$ pairs we take to play the role of $M$, and no matter which bijection $\pi$ we take, we still have that $\pi$ is not an isomorphism between $B_M$ (regardless of any additional changes to part pairs of $B$) and $H[K]$, with high probability. How many possible sets $M$ can we have? As we are selecting at most $\lfloor 10^{-5}h^2 \rfloor$ pairs from a pool of at most $\binom{h}{2} < h^2/2$ possible pairs, the number of possible $M$ is smaller than $$\label{e:num-m} \sum _{i=0}^{\lfloor 10^{-5}h^2 \rfloor} \binom{h^2/2}{i} < h^2 \binom{h^2/2}{\lfloor 10^{-5}h^2 \rfloor} < e^{0.00007h^2}$$ where we have used the standard Stirling approximation. It follows from the union bound and the fact that there are less than $h!$ possible $\pi$, that the probability that $B$ is blowup $10^{-5}h^2$-close to $H[K]$ is at most $$h! e^{0.00007h^2} \cdot e^{-0.002h^2} < e^{-0.001h^2}\;.$$ ------------------------------------------------------------------------ \[l:cond-5\] If $h$ is sufficiently large and $H \sim G(h,1/2)$, then the following holds. Let $K \subset [h]$ and $J \subset [h]$ with $\|K\|=\|J\|=\lceil 0.7h \rceil$ and let $\pi$ be a bijection from $J$ to $K$ having at least $0.1h$ non-stationary points. The probability that $\pi$ is $10^{-5}h^2$-close to an isomorphism between $H[J]$ and $H[K]$ is at most $e^{-0.12h^2}$. [[Proof.]{} ]{}Let $k=\lceil 0.7h \rceil$ be the order of $J$ and $K$. Let $A \subset J$ be a set of $s=\lceil 0.1h \rceil$ non-stationary points and let $B=\pi(A) \subset K$ be the set of their images. Then, $\pi$ restricted to $A$ is a bijection from $A$ to $B$ with no stationary points. By Lemma \[l:non-stationary\], we can pick $t=\lceil s/3 \rceil$ pairs $(a_1,\pi(a_1)),\ldots,(a_t,\pi(a_t))$ such that all $2t$ elements in these pairs are distinct. As in the proof of the previous lemma, let us fix a set $M$ of $m \le \lfloor 10^{-5}h^2 \rfloor$ pairs of vertices of $J$. Let $H[J]_M$ be the graph obtained from $H[J]$ after changing $M$ (making pairs in $M$ that are edges into non-edges and making pairs in $M$ that are non-edges into edges). Call a vertex of $J$ [problematic]{} if it appears in at least $h/30$ elements of $M$. Observe that being problematic has nothing to do with the structure of $H[J]$, as it only depends on the choice of $M$ and the choice of subset $J$. As in (\[e:num-prob\]), the number of problematic vertices is at most $0.0006h$. Now, remove from the set of pairs $(a_1,\pi(a_1)),\ldots,(a_t,\pi(a_t))$ all the pairs for which $a_i$ is problematic. The number of pairs that remain is at least $$t-0.0006h = \lceil s/3 \rceil - 0.0006h \ge 0.1h/3 - 0.0006h \ge 0.03h\;.$$ Without loss of generality, let these pairs be $(a_1,\pi(a_1)),\ldots,(a_\ell,\pi(a_\ell))$ where $\ell = \lceil 0.03h \rceil$. So, all of these $2\ell$ vertices are distinct, and all the $a_1,\ldots,a_\ell$ are non-problematic. Consider the set $S = \{a_1,\pi(a_1),\ldots,a_\ell,\pi(a_\ell)\}$. Notice that $\|J \setminus S\| \ge k - 2\ell$. Now, let $w \in J \setminus S$ be such that $(a_i,w) \notin M$. Since $a_i$ is non-problematic the number of choices for $w$ is at least $$\|J \setminus S\| - h/30 \ge k-2\ell - h/30 \ge 0.7h -2\lceil 0.03h \rceil - h/30 \ge 0.6h\;.$$ consider the two pairs $\{a_i,w\}$ and $\{\pi(a_i),\pi(w)\}$. In order for $\pi$ to be an isomorphism between $H[J]_M$ and $H[K]$ we must have that $\{a_i,w\}$ and $\{\pi(a_i),\pi(w)\}$ agree (both are edges or both are non-edges). Since agreement occurs with probability $\frac{1}{2}$ and since all the choices of $i$ and $w$ result in independent events, since they correspond to distinct pairs, we obtain that $$\label{e:pi-2} \Pr[\pi \text{ is an isomorphism between $H[J]_M$ and $H[K]$}] \le 2^{-\ell \cdot 0.6h} \le 2^{-0.18h^2}\;.$$ We note that a similar idea to the one in the last inequality has been used in the proof of Theorem 3.1 in [@KSV-2002]. To complete the proof of the lemma, we must again, as in the previous lemma, consider all possible choices for $M$. As in (\[e:num-m\]), the number of possible $M$ is smaller than $e^{0.00007h^2}$. It follows from the union bound that the probability that $\pi$ is $10^{-5}h^2$-close to an isomorphism between $H[J]$ and $H[K]$ is at most $$e^{0.00007h^2} \cdot 2^{-0.18h^2} < e^{-0.12h^2}\;.$$ ------------------------------------------------------------------------ Recall that in the proof of Lemma \[l:cond-5\] we have used the following lemma. \[l:non-stationary\] Let $f$ be a bijection between sets $A$ and $B$ of size $s$ and with no stationary points. Then there are at least $t=\lceil s/3 \rceil$ pairs $(a_1,f(a_1)),\ldots,(a_t,f(a_t))$ such that all $2t$ elements in these pairs are distinct. [[Proof.]{} ]{}A [chain]{} of $f$ is a maximal length sequence $(x_1,\ldots,x_\ell)$ such that $f(x_i)=x_{i+1}$ for $i=1,\ldots\ell-1$ and for which $x_\ell \notin A$ or $f(x_\ell) = x_1$ (in the latter case, the chain forms a [cycle]{}). Clearly, $A \cup B$ can be partitioned into chains. Furthermore, no chain is a singleton since there are no stationary points. So, from each chain $(x_1,\ldots,x_\ell)$ we can pick at least $\lfloor \ell/2 \rfloor$ pairs $(x_1,x_2=f(x_1))$, $(x_3,x_4=f(x_3))$ and so on, such that all the elements in the picked pairs are distinct. Thus, if the partition has $c$ odd length chains, we have picked $(\|A \cup B\|-c)/2$ pairs. Now, since $\|A \cup B\| \ge s$ and since $c \le \|A \cup B\|/3$ the lemma follows. ------------------------------------------------------------------------ [Proof of Theorem \[t:phe\].]{} Assume that $h$ is sufficiently large to satisfy all five lemmata \[l:cond-1\] until \[l:cond-5\]. Consider $H \sim G(h,1/2)$. The probability that it does not satisfy the first condition for ${\cal P}_h$ is at most $1/(5h)$ by Lemma \[l:cond-1\]. The probability that it does not satisfy the second condition for ${\cal P}_h$ is at most $1/(5h)$ by Lemma \[l:cond-2\]. The probability that it does not satisfy the third condition for ${\cal P}_h$ is at most $1/h$ by Lemma \[l:cond-3\]. There are less than $2^h$ options for $S \subset [h]$, at most $2^h$ options for $K \subseteq [h]$ and less than $h^h$ options for $B$ to be the blowup of $H[S]$. Hence, the probability that $H$ does not satisfy the fourth condition for ${\cal P}_h$ is, by Lemma \[l:cond-4\] at most $$2^h \cdot 2^h \cdot h^h \cdot e^{-0.001h^2} < \frac{1}{5h}$$ (where here we assume that $h$ is also sufficiently large to satisfy the last inequality). There are less than $2^h$ options for each of $J$ and $K$ in the fifth condition of ${\cal P}_h$, and less than $h!$ possible bijections $\pi$. Hence, the probability that $H$ does not satisfy the fifth condition for ${\cal P}_h$ is, by Lemma \[l:cond-5\] at most $$2^h \cdot 2^h \cdot h! \cdot e^{-0.12 h^2} < \frac{1}{5h}\;.$$ Hence the probability that $H \notin {\cal P}_h$ is at most $$4 \cdot \frac{1}{5h} + \frac{1}{h} < \frac{2}{h}\;.$$ ------------------------------------------------------------------------ Proof of the main result ======================== We will prove a stronger theorem which immediately implies Theorem \[t:main\] by just using $\epsilon = \frac{1}{2}$ in the following statement and by redefining ${\cal P}_h = \emptyset$ for $h \le h_0(\frac{1}{2})$. \[t:super\] For every $\epsilon > 0$ there exists $h_0=h_0(\epsilon)$ such that the following hold for all $h > h_0$ and for all $n \le 2^{h^{1-\epsilon}}$. 1. If $H \in {\cal P}_h$, then $i_H(n)=g(n,h)$. Furthermore, the family of extremal graphs is precisely $H^(n)$. 2. if $G$ has $n$ vertices and $i_H(G) \ge f(n,h)$ then $G$ must be obtained from a balanced blowup of $H$ by adding some edges inside the blowup parts. [[Proof.]{} ]{}The fact that $i_H(n) \ge g(n,h)$ was already shown in Section 2. Thus, we must prove that $i_H(n) \le g(n,h)$ and that the only graphs which attain this bound are the nested balanced blowups of $H$. Throughout the proof we shall assume that $h$ is sufficiently large as a function of $\epsilon$ to satisfy various inequalities, thereby establishing the existence of $h_0$. The proof will be by induction on $n$ where the base case $n \le h$ is trivial since $g(h,h)=1$ and since $g(n,h)=0$ for all $n < h$. Let $n \le 2^{h^{1-\epsilon}}$ and assume that we have already established the result for all values smaller than $n$. Let $G$ be a graph with $n$ vertices. Let ${\mathcal H}$ denote the set of all induced copies of $H$ in $G$. In other words, considering $G$ and $H$ as labeled graphs, each $F \in {\mathcal H}$ is identified with an injective function from $V(H)$ to $V(G)$ which preserves adjacencies and non-adjacencies (thus, we view $F$ as such a function). We must therefore prove that $\|{\mathcal H}\| \le g(n,h)$. Assume that $V(H)=[h]$. Since $H \in {\cal P}_H$, it has a distinguishing set $Q \subset [h]$ with $\|Q\|=q \le 3\log h$. Recall that $Q$ has the following property. For any two distinct vertices $u,v \in [h] \setminus Q$, there exists $w \in Q$ such that exactly one of $u$ or $v$ is adjacent to $w$. For the rest of this proof we fix $Q$. [\[Role consistent\]]{} Let $F_1,F_2 \in {\mathcal H}$ be two induced copies of $H$ in $G$. We say that $F_1$ and $F_2$ are [role-consistent]{} if for all $x \in Im(F_1) \cap Im(F_2)$ we have $F_1^{-1}(x) = F_2^{-1}(x)$. A subset ${\mathcal F}\subseteq {\mathcal H}$ is [role-consistent]{} if any pair of elements of ${\mathcal F}$ are role-consistent. A role-consistent subset ${\mathcal F}$ induces a partition $P({\mathcal F}) = \{R,P_1,\ldots,P_h\}$ of $V(G)$ as follows. For each $v \in V(G)$, if $v$ is in no element of ${\mathcal F}$ (namely, in no image of any $F \in {\mathcal F}$), then assign $v$ to $R$. Otherwise, let $v \in Im(F)$ for some $F \in {\mathcal F}$. Then, place $v$ in $P_j$ where $F^{-1}(v)=j$. Since ${\mathcal F}$ is role-consistent, this assignment is consistent regardless of the choice of $F \in {\mathcal F}$ for which $v \in Im(F)$. Stated otherwise, every copy of $H$ in ${\mathcal F}$ has the property that it contains precisely one vertex from each of $P_1,\ldots,P_h$ where the vertex of the copy that plays the role of $i$ belongs to $P_i$. [\[Role partition; redundant part\]]{} We call $P({\mathcal F})$ the [role partition]{} of ${\mathcal F}$ and call $R$ the [redundant part]{} of the role partition. It immediately follows that if ${\mathcal F}$ is role-consistent and $P({\mathcal F}) = \{R,P_1,\ldots,P_h\}$ is its role partition, then $$\label{e:fnh} \|{\mathcal F}\| \le \Pi_{i=1}^h \|P_i\| \le f(n-\|R\|,h) \le f(n,h)\;.$$ [\[$Q$-partition\]]{} Let ${\mathcal I}$ be the set of all injective functions from $Q$ to $V(G)$. For $I \in {\mathcal I}$ let ${\mathcal H}_I$ be the set of all elements $F \in {\mathcal H}$ which are consistent with $I$. Namely, $F(x)=I(x)$ for all $x \in Q$. Clearly $\{{\mathcal H}_I ~\|~ I \in {\mathcal I}\}$ is a partition of ${\mathcal H}$ which we call the [$Q$-partition]{}. Observe also that since $\|Q\|=q$ we have $\|{\mathcal I}\| < n^q$. Our high level approach is to prove that a delicate modification of the $Q$-partition has the property that most of its elements are small, and that the remaining (not so small) elements of this partition can be absorbed into one large part that is still role-consistent. We first need the following lemma. \[l:main-1\] Each element of the $Q$-partition is role-consistent. Namely, for each $I \in {\mathcal I}$, the set ${\mathcal H}_I$ is role-consistent. [[Proof.]{} ]{}Let $F_1,F_2 \in {\mathcal H}_I$. Let $x \in Im(F_1) \cap Im(F_2)$. Suppose $F_1^{-1}(x)=i$ and $F_2^{-1}(x)=j$. Now, if $i \in Q$ or $j \in Q$, then by the definition of ${\mathcal F}_I$ we must have $i=j$. Otherwise, if $i \notin Q$ and $j \notin Q$, then for every vertex $w \in Q$, either both $i$ and $j$ are adjacent to $w$ or both are non-adjacent. Thus, by the property of $Q$, $i$ and $j$ must be the same vertex of $H$, namely $i=j$. ------------------------------------------------------------------------ By Lemma \[l:main-1\], by (\[e:fnh\]) and by the fact that $\|{\mathcal I}\| < n^q$ we have the following crude upper bound. $$\label{e:crude} \|{\mathcal H}\| \le f(n,h)n^q \le f(n,h) n^{3 \log h}\;.$$ Notice that this bound applies to all $n \ge h$. Recall however, that the upper bound obtained in (\[e:crude\]) is still far from what we require as we would like to essentially eliminate the $n^{3 \log h}$ factor. Consider the following process performed on a role-consistent set ${\mathcal F}$ which we call [core production]{}. Start by defining $X_0=\cal F$. As long as $X_i$ has the property that there is a vertex $v$ of $G$ that appears in at most $f(n,h)n^{-3 -3 \log h}$ elements of $X_i$ ($X_i$ is a subset of ${\mathcal F}$ hence elements of $X_i$ are induced copies of $H$ in $G$), then remove from $X_i$ all elements containing $v$ and denote the remaining set by $X_{i+1}$. When the process ends at some stage $t$ we either have $X_t=\emptyset$ or else $X_t$ is a subset of ${\mathcal F}$ having the property that each vertex appearing in any element of $X_t$ appears in at least $f(n,h)n^{-3 -3 \log h}$ elements of $X_t$. Set $X_t = {\mathcal F}^$. [\[core; leftover; large\]]{} Let ${\mathcal F}$ be role-consistent. We call ${\mathcal F}^$ a [core]{} of ${\mathcal F}$ and call ${\mathcal F}^\# = {\mathcal F}\setminus {\mathcal F}^$ a [leftover]{} of ${\mathcal F}$. We say that a nonempty ${\mathcal F}$ is [large]{} if ${\mathcal F}^={\mathcal F}$. Notice that since ${\mathcal F}^$ is a subset of ${\mathcal F}$, it is also role-consistent. Also observe that $({\mathcal F}^)^={\mathcal F}^$ so every nonempty core is large. Finally, observe that if ${\mathcal F}$ is large, then every vertex of $G$ that appears in an element of ${\mathcal F}$ appears in at least $f(n,h)n^{-3 -3 \log h}$ other elements, so in particular $\|{\mathcal F}\| \ge f(n,h)n^{-3 -3 \log h}$. \[l:main-2\] Suppose we performed core production on each element of the $Q$-partition, thereby obtaining cores $\{{\mathcal H}_I^~\|~ I \in {\mathcal I}\}$ and leftovers $\{{\mathcal H}_I^\#~\|~ I \in {\mathcal I}\}$. Let ${\mathcal H}^\# = \cup_{I \in {\mathcal I}} {\mathcal H}_I^\#$. Then, $$\|{\mathcal H}^\#\| \le \frac{f(n,h)}{n^2}\;.$$ [[Proof.]{} ]{}In the core production of a role-consistent ${\mathcal F}$ we add at each step at most $f(n,h)n^{-3 -3 \log h}$ elements to the leftover set, where all the added elements of the step contain some vertex $v \in V(G)$. Since there are $n$ vertices, the final leftover of ${\mathcal F}$ satisfies $$\|{\mathcal F}^\#\| \le n \cdot f(n,h)n^{-3 -3 \log h} = f(n,h)n^{-2 -3 \log h}\;.$$ Since the number of elements $I \in {\mathcal I}$ is less than $n^q \le n^{3\log h}$, we have $$\|{\mathcal H}^\#\| \le \sum_{I \in {\mathcal I}} \|{\mathcal H}_I^\#\| \le n^{3\log h}f(n,h)n^{-2 -3 \log h}=\frac{f(n,h)}{n^2}\;.$$ ------------------------------------------------------------------------ Our main lemma (whose proof is deferred to Section \[sec:main-lemma\]) is the following. \[l:main-3\] Suppose that ${\mathcal F}$ and ${\mathcal E}$ are two large sets. Then their union is role-consistent. Once we establish Lemma \[l:main-3\] we can complete the proof of Theorem \[t:super\] as follows. Consider the set of cores $\{{\mathcal H}_I^~\|~ I \in {\mathcal I}\}$ and leftovers $\{{\mathcal H}_I^\#~\|~ I \in {\mathcal I}\}$. Recall that if ${\mathcal H}_I^ \neq \emptyset$, then ${\mathcal H}_I^$ is large. Hence, by Lemma \[l:main-3\] used repeatedly, $${\mathcal H}^ = \cup_{I \in {\cal I}} {\mathcal H}_I^$$ is role-consistent (observe that if ${\mathcal F}$ and ${\mathcal E}$ are large and their union is role-consistent then this union is large as well). There are now two cases to consider: [Case 1:]{} $\|{\mathcal H}^\| < f(n,h)$. Consider the role-partition $P({\mathcal H}^)=\{R,P_1,\ldots,P_h\}$. Now, $\|{\mathcal H}^\| < f(n,h)$ could happen for three reasons. (i) Either $R \neq \emptyset$. (ii) Else, the partition is not equitable. (iii) Else, the partition is equitable, but there is some pair of vertices $u,v$ in $G$ such that $u \in P_i$, $v \in P_j$, $i \neq j$, and the following occurs. either $uv \in E(G)$ but $ij \notin E(H)$ or $uv \notin E(G)$ but $ij \in E(H)$. Equivalently, no element of ${\mathcal H}^$ contains both $u$ and $v$. Suppose (i) occurs. In this case, $\|{\mathcal H}^\| \le f(n-1,h)$. But then by Lemma \[l:simple\] and Lemma \[l:main-2\] we obtain: $$\|{\mathcal H}\| \le \|{\mathcal H}^\| + \|{\mathcal H}^\#\| \le f(n-1,h)+\frac{f(n,h)}{n^2} = \left(1-\frac{1}{\lceil n/h \rceil}\right)f(n,h)+\frac{f(n,h)}{n^2} < f(n,h)\;.$$ Suppose (ii) occurs. In this case, $\|{\mathcal H}^\|$ is at most $\Pi_{i=1}^h\|P_i\|$ and the partition is not equitable, hence by Lemma \[l:simple\] and Lemma \[l:main-2\] we obtain: $$\|{\mathcal H}\| \le \|{\mathcal H}^\| + \|{\mathcal H}^\#\| \le f(n,h)\left(1-\frac{1}{\lceil n/h \rceil^2}\right)+\frac{f(n,h)}{n^2} < f(n,h)\;.$$ Suppose (iii) occurs. In this case, no element of ${\mathcal H}^$ contains both $u$ and $v$ and the partition is equitable. Thus, $$\|{\mathcal H}^\| \le f(n,h)\left(1-\frac{1}{\|P_i\|\|P_j\|}\right) \le f(n,h)\left(1-\frac{1}{\lceil n/h \rceil^2}\right)\;.$$ So, again by Lemma \[l:main-2\] we obtain: $$\|{\mathcal H}\| \le \|{\mathcal H}^\| + \|{\mathcal H}^\#\| \le f(n,h)\left(1-\frac{1}{\lceil n/h \rceil^2}\right) + \frac{f(n,h)}{n^2} < f(n,h)\;.$$ [Case 2:]{} $\|{\mathcal H}^\| = f(n,h)$. First observe that the structure of ${\mathcal H}^$ is unique. Clearly we must have $P({\mathcal H}^)=\{\emptyset,P_1,\ldots,P_h\}$ and the partition is equitable. Furthermore, for each choice of $h$ vertices, one from each part (i.e. a transversal), we obtain an element of ${\mathcal H}^$, as this is the only way to get $f(n,h)$ elements in ${\mathcal H}^$. Observe also that this means that the bipartite graph between $P_i$ and $P_j$ is either empty (if $ij \notin E(H)$) or complete (if $ij \in E(H)$). Notice that as a bonus we have now proved the second part of Theorem \[t:super\], since in the previous case where $\|{\mathcal H}^\| < f(n,h)$ we have already shown that $\|{\mathcal H}\| < f(n,h)$. We claim that every element of ${\mathcal H}^\#$ is entirely contained in some $P_i$. Assume otherwise, and let $X \in H^\#$ have its vertices in more than one part. Observe that it cannot have each vertex in a distinct part, as in this case we are repeating a copy of $H$ which is already in ${\mathcal H}^$, while ${\mathcal H}^ \cap H^\# = \emptyset$. So, $X$ has vertices in more than one part but not in all parts. Consider the smallest non-singleton part, say $P_i$. Thus, $\|P_i \cap V(X)\| \ge 2$. If $\|P_i \cap V(X)\| < 0.4h$, then any two vertices from $P_i \cap V(X)$ have all the vertices in $V(X) \setminus P_i$, namely more than $0.6h$ vertices, in their agreement set. This contradicts the assumption that $H \in {\cal P}_H$. If $\|P_i \cap V(x)\| > 0.6h$ then any vertex in $V(X) \setminus P$ has either all the vertices of $P_i \cap V(x)$ as its neighbors, or as its non-neighbors. But this contradicts the assumption that for $H \in {\cal P}_H$, the degree of any vertex is between $0.4h$ and $0.6h$. Finally, if $0.4h \le \|P_i \cap V(X)\| \le 0.6h$ then any three vertices from $P_i \cap V(X)$ have all the vertices in $V(X) \setminus P_i$, namely more than $0.4h$ vertices, in their agreement set, contradicting the assumption that $H \in {\cal P}_H$. Having proved that every element of ${\mathcal H}^\#$ is entirely contained in some $P_i$, we can now use the induction hypotheses to obtain that $$\|{\mathcal H}^\#\| \le \sum_{i=1}^h g(\|P_i\|,h).$$ Therefore, $$\|{\mathcal H}\| \le \|{\mathcal H}^\| + \|{\mathcal H}^\#\| \le f(n,h) + \sum_{i=1}^h g(\|P_i\|,h) = g(n,h)\;.$$ Finally, notice that the only way to obtain an equality is to have each $P_i$ induce a graph with the maximum number of induced copies of $H$, namely, by induction, with $g(\|P_i\|,h)$ induced copies of $H$. Thus, each $P_i$ induces an element of the nested balanced blowup on $\|P_i\|$ vertices. Hence, by the definition of $H^(n)$, $G$ is a nested balanced blowup of $H$ with $n$ vertices, namely $G \in H^(n)$. ------------------------------------------------------------------------ Properties of large sets {#sec:properties} ======================== In order to prove Lemma \[l:main-3\], we first need to establish some properties that hold for every large set. Let ${\mathcal F}$ be a large set and let $P({\mathcal F})=\{R,P_1,\ldots,P_h\}$ be the role partition of ${\mathcal F}$. Our first lemma quantifies the fact that in a large set, a small fraction of parts cannot contain too many elements. \[l:some-parts\] Let $J \subset [h]$ with $\|J\| = \lfloor 0.00001h \rfloor$. Then, $\sum_{j \in J} \|P_j\| \le 0.0004n$. [[Proof.]{} ]{}Consider the following process. Let $n \ge h$. Start with an equitable partition of $n$ elements into $h$ parts. The product of the sizes of the parts is thus $f(n,h)$. Let $0 < k < h$. Designate the largest $k$ parts of the equitable partition (so we prefer take parts of size $\lceil n/h \rceil$ to the designated parts as long as there are enough, and if there are less than $k$ such parts, then the remaining designated parts are of size $\lfloor n/h \rfloor$). Now, suppose $t \ge k \lceil n/h \rceil$. Note that the designated parts contain altogether at most $t$ elements. Move elements from the non-designated parts to the designated parts so that after each move, the designated parts form an equitable partition of $k$ parts and the non-designated parts from an equitable partition into $h-k$ parts. Stop moving after the designated parts contain precisely $t$ elements. The number of moves that we have made is therefore at least $$t- \lceil n/h \rceil \cdot k\;.$$ What happened to the product of the sizes of the parts after each move? Suppose we have moved an element from a non-designated part whose current size is $a$ to a designated part whose current size is $b$, hence $b \ge a$. So the ratio between the product of the part sizes after the move and the product before the move is precisely $$\frac{(a-1)(b+1)}{ab} = \left(1-\frac{b-a+1}{ab} \right) < 1-\frac{1}{a}+\frac{1}{b}\;.$$ Notice that at any point, we have $a \le \lceil n/h \rceil$. Also, after half of the moves are made, each element in a designated set contains at least $\lfloor t/2k \rfloor$ elements. Hence for the last half of the moves, we have $b \ge \lfloor t/2k \rfloor$. In these moves the aforementioned ratio is at most $$1-\frac{1}{\lceil n/h \rceil}+ \frac{1}{\lfloor t/2k \rfloor}\;.$$ So after all moves are completed, the product of the sizes of the parts is now at most $$\label{e:process} f(n,h) \left(1-\frac{1}{\lceil n/h \rceil}+ \frac{1}{\lfloor t/2k \rfloor}\right)^{t/2-\frac{\lceil n/h \rceil \cdot k}{2}}\;.$$ Let us now return to the statement of the lemma. We will consider the product of the sizes of the parts $\Pi_{i=1}^h \|P_i\|$. Let $k=\lfloor 0.00001h \rfloor$ and let $t > 0.0004n$. Suppose the lemma does not hold. Then we have $k$ parts that together contain $t$ elements. By convexity, the maximum of the product of the sizes of the parts subject to this condition is obtained when $R = \emptyset$ (no redundant part), and when the $k$ parts form an equitable partition of $t$ into $k$ parts, and the remaining $h-k$ parts form an equitable partition of $n-t$ into $h-k$ parts. By (\[e:process\]) we have that $$\begin{aligned} \Pi_{i=1}^h \|P_i\| & \le & f(n,h) \left(1-\frac{1}{\lceil n/h \rceil}+ \frac{1}{\lfloor t/2k \rfloor}\right)^{t/2-\frac{\lceil n/h \rceil \cdot k}{2}}\\ & \le & f(n,h) \left(1-\frac{1}{\lceil n/h \rceil}+ \frac{1}{\lfloor 5n/h \rfloor}\right)^{0.0002n-\frac{\lceil n/h \rceil \cdot \lfloor 0.00001h \rfloor}{2}}\\ & \le & f(n,h) \left(1-\frac{h}{4n}\right)^{0.0001n}\\ & \le & f(n,h) e^{-0.000025h}\\ & \le & f(n,h) 2^{-h^{1-\epsilon}4\log h}\\ & < & f(n,h) n^{-4 \log h}\;.\end{aligned}$$ Here we have used $t/(2k) \ge 5n/h$ and $\frac{1}{\lceil n/h \rceil}- \frac{1}{\lfloor 5n/h \rfloor} \ge h/(4n)$ while recalling also our assumptions that $h$ is sufficiently large as a function of $\epsilon$ and that $n \le 2^{h^{1-\epsilon}}$. Now, since $\|{\mathcal F}\| \le \Pi_{i=1}^h \|P_i\|$ we have reached a contradiction to the assumption that ${\mathcal F}$ is large, as large sets have at least $f(n,h) n^{-3-3 \log h}$ elements. ------------------------------------------------------------------------ Call a part $P_i$ [small]{} if it contains at most $\lfloor n/(10h) \rfloor$ vertices. \[l:small-parts\] The number of small parts is less than $\lfloor 0.0002h \rfloor$. [[Proof.]{} ]{}The proof is similar to the previous one, only now we wish to focus on moving elements away from a set of designated parts so the bounds we obtain on the product of the set sizes fits better to this scenario. Consider the following process. Let $n \ge h$. Start with an equitable partition of $n$ elements into $h$ parts. The product of the sizes of the parts is thus $f(n,h)$. Let $0 < k < h$. Designate the smallest $k$ parts of the equitable partition. Now, suppose $2t \le k \lfloor n/h \rfloor$. Note that the designated parts contain altogether at least $2t$ elements. Move elements from designated parts to the non-designated parts so that after each move, the designated parts form an equitable partition of $k$ parts and the non-designated parts from an equitable partition into $h-k$ parts. Stop moving after the designated parts contain precisely $t$ elements. What happened to the product of the sizes of the parts after each move? Suppose we have moved an element from a designated part whose current size is $a$ to a non-designated part whose current size is $b$, hence $b \ge a$. So the ratio between the product of the part sizes after the move and the product before the move is smaller than $1-\frac{1}{a}+\frac{1}{b}$. Notice that at any point, we have $b \ge \lfloor n/h \rfloor$. Consider the last $t$ moves. In each such move, the total number of elements in the designated sets prior to the move is at most $2t$. Hence, for each such move we have $a \le \lceil 2t/k \rceil$. So in these moves the aforementioned ratio is at most $$1-\frac{1}{\lceil 2t/k \rceil}+ \frac{1}{\lfloor n/h \rfloor}\;.$$ So after all moves are completed, the product of the sizes of the parts is now at most $$\label{e:process-2} f(n,h) \left(1-\frac{1}{\lceil 2t/k \rceil}+ \frac{1}{\lfloor n/h \rfloor}\right)^t\;.$$ Let us now return to the statement of the lemma. We will consider the product of the sizes of the parts $\Pi_{i=1}^h \|P_i\|$. First, observe that we may assume that $n \ge 10h$ as otherwise having a small part means having an empty part, meaning that ${\mathcal F}= \emptyset$ which is impossible as we assume that ${\mathcal F}$ is large. Let $k=\lfloor 0.0002h \rfloor$. Suppose the lemma does not hold. Then we have $k$ small parts that together contain $t$ elements. By convexity, the maximum of the product of the sizes of the parts subject to this condition is obtained when $R = \emptyset$ (no redundant part), when the $k$ parts form an equitable partition of $t$ into $k$ parts, the remaining $h-k$ parts form an equitable partition of $n-t$ into $h-k$ parts, and $t$ is maximized namely $t=k\lfloor n/10 h \rfloor$. By (\[e:process-2\]) we have that $$\begin{aligned} \Pi_{i=1}^h \|P_i\| & \le & f(n,h) \left(1-\frac{1}{\lceil 2 \lfloor n/10 h \rfloor \rceil}+ \frac{1}{\lfloor n/h \rfloor}\right)^{\lfloor 0.0002h \rfloor\lfloor n/10 h \rfloor}\\ & \le & f(n,h) \left(1-\frac{h}{n}\right)^{10^{-5}n}\\ & \le & f(n,h) e^{-10^{-5}h}\\ & \le & f(n,h) 2^{-h^{1-\epsilon}4\log h}\\ & < & f(n,h) n^{-4 \log h}\;.\end{aligned}$$ Here we have used $\frac{1}{\lceil 2 \lfloor n/10 h \rfloor \rceil}- \frac{1}{\lfloor n/h \rfloor} \ge h/n$ while recalling also our assumptions that $h$ is sufficiently large as a function of $\epsilon$ and that $n \le 2^{h^{1-\epsilon}}$. Now, since $\|{\mathcal F}\| \le \Pi_{i=1}^h \|P_i\|$ we have reached a contradiction to the assumption that ${\mathcal F}$ is large, as large sets have at least $f(n,h) n^{-3-3 \log h}$ elements. ------------------------------------------------------------------------ Let $u \in V(G) \setminus R$ and let ${\hat {\mathcal F}}_u$ be the set of elements of ${\mathcal F}$ that contain $u$. Let $B(u,{\mathcal F}) \subset V(G)$ be the set of vertices of $G$ that do not appear in any element of ${\hat {\mathcal F}}_u$. Clearly, $R \subseteq B(u,{\mathcal F})$ as vertices in the redundant part do not appear in any element of ${\mathcal F}$. Since ${\mathcal F}$ is large, we have that any vertex of $G$ which is not in $R$, appears in at least $f(n,h)n^{-3 - 3\log h}$ elements of ${\mathcal F}$. Thus, $\|{\hat {\mathcal F}}_u\| \ge f(n,h)n^{-3 - 3\log h}$. Let ${\mathcal F}_u$ be the set of elements of ${\mathcal F}$ which do not contain vertices from $B(u,{\mathcal F})$. Notice that the role partition of ${\mathcal F}_u$ is precisely $P({\mathcal F}_u)=\{R',P_1',\ldots,P_h'\}$ where $R' = B(u,{\mathcal F})$ and $P_i' = P_i \setminus B(u,{\mathcal F})$. Also, clearly, ${\hat {\mathcal F}}_u \subseteq {\mathcal F}_u$ but the inclusion may be proper as there may be elements in ${\mathcal F}$ that do not contain $u$ nor any element from $B(u,{\mathcal F})$. Given this inclusion, we have $$\|{\mathcal F}_u\| \ge \|{\hat {\mathcal F}}_u\| \ge f(n,h)n^{-3 - 3\log h}\;.$$ Our next lemma bounds the order of $B(u,{\mathcal F})$. \[l:num-bad\] $\|B(u,{\mathcal F})\| \le 10^{-5}n$. [[Proof.]{} ]{}Suppose $\|B(u,{\mathcal F})\|=k$. Since each element of ${\hat {\mathcal F}}_u$ contains no vertex of $B(u,{\mathcal F})$, we have that $\|{\hat {\mathcal F}}_u\| \le f(n-k,h)$. Using Lemma \[l:simple\] we have $$\|{\hat {\mathcal F}}_u\|\le f(n-k,h) \le f(n,h) \left(1-\frac{1}{\lceil n/h \rceil} \right)^k\;.$$ We must show that $k \le 10^{-5}n$. Assume otherwise, then $$\frac{\|{\hat {\mathcal F}}_u\|}{f(n,h)} \le \left(1-\frac{1}{\lceil n/h \rceil} \right)^k \le \left(1-\frac{h}{2n} \right)^{10^{-5}n} < e^{-10^{-6}h} < 2^{-h^{1-\epsilon}4\log h} < n^{-4 \log h}$$ contradicting the fact that $\|{\hat {\mathcal F}}_u\| \ge f(n,h)n^{-3 - 3\log h}$. ------------------------------------------------------------------------ Call a part $P'_i$ [very small]{} if it contains at most $n/(15h)$ vertices. \[l:very-small\] The number of very small parts is at most $0.0005h$. [[Proof.]{} ]{}By Lemma \[l:small-parts\], the number of small parts $P_i$ is less than $\lfloor 0.0002h \rfloor$. How many parts $P_i$ that were not small turned out to be very small parts in $P_i' = P_i \setminus B(u,F)$. For this to happen, $P_i$ must lose at least $n/(10h) - n/(15h) = n/(30h)$ of its vertices because they joined $B(u,F)$. But since by Lemma \[l:num-bad\] $\|B(u,{\mathcal F})\| \le 10^{-5}n$, we have that the number of such parts is at most $(n/100000)/(n/(30h)) = 0.0003h$. Hence, overall there are at most $0.0003h+\lfloor 0.0002h \rfloor \le 0.0005h$ very small parts. ------------------------------------------------------------------------ Let $u \in P_i'$ and $v \in P_j'$ with $i \neq j$. We say that the pair is [inconsistent]{} if either $ij \in E(H)$ and $uv \notin E(G)$ or $ij \notin E(H)$ and $uv \in E(G)$. Otherwise, the pair is [consistent]{}. Let $z > 0$ be a real parameter. We define the $z$-consistency graph $Z_{{\mathcal F}_u}(z)$ as follows. The set of vertices of $Z_{{\mathcal F}_u}(z)$ is $[h]$ (namely, it has the same set of vertices as $H$). Two vertices $i$ and $j$ are adjacent in $Z_{{\mathcal F}_u}(z)$ if and only if the number of consistent pairs with one endpoint in $P_i'$ and one endpoint in $P_j'$ is more than $(1-z)\|P'_i\|\|P'_j\|$. Our next lemma shows that for relatively small $z$, $Z_{{\mathcal F}_u}(z)$ has large cliques. \[l:zfz\] $Z_{{\mathcal F}_u}(h^{-\epsilon/2})$ has a clique of size at least $0.99999h+1$. We will actually prove that the complement of $Z_{{\mathcal F}_u}(h^{-\epsilon/2})$ does not have a matching of size at least $0.000005h-1$, as this implies that $Z_{{\mathcal F}_u}(h^{-\epsilon/2})$ has a clique of size at least $h-2\cdot 0.000005h+2=0.99999h+2$. Denote this complement by $C$. Suppose $C$ has a matching of size $k=\lceil 0.000005h \rceil-1$. Denote it by $(x_1,y_1), \ldots, (x_k,y_k)$. Any copy of $H$ in ${\mathcal F}_u$ contains precisely one vertex from each of the parts $P_1',\ldots,P_h'$. By the definition of $Z_{{\mathcal F}_u}(z)$, each such copy must avoid at least $z\|P'_{x_i}\|\|P'_{y_i}\|$ of the pairs of vertices from $P'_{x_i} \times P'_{y_i}$, as it must avoid inconsistent pairs and this holds for all $i=1,\ldots,k$. Hence, for $z=h^{-\epsilon/2}$ we have $$\begin{aligned} \|{\mathcal F}_u\| & \le & (1-z)^k \Pi_{i=1}^h \|P'_i\|\\ & \le & \left( 1-\frac{1}{h^{\epsilon/2}}\right)^{0.000005 h-1} \Pi_{i=1}^h \|P'_i\|\\ & \le & e^{-0.000004h^{1-\epsilon/2}} f(n,h) \\ & \le & f(n,h)2^{-h^{1-\epsilon}4\log h} \\ & \le & f(n,h) n^{-4\log h}\end{aligned}$$ contradicting the fact that $\|{\mathcal F}_u\| \ge f(n,h)n^{-3 - 3\log h}$. ------------------------------------------------------------------------ As we will be interested in the case $z=h^{-\epsilon/2}$, we will denote for simplicity $Z_{{\mathcal F}_u}=Z_{{\mathcal F}_u}(h^{-\epsilon/2})$. As Lemma \[l:zfz\] suggests, we have a clique of order $\lceil 0.99999h \rceil+1$ in $Z_{{\mathcal F}_u}$ so let $K^ \subset [h]$ denote such a clique. By Lemma \[l:very-small\], there are at most $0.0005h$ very small parts. Hence, there is $K(u,{\mathcal F}) \subseteq K^$ with $\|K(u,{\mathcal F})\| \ge (\lceil 0.99999h\rceil +1)- 0.0005h -1\ge 0.999h$ such that $K(u,{\mathcal F})$ is a clique in $Z_{{\mathcal F}_u}$ and for any $j \in K(u,{\mathcal F})$ we have that $\|P'_j\| \ge n/(15h)$. Furthermore, if $u \in P_\ell$ then $\ell \notin K(u,{\mathcal F})$. How many vertices of $G$ appear in parts $P_i'$ for which $i \in K(u,{\mathcal F})$ (namely, in the clique parts that are not very small, and not containing the part of $u$). We claim that these are most of the vertices. Denote this set by $W(u,{\mathcal F}) = \cup_{i \in K(u,{\mathcal F})} P'_i$. \[l:num-good\] $\|W(u,{\mathcal F})\| \ge 0.9995 n$. [[Proof.]{} ]{}A vertex $v \notin W(u,{\mathcal F})$ has to satisfy one of the following conditions. Either $v \in B(u,{\mathcal F})$, else $v \in P_i'$ but $i \notin K^$, else $v \in P_i'$, $i \in K^$ but $P'_i$ is very small, else $v$ is in the same part as $u$. Summing the sizes of these forbidden sets we obtain by Lemmas \[l:some-parts\], \[l:num-bad\], \[l:very-small\] and the fact that $\|[h]\setminus K^\| + 1 \le \lfloor 0.00001h \rfloor$ that it is at most $$\frac{n}{100000}+0.0004n+0.0005h\cdot (n/(15h)) \le 0.0005 n\;.$$ ------------------------------------------------------------------------ Let $C^(u,{\mathcal F})$ denote the set of inconsistent pairs $x,y$ with $x \in P_i'$, $y \in P_j'$ and both $i,j \in K(u,{\mathcal F})$. \[l:num-inconsistent\] $\|C^(u,{\mathcal F})\| \le n^2 h^{-\epsilon/2}$. [[Proof.]{} ]{}Clearly, by the definition of $K(u,{\mathcal F})$, the number of inconsistent pairs with one endpoint in $P_i'$ and the other in $P_j'$ is at most $\|P'_i\|\|P'_j\|h^{-\epsilon/2}$ when $i,j \in K(u,{\mathcal F})$ (in fact, this also holds in the larger set $K^$). Hence, $$\begin{aligned} \|C^(u,{\mathcal F})\| & \le & h^{-\epsilon/2} \sum_{i,j \in K(u,{\mathcal F}), i \neq j} \|P'_i\|\|P'_j\|\\ & \le & h^{-\epsilon/2} \sum_{i=1}^h \sum_{j=i+1}^h \|P_i\|\|P_j\|\\ & \le & h^{-\epsilon/2} h^2 \frac{n^2}{h^2} \le n^2 h^{-\epsilon/2}\;.\end{aligned}$$ ------------------------------------------------------------------------ Proof of Lemma \[l:main-3\] {#sec:main-lemma} =========================== We now return to the proof of Lemma \[l:main-3\]. Recall that we want to prove that if ${\mathcal F}$ and ${\mathcal E}$ are two large sets, then their union is role-consistent. Let $P({\mathcal F})=\{R,P_1,\ldots,P_h\}$ and $P({\mathcal E})=\{T,S_1,\ldots,S_h\}$ be the role partitions of ${\mathcal F}$ and ${\mathcal E}$. We must therefore prove that if $u \in P_i$ and $u \in S_j$, then $i=j$. Assume otherwise, then there exists some $u \in V(G)$ such that $u \in P_i$ and $u \in S_j$ and $i \neq j$. Using the aforementioned notations we focus on the objects $B(u,{\mathcal F})$, ${\mathcal F}_u$, $P({\mathcal F}_u)=\{R',P_1',\ldots,P_h'\}$, $Z_{{\mathcal F}_u}$, $K(u,{\mathcal F})$, $W(u,{\mathcal F})$, $C^(u,{\mathcal F})$ and correspondingly $B(u,{\mathcal E})$, ${\mathcal E}_u$, $P({\mathcal E}_u)=\{T',S_1',\ldots,S_h'\}$, $Z_{{\mathcal E}_u}$, $K(u,{\mathcal E})$, $W(u,{\mathcal E})$, $C^(u,{\mathcal E})$. Also recall that $R' = B(u,{\mathcal F})$ and $T'=B(u,{\mathcal E})$. Call an index $i \in K(u,{\mathcal E})$ [problematic for ${\mathcal F}$]{} if more than $\|S'_i\|/10$ of the vertices of $S'_i$ do not belong to $W(u,{\mathcal F})$. Notice that unlike the definitions in the previous section which only depend on a given large set and a vertex of its partition, the definition of being problematic involves two large sets. Informally, being non-problematic means that $90$ percent of the vertices of the part appear in “good sets” of the other partition. How many problematic parts are there? \[l:num-problematic\] The number of vertices of $K(u,{\mathcal E})$ that are problematic for ${\mathcal F}$ is at most $0.075h$. [[Proof.]{} ]{}Suppose $i \in K(u,{\mathcal E})$. Then, in particular $S'_i$ is not very small so it contains at least $n/(15h)$ vertices. If it were problematic for ${\mathcal F}$, then it would have at least $0.1 n/(15h) = n/(150h)$ vertices not in $W(u,{\mathcal F})$. But the number of vertices not in $W(u,{\mathcal F})$ is at most $0.0005n$ by Lemma \[l:num-good\]. So, the overall number of vertices of $K(u,{\mathcal E})$ that are problematic for ${\mathcal F}$ is at most $0.0005n/(n/(150h)) \le 0.075h$. ------------------------------------------------------------------------ Let $J$ denote the set of vertices of $K(u,{\mathcal E})$ that are non-problematic for ${\mathcal F}$. Recall that $\|K(u,{\mathcal E})\| \ge 0.999h$. Hence, by Lemma \[l:num-problematic\], $\|J\| \ge (0.999-0.075)h \ge 0.9h$. Let $X \subset V(G)$ be a set of $\|J\|$ vertices obtained by [randomly]{} selecting precisely one vertex from each part $S'_j$ for $j \in J$. All $\|J\|$ choices are performed independently. Thus, $G[X]$ is a subgraph of $G$ on $\|J\|$ vertices. We would like to first see how close $G[X]$ is to a subgraph of $H$ on $\|J\|$ vertices. \[l:far-x\] With probability larger than $3/4$, $G[X]$ is $4h^{2-\epsilon/2}$-close to $H[J]$. [[Proof.]{} ]{}Denote $X=\{x_j~\|~j \in J\}$. So $x_j$ is the vertex selected at random from $S'_j$. Consider a pair of distinct indices $i,j$ such that $i,j \in J$. What is the probability that $x_i$ and $x_j$ are consistent? First notice that since $i,j \in J$, then $\|S'_i\|,\|S'_j\| \ge n/(15h)$ as they are not very small. Next, notice that $ij$ is a clique edge in $Z_{{\mathcal E}_u}$, and hence the number of pairs between $S'_i$ and $S'_j$ that are consistent is at least $(1-h^{-\epsilon/2})\|S'_i\|\|S'_j\|$. So the probability that $x_i$ and $x_j$ are inconsistent is at most $h^{-\epsilon/2}$. Thus, the expected number of inconsistent pairs in $X$ is at most $$\binom{\|J\|}{2}h^{-\epsilon/2} < h^{2-\epsilon/2}\;.$$ By Markov’s inequality the probability of the number of inconsistent pairs being more than $4h^{2-\epsilon/2}$ is less than $1/4$. The proof is complete by noticing that changing the inconsistent pairs results in a graph that is isomorphic to $H[J]$ where $x_j$ plays the role of $j$ in $H[J]$. ------------------------------------------------------------------------ \[l:x\\] Let $X^ = X \cap W(u,{\mathcal F})$. Then with probability larger than $3/4$, $\|X^\| > 0.8h$. [[Proof.]{} ]{}Recall that $X=\{x_j~\|~j \in J\}$ where $x_j$ is the vertex selected at random from $S'_j$. Since $S'_j$ is non-problematic for ${\mathcal F}$, at least a fraction of $0.9$ of its vertices are in $W(u,{\mathcal F})$. Hence the probability that $x_j \in W(u,{\mathcal F})$ is at least $0.9$. The expected value of $\|X^\|$ is therefore at least $0.9\|J\| \ge 0.9 \cdot 0.9h = 0.81h$. As $\|X^\|$ is the sum of independent indicator random variables, the probability that it is smaller by a constant factor than its expectation is exponentially small in $h$. Hence, $\Pr[\|X^\| \le 0.8h]$ is exponentially small in $h$. ------------------------------------------------------------------------ We next need to bound the number of elements of $C^(u,{\mathcal F})$ with both endpoints in $X^$. \[l:num-inc-x\\] With probability larger than $3/4$, $\|C^(u,{\mathcal F}) \cap (X^* \times X^)\| \le 1000 h^{2-\epsilon/2}$. [[Proof.]{} ]{}We will actually upper bound the size of the potentially larger set $C^(u,{\mathcal F}) \cap (X \times X)$. Let $x,y$ be a pair of vertices in $C^(u,{\mathcal F})$. If $x \in B(u,{\mathcal E})$, then trivially $x \notin X$. Similarly, if $y \in B(u,{\mathcal E})$, then $y \notin X$. Hence, we can assume that $x \in S'_i$ and $y \in S'_j$. Now if $i=j$ then at least one of them is not selected to $X$, since from each part we selected at most one vertex to $X$. So we can assume $i \neq j$. If $i \notin J$ then $x \notin X$ since we selected no vertex of $S'_i$ to $X$. Likewise, if $j \notin J$ then $y \notin X$ since we selected no vertex of $S'_j$ to $X$. So we can assume that $i \neq j$, $i,j \in J$. But recall that $J \subseteq K(u,{\mathcal E})$, so $\|S_i'\| \ge n/(15h)$ and $\|S_j'\| \ge n/(15h)$ (i.e. they are not very small). Since the vertex from $S'_i$ was chosen at random and the vertex of $S'_j$ was chosen at random, the probability of the pair $x,y$ to be in $X \times X$ is at most $(15h/n)^2$. By Lemma \[l:num-inconsistent\], $\|C^(u,{\mathcal F})\| \le n^2 h^{-\epsilon/2}$. Hence the expected value of $\|C^(u,{\mathcal F}) \cap (X \times X)\|$ is at most $$n^2 h^{-\epsilon/2} \cdot \left(\frac{15h}{n}\right)^2 \le 225 h^{2-\epsilon/2}\;.$$ By Markov’s inequality the probability of being larger than four times the expectation is less than $1/4$. ------------------------------------------------------------------------ We now want to show that the vertices of $X^$ are not too concentrated among some parts of $P_1',\ldots,P_h'$, namely that no part $P_i'$ contains too many vertices of $X^$. \[l:non-concentration\] With probability larger than $3/4$, for all $i =1,\ldots,h$ we have $\|P_i' \cap X^\| \le h/100$. [[Proof.]{} ]{}We will actually prove the slightly stronger statement bounding $\|P_i' \cap X\|$. Consider some part $P'_i$ and the disjoint parts $P'_{i,j} = P'_i \cap S'_j$ for $j \in J$ (notice that these disjoint parts might not cover all the vertices of $P'_i$, but we don’t care about these uncovered vertices since they cannot be selected to $X$). Now, Since we selected one vertex from each $S'_j$, we have that $\|P'_{i,j} \cap X\| \in \{0,1\}$. It is therefore an indicator random variable with probability of success $$\Pr[\|P'_{i,j} \cap X\| = 1] = \frac{\|P'_{i,j}\|}{\|S'_j\|} \le \frac{\|P'_{i,j}\|}{n/(15h)}$$ where we have used the fact that $S'_j$ is not very small. It now follows that the expectation of $\|P_i' \cap X\|$ is at most $$\sum_{j \in J} \frac{\|P'_{i,j}\|}{n/(15h)} \le \frac{15h}{n}\|P_i'\| \le \frac{15h}{n} \cdot 0.0004n \le 0.006h$$ where we have used the fact that no single set $P_i$ (moreover $P_i' \subseteq P_i$) can contain more than $0.0004n$ vertices, by Lemma \[l:some-parts\]. But now observe that $\|P_i' \cap X\|$ is the sum of $\|J\|$ independent indicator random variables (recall, from each $S'_j$ for $j \in J$ we select a single vertex to $X$ independently), so by a Chernoff bound (see appendix in [@AS-2004]), the probability that $\|P_i' \cap X\|$ is larger than its expectation by a constant factor is exponentially small in $\|J\| \ge 0.9h$, so in particular, exponentially small in $h$. So, for $h$ sufficiently large, with probability smaller than $1/(4h)$, $\|P_i' \cap X\| > h/100$. Now, by the union bound, we have that with probability larger than $3/4$, for all $i=1,\ldots,h$ it holds that $\|P_i' \cap X\| \le h/100$. ------------------------------------------------------------------------ Since each of the four lemmas \[l:far-x\], \[l:x\\], \[l:num-inc-x\\] \[l:non-concentration\] states that the corresponding event happens with probability larger than $3/4$, we have that with positive probability, all four events hold. Thus, we can pick a set $X=\{x_j~\|~j \in J\}$ such that $X$ and its subset $X^* = X \cap W(u,{\mathcal F})$ satisfy the following four properties: 1. $G[X]$ is $4h^{2-\epsilon/2}$-close to $H[J]$. In particular, $G[X^]$ is $4h^{2-\epsilon/2}$-close to $H[J^]$ where $J^* \subset J$ satisfies $J^= \{j~\|~x_j \in X^\}$. 2. $\|X^\|=\|J^\| > 0.8h$. 3. $\|C^(u,{\mathcal F}) \cap (X^ \times X^)\| \le 1000 h^{2-\epsilon/2}$. 4. For all $i =1,\ldots,h$ we have $\|P_i' \cap X^\| \le h/100$. We now consider the way the vertices of $X^$ appear in $P_1',\ldots,P_h'$. Let $S \subset [h] = \{i~\|~X^ \cap P'_i \neq \emptyset\}$. So $i \in S$ means that it contains a least one vertex of $X^$ and $i \notin S$ means that it does not contain vertices of $X^$. There are two cases to consider: [Case 1:]{} $\|S\| \le 0.7h$. In this case we obtain by P1 on the one hand, that $G[X^]$ is $4h^{2-\epsilon/2}$-close to $H[J^]$. On the other hand, we obtain by P3 that there is a blowup of $H[S]$ that is blowup $1000 h^{2-\epsilon/2}$-close to $G[X^]$. Furthermore, $\|S\| \le 0.7h$ while $\|X^\| \ge 0.8h$ by P2 and no blowup part contains more than $h/100$ vertices by P4. So, this blowup is blowup $(4h^{2-\epsilon/2}+1000 h^{2-\epsilon/2})$-close to $H[J^]$. Since $4h^{2-\epsilon/2}+1000 h^{2-\epsilon/2} < 10^{-5}h^2$ this contradicts the fourth condition in the definition of ${\cal P}_h$, thus our assumption that $H \in {\cal P}_h$. [Case 2:]{} $\|S\| \ge 0.7h$. Recall that we assume that $u \in P_i$ and $u \in S_j$ where $i \neq j$. So we fix the two indices $i$ and $j$. Let $S^ \subseteq S \cup \{i\}$ have precisely $\lceil 0.7h \rceil$ elements where we force $i \in S^$. Now consider the following set of vertices $X^{}$ of $G$. For each $t \in S^$, if $t \neq i$ then add to $X^{*}$ some vertex of $X^$ which belongs to $P_t'$ (this is always possible since $P'_t \cap X^* \neq \emptyset$ by the definition of $S$). If $t=i$, then add $u$ to $X^{}$. Observe that $\|X^{}\|=\|S^\|=\lceil 0.7h \rceil$ and that $u$ is always in $X^{}$. We will prove that $G[X^{}]$ is very close to $H[S^]$ and that it is also very close to $H[Y^]$ where $Y^ = \{ p ~\|~ x_p \in X^{*}\} \cup \{j\}$. Observe also that $\|Y^\|=\lceil 0.7h \rceil$. Consider the natural bijection $\pi$ from $S^$ to $X^{}$ defined as follows. Let $t \in S^$. If $t \neq i$ let $\pi(t)$ be the vertex of $X^{*}$ which belongs to $P_t'$. For $t=i$ define $\pi(i)=u$. Why is $\pi$ close to an isomorphism between $H[S^]$ and $G[X^{*}]$? Every inconsistent pair has to be in $C^(u,{\mathcal F})$ and also has to be in $(X^* \times X^)$, unless that pair involves $u$. But by the definition of ${\mathcal F}_u$, all the vertices in the parts $P_1',\ldots,P_h'$ are consistent with $u$. Hence, by P3, $\pi$ is $1000 h^{2-\epsilon/2}$-close to an isomorphism between $H[S^]$ and $G[X^{}]$. Consider the natural bijection $\sigma$ from $X^{}$ to $Y^$ defined as follows. Suppose $x_p \in X^{}$ then $\sigma(x_p)=p$. Also define $\sigma(u)=j$. Why is $\pi$ close to an isomorphism between $G[X^{}]$ and $H[Y^]$? Every inconsistent pair that does not involve $u$, has to be in $G[X]$. Also, $Y^* \setminus \{j\} \subset J^$. But by the definition of ${\mathcal E}_u$, all the vertices in the parts $S_1',\ldots,S_h'$ are consistent with $u$. Hence, by P1, $\sigma$ is $4h^{2-\epsilon/2}$-close to an isomorphism between $G[X^{}]$ and $H[Y^]$. It follows that the composition $\pi\sigma$ is $4h^{2-\epsilon/2}+1000 h^{2-\epsilon/2}$ close to an isomorphism between $H[S^]$ and $H[Y^]$. Since $4h^{2-\epsilon/2}+1000 h^{2-\epsilon/2} < 10^{-5}h^2$, we must have by the fifth condition in the definition of ${\cal P}_h$, that $\pi\sigma$ has less than $0.1h$ non-stationary points. There now two options to consider. If $\|S^* \; \triangle \; Y^\| \ge 0.2 h$, then trivially $\pi\sigma$ has at least $\|S^ \triangle Y^\|/2 = 0.1h$ non-stationary points, a contradiction. If $\|S^ \; \triangle \; Y^\| < 0.2 h$, then $\|S^ \cap Y^\| \ge 0.6h$. But since $\pi(i)=u$ and $\sigma(u)=j$ this means that in $H$, the distinct vertices $i$ and $j$ have more than $0.6h-2 > 0.55h$ vertices in their agreement set, contradicting the assumption that $H \in {\cal P}_h$. ------------------------------------------------------------------------ Inducibility ============ In this section we prove Theorem \[t:inducibility\]. Let $H \in {\cal P}_h$ from Theorem \[t:main\]. By Theorem \[t:main\] it holds for $n \le 2^{\sqrt{h}}$ that $i_H(n)=g(n,h)$. It will be profitable to use the largest $k$ such that $n=h^k$ and still $n \le 2^{\sqrt{h}}$. We can therefore assume that $k \ge h^{1/3}+2$ as $$h^{\lceil h^{1/3}+2 \rceil} \le 2^{\sqrt{h}}\;.$$ Now, by (\[e:ghk\]) we have that $$i_H(h^k) = g(h^k,h) = \frac{h^{h(k-1)}(1- h^{k(1-h)})}{1-h^{1-h}}\;.$$ Since $g(h^k,h)/\binom{h^k}{h}$ serves as a trivial upper bound for $i_H$ we obtain that $$i_H \le \frac{h^{h(k-1)}(1- h^{k(1-h)})}{(1-h^{1-h})\binom{h^k}{h}}\;.$$ Clearly if we would have let $k \rightarrow \infty$ we would have obtained $i_H = h!/(h^h-h)$. However we can only assume that $k$ is bounded from below $h^{1/3}+2$. Nevertheless, we have that $$\frac{h^{h(k-1)}(1- h^{k(1-h)})}{(1-h^{1-h})\binom{h^k}{h}} \le \frac{h! h^{h(k-1)}(1- h^{k(1-h)})}{(1-h^{1-h})(h^k-h)^h} \le \frac{h!}{h^h-h} \cdot \frac{1}{(1-h^{1-k})^h}\;.$$ But notice that $$\frac{1}{(1-h^{1-k})^h} \le e^{\frac{2}{h^{k-2}}} \le 1 + \frac{4}{h^{k-2}} \le 1+\frac{4}{h^{h^{1/3}}}\;.$$ It follows that $$i_H \le \frac{h!}{h^h-h} \cdot \left( 1+\frac{4}{h^{h^{1/3}}} \right)\;.$$ ------------------------------------------------------------------------ Concluding remarks and open problems ==================================== We conjecture that the statement of Theorem \[t:super\] and hence of Theorem \[t:main\] can be extended to all $n$. Moreover: \[c:1\] For all strongly asymmetric graphs $H$ it holds that $i_H(n)=g(n,h)$ for all $n \in \mathbb{N}$. In particular, $$i_H= \frac{h!}{h^h-h}\;.$$ It would be extremely interesting to determine all graphs for which $i_H = \frac{h!}{h^h-h}$. As mentioned in the introduction, all cycles of length at least $5$ are conjectured to be in this family. Likewise, determining the set of graphs for which $i_H(n)=g(n,h)$ for, say, all $n \le h^3$, would also be gratifying as such graphs have $i_H= (1+o_h(1))\frac{h!}{h^h-h}$, as can be seen from the proof of Theorem \[t:inducibility\]. Given that this is true (for a much larger range of $n$) for all strongly asymmetric graphs, one might stipulate that Theorem \[t:main\] (or maybe even Conjecture \[c:1\]) holds for all asymmetric graphs. In fact, as the next proposition shows, this is very far from true. For all $h$ sufficiently large, there are asymmetric graphs $H$ on $h$ vertices for which $i_H(n) > g(n,h)$ for $n < h \log h$. [[Proof.*]{} ]{}Recall that a random graph from $G(q,1/2)$ is almost surely asymmetric [@ER-1963]. Furthermore, almost surely a random graph from $G(q,1/2)$ has all its degrees at most $0.55q$ and the number of common neighbors of every two vertices is at least $0.2q$ (see the proofs of Lemma \[l:cond-1\] and Lemma \[l:cond-2\]) . Hence, for $q$ sufficiently large, we can find (much more than) $2^q$ pairwise non-isomorphic graphs $H_1,\ldots,H_{2^q}$ such that each $H_i$ is an asymmetric graph on $q$ vertices, its maximum degree is at most $0.55q$, and the number of common neighbors in $H_i$ of every two vertices is at least $0.2q$. Now let $H$ be the graph on $h=q2^q$ vertices obtained by taking a balanced blowup of the clique $K_{2^q}$ where each blowup part is of size $q$. Suppose the parts are $A_1,\ldots,A_{2^q}$. Now make each $A_i$ induce a copy of $H_i$. We claim that $H$ is asymmetric. Indeed, consider some automorphism $\pi$. We first notice that $\pi(A_i) = A_j$ for some $j$ (possibly $i=j$). Suppose not, then there are $u,v \in A_i$ such that $\pi(u) \in A_j$ and $\pi(v) \in A_k$ for $j \neq k$. Since $\pi$ is an automorphism, the number of common neighbors of $u$ and $v$ must equal the number of common neighbors of $\pi(u)$ and $\pi(v)$. But the number of common neighbors of $u$ and $v$ in $H$ is at least $h-q+z \ge h-q+0.2q$ where $z$ is the number of common neighbors of $u$ and $v$ in $H_i$. On the other hand, the number of common neighbors of $\pi(u)$ and $\pi(v)$ in $H$ is $h-2q+x+y \le h-2q+0.55q+0.55q=h-q+0.1q$ where $x$ is the degree of $\pi(v)$ in $H_k$ and $y$ is the degree of $\pi(u)$ in $H_j$. This contradicts that $\pi$ is an automorphism. Hence, we must have $\pi(A_i)=A_j$ for some $j$. But now we claim that we must have $\pi(A_i)=A_i$. Indeed, if $\pi(A_i)=A_j$, then since $\pi$ is an automorphism, this means that $H_i$ and $H_j$ are isomorphic. Since $H_i$ is not isomorphic to $H_j$ whenever $i \neq j$, this implies that $\pi(A_i)=A_i$ for all $i=1,\ldots, 2^q$. But if $\pi(A_i)=A_i$ this means that $\pi$ restricted to $A_i$ is an automorphism of $H_i$. Since $H_i$ is asymmetric, $\pi$ must be trivial on $A_i$. As this holds for all $i=1,\ldots, 2^q$, we must have that $\pi$ is trivial, proving that $H$ is asymmetric. Designate a set $X \subset V(H)$ order $2^q$ obtained by taking one vertex from each $A_i$ for $i=1,\ldots,2^q$. Notice in fact that $H[X]$ is a clique of order $2^q$. Now consider the graph $G$ on $n=hq$ vertices which is the balanced blowup of $H$. So the parts of this blowup are $\{P_x~\|~x \in V(H)\}$ and $\|P_x\|=q$. Now, replace each independent set $P_x$ with a copy of $H_i$ where $x \in A_i$ (recall that the $A_i$ are a partition of $H$ and each induce a copy of $H_i$ in $H$). Now consider the subgraph of $G$ induced by the parts $\{P_x~\|~x \in X\}$. Notice that this subgraph is isomorphic to $H$. On the other hand, it is not an induced copy of $H$ obtained by taking one vertex from each part of $G$, so it is an additional induced copy to the $f(n,h)=g(n,h)=q^h$ induced copies that can be obtained as such. Hence, $i_H(G) > g(n,h)$ while $n=hq < h \log h$. ------------------------------------------------------------------------ Theorem \[t:super\], and therefore Theorem \[t:main\] have variants which apply to other graph densities and some other combinatorial structures. Indeed, suppose $0 < p < 1$ is given. One can adjust the definition of ${\cal P}_h$ to obtain a graph property ${\cal P}_h(p)$ which contains almost all graphs of $G(h,p)$. Indeed, Condition 1 in the definition could be changed to requiring that the degree be between $p(1-\delta)n$ and $p(1+\delta)n$ for an absolute constant $\delta$, that the agreement sets in Condition 2 would be at most $(1+\delta)(p^2+(1-p)^2)$ and $(1+\delta)(p^3+(1-p)^3)$ respectively, that the distinguishing set in Condition 3 be of size at most $-3\log (p^2+(1-p)^2) \log h$, and that the constant $10^{-5}h^2$ in Conditions 4,5 be replaced with $c\min \{p,1-p\}h^2$ for some absolute constant $c$. The proof of Theorem \[t:super\] and its lemmas stay essentially the same, after adjusting constants everywhere. Similarly, a variant of the theorems applies to the family of tournaments. We can naturally define a tournament to be strongly asymmetric by modifying the definition of ${\cal P}_h$ to suit tournaments (namely, degree requirements are replaced with in-degree and out-degree requirements, changing an edge means flipping its direction, etc.). Likewise, instead of considering $G(h,1/2)$ we consider the probability space of all tournaments on $h$ vertices. Acknowledgment ============== Independently of our work, Jacob Fox, Hao Huang, and Choongbum Lee (private communication) have sent the author a manuscript proving that almost all graphs $H$ have $i_H(n)=g(n,h)$ for all $n \in \mathbb{N}$. Once their paper is publicly available, a reference will be given. [^1]: Department of Mathematics, University of Haifa, Haifa 31905, Israel. Email: [email protected]. This research was supported in part by the Israel Science Foundation (grant No. 1082/16). [^2]: Throughout this paper all logarithms are in base $2$.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We study the connectedness of the planar self-affine sets $T(A,{\mathcal{D}})$ generated by an integer expanding matrix $A$ with $\|\det(A)\|=3$ and a non-collinear digit set ${\mathcal D}=\{0, v, kAv\}$ where $k\in {\mathbb Z}\setminus\{0\}$ and $v\in {\mathbb Z}^2$ such that $\{v, Av\}$ is linearly independent. By checking the characteristic polynomials of $A$ case by case, we obtain a criterion concerning only $k$ to determine the connectedness of $T(A,{\mathcal{D}})$. address: - ' Department of Mathematics and Information Technology, The Hong Kong Institute of Education, Hong Kong' - 'Department of Mathematics, Shantou Uinversity, Shantou 515063, China' author: - 'King-Shun Leung' - Jun Jason Luo title: 'Connectedness of planar self-affine sets associated with non-collinear digit sets' --- [^1] [Introduction]{} Let $M_n({\mathbb{Z}})$ denote the set of $n\times n$ matrices with integer entries, let $A\in M_n({\mathbb{Z}})$ be expanding, i.e., all eigenvalues of $A$ have moduli strictly larger than $1$. Assume $\|\det A\|=q$, and a finite set ${\mathcal{D}}=\{d_1,\dots,d_q\}\subset {\mathbb{R}}^n$ with cardinality $q$, we call it a q-digit set. It is well known that there exists a unique self-affine set $T{\mathrel{\mathop:\!\!=}}T(A,{\mathcal{D}})$ [@LaWa] satisfying: $$T= A^{-1}(T + {\mathcal{D}})=\left\{\sum_{i=1}^{\infty}A^{-i}d_{j_i}: d_{j_i}\in {\mathcal{D}}\right\}.$$ $T$ is called a self-affine tile if such $T$ has a nonvoid interior. The topological structure of $T(A,{\mathcal{D}})$, especially its connectedness, has attracted a lot of attentions in fractal geometry and number theory. It was asked by Gröchenig and Haas [@GrHa] that given an expanding integer matrix $A\in M_n({\mathbb{Z}})$, whether there exists a digit set ${\mathcal{D}}$ such that $T(A,{\mathcal{D}})$ is a connected tile and they partially solved the question in ${\mathbb R}^2$. Hacon et al. [@HaSaVe] proved that any self-affine tile $T(A,{\mathcal{D}})$ with a $2$-digit set is always pathwise connected. Lau and his collaborators ([@HeKiLa], [@KiLa], [@KiLaRa], [@LeLa]) systematically studied connectedness of the self-affine tiles generated by a kind of special digit sets of the form $\{0,1,\dots,q-1\}v$ where $v\in {\mathbb Z}^n\setminus\{0\}$, which are called [consecutive collinear digit sets]{}. They observed a height reducing property (HRP) of the characteristic polynomial of $A$ to determine the connectedness of $T(A,{\mathcal{D}})$, and conjectured that all monic expanding polynomials have the HRP, thus all the tiles generated by consecutive collinear digit sets are connected. Akiyama and Gjini [@AkGj] solved it up to degree $4$. However it is still open for arbitrary degree. In the plane, the disk-likeness (i.e. homeomorphic to a closed unit disk) is an interesting topic, Bandt and Gelbrich [@BaGe], Bandt and Wang [@BaWa], and Leung and Lau [@LeLa] investigated the disk-likeness of self-affine tiles in terms of the neighborhoods of $T$. Deng and Lau [@DeLa], and Kirat [@Ki] concerned themselves about a class of planar self-affine tiles generated by product digit sets. With regard to other types of digit sets, there are few results about the connectedness of $T(A,{\mathcal{D}})$ generated by non-consecutive or non-collinear digit sets. In [@LeLu], by counting the neighborhoods of $T$, the authors made a first attempt to exploit the case of non-consecutive collinear digit sets with $\|\det(A)\|=3$, and obtained a complete characterization for $T(A,{\mathcal{D}})$ to be connected or not. As a subsequent research, in the present paper, we will focus on the non-collinear digit sets. Precisely, we discuss the self-affine sets $T(A,{\mathcal{D}})$ generated by an expanding matrix $A$ with $\|\det(A)\|=3$ and a non-collinear digit set ${\mathcal D}=\{0, v, kAv\}$ for $k\in {\mathbb Z}\setminus\{0\}$ and $v\in {\mathbb R}^2$ such that $\{v, Av\}$ is linearly independent. By checking the characteristic polynomials of $A$ case by case, we obtain a criterion concerning only $k$ to determine the connectedness of $T(A,{\mathcal{D}})$. Let $A$ be a $2\times 2$ integral expanding matrix with $\|\det(A)\|=3$, and let ${\mathcal{D}}=\{0,v, kAv\}$ be a digit set where $k\in {\mathbb Z}\setminus\{0\}$ such that $\{v, Av\}$ is linearly independent. Then the self-affine set $T(A, {\mathcal{D}})$ is connected if and only if $k=\pm 1$. [Preliminaries]{} In this section, we give some preparatory results of self-affine sets which will be used frequently in the paper. Define $${\mathcal{E}}=\{(d_i,d_j):\ (T+d_i)\cap(T+d_j)\ne\emptyset,\ d_i,\ d_j\in {\mathcal{D}}\}.$$ We say that $d_i$ and $d_j$ are [$\mathcal{E}$-connected]{} if there exists a finite sequence $\{d_{j_1},\dots,d_{j_k}\}\subset {\mathcal{D}}$ such that $d_i=d_{j_1},d_j=d_{j_k}$ and $(d_{j_l},d_{j_{l+1}})\in {\mathcal{E}}, 1\leq l \leq k-1.$ It is easy to check that $(d_i,d_j)\in {\mathcal{E}}$ if and only if $$d_i-d_j=\sum_{k=1}^{\infty}A^{-k}v_k \quad \text{where}\quad v_k\in \Delta\mathcal{D}{\mathrel{\mathop:\!\!=}}\mathcal{D}-\mathcal{D}.$$ Then we get the following criterion of connectedness of a self-affine set. ([@Ha], [@KiLa]) \[e-connected prop\] A self-affine set $T$ with a digit set $\mathcal{D}$ is connected if and only if any two $d_i, d_j\in {\mathcal{D}}$ are $\mathcal{E}$-connected. In the following, we mainly consider the planar self-affine set $T(A,{\mathcal{D}})$ generated by a $2\times 2$ integral expanding matrix $A$ with $\|\det A\|=3$ and a digit set ${\mathcal{D}}=\{0,v, kAv\}$ such that $\{v, Av\}$ is linearly independent, where $k\in {\mathbb Z}\setminus\{0\}$. Denote the characteristic polynomial of $A$ by $f(x)=x^2+px+q$. Define $\alpha_i,\beta_i$ by $$A^{-i}v=\alpha_iv+\beta_iAv, \quad i=1,2,\dots.$$ Applying the Hamilton-Cayley theorem $f(A)=A^2+pA+qI=0$, it follows a lemma. ([@LeLa]) \[evaluation\] Let $\alpha_i,\beta_i$ be defined as the above. Then $q\alpha_{i+2}+p\alpha_{i+1}+\alpha_i=0$ and $q\beta_{i+2}+p\beta_{i+1}+\beta_i=0$. Especially, $\alpha_1=-p/q,\ \alpha_2=(p^2-q)/q^2;\ \beta_1=-1/q,\ \beta_2=p/q^2$. Moreover for $\Delta=p^2-4q\ne 0$, we have $$\alpha_i=\frac{q(y_1^{i+1}-y_2^{i+1})}{\Delta^{1/2}} \quad\text{and}\quad \beta_i=\frac{-(y_1^i-y_2^i)}{\Delta^{1/2}}$$ where $y_1=\frac{-p+\Delta^{1/2}}{2q},\ y_2=\frac{-p-\Delta^{1/2}}{2q}$ are the two roots of $qx^2+px+1=0$. Let $$\tilde{\alpha}{\mathrel{\mathop:\!\!=}}\sum_{i=1}^{\infty}\|\alpha_i\|,\quad \tilde{\beta}{\mathrel{\mathop:\!\!=}}\sum_{i=1}^{\infty}\|\beta_i\|.$$ \[cor.ref\] Assume $f(x)=x^2+px+q$ and $g(x)=x^2-px+q$ be the characteristic polynomials of expanding matrices $A$ and $B$, respectively. Let $\alpha_i, \beta_i, \tilde{\alpha}, \tilde{\beta}$ for $f(x)$ be as before; let $\alpha_i', \beta_i', \tilde{\alpha'}, \tilde{\beta'}$ be the corresponding terms for $g(x)$. Then $$\alpha_{2j}'=\alpha_{2j}, \ \alpha_{2j-1}'=-\alpha_{2j-1}, \ \beta_{2j}'=-\beta_{2j}, \ \beta_{2j-1}'=\beta_{2j-1},$$ and hence $\tilde{\alpha}=\tilde{\alpha'}, \ \tilde{\beta}=\tilde{\beta'}$. When $\|\det A\| =3$, it is known by [@BaGe] that there are $10$ eligible characteristic polynomials of $A$: $$x^2\pm 3;\quad x^2\pm x+ 3;\quad x^2\pm 2x + 3;\quad x^2\pm 3x + 3;\quad x^2\pm x - 3.$$ Following [@LeLu], together with Corollary \[cor.ref\], we obtained the estimates or values of the corresponding $\tilde{\alpha}$ and $ \tilde{\beta}$ as follows: $$\begin{aligned} && f(x)=x^2 \pm x+3: \quad \tilde{\alpha}< 0.88,\ \tilde{\beta}<0.63;\label{estimate1} \\ && f(x)=x^2 \pm 2x+3: \quad \tilde{\alpha}< 1.17,\ \tilde{\beta}<0.73;\label{estimate2}\\ && f(x)=x^2 \pm 3x+3: \quad \tilde{\alpha}< 2.24,\ \tilde{\beta}<1.08;\label{estimate3}\\ && f(x)=x^2 \pm x-3: \quad \tilde{\alpha}=2,\ \tilde{\beta}=1.\label{estimate4}\end{aligned}$$ [Main results]{} For a digit set ${\mathcal{D}}=\{0,v, kAv\}$, we denote by $\Delta{\mathcal D}=\{0,\pm v, \pm(kAv-v), \pm kAv\}$ the difference set. First we show the following simplest case according to the characteristic polynomials of $A$. Let $A$ be a $2\times 2$ integral expanding matrix with characteristic polynomial $f(x)= x^2 \pm 3$ and ${\mathcal{D}}=\{0,v, kAv\}$ be a digit set where $k\in {\mathbb Z}\setminus\{0\}$ such that $\{v, Av\}$ is linearly independent. Then the self-affine set $T(A, {\mathcal{D}})$ is connected if and only if $k=\pm 1$. Since the case of $f(x)=x^2-3$ is more or less the same as that of $f(x)=x^2+3$, it suffices to show the last one. If $k=1$, then $\Delta{\mathcal D}=\{0,\pm v, \pm(Av-v), \pm Av\}$. From $f(A)=A^2+3I=0$, we have $$\label{eq3.00} I=-2A^{-2}(I+A^{-2})^{-1}=2\sum_{n=1}^{\infty}(-1)^n A^{-2n}$$ and $$\label{eq3.01} v=2\sum_{n=1}^{\infty}(-1)^n A^{-2n}v=\sum_{n=0}^{\infty}A^{-4n}\big(A^{-2}(-v)+A^{-3}(-Av)+A^{-4}v+A^{-5}(Av)\big).$$ Hence $v\in T-T$, or equivalently $T\cap (T+v)\ne\emptyset$. Moreover, $$\label{eq3.02} Av=\sum_{n=0}^{\infty}A^{-4n}\big(A^{-1}(-v)+A^{-2}(-Av)+A^{-3}v+A^{-4}(Av)\big)$$ which implies $T\cap(T+Av)\ne\emptyset$. Consequently, by Proposition \[e-connected prop\], $T$ is connected (see Figure \[fig1\](a)). If $k=-1$, then $\Delta{\mathcal D}=\{0,\pm v, \pm(Av+v), \pm Av\}$, and (\[eq3.00\]), (\[eq3.01\]), (\[eq3.02\]) still hold. Hence $T$ is also connected. If $\|k\|>1$, let $k_i Av+l_iv\in \Delta{\mathcal D}$ for $i\geq 1$, then a point of $T-T$ can be written as $$\begin{aligned} &&\sum_{i=1}^{\infty}A^{-i}(k_iAv+l_iv)=\sum_{i=1}^{\infty}A^{-2i}(k_{2i}Av+l_{2i}v)+ \sum_{i=1}^{\infty}A^{-2i+1}(k_{2i-1}Av+l_{2i-1}v)\\ &=& \sum_{i=1}^{\infty}(-\frac{1}{3})^{i}(k_{2i}Av+l_{2i}v)+ \sum_{i=1}^{\infty}(-\frac{1}{3})^{i}(-3k_{2i-1}v+l_{2i-1}Av)\\ &=& \left(k_1+\sum_{i=1}^{\infty}(-\frac{1}{3})^i(l_{2i}+k_{2i+1})\right)v+ \left(\sum_{i=1}^{\infty}(-\frac{1}{3})^i(l_{2i-1}+k_{2i})\right)Av\\ &:=& Lv+KAv.\end{aligned}$$ As $\|l_i+k_{i+1}\|\leq 1+ \|k\|$, it follows that $\|K\|\leq (1+ \|k\|)\sum_{i=1}^{\infty}(\frac{1}{3})^i=(1+\|k\|)/2<\|k\|$. Hence $T\cap(T+kAv)=\emptyset$ and $(T+v)\cap(T+kAv)=\emptyset$, which imply that $T$ is disconnected (see Figure \[fig1\](b)). \[th2\] Let $A$ be a $2\times 2$ integral expanding matrix with characteristic polynomial $f(x)= x^2 + px \pm 3$ where $p>0$, and let ${\mathcal{D}}=\{0,v, kAv\}$ be a digit set where $k\in {\mathbb Z}\setminus\{0\}$ such that $\{v, Av\}$ is linearly independent. Then the self-affine set $T(A, {\mathcal{D}})$ is connected if and only if $k=\pm 1$. For the cases of $f(x)=x^2+px+3$ with $0<p< 3$, by using $0=f(A)=f(A)(A-I)=A^3+(p-1)A^2+(3-p)A-3I$, we obtain $$\label{eq3.1} I = \sum_{i=1}^{\infty}A^{-3i}\big((1-p)A^2-(3-p)A+2I\big).$$ $f(x)=x^2+x+3$: For $k=1$, then $\Delta{\mathcal D}=\{0,\pm v, \pm(Av-v), \pm Av\}$. By (\[eq3.1\]), $I=\sum_{i=1}^{\infty}A^{-3i}\big(-2A+2I\big)$ and $$\label{equa3.5} v=\sum_{i=1}^{\infty}A^{-3i}\big(-2Av+2v\big)=\sum_{i=0}^{\infty}A^{-3i}\big(A^{-2}(-v)+ A^{-3}(v-Av)+A^{-4}(Av)\big).$$ Hence $T\cap (T+v)\ne\emptyset$. Moreover, $$\label{equa3.6} Av=\sum_{i=0}^{\infty}A^{-3i}\big(A^{-1}(-v)+ A^{-2}(v-Av)+A^{-3}(Av)\big)$$ which implies $T\cap(T+Av)\ne\emptyset$. Consequently, $T$ is connected (see Figure \[fig2\](a)). \ For $k=-1$, then $\Delta{\mathcal D}=\{0,\pm v, \pm(Av+v), \pm Av\}$. From $f(A)=0$, we deduce that $I=(-A-2I)(A^2+I)^{-1}$, which in turn gives $$\begin{aligned} v &=& -A^{-1}v-2A^{-2}v + A^{-3}v+ 2A^{-4}v-A^{-5}v-2A^{-6}v+A^{-7}v + 2A^{-8}v-\cdots \\ &=& A^{-2}(-Av-v)+ A^{-3}(-Av) + A^{-4}(Av+v)+A^{-5}(Av)+ A^{-6}(-Av-v) \\ & & + A^{-7}(-Av)+ A^{-8}(Av+v) + A^{-9}(Av)+\cdots \\ &\in & T-T.\end{aligned}$$ Hence $T\cap (T+v)\ne \emptyset$. Multiplying the above expression by $A$, we have $$\begin{aligned} Av &=& A^{-1}(-Av-v)+ A^{-2}(-Av) + A^{-3}(Av+v)+A^{-4}(Av) \\ & & +A^{-5}(-Av-v)+ A^{-6}(-Av)+ A^{-7}(Av+v) + A^{-8}(Av)+\cdots \\ &\in & T-T\end{aligned}$$ which implies $T\cap (T+Av)\ne \emptyset$. It follows that $T$ is connected. For $\|k\|>1$. A point of $T-T$ can be written as $$\sum_{i=1}^{\infty}A^{-i}(k_iAv+l_iv)$$ where $k_i Av+l_iv\in \Delta{\mathcal D}$ for $i\geq 1$. By using the relation $A^{-i}v=\alpha_i v + \beta_i Av$, $$\begin{aligned} \label{eq3.2} &&\sum_{i=1}^{\infty}A^{-i}(k_iAv+l_iv)=\sum_{i=1}^{\infty}(k_iA^{-i+1}v+ l_iA^{-i}v) \nonumber\\ &=& \sum_{i=1}^{\infty}k_i(\alpha_{i-1}v + \beta_{i-1}Av) + \sum_{i=1}^{\infty}l_i(\alpha_iv + \beta_i Av) \nonumber\\ &=&\left(k_1+ \sum_{i=1}^{\infty}(k_{i+1}+l_i)\alpha_i\right)v + \left( \sum_{i=1}^{\infty}(k_{i+1}+l_i)\beta_i \right)Av \nonumber\\ &:=& Lv+KAv.\end{aligned}$$ As $\|l_i+k_{i+1}\|\leq 1+ \|k\|$ and $\tilde{\beta}<0.63$ (\[estimate1\]), we conclude $\|K\|\leq 0.63(1+ \|k\|)< \|k\|$, which yields $T\cap(T+kAv)=\emptyset$ and $(T+v)\cap(T+kAv)=\emptyset$. Hence $T$ is disconnected (see Figure \[fig3\](a)). \ $f(x)=x^2+2x+3$: For $k=1$. By (\[eq3.1\]), $I=\sum_{i=1}^{\infty}A^{-3i}\big(-A^2-A+2I\big)$ and $$v=\sum_{i=1}^{\infty}A^{-3i}\big(-A^2v-Av+2v\big)=A^{-1}(-v)+\sum_{i=0}^{\infty}A^{-3i}\big(A^{-2}(-v)+ A^{-3}v+A^{-4}(Av-v)\big).$$ Hence $T\cap (T+v)\ne\emptyset$. Moreover, $$Av=A^{-1}(-Av)+\sum_{i=0}^{\infty}A^{-3i}\big(A^{-1}(-v)+ A^{-2}v+A^{-3}(Av-v)\big)$$ which implies $T\cap(T+Av)\ne\emptyset$. Consequently, $T$ is connected (see Figure \[fig2\](b)). For $k=-1$. We obtain from (\[eq3.1\]) that $$\begin{aligned} v &=& -A^{-1}v-A^{-2}v+2A^{-3}v-A^{-4}v-A^{-5}v \\ && +2A^{-6}v-A^{-7}v -A^{-8}v+2A^{-9}v+\cdots\\ &=& A^{-1}(-v)+A^{-2}(-v)+A^{-3}v +A^{-4}(Av)+A^{-5}(-Av-v) \\ && +A^{-6}v+A^{-7}(Av)+ A^{-8}(-Av-v)+\cdots \\ &\in& T-T,\end{aligned}$$ implying $T\cap (T+v)\ne \emptyset$. Multiplying the above expression by $A$, we have $$\begin{aligned} Av+v &=& A^{-1}(-v)+A^{-2}v +A^{-3}(Av)+A^{-4}(-Av-v)\\ && +A^{-5}v+A^{-6}(Av)+A^{-7}(-Av-v)+\cdots\\ &\in & T-T,\end{aligned}$$ implying $(T+v)\cap (T-Av)\ne \emptyset$. Hence $T$ is connected. For $\|k\|>1$. By (\[eq3.2\]) and $\tilde{\beta}<0.73$ (\[estimate2\]), we have $\|K\|\leq (1+\|k\|)\tilde{\beta}<0.73(1+\|k\|)$. When $\|k\|\geq 3$, $\|K\|< 0.73(1+\|k\|)<\|k\|$, which yields $T\cap(T+kAv)=\emptyset$ and $(T+v)\cap(T+kAv)=\emptyset$. Hence $T$ is disconnected. When $k=2$, suppose $(T+2Av)\cap T \ne\emptyset$ or $(T+2Av)\cap(T + v)\ne\emptyset$, i.e., $2Av+lv\in T-T$ for $l=0$ or $-1$. By (\[eq3.2\]), we obtain $$\label{eq3.6} (k_2+l_1)\beta_1=2-\sum_{i=2}^{\infty}(k_{i+1}+l_i)\beta_i\geq 2- (1+2)(\tilde{\beta}-\|\beta_1\|)> 0.8$$ where $\beta_1=-1/3$. It follows that $l_1=-1,\ k_2=-2$. Then using $f(A)=0$, we have $$A(2Av+lv)-k_1Av-l_1v=(-4+l-k_1)Av-(6+l_1)v\in T-T.$$ It follows from $l_1=-1$ that $k_1=0$ or $2$. Hence $-4+l-k_1\leq -4+0+0=-4$, which contradicts the inequality $\|K\|\leq (1+\|k\|)\tilde{\beta}<3\times 0.73=2.19$. So $T$ is disconnected for $k=2$ (see Figure \[fig3\](b)). When $k=-2$, suppose $2Av+lv\in T-T$ where $l=0$ or $1$. Similarly, it yields from (\[eq3.6\]) that $l_1=-1,\ k_2=-2$ and $k_1=-2$ or $0$. If $k_1=-2$, then $(l-2)Av-5v\in T-T$. Multiplying the expression by $A$ and using $f(A)=0$, we obtain $$(l-2)A^2v-5Av-(k_2Av+l_2v)=(1-2l)Av+(6-3l-l_2)v\in T-T$$ and $6-3l-l_2\geq 6-3-1=2$. On the other hand, by (\[eq3.2\]) and (\[estimate2\]), $-5.81=-2-3\tilde{\alpha}\leq L\leq -2+3\tilde{\alpha}<1.81$. This is ridiculous. If $k_1=0$, then $(l-4)Av-5v\in T-T$ and $\|l-4\|\geq 3$ contracts $\|l-4\|=\|K\|< 2.19$. Therefore $T$ is disconnected for $k=-2$. $f(x)=x^2+3x+3$: For $k=1$. From $(A-I)f(A)=0$, we get $A^2+A-I=2(A+I)^{-1}$, which yields $I=-A^{-1}+A^{-2}+2\sum_{i=3}^{\infty}(-1)^{i+1}A^{-i}$ and $$\begin{aligned} v&=&-A^{-1}v+A^{-2}v+2\sum_{i=3}^{\infty}(-1)^{i+1}A^{-i}v\\ &=&A^{-2}(v-Av)+ A^{-3}v + \sum_{i=0}^{\infty}A^{-2i}\big(A^{-4}(Av-v)+ A^{-5}(v-Av)\big).\end{aligned}$$ Hence $T\cap (T+v)\ne\emptyset$. Moreover, $$Av=A^{-1}(v-Av)+ A^{-2}v + \sum_{i=0}^{\infty}A^{-2i}\big(A^{-3}(Av-v)+ A^{-4}(v-Av)\big)$$ which implies $T\cap(T+Av)\ne\emptyset$. Consequently, $T$ is connected (see Figure \[fig2\](c)). For $k=-1$. From $f(A)=0$ we get $A+2I=-(A+I)^{-1}$. It follows that $I= -2A^{-1}-A^{-2}+A^{-3}-A^{-4}+A^{-5}+\cdots$ and $$\begin{aligned} v &=& -2A^{-1}v-A^{-2}v+A^{-3}v-A^{-4}v+A^{-5}v+\cdots \\ &=& A^{-1}(-v) +A^{-2}(-Av-v)+A^{-3}v+ A^{-4}(-v)+A^{-5}v+A^{-6}(-v)+\cdots \\ &\in & T-T,\end{aligned}$$ then $T\cap (T+v)\ne\emptyset$. Also we can deduce immediately that $$\begin{aligned} Av+v &=& A^{-1}(-Av-v)+ A^{-2}v+ A^{-3}(-v)+A^{-4}v + A^{-5}(-v)+\cdots\\ &\in& T-T,\end{aligned}$$ which yields $(T+v)\cap (T-Av)\ne\emptyset$. As a result, $T$ is connected. For $k>1$. By (\[eq3.2\]) and (\[estimate3\]), we have $\|L\|\leq k+(1+k)\tilde{\alpha}<2.24+3.24k$. Suppose $kAv+lv\in T-T$ for $l=0$ or $-1$. Multiplying (\[eq3.2\]) by $A$ and then subtracting $k_1Av + l_1v$ from both sides, we see that $$\label{equa3.9} (-3k+l-k_1)Av -(3k+l_1)v\in T-T.$$ Repeating the process, we obtain $$\label{equa3.10} (6k-3l+3k_1-l_1-k_2)Av + (9k-3l+3k_1-l_2)v\in T-T.$$ Since $9k-3l+3k_1-l_2\geq 9k-0-3k-1=6k-1 > 2.24+3.24k$, which exceeds the upper bound of $\|L\|$. It concludes that $T\cap(T+kAv)=\emptyset$ and $(T+v)\cap(T+kAv)=\emptyset$, that is, $T$ is disconnected (see Figure \[fig3\](c)). For $k\leq -3$, then $\|L\|\leq -k+(1-k)\tilde{\alpha}<2.24-3.24k$. It follows from (\[equa3.10\]) that $\|9k-3l+3k_1-l_2\|\geq -9k-3+3k-1=-6k-4 > 2.24-3.24k$, which also exceeds the upper bound of $\|L\|$. For $k=-2$, then $\|K\|< 3.24$ and $\|L\|< 8.72$. From (\[equa3.9\]), we have $6+l-k_1< 3.24$, hence $l=-1$ and $k_1=2$. From (\[equa3.10\]), we have $l_2=-1$ and $3+l_1+k_2< 3.24$, it follows that $l_1=0, k_2=-2$, or $l_1=0, k_2=0$, or $l_1=1, k_2=-2$. When $l_1=0, k_2=-2$. Multiplying (\[equa3.10\]) by $A$ and then subtracting $k_3Av + l_3v$, we get $$(-5-k_3)Av+(3-l_3)v\in T-T.$$ By $\|5+k_3\|\leq 3.24$, it yields $k_3=-2$. Repeating this process, we obtain $$(12-l_3-k_4)Av+(9-l_4)v\in T-T.$$ Hence we get a contradiction $\|12-l_3-k_4\|\geq 9>3.24$. When $l_1=0, k_2=0$. Multiplying (\[equa3.10\]) by $A$ and then subtracting $k_3Av + l_3v$, we get $$(1-k_3)Av+(9-l_3)v\in T-T.$$ It yields $l_3=1$ and $k_3=0$ or $2$. Repeating this process, we obtain $$(3k_3+5-k_4)Av+(3k_3-3-l_4)v\in T-T.$$ If $k_3=0$, then $(5-k_4)Av+(-3-l_4)v\in T-T$, and $k_4=2$. Finally we get $(-12-l_4-k_5)Av+(-9-l_5)v\in T-T$ and a contradiction $\|12+l_4+k_5\|\geq 9>3.24$. If $k_3=2$, then $(11-k_4)Av+(3-l_4)v\in T-T$, and also $\|11-k_4\|\geq 9>3.24$. When $l_1=1, k_2=-2$. By the same argument as above, we first get $$(-2-k_3)Av+(6-l_3)v\in T-T.$$ It yields $k_3=0$. Repeating this process, we obtain $(12-l_3-k_4)Av+(6-l_4)v\in T-T$ and $\|12-l_3-k_4\|\geq 9>3.24$ follows. Therefore $T$ is disconnected for $\|k\|>1$. $f(x)=x^2+x-3$: For $k=1$. We deduce from $f(A)=0$ that $I= (A^2-I)^{-1}(-A+2I)$, which yields $I=\sum_{i=0}^{\infty}A^{-2i}(-A^{-1}+2A^{-2})$ and $$\begin{aligned} v&=&\sum_{i=0}^{\infty}A^{-2i}(-A^{-1}+2A^{-2})v\\ &=&A^{-2}(v-Av)+ \sum_{i=1}^{\infty}A^{-2i}\big(A^{-1}(Av-v)+ A^{-2}v\big).\end{aligned}$$ Hence $T\cap (T+v)\ne\emptyset$. Moreover, $$Av=A^{-1}(v-Av)+ \sum_{i=1}^{\infty}A^{-2i}\big((Av-v)+ A^{-1}v\big)$$ which implies $T\cap(T+Av)\ne\emptyset$. Consequently, $T$ is connected (see Figure \[fig2\](d)). For $k=-1$. It follows from $v=A^{-1}(Av)\in T-T$ that $T\cap (T+v)\ne \emptyset$. Moreover, we can get $A+I= -I+(A-I)^{-1}$ from $f(A)=0$. This implies $$\begin{aligned} Av+v &=& A^{-1}(-Av)+A^{-2}(Av)+A^{-3}(Av)+A^{-4}(Av)+\cdots\\ &\in & T-T,\end{aligned}$$ that is, $(T+v)\cap (T-Av)\ne \emptyset$. Hence $T$ is connected. For $k>1$. By (\[eq3.2\]) and (\[estimate4\]), we have $\|K\|\leq (1+k)\tilde{\beta}=1+k$ and $\|L\|\leq k+(1+k)\tilde{\alpha}=2+3k$. Suppose $kAv+lv\in T-T$ for $l=0$ or $-1$. Multiplying (\[eq3.2\]) by $A$ and then subtracting $k_1Av + l_1v$ from both sides, we have $$(-k+l-k_1)Av + (3k-l_1)v\in T-T.$$ Repeating the process, we obtain $$\label{equa3.11} (4k-l+k_1-l_1-k_2)Av + (-3k+3l-3k_1-l_2)v\in T-T.$$ Note $4k-l+k_1-l_1-k_2\geq 2k-1$. When $k\geq 3$, $2k-1>k+1$ which contradicts the upper bound of $\|K\|$, hence $T$ is disconnected; when $k=2$, it forces $l=0,\ k_1=-2,\ l_1=1,\ k_2=2 $ and $3Av-l_2v\in T-T$, similarly which implies $$(-3-l_2-k_3)Av+ (9-l_3)v\in T-T.$$ It is required that $\|9-l_3\|\leq 8$, hence $l_3=1$. Furthermore, from $(-3-l_2-k_3)Av+ 8v\in T-T$, we can deduce that $$(11+l_2+k_3-k_4)Av + (-9-3l_2-3k_2-l_4)v\in T-T.$$ Since $11+l_2+k_3-k_4\geq 6 >3$, we also get a contradiction, and $T$ is disconnected. Consequently, $T$ is disconnected for all $k>1$ (see Figure \[fig3\](d)). For $k<-1$, then $\|K\|\leq 1-k$. From (\[equa3.11\]), it follows that $\|K\|\geq -4k+l-k_1+l_1+k_2\geq -4k-1+k+1+k=-2k>1-k$, which is impossible. Therefore $T$ is disconnected for $\|k\|>1$. \[th3\] Let $B\in M_2({\mathbb{Z}})$ be an expanding integral matrix with characteristic polynomial $g(x)= x^2-px\pm 3$ where $p> 0$, and let ${\mathcal{D}}=\{0,v, kBv\}$ be a digit set where $k\in {\mathbb Z}\setminus\{0\}$ such that $\{v, Bv\}$ is linearly independent. Then the self-affine set $T(B, {\mathcal{D}})$ is connected if and only if $k=\pm 1$. The characteristic polynomial of $A$ is $f(x)=x^2+px+q$ if and only if that of $-A$ is $g(x)=x^2-px+q$. Since only the radix expansions matter, we may assume $B=-A$. The proof for the disconnectedness of $T(B, {\mathcal{D}})$ when $\|k\|>1$ can be adapted easily from the proof of Theorem \[th2\] by applying Corollary \[cor.ref\]. Let ${\mathcal{D}}=\{0,v, Av\}=\{0,v, -Bv\}$ and ${\mathcal{D}}'=\{0,v, -Av\}=\{0,v, Bv\}$. For $\|k\|=1$, we deduce the connectedness of $T_1:= T(B, {\mathcal{D}})$ (respectively, $T_1'=T(B, {\mathcal{D}}')$) from that of $T=T(A, {\mathcal{D}})$ (respectively, $T'=T(A, {\mathcal{D}}')$). We only show the case of $f(x)= x^2+x+3$. In Case 1 of the proof of Theorem \[th2\], from (\[equa3.5\]) we have $$v=\sum_{i=0}^{\infty}(-B)^{-3i}\big(B^{-2}(-v)-B^{-3}(v+Bv)+B^{-4}(-Bv)\big)\in T_1-T_1;$$ and from (\[equa3.6\]) we have $$Bv=\sum_{i=0}^{\infty}(-B)^{-3i}\big(B^{-1}(-v)-B^{-2}(v+Bv)+B^{-3}(-Bv)\big)\in T_1-T_1.$$ Hence $T_1$ is connected. Similarly, it can be verified that $v, Bv\in T_1'-T_1'$ and $T_1'$ is also connected. The above proof is indeed an application of the following more general result. Let $A, B\in M_2({\mathbb{Z}})$ be two expanding integral matrices with characteristic polynomials $f(x)= x^2+px+q$ and $g(x)= x^2-px+q$ respectively. Let $v,w$ be two non-zero vectors such that the two sets $\{v, Av\}$ and $\{w, Bw\}$ are both linearly independent. Denote by $L$ and $M$ the lattices generated by $\{v, Av\}$ and $\{w, Bw\}$ respectively. Let ${\mathcal{D}}=\{c_i Av+d_iv\in L: i=0,1,\dots, \|q\|-1\}$ and ${\mathcal{D}}'=\{-c_i Bw+d_iw\in M: i=0,1,\dots, \|q\|-1\}$ be two digit sets. Then $T_1=T(A, {\mathcal{D}})$ is connected if and only if $T_2=T(B, {\mathcal{D}}')$ is connected. We may assume $B=-A$ and $w=-v$. Then ${\mathcal{D}}={\mathcal{D}}'$ and $L=M$. Consider $a_iAv+b_iv=-a_iBv+b_iv\in \Delta {\mathcal{D}}=\Delta {\mathcal{D}}'$, it follows that $a_iAv+b_iv= A^{-1}(c_1'Av+d_1'v)+ A^{-2}(c_2'Av+d_2'v)+\cdots\in T_1-T_1$ if and only if $-a_iBv+b_iv= B^{-1}(c_1'Bv-d_1'v)+ B^{-2}(-c_2'Bv+d_2'v)+\cdots\in T_2-T_2$. Hence $c_iAv+d_iv, c_jAv+d_jv\in {\mathcal D}$ are ${\mathcal E}$-connected if and only if $-c_iBv+d_iv, -c_jBv+d_jv\in {\mathcal D}'$ are ${\mathcal E}$-connected, then the theorem is proved by Proposition \[e-connected prop\]. We can also deduce from Theorems \[th2\] and \[th3\] that Let $A$ be a $2\times 2$ integral expanding matrix with characteristic polynomial $f(x)=x^2+px+q$ where $\|q\|=3$. Let $v\in {\mathbb R}^2$ such that $\{v, Av\}$ is linearly independent. Then the self-affine set $T(A, {\mathcal{D}})$ is connected for ${\mathcal{D}}=\{0, v, Av+v\}$ or $\{0, v, -Av+v\}$. Notice that the difference set $\Delta\{0,v, Av+v\}=\{0, \pm v, \pm Av, \pm(Av+v)\}=\Delta\{0,v, -Av\}$ and $\Delta\{0,v, -Av+v\}=\{0, \pm v, \pm Av, \pm(Av-v)\}=\Delta\{0,v, Av\}$. Hence the result follows from Proposition \[e-connected prop\]. [Remarks:]{} The connectedness of self-affine sets or self-affine tiles is far from known extensively. Even for the planar case, there are still a lot of unsolved questions. The following may be some interesting topics related to the paper. [Q1.]{} Can we characterize the connectedness of $T(A, {\mathcal D})$ with $\|\det(A)\|=3$ and ${\mathcal D}=\{0, v, kAv+lv\}$? [Q2.]{} For a two dimensional digit set ${\mathcal D}=\{0, v,\dots, (l-1)v, Av,\dots, kAv\}$ with $l+k=\|\det(A)\|>3$, can we apply the same method to study the connectedness of $T(A, {\mathcal D})$? Acknowledgments: {#acknowledgments .unnumbered} ---------------- The authors would like to thank Professor Ka-Sing Lau for suggesting a related question and advice on the work. [99]{} S. Akiyama and N. Gjini, [Connectedness of number-theoretic tilings]{}, Discrete Math. Theoret. Computer Science 7 (2005), no. 1, 269-312. C. Bandt and G. Gelbrich, [Classification of self-affine lattice tilings]{}, J. London Math. Soc. [50]{} (1994), 581-593. C. Bandt and Y. Wang, [Disk-like self-affine tiles in ${\mathbb{R}}^2$]{}, Discrete Comput. Geom. 26 (2001), no.4, 591-601. Q.R. Deng and K.S. Lau, [Connectedness of a class of planar self-affine tiles]{}, J. Math. Anal. Appl. 380 (2011) 493-500. K. Gröchenig and A. Haas, [Self-similar lattice tilings]{}, J. Fourier Anal. Appl. 1 (1994), 131-170. D. Hacon, N.C. Saldanha and J.J.P. Veerman, [Remarks on self-affine tilings]{}, Experiment. Math. 3 (1994), 317-327. M. Hata, [On the structure of self-similar sets]{}, Japan J. Appl. Math. 2 (1985), no.2, 381-414. X.G. He, I. Kirat and K.S. Lau, [Height reducing property of polynomials and self-affine tiles]{}, Geom. Dedicata (2011) 152:153-164. I. Kirat, [Disk-like tiles and self-affine curves with non-collinear digits]{}, Math. Comp. 79 (2010), 1019-1045. I. Kirat and K.S. Lau, [On the connectedness of self-affine tiles]{}, J. London Math. Soc. [6227]{} (2000), 291-304. I. Kirat, K.S. Lau and H. Rao, [Expanding polynomials and connectedness of self-affine tiles]{}, Discrete Comput. Geom. 31 (2004), 275-286. J.C. Lagarias and Y. Wang, [Self-affine tiles in ${\mathbb R}^n$ ]{}, Adv. Math. 121 (1996) 21-49. K.S. Leung and K.S. Lau, [Disk-likeness of planar self-affine tiles]{}, Trans. Amer. Math. Soc. [359]{} (2007), 3337-3355. K.S. Leung and J.J. Luo, [Connectedness of planar self-affine sets associated with non-consecutive collinear digit sets]{}, J. Math. Anal. Appl. 395 (2012) 208-217. [^1]: The research is supported by STU Scientific Research Foundation for Talents.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a new quantum communication complexity protocol, the promise–Quantum Random Access Code, which allows us to introduce a new measure of unbiasedness for bases of Hilbert spaces. The proposed measure possesses a clear operational meaning and can be used to investigate whether a specific number of mutually unbiased bases exist in a given dimension by employing Semi–Definite Programming techniques.' author: - 'Edgar A. Aguilar' - 'Jakub J. Borka[ł]{}a' - Piotr Mironowicz - 'Marcin Paw[ł]{}owski' bibliography: - 'QRACMUBrefs.bib' title: Connections Between Mutually Unbiased Bases and Quantum Random Access Codes --- Introduction.- Mutually unbiased bases (MUBs) play a special role in the formalism of quantum mechanics. In particular they serve as complementary quantum tests, and find wide applicability in many fields of quantum information science such as quantum state tomography [@Ivo; @WF], quantum key distribution [@BB84], quantum teleportation and dense coding [@BBELTZ]. Hence, a general understanding of MUBs is well motivated and of general interest, see [@DEBZ_review] for an extensive review and further references. Explicitly, two orthonormal bases $\{\|\psi^1_i{\rangle}\}_i$ and $\{\|\psi^2_j{\rangle}\}_j$ of ${\mathbbm{C}}^d$ are said to be mutually unbiased if \[eq:mub\] [ \| [\^1\_i \| \^2\_j ]{} \|]{}\^2 = , i,j , where $[d] \equiv \{1,2,\dots,d\}$. The term unbiased is used because if we pick any basis vector $\|\psi^1_i{\rangle}$, then performing a measurement in the $\{\|\psi^2_j{\rangle}\}_j$ basis will yield a completely random result (i.e. each outcome $\|\psi^2_j{\rangle}$ will have equal detection probability $1/d$). A set of MUBs in dimension $d$ is said to be maximal, if there are $d+1$ bases which are all pairwise mutually unbiased. The construction of maximal sets when $d=p$, a prime number, was described by Ivonovic [@Ivo], and later by Wootters and Fields when $d=p^k$, a prime power [@WF]. The general problem of whether $d+1$ bases exist for arbitrary dimensions remains open for at least the past 29 years. In particular it is an open question whether a complete set of MUBs exist even in the simplest case, namely in dimension 6. Zauner’s conjecture states that no more than three MUBs exist in dimension 6 [@Zauner]. The task of proving the conjecture is a research field on its own, see e.g. [@JMMSW; @Grassl] for partial analytical results supporting the conjecture. Numerical approaches have also failed to be conclusive, [@BW]. In this paper we introduce a novel protocol named promise-Quantum Random Access Code (pQRAC). The main idea of this protocol is to use the so-called $n^d \rightarrow 1$ Quantum Random Access Codes (QRACs) with certain constraints. Our main technical result shows that a specific average success probability of the protocol can be achieved if and only if $n$ MUBs exist in dimension $d$. The protocol allows us to create a new measure of unbiasedness, which quantifies the amount by which two (or more) bases are mutually unbiased. Other measures currently exist and are in use [@distance], yet the presented one possesses a direct operational interpretation as the success probability of a well–defined communication task. Furthermore, the pQRAC game is suitable for numerical optimization techniques like Semi–Definite Programming (SDP)[@sdp]. In particular, one may use the see-saw method [@seesaw1] to search for $n$ MUBs in dimension $d$. What is more, pQRACs may be used together with the Navascues and Vertesi method [@PhysRevLett.115.020501] to discard the existence of $n$ MUBs in a particular dimension. This exclusion is a rigorous statement, in contrast to drawing the conclusion out of the failure of trying to find them. As a proof of principle, we have applied our method to exclude the existence of 5 MUBs in dimension 3, and 6 MUBs in dimension 4. We have been unable to rule out the existence of 4 MUBs in dimension 6, but argue that the problem is now at arm’s length for future researchers. Methods.- We begin by introducing Random Access Codes (RACs)[@ALMO_qracs]. An $n^d{\rightarrow}1$ RAC is a protocol in which Alice tries to compress an $n$-dit string into $1$ dit, such that Bob can recover any of the $n$ dits with high probability. More precisely, Alice receives a uniformly distributed random input string $\mathbf{x}=x_1x_2\cdots x_{n}$, $x_i\in [d]$. She then uses an encoding function ${\mathcal{E}}_c:[d]^n{\rightarrow}[d]$ (possibly classically probabilistic), and is allowed to send one dit $a={\mathcal{E}}_c(\mathbf{x})$ to Bob. On the other side, Bob receives an input $y\in [n]$ (uniformly distributed), and together with Alice’s message $a$ uses one of $n$ (possibly classically probabilistic) decoding functions ${\mathcal{D}}_c^y:[d]\rightarrow [d]$, to output $b={\mathcal{D}}_c^y(a)$ as a guess for $x_y$. If Bob’s guess is correct (i.e. $b=x_y$) then we say that they are successful, otherwise we say that they are unsuccessful or fail. Similarly, we define $n^d{\rightarrow}1$ Quantum Random Access Codes (QRACs) where Alice encodes her input $n$-dit string into a $d$-dimensional quantum system (qudit) via ${\mathcal{E}}_q:[d]^n{\rightarrow}{\mathbbm{C}}^d$, and sends the qudit $\rho_{\mathbf{x}} = {\mathcal{E}}_q(\mathbf{x})$ to Bob. He then performs one of his decoding functions ${\mathcal{D}}_q^y:{\mathbbm{C}}^d {\rightarrow}[d]$ to output his guess $b$ for $x_y$. The decoding function is simply a quantum measurement, i.e. he outputs his guess $b$ with probability $\mathbb{P}(b=x_y)={\text{tr}}[\rho_{\mathbf{x}} M^y_b]$, where the operators $M^y_b$ are POVMs (i.e. positive and $\forall y {\text{ }}\sum_b M^y_b = \openone$). As a figure of merit, we employ the optimal average success probability for both RACs and QRACs: \[eq:qasp\] \|[P]{}\_[c,q]{} (n,d) = \_[{,}]{} \_\_y (b=x\_y) . The maximization is over encoding-decoding strategies $\{{\mathcal{E}}_{c,q},{\mathcal{D}}_{c,q}\}$ (classical or quantum respectively), and the average is taken over all possible inputs $(\mathbf{x},y)$ of Alice and Bob. In the quantum case, the optimal average success probability $\bar{P}_q$, can be achieved with pure states, $\rho_{\mathbf{x}} = \|\mathbf{x}{\rangle}{\langle}\mathbf{x}\|$ [@ALMO_qracs], where $\|\mathbf{x}{\rangle}$ is the eigenvector of $\sum_y M^y_{x_y}$ with largest eigenvalue. In [@Farkas], it was shown that for $2^d{\rightarrow}1$ QRACs this maximum is achieved when the operators $M^y_b$ are (rank 1) projective measurements. Therefore, throughout the rest of this letter we will be considering only pure-state encoding and von-Neumann measurements. RACs and QRACs have increasingly become an experimental tool to test the “quantumness" or non-classical behavior of a system [@THMB_expqrac; @QRACexp]. For fixed $n$ and $d$, we have $\bar{P}_c < \bar{P}_q$, and a gap is exploited to show that a system is behaving non-classically . For example a $2^2{\rightarrow}1$ RAC has $\bar{P}_c=0.75$, while the corresponding QRAC has an optimal average success probability of $\bar{P}_q=(2+\sqrt{2})/4\approx 0.8536$ [@ANTV]. Thus for a system of dimension 2, observing an average success probability greater than $0.75$ indicates non-classical behaviour. The quantum advantage comes from encoding Alice’s state as a superposition of the bases $\{\|\psi^1_i{\rangle}\}_i$ and $\{\|\psi^2_j{\rangle}\}_j$, namely $\|\mathbf{x}{\rangle}= \alpha \|\psi^1_{x_1}{\rangle}+ \beta \|\psi^2_{x_2}{\rangle}$, while Bob measures in the $\{\|\psi^y_i{\rangle}\}_i $ basis. We have the following: \[result:1\] For a $2^d{\rightarrow}1$ QRAC, the optimal average success probability \[eq:21optimal\] \|[P]{}\_q (2,d) = (1+) is obtained if and only if Bob’s measurement bases $\{\|\psi^1_i{\rangle}\}_i , \{\|\psi^2_j{\rangle}\}_j$ are mutually unbiased. The proof is given in Supplementary Material \[sec:2d1\]. We find it interesting to note here an observation that Lemma \[result:1\] cannot be generalized to the case of $n^d{\rightarrow}1$ QRACs for $n \geq 3$, as stated below: The MU condition on Bob’s measurement bases is not sufficient for obtaining the optimal average success probability in $n^d{\rightarrow}1$ QRACs when $n \geq 3$. The proof of this result is by direct calculation (See Supplementary Material \[sec:anomalies\] for details). This occurs since there are inequivalent subsets of MUBs (i.e. not related by unitary transformations) in higher dimensions. As an example, let us consider the case $n=3, d=5$. Bob must choose 3 different measurement bases, and he can do so in ${{6}\choose{3}} =20$ ways. Half of those selections lead to an average success probability of 0.610855, while the other half give 0.596449. Hence, the choice of the subset of MUBs matters. This feature occurs also for other choices of $n$ and $d$. However, we conjecture that the optimal average success probability for $n^d{\rightarrow}1$ QRACs is indeed achieved with a suitable choice of MUBs. Next we define a $(n,m)^d{\rightarrow}1$ promise-QRAC (pQRAC), $m\leq n$, as an $n^d{\rightarrow}1$ QRAC with an extra promise. Let $S^n_m$ be the set of all possible subsets of $[n]$ of size $m$. Then in a pQRAC, Alice receives an additional input $z \in S^n_m$, with the promise that $y \in z$. That is, Alice knows that Bob will not be questioned over some of Alice’s inputs, see Fig. \[fig:pQRACplot\] for an illustration of a pQRAC. ![Schematic representation of a $(n,m)^d{\rightarrow}1$ promise–Quantum Random Access Code. Here $x_i\in [d],y\in[n]$, and $z$ is a subset of $[n]$ with $m$ elements. The bold inputs $x_k$ depict $k\in z$. $\rho$ is the quantum state that Alice sends to Bob.[]{data-label="fig:pQRACplot"}](pQRAC.pdf) Hence, the optimal average success probability , is modified in the case of $(n,m)^d{\rightarrow}1$ pQRACs to: \[pQRACasp\] \_q (n,m,d) =\_[{,{M}}]{} \_[zS\^n\_m]{} \_[\_z]{} \_[yz]{} \[\_[,z]{} M\^y\_[x\_y]{}\], where the summation over $\mathbf{x}_z$ indicates a summation over $x_{i_1}, x_{i_2} , \dots, x_{i_m}$ such that $\{i_1, i_2, \dots , i_m\}= z$ , and the maximization is taken over all quantum encoding and decoding strategies $\{\rho,\{M\}\}$. Now, we are able to prove our main technical result: \[result:2\] For a $(n,2)^d{\rightarrow}1$ pQRAC, the following holds: \_q(n,2,d) (1+) with equality iff at least $n$ MUBs exist in dimension $d$. We begin by writing the optimal average success probability of the $(n,2)^d{\rightarrow}1$ pQRAC. $$\begin{aligned} \tilde{P}_q (n,2,d) &=\max_{\{\rho,\{M\}\}} \frac{1}{\binom{n}{2} 2d^2} \sum_{z\in S^n_2} \sum_{\mathbf{x}_z} \sum_{y\in z} {\text{tr}}[\rho_{\mathbf{x},z} M^y_{x_y}] \\ &\leq \frac{1}{\binom{n}{2}} \sum_{z\in S^n_2} \left( \max_{\{\rho,\{M\}\}} \frac{1}{2d^2} \sum_{\mathbf{x}_z} \sum_{y\in z} {\text{tr}}[\rho_{\mathbf{x},z} M^y_{x_y}] \right). \end{aligned}$$ The inequality follows, since the strategies to maximize the summands might not be compatible with each other globally. In fact, we recognize the term in parenthesis as $\bar{P}_q (2,d)$, the optimal success probability for a $2^d {\rightarrow}1$ QRAC . From Lemma \[result:1\], this maximization occurs if and only if the measurement bases corresponding to the set $z$ are mutually unbiased. It follows that it is possible to simultaneously satisfy all of these maximization constraints iff there exists $n$ MUBs in dimension $d$. The intuition behind Lemma \[result:2\], is that Bob must be ready to measure in all $n$ bases. If there exist $n$ bases which are all pairwise mutually unbiased, then essentially they are just playing a more complicated version of the usual $2^d {\rightarrow}1$ QRAC. If these bases do not exist, then for some $z \in S^n_2$ , the protocol will not be able to achieve the optimal value , dropping the entire average. Results.- In the context of MUBs, reference [@BBELTZ] has introduced a distance measure between two bases $\{\|\psi^1_i{\rangle}\}_i$ and $\{\|\psi^2_j{\rangle}\}_j$ which quantifies unbiasedness: D\^2\_[\^1 \^2]{} = 1 - \_[i,j]{} (\|\^1\_i \|\^2\_j\|\^2 -)\^2. The measure is symmetric ($D^2_{\psi^1 \psi^2} = D^2_{\psi^2 \psi^1}$). If the bases are the same, then $D^2_{\psi^1 \psi^1} = 0$. The maximum $D^2_{\psi^1 \psi^2} = 1$ is obtained iff the bases are mutually unbiased. For a set of $n$ bases in ${\mathbbm{C}}^d$ ($\{\psi^j\} = \{\|\psi^j_i{\rangle}\}_i$, $j\in[n]$), one can analyze the average square distance between all possible pairs of bases [@distance]: \[eq:dsq\] \|[D]{}\^2({\^i}\_i) = \_[{a,b}S\^n\_2]{} D\^2\_[\^a \^b]{}. Likewise, $\bar{D}^2=1$ iff all bases are pairwise mutually unbiased. However, is an abstract distance measure, lacking an operational interpretation. Lemma \[result:2\] immediately leads us to our first result. Given a set of $n$ bases in dimension $d$ we define as their unbiasedness measure the average success probability in a $(n,2)^d {\rightarrow}1$ pQRAC if the bases are used as Bob’s measurement bases. This measure is thus defined operationally and has the following properties: (1) The maximum of $\bar{P}_q(2,d) = \frac{1}{2}(1+d^{-1/2})$ is attainable iff all bases are pairwise unbiased, (2) It is symmetric under permutation of bases, and (3) The minimal value of $\bar{P}_c(2,d) = \frac{1}{2}(1+d^{-1})$ is achieved iff all bases are the same. The optimal classical success probability of $n^d{\rightarrow}1$ RACs is shown in [@Czechlewski18]. Explicitly, given $n$ bases of ${\mathbbm{C}}^d$, $\{ \psi^i \}_i$, the maximum attainable average success probability of the pQRAC, $\bar{p}$, is: \[eq:Pmeasure\] \|[p]{}({\^i}\_i) = \_[{a,b}S\^n\_2]{} ( + \_[i,j]{} \|\^a\_i \| \^b\_j \| ), which comes as a direct conclusion of Lemma \[result:1\]. We may normalize such that the minimum value is $0$ (obtained iff all bases are the same), and the maximum value is $1$ (obtained iff all bases are pairwise MU) and get the expression: \[eq:Qmeasure\] \|[Q]{}({\^i}\_i) = . See Supplementary Material \[sec:Qbar\] to see a direct comparison between and . For illustrative purposes, we have optimized the value of the $(4,2)^6 {\rightarrow}1$ pQRAC game expression using the see-saw method [@seesaw1]. This allows us to show how the optimization of may be used to construct MUBs in a particular dimension, as well as providing numerical examples of how $\bar{D}^2$ and $\bar{Q}$ compare. With this method, the maximal value of $\bar{D}^2$ of four MUBs in dimension $6$ we obtained is $0.998284$ with $\bar{Q}=0.998045$. On the other hand, the bases from [@BH; @RLE] have $\bar{D}^2 = 0.998292$, and $\bar{Q} = 0.998036$. With this result one sees that the two measures, and are not equivalent, and induce different partial orderings on the sets of bases. See Supplememntary Material \[sec:seesaw\] for more details. Our second result is another direct application of Lemma \[result:2\], and deals with ruling out if there are $n$ mutually unbiased bases in dimension $d$. Explicitly, if it is possible to show that there are no sets of encoded states and measurement bases that would obtain a success probability of $\bar{P}_q(2,d)$, then one immediately concludes that there does not exist $n$ MUBs in the given dimension. Thus one may use the SDP hierarchy of relaxations proposed by Navascues and Vertesi (NV) [@PhysRevLett.115.020501]. The method defines a sequence of SDP problems yielding upper bounds to optimization tasks over quantum probability distributions with dimensional constraints. One can show that the method converges to the accurate quantum values [@PhysRevA.92.042117]. If at a given level of the hierarchy the upper bound falls below the threshold $\bar{Q}=1$, then the conclusion follows. We emphasize that if $n$ MUBs do not exist in a particular dimension, then applying the SDP hierarchy to the $(n,2)^d$ pQRAC gives an algorithmic way of proving their non-existence. On the other hand, if $n$ MUBs do exist, the proposed method will fail to draw a conclusion. Implementing the hierarchy.- Let us try to directly apply the NV hierarchy to the $(n,2)^d {\rightarrow}1$ pQRAC. To implement the $k$-th level of the hierarchy, $Q^k$, the set ${\mathbb{S}}^k_d$ of all feasible moment matrices of order $2k$ arising from quantum systems of dimension $d$ must be calculated. For this, moment matrices $\Gamma^j_k$ are randomly generated from this set until $\text{span}(\{\Gamma^j_k\}_j) = {\mathbb{S}}^k_d$. In practice, the algorithm keeps creating new moment matrices $j=\{1,2,\ldots,v_k\}$ and stops when $\Gamma^{v_k+1}_k \in \text{span}(\{\Gamma^j_k\}_{j=1}^{v_k})$. The method requires an assumption on the rank of the projectors $\{M^y_b\}$, but in our scenario Bob’s optimal strategy is to implement $d$-dimensional von Neumann measurements, therefore all operators are rank 1. In order to generate $\Gamma^j_k$, we randomly choose $A=\binom{n}{2} d^2$ states for Alice to encode and $B=nd$ measurement operators for Bob ($n$ bases of ${\mathbbm{C}}^d$). Then, $\Gamma^j_k$ contains the traces of all strings of size less than or equal to $2k$ constructed from Alice’s states and Bob’s operators. For example, typical matrix elements of $\Gamma^j_1$ include ${\text{tr}}[\rho^j_{\mathbf{x},z}\rho^j_{\mathbf{x}',z'}]$, ${\text{tr}}[\rho^j_{\mathbf{x},z}M^{y,j}_b]$, and ${\text{tr}}[M^{y,j}_b M^{y',j}_{b'} ]$. While in $\Gamma^j_3$, we can find ${\text{tr}}[\rho^j_{\mathbf{x},z} M^{y,j}_{b} M^{y',j}_{b'} M^{y'',j}_{b''} \rho^j_{\mathbf{x}',z'} M^{y''',j}_{b'''}]$, etc. We write the $k$-th order relaxation to our problem as the following semidefinite program [@PhysRevLett.115.020501]: $$\begin{split} \label{eq:SDPklevel} &\tilde{P}_q(n,2,d) = \max {\text{tr}}[ \hat{\mathcal{B}} \text{ }\Gamma_k] \\ &\text{s.t. } \Gamma_k \in {\mathbb{S}}^k_d , \text{ }\left(\Gamma_k\right)_{1,1} = 1, \text{ }\Gamma_k\geq 0, \end{split}$$ where we call $\hat{\mathcal{B}}$ the pQRAC game matrix, and construct it to “pick out" the values ${\text{tr}}[\rho_{\mathbf{x},z} M^y_b]$ from $\Gamma_k$ such that $b=x_y$ and $y\in z$. Roughly $\frac{1}{2}(A+B)^{4k}$ real-valued numbers need to be stored in a computer’s RAM in order to describe the set of all feasible moment matrices ${\mathbb{S}}^k_d$. Below, we describe a potentially quadratic reduction in the problem’s memory requirements. See Supplementary Material \[sec:estimate\] for details. Note that $\hat{\mathcal{B}}=\hat{\mathcal{B}}^T$, and is a sparse matrix with a lot of symmetries. In this case, we employ the symmetries corresponding to relabeling measurement device outputs, and the ones corresponding to permuting the labels of the measurement devices themselves. This approach has been followed on the NPA hierarchy in the Bell-test scenario [@BellSymmetries]. Let $\hat{\mathcal{B}}$ be invariant under the group of transformations $\mathcal{G}$. In other words, for every representation $G$ of an element $g\in\mathcal{G}$, $G \hat{\mathcal{B}} G^T = \hat{\mathcal{B}}$. Then, if we apply a group action on the game matrix inside the objective function , this would be equivalent to applying a group action on $\Gamma_k$. Namely, ${\text{tr}}[ \hat{\mathcal{B}} {\text{ }}\Gamma_k] = {\text{tr}}[ \hat{\mathcal{B}} {\text{ }}G^T \Gamma_k G]$. Therefore, it is unnecessary to consider the full space of feasible moment matrices ${\mathbb{S}}^k_d$ and can simplify into: $$\begin{split} \label{eq:SDPklevelsymm} &\tilde{P}_q(n,2,d) = \max {\text{tr}}[ \hat{\mathcal{B}} \text{ }\hat{\Gamma}_k] \\ &\text{s.t. } \hat{\Gamma}_k \in \mathcal{G}({\mathbb{S}}^k_d) , \text{ }\left(\hat{\Gamma}_k\right)_{1,1} = 1, \text{ }\hat{\Gamma}_k\geq 0 , \end{split}$$ where we denote $\mathcal{G}({\mathbb{S}}^k_d)$ as the set of feasible moment matrices which are $\mathcal{G}$-invariant. In order to implement this, we generate random invariant moment matrices $\hat{\Gamma}^j_k$ by first creating a moment matrix $\Gamma^j_k$ and averaging it out over all of the group elements: \[eq:avgsymm\] \^j\_k = \_ G\^j\_k G\^T. Clearly $\hat{\Gamma}^j_k \in \mathcal{G}({\mathbb{S}}^k_d)$, and this is repeated until $\text{span}(\{\hat{\Gamma}^j_k\}_j) = \mathcal{G}({\mathbb{S}}^k_d)$. To illustrate the power of the proposed method, we report that for a $(4,2)^5{\rightarrow}1$ pQRAC, $\text{dim}({\mathbb{S}}^1_5)=13672$ and the SDP running time was 22.5h on a desktop computer, whereas $\text{dim}(\mathcal{G}({\mathbb{S}}^1_5))=7$ and had a total run-time of 50s. Using this, we have implemented $Q^1$ and a subset of the “almost quantum" level [@Nav15q1ab] ($Q^{1+\textit{succ}}$) for some relevant pQRAC cases, see Table \[tab:tabSDP\]. The level $Q^{1+\textit{succ}}$ includes traces of strings of length $\leq 2$ from the set of operators $\{ \{\rho^j_{\mathbf{x},z}\}, \{M^{y,j}_b\}, \{ \rho^j_{x_{z_1},x_{z_2},\{z_1,z_2\}} M^{z_i,j}_{x_{z_i}}\}\}$. That is, we also included pairs of states and measurements which lead to successful trials. The details of the implementation are found in Supplementary Material \[sec:Symmetries\]. As a proof of principle, we note that we have been able to rigorously rule out the existence of $d+2$ MUBs in dimensions $d=3$ and $d=4$. However, with our numerical precision and at this hierarchy level we have been unable to rigorously exclude the existence of 4 MUBs in dimension 6. We notice that the hierarchy level $Q^{1+\textit{succ}}$ was also unable to rule out the existence of 4 MUBs in $d=2$. If these four bases existed, together with the $d=3$ MUBs one could create four MUBs in dimension 6. We conjecture that in order for a level of the hierarchy to be able to rule out the existence of 4 bases in dimension 6, it must first rule out the existence of 4 MUBs in $d=2$. In future work, we wish to find more efficient ways of calculating , and higher levels of the hierarchy. case $(n,2)^d$ $\bar{Q}$ for $Q^1$ $\bar{Q}$ for $Q^{1+\textit{succ}}$ ---------------- ----------------------- ------------------------------------- -- $(3,2)^2$ $\textbf{0.999999}99$ $\textbf{0.999999}99$ $(4,2)^2$ $\textbf{0.999999}99$ $\textbf{0.999999}92$ $(4,2)^3$ $\textbf{1.131652}47$ $\textbf{0.999999}95$ $(5,2)^3$ \* $\textbf{1.131652}49$ $\textbf{0.999898}95$ $(5,2)^4$ \* $\textbf{1.242640}66$ $\textbf{0.999999}96$ $(6,2)^4$ \* $\textbf{1.242640}66$ $\textbf{0.999916}55$ $(3,2)^6$ $\textbf{1.428825}41$ $\textbf{0.999999}98$ $(4,2)^6$ \* $\textbf{1.428825}38$ $\textbf{0.999999}97$ : Results of implementing hierarchy levels $Q^1$ and $Q^{1+\textit{succ}}$ to $(n,2)^d{\rightarrow}1$ pQRACs. The \* indicates an optimization over $\mathcal{G}({\mathbb{S}}^k_d)$. Other cases were executed for both ${\mathbb{S}}^k_d$ and $\mathcal{G}({\mathbb{S}}^k_d)$ to test our code implementation. \[tab:tabSDP\] Conclusions.- In this paper we give a new class of quantum games, pQRACs, which serve as an operational way of testing unbiasedness. It also enables one to reformulate the problem of searching for a given number of MUBs in a particular dimension as a problem of optimizing the strategy of the pQRAC game. In particular if one is able to get a proper upper bound on the value of the game, then our formulation allows to exclude the existence of a given number of MUBs in the considered dimension. We have exploited the symmetries of the pQRAC game matrix in the Navascues and Vertesi hierarchy. We hope this will lead to rigorously proving Zauner’s conjecture by considering higher levels. Acknowledgments.- The paper was supported by EU grant RAQUEL, ERC AdG QOLAPS, National Science Centre (NCN) Grant No. 2014/14/E/ST2/00020 and DS Programs of the Faculty of Electronics, Telecommunications and Informatics, Gdańsk University of Technology. EA acknowledges support from CONACyT. The SDP optimizations have been performed using OCTAVE [@octave] with SDPT3 solver [@SDPT3a; @SDPT3b], SeDuMi [@Sturm98usingsedumi], and YALMIP toolbox [@yalmip]. We thank D. Saha and M. Farkas for discussions, and I. Bengtsson for guidance with the literature. Optimal average success probability of $2^d\rightarrow 1$ QRAC {#sec:2d1} ============================================================== Let $\{\|\psi_i{\rangle}\}_i$ and $\{\|\phi_j{\rangle}\}_j$ be two orthonormal bases used on Bob’s side to perform the von Neumann measurement, and we write their projectors as $\|\psi_i{\rangle}{\langle}\psi_i\|=M^0_i$ and $\|\phi_j{\rangle}{\langle}\phi_j\|=M^1_j$. The optimization of the average success probability is simply a maximization over the measurement bases $\{M^0_i\},\{M^1_j\}$ of the expression: \|[p]{}\_[q]{} = \_[x\_0,x\_1=0]{}\^[d-1]{} \[\_[x\_0,x\_1]{} (M\^0\_[x\_0]{}+M\^1\_[x\_1]{})\]. The maximum is achieved when $\rho_{x_0,x_1}$ (a pure state) is an eigenvector, corresponding to the largest eigenvalue of $(M^0_{x_0}+M^1_{x_1})$. This can be seen by writing $\rho_{x_0,x_1} = \|\xi{\rangle}{\langle}\xi\|$ , and expressing the pure state as $\|\xi{\rangle}= \sum c_k \|k{\rangle}$, where $\{\|k{\rangle}\}_k$ is the eigenbasis of the sum of the operators ($M^0_{x_0}+M^1_{x_1} = \sum \lambda_k \|k{\rangle}{\langle}k\|$). Then ${\text{tr}}[\rho_{x_0,x_1} (M^0_{x_0}+M^1_{x_1})] = \sum \|c_k\|^2 \lambda_k $, which is clearly maximal when $\rho_{x_0,x_1} = \|\lambda_{\text{max}}{\rangle}{\langle}\lambda_{\text{max}}\|$. Thus we are trying to maximize the sum $\sum_X \lambda_{\max}(M^0_{x_0}+M^1_{x_1})$. Direct calculations considering only two dimensional subspaces at a time yields: \[eq:21qracmax\] \|[P]{}\_[q]{} = \_[x\_0,x\_1=0]{}\^[d-1]{} (1+\_[x\_0,x\_1]{} ), where $\alpha_{x_0,x_1}=\sqrt{{\text{tr}}[M^0_{x_0}M^1_{x_1}]}=\|{\langle}\psi_{x_0}\|\phi_{x_1}{\rangle}\|$ is simply the modulus of the inner product between two elements of the different bases. To see this, we first fix $x_0$ and $x_1$. Then, any 2 linearly independent vectors ${ \| \, \psi_{x_0} \rangle},{ \| \, \phi_{x_1} \rangle}$ span a 2 dimensional subspace of $\mathbb{C}^d$. Let ${ \| \, 0 \rangle},{ \| \, 1 \rangle}$ be the basis vectors of such a subspace and for simplicity we choose them such that: $$\begin{aligned} &{ \| \, \psi_{x_0} \rangle}={ \| \, 0 \rangle} \\ &{ \| \, \phi_{x_1} \rangle}=\alpha{ \| \, 0 \rangle} +\beta{ \| \, 1 \rangle}, \end{aligned}$$ where $\|\alpha\|^2+\|\beta\|^2=1$, and $\alpha=\alpha_{x_0,x_1}=\|{\langle}\psi_{x_0}\|\phi_{x_1}{\rangle}\|$. Then, we solve the characteristic equation for the measurement operator, $0=\det({ \| \, \psi_{x_0} \rangle}{ \langle \psi_{x_0} \, \|}+{ \| \, \phi_{x_1} \rangle}{ \langle \phi_{x_1} \, \|}-\lambda\mathbb{I})$, to get the following two solutions: $$\lambda_{\pm}=1 \pm \sqrt{1-\|\beta\|^2}=1\pm\alpha,$$ where $\lambda_+$ is the maximum eigenvalue used in . Before being able to maximize , we briefly introduce the concepts of majorization and of Schur concavity. We say that a probability distribution $\mathbf{p}=(p_1, p_2 , \ldots , p_N)$ majorizes the probability distribution $\mathbf{q}=(q_1, q_2, \ldots, q_N)$, denoted by $\mathbf{p} \succ \mathbf{q}$, if: \_[i=1]{}\^k p\_i\^ \_[i=1]{}\^k q\_i\^ k, where $\mathbf{p}^\downarrow $ $(\mathbf{q}^\downarrow)$ is a vector with the same components as $\mathbf{p}$ $(\mathbf{q})$, but written in descending order. A function $F:{\mathbbm{R}}^N {\rightarrow}{\mathbbm{R}}$ is said to be Schur concave if $\mathbf{p} \succ \mathbf{q}$ implies $F(\mathbf{q}) \geq F(\mathbf{p})$. In particular, if $f:{\mathbbm{R}}{\rightarrow}{\mathbbm{R}}$ is a concave function , then $F = \sum_i f(x_i)$ is Schur concave. Finally, note that $\sum_x \alpha_x^2=d$, is a constant. Let us define the probability distribution $\mathbf{a} = (a_1, a_2 ,\ldots , a_{d^2})$, with $a_i = \frac{1}{d} \alpha^2_i , \forall i$. Then may be rewritten as: \|[P]{}\_q = \_[x]{} 1 + . By the remark above this function is Schur concave, and is therefore maximized by the uniform distribution $\mathbf{u} = \frac{1}{d^2} (1 , 1 , \ldots , 1)$ [@inequalities]. This is because $\mathbf{u}$ is majorized by all probability distributions. Explicitly, we obtain $\alpha_x^2 = d^{-1} , ~ \forall x$, which is precisely the condition of unbiasedness . Insufficiency of MUBs for optimality in $3^d{\rightarrow}1$ QRACs {#sec:anomalies} ================================================================= In this section we would like to find the maximum success probability for $3^d{\rightarrow}1$ QRACs, and prove that the maximum is achieved when using MUBs. To do this we need to find the optimal encoding/decoding strategy for Alice and Bob.\ For a given encoding/decoding strategy, the average success probability is given by , which we write out here as: \|[p]{}=\_[i=0]{}\^[d-1]{} \_[j=0]{}\^[d-1]{} \_[k=0]{}\^[d-1]{} \[\_[ijk]{} (\_[i]{} + \_[j]{}+ \_[k]{})\], where $\Psi_i = \|\psi_i{\rangle}{\langle}\psi_i\|$, $\Phi_j=\|\phi_j{\rangle}{\langle}\phi_j\|$, $\Theta_k=\|\theta_k{\rangle}{\langle}\theta_k\|$ are three projectors which are used by Bob to measure the message from Alice (encoded as $\rho_{ijk}$). To maximize the above expression, we will let Alice’s encoding $\rho_{ijk}$ be a pure state corresponding to the eigenvector of the largest eigenvalue of the measurement operators $\lambda_{\text{max}}(\Psi_{i} + \Phi_{j}+ \Theta_{k})$. For the moment we analyze the eigenvalues of such operators, and drop the $i,j,k$ subscripts. In short, we want to solve the following characteristic equation: $$\det({ \| \, \psi \rangle}{ \langle \psi \, \|}+{ \| \, \phi \rangle}{ \langle \phi \, \|}+{ \| \, \theta \rangle}{ \langle \theta \, \|}-\lambda\mathbb{I})=0,$$ which leads to the following third degree polynomial for the eigenvalues: $$\begin{aligned} &-\lambda^3+3\lambda^2+m\lambda+n=0, \\ &m=\lvert{ \langle \psi \, \|}\phi{\rangle}\rvert^2 + \lvert{ \langle \psi \, \|}\theta{\rangle}\rvert^2 + \lvert{ \langle \phi \, \|}\theta{\rangle}\rvert^2-3,\\ &n = \text{det}({ \| \, \psi \rangle}{ \langle \psi \, \|}+{ \| \, \phi \rangle}{ \langle \phi \, \|}+{ \| \, \theta \rangle}{ \langle \theta \, \|}). \end{aligned}$$ \[eq:char\] The 3 vectors ${ \| \, \psi \rangle},{ \| \, \phi \rangle},{ \| \, \theta \rangle}$ span at most a 3 dimensional subspace of $\mathbb{C}^d$. Let ${ \| \, 0 \rangle},{ \| \, 1 \rangle},{ \| \, 2 \rangle}$ be the basis vectors of such a subspace and write: $$\begin{aligned} &{ \| \, \psi \rangle}={ \| \, 0 \rangle}, \\ &{ \| \, \phi \rangle}=\alpha{ \| \, 0 \rangle} +\beta{ \| \, 1 \rangle},\\ &{ \| \, \theta \rangle}=a{ \| \, 0 \rangle} + b{ \| \, 1 \rangle}+ c{ \| \, 2 \rangle}. \end{aligned}$$ Then $n=\|c\beta\|^2$. We observe that the parameter $m\in [-3,0 ]$ and $n\in [0,1]$ but since ${ \| \, \psi \rangle}{ \langle \psi \, \|}+{ \| \, \phi \rangle}{ \langle \phi \, \|}+{ \| \, \theta \rangle}{ \langle \theta \, \|}$ is a Hermitian operator, we need to focus only on the subdomain of $m$ and $n$ where all eigenvalues are real. Solving equation gives three eigenvalues $\lambda_1,\lambda_2,\lambda_3$ as functions of $m,n$. Let $\lambda_1$ be the largest of the eigenvalues. The maximal average success probability of the $3^d{\rightarrow}1$ QRAC is given by the expression \[eq:sumlambda\] \_[i,j,k]{}\_1(m\_[ijk]{},n\_[ijk]{}). From this formula one can easily see that this depends not only on unbiasedness of the basis, $\{m_{ijk}\}$, but also on the additional parameters $\{n_{ijk}\}$. Now, we show that in order to maximize the success probability of a $3^d{\rightarrow}1$ QRAC in general it is not sufficient to take arbitrary MUBs as measurements. In a $3^d{\rightarrow}1$ QRAC protocol, let us use 3 out of the $d+1$ mutually unbiased bases. There are ${d+1}\choose{3}$ such subsets. From numerical results we observe that the choice of bases subset of MUBs affects the final average success probability . As an example there are 20 distinct subsets of 3 MUBs in $d=5$. By direct calculation, using 10 of those subsets we obtain an average success probability of $0.610855$, and with the other 10 subsets $0.596449$. Hence the MUB condition alone does not guarantee the highest average success probability. We call this rather surprising behaviour, an anomaly. We observe this effect also in higher dimensions. We present our numerical observations for higher dimensions in Tab. \[fig:anomalytable\]. \[table:tabmub\] [cc\|c\|c\|c\|c\|c\|c\|c\|l]{} & &\ & & 5 & 7 & 8 & 9 & 11 & 13 & 16\ & & & & & & & &\ & && && & & &\ & & & & & & & &\ From our numerical analysis we have observed something else which might be of interest. Consider the subset of MUBs which give the highest observed average success probability. Here, for every pure state that Alice encodes, the success probability, $p(b=x_y)$, does not depend on the basis $y$ Bob uses to measure. In contrast, for the cases where a lower success probability was observed, the successful collapse probability of the encoded state depended on the measurement basis (and was therefore not uniform). We believe that this effect is somehow related to the complexity of higher dimension ${\mathbbm{C}}^d$ spaces. In general given $n$ vectors $\{\|v_i{\rangle}\}_i$ in ${\mathbbm{C}}^d$, it is not possible to find another normalized vector $\|\psi{\rangle}$ with equal overlap $\|{\langle}v_i \| \psi{\rangle}\|^2$ to all $n$ vectors. There is no obvious pattern which could suggest in which case those anomalies are present. It is surprising for example that in 8 dimensions we did not find this effect. One of the research directions in this topic is answering the question: are there other protocols which will be affected by this sensitivity for choosing different subsets of MUBs. Another question is: is there any pattern which could help us predict which cases of $n,d$ have a so-called MUB anomaly.\ The Unbiasedness Measure $\bar{Q}$ {#sec:Qbar} ================================== To obtain , we simply substitute the expressions $\bar{P}_{c,q}(2,d)$ of the optimal classical and quantum $2^d{\rightarrow}1$ Random Access Codes. Dropping the argument $\{\psi^i\}_i$, this yields: \|[Q]{} = \_[{a,b}S\^n\_2]{} ( \_[i,j]{} ( \|\^a\_i \| \^b\_j \| - ) ). \[eq:Qmeasure2\] We rewrite , to have a side-by-side comparison. \|[D]{}\^2 = \_[{a,b}S\^n\_2]{} ( 1 - \_[i,j]{} (\|\^1\_i \|\^2\_j\|\^2 -)\^2 ). Notice that $\bar{D}^2$ is a function of the ”probabilities" $\|{\langle}\psi^a_i \| \psi^b_j{\rangle}\|^2$, and tries to measure how much these values differ from the uniform distribution. This has a very aesthetic mathematical interpretation. However, it lacks a clear operational meaning which our proposed measure $\bar{Q}$ posseses. Finally, we point out that $\bar{D}^2$ is not a function of $\bar{Q}$ or vice-versa, except for the trivial cases ($0$ and $1$). See-Saw for $(n,2)^6 {\rightarrow}1$ pQRACs {#sec:seesaw} =========================================== Here we use the see-saw method and our unbiasedness measure ($\bar{Q}$ or $\bar{p}$) to add to the pre-existing evidence that $4$ mutually unbiased bases do not exist in dimension 6. In the see-saw method one interlaces two steps of SDP optimizations: For every other iteration one optimizes Alice’s preparation states for a given set of Bob’s measurements, and for the remaining steps one optimizes Bob’s measurements for a given set of Alice’s preparation states [@seesaw1]. Even though this method does not guarantee the convergence to the global maximum, it has proved to be efficient in Bell-type scenarios, see e.g. [@seesaw2]. The efficiency of the see-saw method can be improved in the following way. Let $\{ M_y^b\}$ be the set of measurements of Bob. Note that when Alice gets as input the information that Bob has either $y_1$ or $y_2$ setting, her strategy should be to send a state $\rho$ maximizing the value of $\operatorname{Tr}\left[ \rho \left( M_{y_1}^{x_{y_1}} + M_{y_2}^{x_{y_2}} \right) \right]$. This is obtained if the state is in the subspace of vectors with maximal eigenvalue of the operator $M_{y_1}^{x_{y_1}} + M_{y_2}^{x_{y_2}}$. Thus the see-saw step optimizing the states can be replaced with a construction of the states basing on eigenvectors decomposition of the measurement, cf. remark below . We have used the see-saw optimization for several cases of $(n,k)^d \rightarrow 1$ pQRACs. The results are given in Tab. \[tab:seesaw\]. case $(n,m)^d$ $\bar{D}^2$ $\bar{Q}$ ---------------- ----------------------- ----------------------- $(4,2)^3$ $\mathbf{0.999999}59$ $\textbf{0.999999}72$ $(5,2)^4$ $\mathbf{0.999999}22$ $\textbf{0.999999}39$ $(6,2)^5$ $\mathbf{0.999999}04$ $\textbf{0.999999}10$ $(3,2)^6$ $\mathbf{0.999998}69$ $\textbf{0.986390}54$ $(4,2)^6$ $\mathbf{0.998283}88$ $\textbf{0.998046}89$ $(4,2)^7$ $\mathbf{0.992371}97$ $\textbf{0.977929}16$ : The results of execution of the see-saw optimization for different pQRACs in search of $n$ MUBs in dimension $d$. The table shows the value of $D^2$ parameter for measurements obtained in optimizations and the success probability of the pQRAC with these measurement basis. We conclude that for $d = 3,4,5$ we have found the maximal amount of MUBs using the method. \[tab:seesaw\] Above we observed that for the $(4,2)^6$ pQRAC game to obtain the maximal guessing probability , the measurements of Bob has to be unbiased, and the value of is $1$. We will now use the Monte Carlo method to investigate the relation between the quantities and the games’ average success probability . Our results show that the majority of highly unbiased measurements in the meaning of gives large values of the $(4,2)^6$ pQRAC game, and vice versa. We have randomized instances of measurements of Bob in the following way. For each measurement we randomize a $6 \times 6$ complex matrix $A$ with each entry given by uniform distribution on the set $\{ x + {{i\mkern1mu}}y: x,y \in [0,1] \}$; then the SVD decomposition is performed giving matrices $U$, $S$ and $V$ satisfying $U S V^{\dagger} = A$, with $U$ and $V$ being unitary. We form the basis for the measurement by taking subsequent columns of the $U$ matrix. In our Monte Carlo experiment we randomized $10 000$ instances of random sets of measurements, calculated optimal states and the value of the $(4,2)^6$ pQRAC game, $\bar{p}$, , and the value of $\bar{D}^2$, . ![The values represents relative densities of the number of different basis giving similar values of the pair $(\bar{p}, D^2)$. \[fig:HSplot\]](HSplot.pdf) The meaning of this result is the following. If one considers a set of basis constructed in a natural way, then he can expect that the higher the unbiasedness, $D^2$, the higher the success probability $\bar{p}$ in the $(4,2)^6$ pQRAC game, and vice versa. The Spearman [@Spearman1; @Spearman2] rank correlation coefficient is a way of measuring the rank correlation between two (random) variables. It is a generalization of the Pearson coefficient, where the latter measures only linear dependence, and the former is able to express any monotonic relationship, with $0$ meaning no correlation and $\pm1$ meaning perfect monotonicity. We have calculated the value of Spearman rank correlation coefficient for the randomized data points and obtained that it’s value is about 0.91 meaning very strong dependence, but also displaying that these measures are subtly different. Estimating the Memory Requirements of the Navascues Vertesi hierarchy {#sec:estimate} ===================================================================== The random moment matrices $\Gamma^j_k$ are created by generating $A=\binom{n}{2} d^2$ random pure states $\{\rho^j_{\mathbf{x},z}\}$, and $B=nd$ projective measurement operators $\{M^{y,j}_b\}$. Let $w_k = w_k(\{ \{\rho^j_{\mathbf{x},z}\},\{M^{y,j}_b\}\})$ be the vector containing all strings of operators of length smaller than or equal to $k$. For convenience we relabel the indicess of the states $\{\rho^j_\alpha\}_{\alpha =1}^A $, and of the measurements $\{M^j_\beta\}_{\beta =1}^B$. Then, e.g. $w_1 = (\mathbbm{1}, \rho^j_1 , \ldots , \rho^j_{A} , M^{j}_1,\ldots , M^{j}_B)$, and $w_2$ would further contain the elements ($\rho_{\alpha_1}^j \rho^j_{\alpha_2}$, $\rho^j_{\alpha_1}M^j_{\beta_1}$, $M^j_{\beta_1}\rho^j_{\alpha_1}$, $M^j_{\beta_1}M^j_{\beta_2}$), $\forall \alpha_{1,2}\in [A] , \beta_{1,2}\in [B]$. The moment matrix is constructed as follows: ( \^j\_k )\_[i,j]{} = , with $\left( \cdot \right)_{i,j}$ corresponding to the components of the matrix, and similarly for vectors. We note that by construction $\left( \Gamma^j_k \right)_{i,j} = \left( \Gamma^j_k \right)_{j,i}$, and $\Gamma^j_k \geq 0$. We wish to estimate the size of ${\mathbb{S}}^k_d$. The moment matrices are $\|w_k\|$-by-$\|w_k\|$, and by construction we see that the length of the vectors $\|w_k\| = \sum_{i=0}^{k} (A+B)^i = O \left( (A+B)^k \right)$. In general, it is unknown how many matrices $\Gamma^j_k$ need to be generated until the set $\{\Gamma^j_k\}_j$ spans the full space ${\mathbb{S}}^k_d$. We can upper bound this number by $\frac{1}{2} \|w_k\|^2$ trivially (since the $\Gamma^j_k$ are real-symmetric). We note that e.g. for a $(4,2)^5{\rightarrow}1$ pQRAC the actual number of linearly independent matrices was just 6.5% lower than this trivial bound. Putting everything together, we see that the space ${\mathbb{S}}^k_d$ is described by roughly $O \left( (A+B)^{4k}\right)$ real parameters. For the case of 4 MUBs in dimension 6, $(A+B)=240\approx 2^8$. The $k$-th level of the hierarchy thus requires roughly $2^{32k}$ bits of memory just for indices and a further $2^6$ bits if each number is stored in double precision floating-point format. This exponential growth on memory requirements means that a typical 32-bit operating system (or mathematical software) cannot compute the first level of the hierarchy since it will run out of available memory indices. In fact, we encountered this problem when executing the program without symmetries in 32-bit OCTAVE. More drastically, a 64-bit system would be unable to calculate the second level, and all of the world’s storage space is insufficient for the third level (approximately $2^{74}$ bits of information were generated in 2016 [@worlddata]). Requirements considering symmetries.- Even if we restrict our optimization to the symmetric subspace $\mathcal{G}({\mathbb{S}}^k_d)$, the moment matrices are still of size $\|w_k\|$-by-$\|w_k\|$. Just as we do not know a priori the dimension of the space of feasible moment matrices, $\text{dim}({\mathbb{S}}^k_d)$, we cannot predict the dimension of the symmetric subspace, $\text{dim}(\mathcal{G}({\mathbb{S}}^k_d))$. Therefore, our proposal can at most be a quadratic improvement, i.e. of $O \left( (A+B)^{2k}\right)$. However, as we have portrayed, this reduction is sufficient to calculate the first level of the hierarchy as well as a subset of the $Q^{1+AB}$ level. Symmetries {#sec:Symmetries} ========== We find it easier to describe the symmetry group $\mathcal{G}$ by analyzing that there are two different types of possible relabelings. Let Bob have $n$ different work stations, and on each station he has a quantum measurement device. The success probability should not depend on which physical device is situated at which work station, so long as Alice is aware of which measurement basis is being used for every input $y$ of Bob. Likewise, Bob is free to relabel the outputs of every work station at will without dropping the success rate if Alice applies the same permutation to her inputs. In order to properly describe the group, we will use the picture where Bob first moves the measurement apparatuses around with a permutation $\omega \in \mathcal{S}_n$, and afterwards he relabels the device in work station $i$ with the permutation $\pi_i \in \mathcal{S}_d$. We stress that the permutations $\pi_1,\pi_2,\ldots$ address a specific work station - as opposed to a specific measurement device. Then, any relabeling of device outputs, and devices themselves can be achieved with these operations. We symbolically denote this abstract group element $g\in \mathcal{G}$ as: $$g = \pi_1 , \pi_2 , \ldots , \pi_n ; \omega .$$ If, $\tilde{g} = \tilde{\pi}_1,\ldots,\tilde{\pi_n};\tilde{\omega} \in \mathcal{G}$, then $\tilde{g} g \in \mathcal{G}$: $$\tilde{g}g = \tilde{\pi}_1\pi_{\tilde{\omega}^{-1}(1)} , \ldots , \tilde{\pi}_n\pi_{\tilde{\omega}^{-1}(n)} ; \tilde{\omega}\omega .$$ The inverse of $g$ is: $$g^{-1} = \pi^{-1}_{\omega^{-1}(1)}, \pi^{-1}_{\omega^{-1}(2)} , \ldots , \pi^{-1}_{\omega^{-1}(n)} ; \omega^{-1}.$$ More importantly though, is to see the action of $g$ on Bob’s measurements and Alice’s encoded states: $$\begin{aligned} g(M^y_b) &\mapsto M^{y'}_{b'} \\ g(\rho_{x_{z_1},x_{z_2},\{z_1,z_2\}}) &\mapsto \rho_{x'_{z_1'},x'_{z_2'},\{z_1',z_2'\}}) ,\end{aligned}$$ such that $$\begin{aligned} b' &= \pi_{y'} (b) \\ y' &= \omega(y) \\ x'_{z_i'} &= \pi_{z_i'} (x_{z_i'}) &{\text{ }}\forall i=1,2 \\ z_i' &= \omega(z_i) &{\text{ }}\forall i=1,2.\end{aligned}$$ It is therefore clear, that if we apply an element $g$ to the vector of operator strings $w_k$, it just acts as a permutation. The operators are the same, just re-indexed. Hence, we can represent the elements $g$ as the permutation matrices $G$. What is more, by construction we know that there is an order in which these elements are applied, and $G$ can be seen as a product of permutation matrices: $$G = \Pi_1 \Pi_2 \cdots \Pi_n \Omega ,$$ where $\Pi_i$ is the matrix representation of $\pi_i$, and similarly $\Omega$ is the matrix representation of $\omega$. When this representation acts on a moment matrix, it will first shift the indices of the measurement devices with $\Omega \Gamma^j_k \Omega^T$, and later it will reshuffle the output labels of each work station. Notice that the representations of the output labels commute amongst each other ($[\Pi_i,\Pi_j]=0$), but not with $\Omega$. We are now in a position to calculate the average of the moment matrix over the group representation . $$\label{eq:avgsymmDecomp} \hat{\Gamma}^j_k = \frac{1}{\|\mathcal{S}_n\|\|\mathcal{S}_d\|^n } \sum_{\Pi_1}\cdots \sum_{\Pi_n} \sum_{\Omega} \Pi_1 \cdots \Pi_n \Omega \Gamma^j_k \Omega^T \Pi_n^T \cdots \Pi_1^T$$ Crucially for implementations, the most important thing about is that the sum may be rewritten as: $$\sum_{\Pi_1} \Pi_1 \left( \cdots \left(\sum_{\Pi_n} \Pi_n \left( \sum_{\Omega} \Omega \Gamma^j_k \Omega^T \right)\Pi_n^T \right) \cdots \right) \Pi_1^T .$$ This reduces the amount of operations needed to perform the average from $\|\mathcal{S}_n\|\|\mathcal{S}_d\|^n$ to $\|\mathcal{S}_n\|+n\|\mathcal{S}_d\|$. These methods apply to a wide variety of SDP relaxation problems, and will be further discussed in [@symEP] in a more rigorous fashion.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We use the newly published 28 observational Hubble parameter data ($H(z)$) and current largest SNe Ia samples (Union2.1) to test whether the universe is transparent. Three cosmological-model-independent methods (nearby SNe Ia method, interpolation method and smoothing method) are proposed through comparing opacity-free distance modulus from Hubble parameter data and opacity-dependent distance modulus from SNe Ia . Two parameterizations, $\tau(z)=2\epsilon z$ and $\tau(z)=(1+z)^{2\epsilon}-1$ are adopted for the optical depth associated to the cosmic absorption. We find that the results are not sensitive to the methods and parameterizations. Our results support a transparent universe.' author: - 'Kai Liao[^1], Zhengxiang Li, Jing Ming, Zong-Hong Zhu' title: 'Testing cosmic opacity from SNe Ia and Hubble parameter through three cosmological-model-independent methods' --- Introduction ============ The unexpected dimming of Type Ia supernova (SNe Ia) is thought to be the evidence of acceleration of the universe [@acceleration]. In the frame of General Relativity (GR), the most famous explanation is the existence of dark energy with a negative pressure [@dark; @energy]. However, there are some issues on this plausible mechanism for observed SNe Ia dimming. The photon number conservation may be deviated. For example, it is due to the dust in our galaxy and oscillation of photons propagating in extragalactic magnetic fields into very light axions. These absorption, scattering or axion-photon mixing may lead to dimming [@issue]. Other mechanisms are widely proposed including modified gravity [@modified], dissipative processes [@dissipative], evolutionary effects in SNe Ia events [@evolution], violation of cosmological principle [@LTB] and so on. On the other hand, the deviation of photon number conservation is related to the correction of Tolman test [@Tolman] which is equivalent to measurements of the well-known distance-duality relation (DDR) [@DDR] $$\frac{D_{\scriptstyle L}}{D_{\scriptstyle A}}{(1+z)}^{-2}=1\;, \label{rec}$$ where $D_{\scriptstyle L}$ is luminosity distance, $D_{\scriptstyle A}$ is angular diameter distance and $z$ is redshift. The DDR is in fact a particular form of presenting the general theorem proved by Etherington known as the “reciprocity law” or “Etherington reciprocity theorem”. DDR holds for general metric theories of gravity in any cosmic background and it is valid for any cosmological models based on the Riemannian geometry. It is independent of gravity equation and the universe components. However, DDR may be not valid in the case that photons do not travel along null geodesics or the cosmic opacity exists. Many efforts have been done to test DDR though astronomical observations [@DDRobs]. Usually, they assume the form $\frac{D_{\scriptstyle L}}{D_{\scriptstyle A}}{(1+z)}^{-2}=\eta(z)\;$, where $\eta(z)=1+\eta_0z$ or $\eta(z)=1+\eta_0\frac{z}{1+z}$. Compared to conservation of photon number, the assumptions that the mathematical tool used to describe the space-time of universe is Riemannian geometry and photon travels along null geodesic are more fundamental and unassailable, thus the deviation of DDR most possibly indicates cosmic absorption. In this case, the flux received by the observer will be reduced by a factor $e^{-\tau(z)}$, and observed luminosity distance can be obtained by [@depth] $$D_{L,obs}=D_{L.true}e^{\tau(z)/2}, \label{rec}$$ where $\tau(z)$ is the optical depth related to the cosmic absorption. The relation between $\tau(z)$ and $\eta(z)$ is $e^{\tau(z)/2}=\eta(z)$ [@depth2]. Following this assumption, Avgoustidis et al. [@Avgoustidis] studied the difference between SNe Ia observations and Hubble parameter data. $H(z)$ data are mainly obtained through the measurements of differential ages of red-envelope galaxies known as “differential age method”. The aging of stars can be regarded as an indicator of the aging of the universe. The spectra of stars can be converted to the information of their ages, as the evolutions of stars are well known. Since the stars cannot be observed one by one at cosmological scales, people usually take the spectra of galaxies which contain relatively uniform star population. Moreover, $H(z)$ data can be obtained from the BAO scale as a standard ruler in the radial direction known as “Peak Method”. These methods are apparently independent of galaxy luminosity so that it will not be affected by cosmic opacity. However, SNe Ia observations are affected by many sources of opacity, such as the hosting galaxy, intervening galaxies, intergalactic medium, the Milky Way and exotic physics which affect photon conservation. Under the assumption $D_L=D_A(1+z)^{2+\epsilon}$, they investigated the cosmic opacity by confronting the standard luminosity distance in spatially flat $\Lambda$CDM model with the observed one from SNe Ia observations. Combining with the $H(z)$ data, which is not affected by transparency but yields constraints on $\Omega_m$, and marginalizing over all other parameters except $\epsilon$, they got $\epsilon=-0.04_{-0.07}^{+0.08}$ (2$\sigma$). Noticing that their method depends on cosmological models, Holanda et al. [@Holanda] further proposed a model-independent estimate of $D_L$ which are obtained from a numerical integration of $H(z)$ data. They also explored the influence of different SNe Ia light-curve fitters (SALT2 and MLCS2K2) and found a significant conflict. Based on Holanda et al., we proposed three model-independent methods that are different from theirs to explore the cosmic opacity. They firstly got the luminosity distances from $H(z)$ data (12 data) at corresponding redshifts (at $H(z)$ data) and gave a polynomial fit based on these 12 luminosity distances data with their errors, then calculated the values at the redshifts corresponding to SNe Ia through this polynomial fit curve. On the contrary, we use SNe Ia data to get the luminosity distances at the redshifts corresponding to $H(z)$ data through interpolation method, smoothing method and nearby SNe Ia method. Our data sets contains 28 available $H(z)$ data and the largest SNe Ia samples Union2.1 [@Union2.1]. The Letter is organized as follows: In Section 2, we introduce the method of obtaining luminosity distance from $H(z)$ data. In Section 3, we give the three methods that can convert SNe Ia luminosity distances to the luminosity distances at the redshifts corresponding to $H(z)$ data. In Section 4, the results are performed. Finally, we make a conclusion in Section 5. Luminosity distance from observational Hubble parameter data ============================================================ In this section, we introduce the method proposed by Holanda et al. [@Holanda]. The expression of the Hubble parameter can be written in this form $$H(z)=-\frac{1}{1+z}\frac{dz}{dt},$$ which depends on the differential age as a function of redshift. Based on Jimenez et al. [@Jimenez], Simon et al. [@Simon] used the age of evolving galaxies and got nine $H(z)$ data. Stern et al. [@Stern] revised these data at 11 redshifts from the differential ages of red-envelope galaxies. Gaztañaga et al. [@Gaztanaga] took the BAO scale as a standard ruler in the radial direction, obtained two data. Recently, Moresco et al. [@Moresco] obtained 8 data from the differential spectroscopic evolution of early-type galaxies as a function of redshift. Blake et al. [@Blake] obtained 3 data through combining measurements of the baryon acoustic peak and Alcock-Paczynski distortion from galaxy clustering in the WiggleZ Dark Energy Survey. Zhang et al. [@zhang] obtained another 4 data. Totally, we have 28 available data summarized in Table 1. Following Holanda et al. [@Holanda], we transform these 28 $H(z)$ data into luminosity distance. Using a usual simple trapezoidal rule, the comoving distance can be calculated by $$r=c \int_0^z{dz^\prime \over H(z^\prime)}\approx {c\over 2}\sum_{i=1}^{n} (z_{i+1}-z_i)\left[ {1\over H(z_{i+1})}+{1\over H(z_i)} \right].$$ Since we use much more data than Holanda et al. (they used only 12 data), this trapezoidal rule will work much better. In Fig. 1, we show the relative error with respect to the data number at a characteristic redshift $z=1$. We assume a standard $\Lambda$CDM model with $\Omega_m=0.3,\Omega_\Lambda=0.7$, and divide $z=1$ into different numbers of intervals averagely, then calculate the relative errors according to Eq. (4). We find that the relative errors decrease remarkably when the numbers of intervals increase from 12 to 28. With the standard error propagation formula, the error associated to the $i^{th}$ bin is given by $$s_i={c\over 2}(z_{i+1}-z_i)\left({\sigma_{H_{i+1}}^2\over H_{i+1}^4} + {\sigma_{H_{{i}}}^2\over H_{i}^4}\right)^{1/2}\;,$$ where $\sigma_{H_{i}}$ is the error of $H(z)$ data. The error corresponding to certain redshift is the sum of $s_i$. The Hubble constant $H_0=73.8 \pm 2.4$ km/s/Mpc [@h0] is used in our study. The 28 luminosity distance data from $H(z)$ data are shown in Fig. 2, as well as the $D_L$ from Union2.1 SNe Ia samples. ![Relative error with respect to the data number at $z=1$. $\Omega_m=0.3$, $\Omega_\Lambda=0.7$. ](simu.eps){width="12cm"} ![$D_L(z)$ obtained from measurements of the Hubble parameter data and the Union2.1 samples, respectively. []{data-label="DL"}](DL.eps){width="12cm"} Dealing with SNe Ia samples =========================== In this section, we introduce three methods through which we can obtain the luminosity distance of one certain SNe Ia point at the same redshift of the corresponding $H(z)$ data. nearby SNe Ia method -------------------- Since the SNe Ia Samples is much larger than $H(z)$ data, the nearby SNe Ia can be substituted for the one at the redshift corresponding to $H(z)$ data. Points $z_{SNe~Ia}-z_{H}$ are centered around the line $\Delta z=0$, as shown in Fig. 3 which plots the subtractions of redshifts between $H(z)$ data and the associated SNe Ia. Similar with the DDR test [@DDRobs], we have to choose a criterion based on the data. Our selection criterion is $\Delta z=\left\|z_{{H}}-z_{{SNe~Ia}}\right\|<0.003$. This selection criteria can be satisfied for most of the $H(z)$ data except for the points at $z=0.9$ and $z=1.037$ ($z=1.43, 1.53, 1.75$ are obviously ruled out) and small enough to reduce the systematic errors and guarantee the accuracy. ![Subtractions of redshifts between $H(z)$ data and the associated SNe Ia. ](zdistr.eps){width="12cm"} interpolation method -------------------- In order to avoid any bias brought by redshift incoincidence between $H(z)$ data and SNe Ia, as well as to ensure the integrity of the $H(z)$ data, we can use the nearby SNe Ia points to obtain the luminosity distance of SNe Ia point at the same redshift of the corresponding $H(z)$ data. This situation is similar with the cosmology-independent calibration of GRB relations directly from SNe Ia [@Interpolation]. When the linear interpolation is used, the distance modulus and the error can be calculated by $$\mu = \mu_i+(\mu_{i+1}-\mu_i)[(z-z_{i})/(z_{i+1}-z_i)],$$ $$\sigma_{\mu}=([(z_{i+1}-z)/(z_{i+1}-z_i)]^2\sigma_{\mu,i}^2+[(z-z_{i})/(z_{i+1}-z_i)]^2\sigma_{\mu,i+1}^2)^{1/2},$$ where subscripts $i$ and $i+1$ stand for the nearby data. We use the same method as used in [@Schaefer] to obtain the best estimate for each SNe Ia which is weighted average of all available distance moduli at the same redshift. The derived distance modulus for each SNe Ia is $$\mu = (\sum_i \mu_{\rm i} / \sigma_{\mu_{\rm i}}^2)/(\sum_i \sigma_{\mu_{\rm i}}^{-2}),$$ with its uncertainty $ \sigma_{\mu} = (\sum_i \sigma_{\mu_{\rm i}}^{-2})^{-1/2}$. smoothing method ---------------- We introduce a non-parametric method of smoothing supernova data over redshift using a Gaussian kernel in order to reconstruct luminosity distance [@smooth]. This procedure was initially used in the analysis of the cosmic large scale structure [@structure]. Through this model-independent method, we can extract information of various cosmological parameters, such as Hubble parameter, the dark energy equation of state, the matter density. Wu and Yu [@generalized] generalized this method to eliminate the impact of $H_0$ and obtained the evolutionary curve of luminosity distance using SNe Ia Constitution set and Union2 set. In this Letter, we follow this generalized method to get the luminosity distance curve using Union2.1 set. Firstly, we obtain the variable $\ln f(z)=\ln D_L(z)-\ln h$ through iterative method $$\ln f(z)_n^s=\ln\ f(z)_{n-1}^s+ N(z) \sum_i \left [ \ln f^{obs}(z_i)-\ln\ f(z)_{n-1}^s \right] {\large \times} \ {\rm exp} \left [- \frac{\ln^2 \left ( \frac{1+z}{1+z_i} \right ) }{2 \Delta^2} \right ],$$ where the reduced Hubble constant $h=H_0/100$ and the normalization factor $$N(z)^{-1}=\sum_i {\rm exp} \left [- \frac{\ln^2 \left ( \frac{1+z}{1+z_i} \right ) }{2 \Delta^2} \right ].$$ The value of parameter $\Delta$ was discussed by Shafieloo et al. [@smooth]. The results are not sensitive to $\Delta$. Following Shafieloo et al. [@smooth] and Wu and Yu [@generalized], $\Delta=0.6$ is used in this Letter. Eq. (9) can give the smoothed luminosity distance at any redshift z after n iterations. When $n=1$ $$\begin{split} \ln f(z)_1^s=\ln\ f(z)_0^s+ N(z) \sum_i \left [ \ln f^{obs}(z_i)-\ln\ f(z)_0^s \right] {\large \times} \ {\rm exp} \left [- \frac{\ln^2 \left ( \frac{1+z}{1+z_i} \right ) }{2 \Delta^2} \right ]\\ =\ln\ D_L(z)_0^s+ N(z) \sum_i \left [ \ln f^{obs}(z_i)-\ln\ D_L(z)_0^s \right] {\large \times} \ {\rm exp} \left [- \frac{\ln^2 \left ( \frac{1+z}{1+z_i} \right ) }{2 \Delta^2} \right ], \end{split}$$ where$D_L(z)_0^s$ depends on the suggested background cosmological model. Following Wu an Yu [@generalized], we adopt $\omega$CDM model with $\omega=-0.9$ and $\Omega_{m0}=0.28$ as the background model. The relation between $\ln f^{obs}(z_i)$ and $\ln D^{obs}_L(z_i)$ which is obtained from observed SNe Ia is $$\ln f^{obs}(z_i)={\ln 10\over 5} [\mu^{obs}(z_i)-42.38]=\ln D^{obs}_L(z_i)-\ln h,$$ $\mu^{obs}(z_i)$ is the observed distance modulus from SNe Ia. In order to know whether we get a best-fit value after some iterations, we calculate, after each iteration, $\chi^2_s$ $$\chi^2_{s,n}=\sum_i {(\mu(z_i)_n-\mu^{obs}(z_i))^2 \over \sigma^2_{\mu_{obs,i}}}\;.$$ The best-fit result is corresponding to the minimum value of $\chi^2_{s,n}$. The 1 $\sigma$ corresponds to $\Delta \chi^2_s=1$. The results For Union2.1 samples are shown in Fig. 4 and Fig. 5, the minimum value $n=32$. ![Computed $\chi_s^2$ for the reconstructed results at each iteration. ](chi2.eps){width="12cm"} ![Smoothed luminosity distance for Union2.1 samples. ](s.eps){width="12cm"} Constraints on cosmic opacity ============================= The observed distance modulus can be expressed as [@depth] $$\mu_{obs}(z)=\mu_{true}(z)+2.5[\log_{10} e] \tau(z) \, .$$ To examine the sensitivity of test results on the parametric form, we adopt two parameterizations, $\tau(z)=2\epsilon z$ and $\tau(z)=(1+z)^{2\epsilon}-1$ which are not strongly wavelength dependent on the optical band [@form]. $\epsilon$ here describes the cosmic opacity. The former one is linear and it can be derived from the DDR parameterization $D_L=D_A(1+z)^{2+\epsilon}$ for small $\epsilon$ and redshift. To constrain the value of $\epsilon$, we use the usual maximum likelihood method of $\chi^2$ fitting $$\chi^{2} = \sum_{i}\frac{(\mu_{obs}(i) - \mu_{true}(i)-2.5[\log_{10} e] \tau(i))^2} {\sigma^2_{\mu(obs)}+ \sigma^2_{\mu(true)}},$$ the subscript i stands for the data at the redshifts corresponding to $H(z)$ data. Our results are shown in Fig. 6, Fig. 7 and Fig. 8, respectively. For smoothing method, we consider two cases: $i_{max}=25$ and $i_{max}=28$ (containing the data at $z=1.43, 1.53, 1.75$). From the likelihood of $\epsilon$ using different methods, we can see current SNe Ia samples and $H(z)$ data support a transparent universe. These results are slightly different from Holanda et al. [@Holanda] while their results seem a little prone to a non-transparent universe especially with MLCS2K2 compilation. ![Likelihood function for $\epsilon$ using nearby SNe Ia method. ](nearby.eps){width="12cm"} ![Likelihood function for $\epsilon$ using interpolation method. ](Interpolation.eps){width="12cm"} ![Likelihood function for $\epsilon$ using smoothing method. ](smooth.eps){width="12cm"} Conclusion ========== Until now, modern cosmology has discovered many interesting phenomena behind which there exists underlying physical mechanisms. In principle, since the current astronomical observations are not precise enough to distinguish between different cosmological models, for example, various dark energy models are consistent with observations, we can explore all the possibilities among which the true one exists. Though the matter component in the universe is so diluted, photons will get though a huge space to observers, photon conservation can be violated by simple astrophysical effects or by exotic physics. Amongst the former, the attenuation is due to interstellar dust, gas, plasmas and so on. More exotic sources of photon conservation violation involve a coupling of photons to particles beyond the standard model of particle physics. Therefore, the concept of cosmic opacity should be considered naturally. In this Letter, we use the current observational Hubble parameter data which is opacity-free and SNe Ia observations which depends on cosmic opacity to test whether the universe is transparent. The results from three model-independent methods converge to a point that the effects of cosmic opacity is vanished. For future study on this problem, we think the wavelength is a considerable factor and more independent methods of testing cosmic opacity will confirm the conclusion. ------- -------- -------------- $z$ $H(z)$ $\sigma_{H}$ 0.07 69 19.6 0.09 69 12 0.12 68.6 26.2 0.17 83 8 0.179 75 4 0.199 75 5 0.2 72.9 29.6 0.24 79.69 3.32 0.27 77 14 0.28 88.8 36.6 0.352 83 14 0.4 95 17 0.43 86.45 3.27 0.44 82.6 7.8 0.48 97 62 0.593 104 13 0.6 87.9 6.1 0.68 92 8 0.73 97.3 7 0.781 105 12 0.875 125 17 0.88 90 40 0.9 117 23 1.037 154 20 1.3 168 17 1.43 177 18 1.53 140 14 1.75 202 40 ------- -------- -------------- : Current published observational Hubble parameter data ($km \cdot s^{-1} \cdot Mpc^{-1}$). []{data-label="tab:Hz"} Acknowledgments This work was supported by the National Natural Science Foundation of China under the Distinguished Young Scholar Grant 10825313, the Ministry of Science and Technology National Basic Science Program (Project 973) under Grant No.2012CB821804, the Fundamental Research Funds for the Central Universities and Scientific Research Foundation of Beijing Normal University. [\*]{} A.G. Riess et al., Astron. J. 116 (1998) 1009; S. Perlmutter et al, Astrophys. J. 517 (1999) 565. M. Tegmark et al., Phys. Rev. D, 69 (2004) 103501. P.J.E. Peebles and B. Ratra, ApJL, 325 (1988) 17; Z.K. Guo, Y.S. Piao, X. Zhang and Y.Z. Zhang, Phys. Lett. B 608 (2005) 177; M. Li, Phys. Lett. B, 603 (2004) 1. A. Aguirre, ApJ, 525 (1999) 583; C. Csaki, N. kaloper and J. Terning, Phys. Rev. Lett. 88 (2002) 161302. L. Randall and R. Sundrum, PRL, 83 (1999) 3370; J. S. Alcaniz, PRD, 65 (2002) 123514; R. Kerner, Gen. Rel. Gra., 14 (1982) 453; S. Capozziello, et al., PRD, 71 (2005) 043503; K. Liao and Z.H. Zhu, PLB, 714 (2012) 1. J.A.S. Lima, J.S. Alcaniz, ApJ, 566 (2002) 15. P.S. Drell., T.J. Loredo and I. Wasserman, ApJ, 530 (2000) 593; F. Combes, New Astronomy Rev, 48 (2004) 583. G. Lemaitre, Gen. Rel. Grav. 29 (1997) 641; G. Lemaitre, Ann. Soc. Sci. Bruxelles, A53 (1933) 51; R.C. Tollman, Proc. Nat. Acad. Sci. 20 (1934) 169; H. Bondi, Mon. Not. Roy. Astron. Soc. 107 (1947) 410. R.C. Tolman, Proc. Natl. Acad. Sci. 16 (1930) 511. I.M.H. Etherington, Phil. Mag. 15 (1933) 761; I.M.H. Etherington, Gen. Rel. Grav. 39 (2007) 1055; G.F.R. Ellis, Gen, Rel, Grav. 39 (2007) 1047. R.F.L. Holanda, J.A.S. Lima and M.B. Ribeiro, ApJ, 722 (2010) L233; Z. Li, P.X. Wu and H.W. Yu, ApJ, 729 (2011) L14; X. Meng, T. Zhang, H. Zhan and X. Wang, ApJ, 745 (2012) 98; R.F.L. Holanda, J.A.S. Lima and M.B. Ribeiro, Astron. Astrophys. 538 (2012) A131. S. More, J. Bovy and D.W. Hogg, ApJ, 696 (2009) 1727; B. Chen and R. kantowski, PRD, 79 (2009) 104007; B. Chen and R. kantowski, PRD, 80 (2009) 044019. J.A.S. Lima, J.V. Cunha and V.T. Zanchin, ApJ, 741 (2011) L26. A. Avgoustidis, C. Burrage, J. Redondo, L.Verde and R. Jimenez, JCAP, 1010 (2010) 024. R.F.L. Holanda et al., 2012, arXiv: 1207.1694. N. Suzuki et al., ApJ, 746 (2012) 85. R. Jimenez, et al., ApJ, 593 (2003) 622. J. Simon, et al., PRD, 71 (2005) 123001. D. Stern, et al., JCAP, 02 (2010) 008. E. Gaztañaga, A. Cabré, and L. Hui. MNRAS, 399 (2009) 1663. M. Moresco et al., JCAP, 08 (2012) 006. C. Blake et al., MNRAS, 425 (2012) 405. T. Zhang et al., 2012, arXiv:1207.4541. A.G. Riess et al., ApJ, 730 (2011) 119. N. Liang et al., ApJ, 685 (2008) 354; N. Liang, P. Wu and S.N. Zhang, PRD, 81 (2010) 083518; B.E. Schaefer, ApJ, 660 (2007) 16. A. Shafieloo, MNRAS, 380 (2007) 1573; A. Shafieloo et al., MNRAS, 366 (2006) 1081. V.J. Martinez and E. Saar, 2002, Statistics of the Galaxy Distribution (London: Chapman and Hall). P. Wu and H. Yu, JCAP, 2 (2008) 19; X.Y. Fu, P. Wu, H. Yu and Z.X. Li, Research in Astron. Astrophys. 8 (2011) 895. A. Avgoustidis et al., JCAP, 0906 (2009) 012. [^1]: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'Calculation of the Floquet quasi-energies of a system driven by a time-periodic field is an efficient way to understand its dynamics. In particular, the phenomenon of dynamical localization can be related to the presence of close approaches between quasi-energies (either crossings or avoided crossings). We consider here a driven two-level system and study how the locations of crossings in the quasi-energy spectrum alter as the field parameters are changed. A perturbational scheme provides a direct connection between the form of the driving field and the quasi-energies which is exact in the limit of high frequencies. We firstly obtain relations for the quasi-energies for some common types of applied field in the high-frequency limit, and then show how the locations of the crossings drift as the frequency is reduced. We find a simple empirical formula which describes this drift extremely well in general, and which we conjecture is exact for the specific case of square-wave driving.' author: - 'C.E. Creffield' title: 'Location of crossings in the Floquet spectrum of a driven two-level system.' --- [^1] Introduction ============ The two-level system is a simple model which has been applied to a great variety of physical problems. One application of growing importance is that of quantum computing [@divincenzo], since any quantum two-level system has the potential to act as a quantum bit. For this reason the coherent control of quantum states in these systems has recently become the focus of intense investigation [@vion; @cole]. A concrete example of such a two-level system is provided by a particle tunneling between two potential wells, which can be experimentally realized by confining an electron to a pair of coupled quantum dots [@klitzing]. One method of controlling such a system, without destroying its coherence, is to apply oscillatory electric fields [@leo_dots]. Such fields are able to produce the phenomenon known as [coherent destruction of tunneling]{} (CDT), in which the tunneling of the particle is suppressed when the parameters of the field are tuned to various “magic” values. As the applied field is time-periodic, Floquet analysis [@shirley] has been applied to explain this non-intuitive result, and it has been shown [@hanggi_1; @hanggi_2] that CDT is closely related to the presence of crossings or avoided crossings in the spectrum of Floquet quasi-energies. The driving field most frequently considered is of sinusoidal form, and studies using CDT as a means of quantum-control have generally concentrated on varying either the envelope [@holthaus_pulse] or frequency [@holthaus_chirp] of a sinusoidal signal. In this work, however, we instead consider the effect of altering the signal’s [waveform]{}. By using a perturbational method we first show how the waveform can be directly related to the quasi-energy spectrum, and give analytic results for sinusoidal, square-wave and triangular waveforms. These results are precise in the limit of high frequency. As the frequency is reduced, however, the locations of the crossings drift away from these values. This effect is extremely difficult to treat analytically, and such efforts [@barata; @frasca; @delgado] produce complicated results which are difficult to interpret. Empirically, however, we find a simple formula which describes the drifting with good accuracy for many waveforms, and appears to be [exact]{} for the case of the square-wave. We thus provide a means for predicting the locations of quasi-energy crossings for a given driving field in both high and low frequency regimes. Method ====== Model Hamiltonian ----------------- We consider a charged particle confined to a double quantum-dot system, described by the Hamiltonian: $$H = {\tilde t} \left( c_L^{\dagger} c_R^{ } + h.c. \right) + \left( E_L(t) n_L + E_R(t) n_R \right) \label{ham_tunn}$$ where the subscript $L/R$ denotes the left/right quantum dot, $c_j^{\dagger}$ and $c_j^{ }$ are creation and annihilation operators for a particle in dot $j$, and $n_j =c_j^{\dagger}c_j^{ } $ is the usual number operator. The tunneling between the two dots is described by the hopping parameter $\tilde t$, and $E_j(t)$ is the electrical potential of the external driving field. Clearly only the potential difference between the two dots is physically of importance, and so we can use the symmetric parameterization: $$E_L = \frac{E}{2} f(t), \quad E_R = -\frac{E}{2} f(t)$$ where $E$ is the potential of the driving field and $f(t)$ is a $T$-periodic function describing its waveform. The Hamiltonian (\[ham\_tunn\]) has been written using a basis of [localized]{} states, but it may be easily transformed to the standard two-level form via a SU(2) rotation, yielding the result: $$H = \frac{\Delta}{2} \sigma_z \ + \ \frac{E}{2} f(t) \sigma_x, \label{ham_2lev}$$ where $\sigma_i$ are the standard Pauli matrices. In this representation the basis states used are [extended]{} states, formed by symmetric and anti-symmetric combinations of the localized states. In the absence of a driving field ($E=0$) it is clear that the two eigenstates of this Hamiltonian consist of a symmetric ground state, and an excited anti-symmetric state. The splitting between these two levels is given by $\Delta$, which is related to the inter-dot tunneling via $\Delta = 2 {\tilde t}$. \[proceed\] Floquet theory -------------------------- As the function $f(t)$ is periodic in time, the Floquet theorem may be used to write solutions of the time-dependent Schrödinger equation as $\psi(t) = \exp[-i \epsilon_j t] \phi_j(t)$, where $\phi_j(t)$ is a function with the same periodicity as $f(t)$ and is called a Floquet state, and $\epsilon_j$ is termed the quasi-energy. Although the Floquet states have an explicit time-dependence, their periodicity means that the dynamics of the system on time-scales much larger than the period of the driving field is effectively given [only]{} by the quasi-energies. In particular, if the two quasi-energies approach degeneracy, the dynamics of the system on this time-scale will appear to be frozen, producing the effect of CDT. Consequently, determining the quasi-energies provides a simple and direct way of studying the long time-scale behavior of the system, and indicates whether CDT can occur [@caveat]. In this work we restrict our attention to driving functions which possess the symmetry $f(t) = - f(t + T/2)$. Imposing this restriction means that the Hamiltonian (\[ham\_tunn\]) is invariant under the generalized parity operation $x \rightarrow -x, t \rightarrow t + T/2$, and as a consequence the two Floquet states will also possess this symmetry, one being even and the other being odd. The von Neumann-Wigner theorem [@neumann] thus allows the two quasi-energies to cross as an external parameter, such as the field strength, is varied. Breaking this symmetry by choosing an alternative form for the driving field would mean that the quasi-energies would be forbidden to cross, and thus close approaches between the quasi-energies could only consist of avoided crossings. The Floquet states and their quasi-energies may be conveniently obtained from the eigenvalue equation: $$\left[ H(t) - i \frac{\partial}{\partial t} \right] \phi_j(t) = \epsilon_j \phi_j(t) . \label{floquet}$$ To obtain approximate solutions to this equation we follow a perturbation scheme introduced originally by Holthaus [@holthaus] to treat both the two-level system and driven superlattices, and which was generalized recently to also include the effects of inter-particle interactions [@creffield]. In this approach the Hamiltonian (\[ham\_tunn\]) is divided into two parts: $H_t$ which contains the tunneling terms, and $H_I$ which holds the electric field terms. We then find the eigensystem of the operator ${\cal H}_I(t) = H_I - i {\partial}/{\partial t}$ by working in an [extended]{} Hilbert space of time-periodic functions [@sambe], and apply the tunneling Hamiltonian as a perturbation. A consequence of dividing the Hamiltonian in this way is that the perturbation theory works well in the high-frequency limit $\omega \gg {\tilde t}$, but breaks down in the opposite limit when the tunneling provides the dominant energy-scale of the problem [@hanggi_2]. For the Hamiltonian given in Eq.\[ham\_tunn\], the problem of finding the eigensystem of ${\cal H}_I(t)$ simply requires the solution of two uncoupled differential equations: $$\begin{aligned} \left( -\frac{E}{2} f(t) - i \frac{d}{dt} \right) \phi_+(t) &=& \epsilon_+ \phi_+(t), \\ \left( \frac{E}{2} f(t) - i \frac{d}{dt} \right) \phi_-(t) &=& \epsilon_- \phi_-(t).\end{aligned}$$ These can be integrated immediately, giving the solutions: $$\begin{aligned} \phi_{\pm}(t) &=& \exp[\pm i E F(t)/2] \exp[i \epsilon_{\pm} t], \\ \mbox{where} \ \ F(t) &=& \int_0^{t} f(t') dt' . \label{sols}\end{aligned}$$ The periodicity of the Floquet states clearly requires that $\epsilon_{\pm} = 0 \ \mbox{mod} \ \omega$. Without loss of generality we can restrict the quasi-energies to lie in the “first Brillouin zone” ($- \omega/2 \leq \epsilon < \omega / 2$), and thus to lowest order in the perturbation theory they are degenerate and zero. Standard degenerate perturbation theory can now be used to evaluate the first-order correction to the quasi-energies, requiring only that we work in the the extended Hilbert space of $T$-periodic functions by defining an appropriate scalar product: $$\langle \langle \phi_m \| \phi_n \rangle \rangle_T = \frac{1}{T} \int_0^{T} \langle \phi_m(t') \| \phi_n(t') \rangle dt'$$ where $\langle \cdot \| \cdot \rangle$ is the usual scalar product for the spatial component of the wavefunctions, and $\langle \cdot \| \cdot \rangle_T$ denotes the integration over the compact time coordinate. As the tunneling component of the Hamiltonian $H_t$ is acting as the perturbation, the first-order approximation to the quasi-energies is given by the eigenvalues of the perturbing matrix: $$\langle \langle H_t \rangle \rangle_T = \pmatrix{ &0 \ & {\tilde t} \ \langle \phi_-^2 \rangle_T \cr & {\tilde t} \ \langle \phi_+^2 \rangle_T \ & 0} \label{effective}$$ Comparing this expression with the original tunneling Hamiltonian (\[ham\_tunn\]) reveals that the action of the applied field is to renormalize the tunneling terms by the factors $\langle \phi_{\pm}^2 \rangle_T$. As $\phi_+$ is the complex conjugate of $\phi_-$, the quasi-energies take the simple form: $$\epsilon_{\pm} = \pm \frac{\Delta}{2} \left\| \langle \phi_+^2 \rangle_T \right\| , \label{quasi}$$ where $$\langle \phi_+^2 \rangle_T = \frac{1}{T} \int_0^T \exp[i E F(t)] dt \label{central}$$ and $F(t)$ is defined in Eq.\[sols\]. Clearly the quasi-energies can only become degenerate when they are both equal to zero, and we can note from Eq.\[effective\] that this corresponds, as expected, to the destruction of the effective tunneling. Results ======= To obtain the Floquet quasi-energies for comparison with the prediction of Eq.\[central\], the numerical technique described in Ref.[@creffield] was used. This involves evaluating the unitary evolution operator for one period of the field $U(T,0)$ and obtaining its eigenvalues, which are related to the quasi-energies via $\lambda_j = \exp[-i \epsilon_j T]$. Using this method to obtain the quasi-energies, a standard bisection algorithm could then be used to find the location of the quasi-energy crossings to a high degree of accuracy. The dynamical behavior of the system was also examined directly by integrating it over long time-periods, with the particle initially located in the left quantum dot. To quantify to what extent the tunneling between the left and right quantum dots was destroyed, the probability that the particle was in the left quantum dot ($P_L(t)$) was measured throughout the time evolution. We denote the minimum value of $P_L$ attained during this period to be the “localization”, and thus high values of localization correspond to the presence of CDT, while low values reveal that the particle is able to tunnel from one side to the other, and is therefore delocalized. Sinusoidal driving ------------------ We begin with the most familiar case, when the driving field has the form $f(t) = \cos \omega t$. The procedure outlined in Section \[proceed\] can be followed straightforwardly, leading to the result that: $$\langle \phi_{+}^2 \rangle_T = \frac{1}{T} \int_0^T \exp[i E \sin (\omega t) / \omega] \ dt .$$ By making use of the standard identity: $$\exp[i E \sin (\omega t) / \omega] = \sum_{m=-\infty}^{\infty} J_m \left( E / \omega \right) \exp[i m \omega t]$$ this expression can be substantially simplified, yielding the final result that $\epsilon_{\pm} = \pm (\Delta / 2) \ J_0(E / \omega)$. This reproduces the well-known result that for sinusoidal driving CDT occurs when the ratio of the field strength to its frequency is equal to a root of the Bessel function $J_0$. In Fig.\[j0\]a the locations of the quasi-energies are shown for a fixed frequency $\omega = 8$ as a function of $E/\omega$. It can be seen that the perturbative result works extremely well in this regime (high-frequency). Fig.\[j0\]b shows the localization produced by the field, as defined above. As expected, at the points where the quasi-energies cross the tunneling dynamics of the system is blocked, producing sharp spikes in the localization, centered on the crossings. To investigate how the crossings move away from these points as the driving frequency is reduced, their locations are shown as a function of $1/\omega$ in Fig.\[drifts\]a. In accordance with the von Neumann-Wigner theorem [@neumann; @hanggi_2] we can readily see that the set of crossings form one-dimensional manifolds. As $\omega$ tends to infinity the crossings occur at the roots of $J_0$, as predicted by the perturbation theory, and this remains a good approximation for frequencies as low as $\omega = \Delta$. Below this value, however, the crossings smoothly drift away from these locations, and evolve towards the points $\Delta / \omega = 2 n$ (where $n$ is a positive integer), as was seen earlier in Ref.[@hanggi_2]. This limiting behavior in the low-frequency regime was also predicted in Ref.[@ulloa], where a similar pattern of crossing-drift was observed in an investigation of a related model. The form of Fig.\[drifts\]a immediately suggests fitting the manifolds of crossings with quadrants of ellipses: $$\left( \frac{E / \omega}{y_n} \right)^2 + \left( \frac{\Delta / \omega}{2 n} \right)^2 = 1 , \label{fitting}$$ where $y_n$ is the $n$-th root of $J_0(y)$. It can be seen in Fig.\[deviation\] that this simple parameterization fits the results extremely well for the first crossing-manifold, and that the difference between the exact location of the crossing and the fitting function ($E_{fit}/\omega - E_{exact}/\omega$) never exceeds a value of $0.02$. The degree of deviation becomes larger as the order of the crossing increases, but nonetheless is only visible in Fig.\[drifts\]a for the fourth and fifth crossing-manifolds. In Fig.\[arc\] the localization is plotted as $\omega$ is reduced from a high value towards zero, with $E$ set to hold the ratio $E/\omega$ on a crossing-manifold. For each point the system was evolved over 200 periods of the driving field to study how effectively the field could maintain a localized state. For the high-frequency regime, $\omega \geq \Delta$, the localization is excellent at all the crossings, with less than 0.1 of the particle density leaking across to the right-side dot during the time-evolution. As can be expected, the high-order crossings, which occur at higher values of $E$, can maintain better levels of localization than the low-order crossings [@hanggi_2]. This difference becomes more pronounced as the frequency is reduced, and although the localization in all cases decays smoothly to zero, the localization at the higher-order crossings decays much more slowly. For frequencies as low as $\omega=0.4 \Delta$, however, the inhibiting effect of CDT is still evident for all the crossings, indicating that even low-frequency fields may serve a useful role in stabilizing electron-leakage from quantum dot devices. Square-wave driving ------------------- Square-wave driving has been considered to a lesser extent than the sinusoidal case, although it is also an easily realizable waveform in experiment. Ref.[@zhao_square] investigated the case of a superlattice driven by a square-wave field, and found that for suitable choices of parameters CDT would indeed occur, while sinusoidal driving of this system could only produce partial CDT [@zhao_bloch]. Recently in Ref.[@dignam] it has been shown that in a superlattice CDT can only be produced if the crossings of the quasi-energies are [equally]{} spaced, which clearly does not occur for sinusoidal driving. For this reason it is of interest to derive the behavior of the quasi-energies for square-wave driving to see explicitly how this condition is fulfilled. We consider the square-wave driving field, $f(t) = \Theta(t) - 2 \Theta(t - T/2)$, defined over the interval $0 \leq t < T$. The integrations required to obtain the quasi-energies may again be done straightforwardly, giving the result that: $$\epsilon_{\pm} = \pm \frac{\Delta}{2} \ \frac{ \sin \left( \pi E / 2 \omega \right) }{\pi E / 2 \omega} . \label{pert_sq}$$ From this it is immediately clear that the crossings are equally spaced as required, being given by the condition $E / \omega = 2 n$ where $n$ is a positive integer. In Fig.\[sinc\]a the quasi-energies obtained for a frequency of $\omega=8$ are shown in comparison with the above result, and it can be clearly seen that the agreement is excellent. Below this figure is plotted the localization produced by the field, and as for the case of sinusoidal driving, the crossings of the quasi-energies correspond to sharp spikes in the localization, verifying that CDT is indeed occurring. In Fig.\[drifts\]b the drifting of the crossings as the frequency is reduced is shown. The behavior is strikingly similar to that observed for sinusoidal driving, and accordingly we use the same functional form (\[fitting\]) to fit the crossing-manifolds, with the y-intersections now given by $y_n = 2 n$. The fit is so good that on this plot no differences can be seen between the exact results and the fits. This is corroborated by Fig.\[deviation\], where the deviation from the exact result for the lowest manifold can be seen to be negligible in comparison with the sinusoidal case, and within the accuracy of the numerical procedures the fit is [identical]{} with the exact result. We therefore conjecture that this fitting is, in fact, exact for the case of square-wave driving. We also show on this plot the result obtained for a bandwidth-limited square-wave, obtained by truncating its Fourier expansion at two terms. We see that the addition of just the second term to the sinusoidal driving already reduces the deviation of the fit from the exact result considerably. Truncating the series at higher points produces steady improvements in the fit, strongly supporting the conjecture that the fit is exact when all terms are included. Triangular driving ------------------ We now consider another easily obtainable form of driving, the triangular waveform: $$f(t) = \left\{ \begin{array}{r@{\quad \quad}l} 1 - 4 t / T & \mbox{for} \ 0 \leq t \leq T/2 \\ -3 + 4 t/T & \quad \ T/2 < t \leq T . \end{array} \right.$$ For this case a closed form solution can again be obtained for the behavior of the quasi-energies, involving the Fresnel sine and cosine functions, $S(x)$ and $C(x)$. The full expression for the quasi-energies is given by: $$\epsilon_{\pm} = \frac{\Delta}{\sqrt{2 x}} \left[ \cos(x \pi/4) C(\sqrt{x/2}) + \sin(x \pi/4) S(\sqrt{x/2}) \right] \label{monster}$$ where $x = E / \omega$. In Fig.\[tri\] it can be seen that this function is indeed an excellent approximation to the true quasi-energies, and that CDT again occurs at the points of quasi-energy crossings. The roots of Eq.\[monster\] may be found numerically, yielding the result that the first three crossings occur when $E/\omega = 2.92519, \ 7.02525$ and 10.9864. Observing the behavior of the Fresnel functions [@abram] reveals that for $x > 1$ they both make small amplitude, decaying oscillations about a value of 0.5, which allows the condition for crossings to be written in the simpler, though approximate, form $\tan (x \pi/4) \simeq -1$. The crossing condition therefore reduces to the simple result $E / \omega \simeq 4 n - 1$, as may be seen from the exact values given above, which becomes increasingly accurate for larger values of $E / \omega$. In Fig.\[drifts\]c it can be seen that the crossing-manifolds for this form of driving have a similar elliptical form to the previous cases. Using the same fitting function (\[fitting\]) as before, with the y-intercepts given by the roots of Eq.\[monster\], gives an accurate description of their behavior, as may be seen in Fig.\[deviation\]. Although the fit is not as good as for the sinusoidal case, the maximum deviation is still less than $0.04$. As seen previously, the fit is best for the lowest-order manifolds, with small deviations being visible in the higher-order manifolds. Nonetheless, in all cases the fitting function gives an impressively accurate approximation to the true result. Conclusions =========== In summary, it has been shown how changing the waveform of a periodic driving-field can be used to modify the location of the quasi-energies of a two-level system. A procedure has been given which relates the waveform explicitly to the quasi-energy spectrum, allowing the positions of the quasi-energy crossings to be located exactly in the limit of high frequency. For various driving fields, including the cases we consider here, an analytic form can be obtained for the quasi-energies, and in other cases they may be obtained numerically with little difficulty. This gives the prospect of designing the waveform to create a desired behavior of the quasi-energy spectrum in a direct and straightforward way. It has also been shown how the positions of the quasi-energy crossings drift as the frequency is reduced from the high-frequency limit. For the driving fields considered here, the crossings fall approximately onto elliptical manifolds, and for the case of square-wave driving it appears that this description is exact. We have examined this behavior for many other waveforms, and we conclude that this form of the crossing-manifolds is very general. Using the perturbation theory to find the crossings in the high-frequency limit, and then making use of this drifting behavior, allows the positions of the quasi-energy crossing to be accurately located in [all]{} regimes of driving. This gives more flexibility in experiment, as the high-field regime may either be difficult to attain, or may induce undesirable transitions to higher energy levels, breaking the two-level approximation. Although the degree of localization that the field can maintain is reduced in the low-frequency regime, it can still produce a useful reduction of the leakage from quantum dot devices, and thereby enhance their decoherence time, which has many possible applications to the coherent control of mesoscopic systems. This research was supported by the EU through the TMR programme “Quantum Electron Transport in the Frequency and Time Domains”. The author thanks Gloria Platero for discussions, and acknowledges the hospitality of the International Institute for Applied Systems Analysis (IIASA) in Vienna, where part of this work was carried out. [^1]: Current address: Dipartimento di Fisica, Università di Roma “La Sapienza”, Piazzale Aldo Moro 2, I-00185, Roma, Italy.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We measure the faint end slope of the galaxy luminosity function (LF) for cluster galaxies at $1 < z < 1.5$ using [Spitzer]{} IRAC data. We investigate whether this slope, $\alpha$, differs from that of the field LF at these redshifts, and with the cluster LF at low redshifts. The latter is of particular interest as low-luminosity galaxies are expected to undergo significant evolution. We use seven high-redshift spectroscopically confirmed galaxy clusters drawn from the IRAC Shallow Cluster Survey to measure the cluster galaxy LF down to depths of $M^* + 3$ (3.6$\mu$m) and $M^* + 2.5$ (4.5$\mu$m). The summed LF at our median cluster redshift ($z=1.35$) is well fit by a @schechter distribution with $\alpha_{3.6\mu\text{m}} = {-0.97}\pm {0.14}$ and $\alpha_{4.5\mu\text{m}} = {-0.91}\pm {0.28}$, consistent with a flat faint end slope and is in agreement with measurements of the field LF in similar bands at these redshifts. A comparison to $\alpha$ in low-redshift clusters finds no statistically significant evidence of evolution. Combined with past studies which show that $M^$ is passively evolving out to $z \sim 1.3$, this means that the shape of the cluster LF is largely in place by $z \sim 1.3$. This suggests that the processes that govern the build up of the mass of low-mass cluster galaxies have no net effect on the faint end slope of the cluster LF at $z \la 1.3$.' author: - 'Conor L. Mancone, Troy Baker, Anthony H. Gonzalez, Matthew L. N. Ashby, Spencer A. Stanford$^,$, Mark Brodwin, Peter R. M. Eisenhardt, Greg Snyder, Daniel Stern, Edward L. Wright' bibliography: - 'galaxy\_lf.bib' title: The Faint End of the Cluster Galaxy Luminosity Function at High Redshift --- Introduction {#sec:intro} ============ Many studies have shown that the low-mass cluster-galaxy population evolves substantially at low redshift. For instance, @cowie96 first recognized that star formation happens primarily in high-mass systems at high redshift and low-mass systems at low redshift, a fact which has been studied extensively since (see, for example, @panter07; @mobasher09; @chen09; @villar11). Moreover, it is well known that cluster galaxies undergo morphological transformation at low redshift, with many cluster members transforming to lenticular galaxies at low redshift [@dressler97; @desai07; @wilman09]. There is also substantial evidence that the low-luminosity red-sequence galaxy population grows substantially in clusters since at least $z \sim 1$ [@stott07; @lu09; @rudnick09; @lemaux12]. Taken together, these facts demonstrate that the low-mass cluster galaxies are actively evolving and forming since $z \sim 1$. Therefore, by comparing low mass, $z=0$ galaxies with their high-redshift progenitors we can potentially constrain the processes important in galaxy formation and evolution. This can be done by studying individual galaxies (through their star formation rates, stellar masses, morphological types, and structural properties) or by studying galaxy populations (through their luminosity and mass functions). In particular, the near-infrared luminosity function (NIR LF) can be used to study the stellar mass growth of a galaxy population, as the rest-frame NIR is a good proxy for stellar mass [@muzzin08]. In clusters, the NIR LF has been used extensively to study the assembly of the most massive cluster galaxies. Such studies have found that the massive end of the NIR LF evolves passively out to $z \sim 1.3$, suggesting that the bulk of the stellar mass of these galaxies is in place at high redshift [@andreon06; @strazzullo06; @depropris07; @muzzin08; @mancone10]. In addition, in @mancone10 we found statistically significant deviations from passive evolution at $z > 1.3$ which we could only explain with ongoing stellar mass assembly at these redshifts. [ccccc]{} ISCS J1432.4+3332 & 14:32:29.18 & 33:32:36.0 & 1.112 & 26\ ISCS J1434.5+3427 & 14:34:30.44 & 34:27:12.3 & 1.238 & 19\ ISCS J1429.3+3437 & 14:29:18.51 & 34:37:25.8 & 1.261 & 18\ ISCS J1432.6+3436 & 14:32:38.38 & 34:36:49.0 & 1.351 & 12\ ISCS J1433.8+3325 & 14:33:51.13 & 33:25:51.1 & 1.369 & 6\ ISCS J1434.7+3519 & 14:34:46.33 & 35:19:33.5 & 1.374 & 10\ ISCS J1438.1+3414 & 14:38:08.71 & 34:14:19.2 & 1.414 & 16\ Most attempts to probe the faint end of the cluster luminosity function (LF) at high redshift have been limited to studying the red sequence. Such studies have found a deficit of faint and red cluster members at high redshift when compared to their low-redshift counterparts [@delucia04; @stott07; @lu09; @rudnick09; @lemaux12]. This could mean that low-mass cluster galaxies undergo substantial mass growth at low redshift, or simply that low-mass cluster galaxies are still blue at high redshift and have not finished transitioning onto the red sequence [@lemaux12]. Differentiating between these two cases requires measuring the LF of all faint cluster members. Previously, @strazzullo10 was the only study to do this, finding a faint-end slope consistent with flat. However, they did not compare their results to low-redshift clusters to determine the implications for the stellar mass growth of low-mass cluster galaxies. In this paper we measure the 3.6 and 4.5 $\mu$m LF of high redshift ($1 < z < 1.5$) galaxy clusters. Our measurements trace the rest-frame NIR, where the LF is known to correlate well with stellar mass. Most importantly, our data are deep enough to constrain $\alpha$, the faint-end slope of the LF. This, combined with low-redshift results from the literature, allows us to measure the stellar mass buildup of the low-mass cluster galaxy population over a redshift range when these galaxies are known to be actively evolving. This paper is structured as follows. Section \[sec:data\] describes our data. Section \[sec:lf\] presents our method for measuring the galaxy cluster luminosity function and gives our results. In Section \[sec:discussion\] we compare our results to low-redshift clusters and the field. Our conclusions are presented in Section \[sec:conclusions\]. All magnitudes are on the Vega system, and we assume a WMAP 7 cosmology (@komatsu11; $\Omega_m=0.272, \Omega_\Lambda=0.728, h=0.704$) throughout. All SPS model predictions are generated using [EzGal]{} [@mancone12]. Data {#sec:data} ==== Cluster Sample -------------- The clusters from this study are part of the IRAC Shallow Cluster Survey (ISCS) [@stanford05; @elston06; @brodwin06; @eisenhardt08], a catalog of clusters identified as 3-D overdensities using photometric redshifts in the $8.5$ deg$^2$ Boötes field. Further work with the high-redshift ($1 < z < 1.5$) clusters in the ISCS has included deep (1000s) IRAC imaging, spectroscopic followup, and [HST]{} imaging. Seven of the spectroscopically confirmed, high-redshift ISCS clusters have both deep IRAC imaging and ACS F775W imaging. It is this subsample of ISCS clusters that we use to study the LF of high-redshift galaxy clusters. We supplement the 1000 seconds of targeted IRAC observations for each cluster with imaging from the [Spitzer]{} Deep, Wide-Field Survey (SDWFS, @ashby09) which has a median exposure time of 420 seconds throughout the Boötes field. This gives a total observing time of roughly 1400s per cluster in all four IRAC bands. We list in Table \[tbl:cluster\_summary\] our clusters along with their positions, redshifts and number of spectroscopic members. ISCS J1438.1+3414 has a published X-ray mass estimate of $\log(M_{200}^{L_X}/M_\odot) = 14.35^{+0.11}_{-0.14}$ which comes from a 143 ks Chandra exposure [@andreon11; @brodwin11]. All of these clusters (with the exception of ISCS J1433.8+3325) have a weak-lensing mass estimate from @jee11, with masses in the range of $14.40 \le \log(M^{WL}_{200}/M_\odot) \le 14.73$. Comparison Fields ----------------- Given the limited spectroscopic redshifts in these fields, statistical background subtraction is required to recover the underlying LF. A statistical background subtraction involves measuring the number counts of galaxies in a field region and subtracting it from the number counts of galaxies near the cluster. This technique has been used successfully by @depropris98, @lin04, @muzzin08, and @mancone10. This requires a survey with IRAC imaging of at least the same depth as our cluster images as well as ACS F775W imaging. For this purpose we select the GOODS North and South [@goods] fields. We downloaded the latest fully reduced 3.6 and $4.5\mu$m [Spitzer]{} IRAC images taken of the GOODS fields. We also retrieved the latest [HST]{} ACS F775W catalogs from the GOODS survey. Throughout this paper we refer to the GOODS fields as our control fields. Data Reduction and Processing {#sec:reduction} ----------------------------- We produced IRAC mosaics of all seven clusters by combining data from our own programs (PID78, PID30950) with that from SDFWS, following procedures identical to those described in @ashby09. This included the manner in which outliers were rejected and in the way the individual IRAC frames were prepared for mosaicing by first removing the residual images arising from earlier exposure to bright sources. We generated catalogs by running our fully reduced 3.6 and 4.5$\mu$m IRAC images (for the clusters and the control fields) through Source Extractor [@bertin96] in single-image mode. We used 4$\prime\prime$ diameter aperture mags that were aperture-corrected to total mags by comparing 4$\prime\prime$ and 24$\prime\prime$ diameter aperture magnitudes for bright, unsaturated stars in our images. We used stars to measure the aperture corrections because galaxies at these redshifts are typically unresolved in IRAC imaging. This gave aperture corrections of $-0.32$ ($-0.34$) magnitudes in 3.6 (4.5)$\mu$m for our cluster images and $-0.31$ ($-0.32$) mags for our control images. For reference, the difference between the aperture corrections of the cluster and control fields is smaller than the uncertainty of the absolute flux calibration for IRAC images [@reach05]. To verify our calculated aperture corrections we compared our 4$\prime\prime$ aperture corrected magnitudes to the 4$\prime\prime$ aperture corrected magnitudes from SDWFS, and found very small systematic offsets ($<0.03$ mags). The ACS imaging for our clusters was obtained as part of the [HST]{} Cluster Supernova Survey, and the reduction of the images is described in detail in @suzuki12. We ran the reduced ACS F775W images through Source Extractor and used MAG\_AUTO to calculate the F775W magnitude of our galaxies. We used Our F775W photometry to perform an optical$-$NIR color cut to remove contaminants from our LF (Section \[sec:color\_cut\]). For our control we used the ACS F775W MAG\_AUTO values from the GOODS catalogs, which were also generated with Source Extractor. Next we calculated completeness as a function of magnitude at 3.6 and $4.5\mu$m for each cluster image and each control image separately. We approximated our galaxies as point sources due to the coarse IRAC point spread function (PSF). We generated 24,000 artificial point sources for each image, uniformly distributed between 13 and 25 mags. Our artificial point sources were simply copies of the PSF for each image, which we generated by median combining unsaturated stars taken directly from each image. We added these sources to the original images ten at a time, ran Source Extractor again for each new image, and finally calculated the recovery rate as a function of magnitude for a given image and filter. Figure \[fig:complete\] shows a histogram of the measured 50% point source completeness limits for each of our cluster images and control images at 3.6 (left) and 4.5$\mu$m (right). For comparison the vertical black line denotes the 50% completeness limit in each band from the SDWFS survey. We only want to fit for the cluster galaxy LF when the completeness of all galaxies in all clusters is at least 50%. Therefore we limit our LF fitting procedure to the the brightest 50% point source completeness limit for all our clusters, which is 20.37 (19.60) mags in 3.6 (4.5)$\mu$m. The 50% completeness limit for our cluster images is $\sim$0.75 mags fainter than for the SDWFS images and $\sim$1 mag brighter than our control images. Observed Luminosity Function {#sec:lf} ============================ Optical$-$NIR Color Cut {#sec:color_cut} ----------------------- We use a simple color cut to remove stars and low-redshift galaxies from our catalogs and increase the signal-to-noise ratio of our high-redshift cluster galaxies. Our color cut is designed to include the blue cloud, as excluding part of it would induce a systematic bias in our measurement of the faint-end slope. We choose our color cut by using the @bruzual03 stellar models to create a model of a star-forming galaxy with a color on the blue side of the blue cloud, consistent with @lemaux12. We then use this same model to estimate the color of the bluest star forming galaxies in our clusters, finding F775W$-$\[3.6\] $\ge$ 3.5 and F775W$-$\[4.5\] $\ge$ 3.75. We use these values for our color cut, and note that our final results are not sensitive to our exact choice because our results change by less than our random errors for a wide range of color cuts ($2.5<\text{F775W}-\text{[3.6]}<5$). Stellar contamination is a potential issue for our sample as our cluster and control fields are at different galactic latitudes. A \[3.6\]-\[4.5\] cut could effectively remove stars from our sample, but is not possible because there is little overlap between the 3.6 and 4.5 $\mu$m imaging of our control regions. Limiting our sample to areas with 3.6 and 4.5 $\mu$m imaging would remove about 75% of our control region. However, stellar contamination is effectively removed via our optical-NIR color cut. To verify that stellar contamination is not an issue for our results we calculate the expected colors for local stellar populations. To do this we download stellar isochrones using the CMD 2.3 software[^1] which includes the latest stellar modeling details from a number of sources [@bonatto04; @girardi08; @marigo08; @girardi10]. For a low metallicity ($Z=0.008$) model, which is relevant to the Galactic halo, no star of any age has F775W$-$\[3.6\] $\ga$ 4. Solar metallicity stars have F775W$-$\[3.6\] $\sim 5$, at the reddest. Therefore, only the tip of the RGB and AGB for old stars extend redder than the color cut. As such, only a small fraction of stars might remain after the cut, meaning that stellar contamination is not an issue. This is confirmed by the fact that our results do not change even when using color cuts as red as F775W$-$\[4.5\] = 5, which removes all stellar contamination. We additionally remove from our catalogs all objects with CLASS\_STAR $>$ 0.8 in the ACS catalogs. We find that this has a negligible impact on our results, again showing that stellar contamination is not an issue for our sample. LF Fitting Procedure {#sec:bgsubtract} -------------------- We design our methodology so that we can perform an unbinned fit to the cluster member LF, and so that the background subtraction is done in a way equivalent to a subtraction in observed space. We generate an individual cluster LF and control LF for every cluster. The cluster LF is simply composed of the galaxies in the cluster image which passed our various cuts (Section \[sec:color\_cut\]), are within 1.5 Mpc of the cluster center, and are outside of the heavily blended cluster cores (typically $\sim$100kpc). The latter restriction also removes the BCGs from our sample, which are known to not follow a Schechter distribution. For each cluster we use a @bruzual03 SPS model to calculate the $k$-correction and distance modulus correction needed to move a passively evolving galaxy ($z_f=3$, Chabrier IMF, solar metallicity) from the cluster redshift to the median redshift of our cluster sample ($z$=1.35). We then apply this $k$-correction and distance modulus correction to all galaxies in the cluster LF. Next we build a control LF for each cluster in a similar fashion. We select all galaxies in the control images which pass our cuts, weight them according to the relative area of the cluster and field images, and apply the exact same $k$-correction and distance modulus correction that we applied to the cluster LF to all the galaxies in the control LF. We do this so that the same transformation has been applied in the same way to the cluster and control galaxies, and therefore when we subtract the control LF from the cluster LF the subtraction is effectively done in observed space. This procedure gives us an unbinned cluster and control LF for each cluster. We then combine the individual cluster and control LFs into a composite cluster and composite control LF, which we use to measure the LF of cluster members. We parameterize the luminosity function of cluster members as a @schechter luminosity function and measure the best fitting Schechter parameters with maximum likelihood fitting, similar to the procedure used in @mancone10. This procedure requires an analytical representation for the contribution from the control region so we bin our composite control LF by magnitude, correct for photometric incompleteness, and fit a third order polynomial to it in log space. We then use maximum likelihood fitting to fit the sum of a Schechter luminosity function (the cluster member LF) and the fitted composite control LF to the composite cluster LF. We use a downhill simplex algorithm [@nr] to maximize the likelihood as a function of $\Phi^$, $m^$, and $\alpha$, and fit all galaxies brighter than the 50% completeness limit for the clusters (Section \[sec:reduction\]). [cccccc]{} 1.35 & 7 & ${17.44}\pm {0.30}$ & ${-0.97}\pm {0.14}$ & ${17.10}\pm {0.43}$ & ${-0.91}\pm {0.28}$\ [ccc]{} 15.25 & $ -1.32 \pm 0.50$ & $ 7.09 \pm 7.02$\ 15.75 & $ 10.64 \pm 9.10$ & $ -3.31 \pm 10.12$\ 16.25 & $ 0.87 \pm 13.97$ & $ 24.34 \pm 20.76$\ 16.75 & $ 32.86 \pm 23.12$ & $ 88.94 \pm 30.76$\ 17.25 & $ 95.61 \pm 32.39$ & $114.71 \pm 37.00$\ 17.75 & $128.29 \pm 37.82$ & $118.11 \pm 41.72$\ 18.25 & $136.13 \pm 41.68$ & $143.18 \pm 45.37$\ 18.75 & $181.98 \pm 48.57$ & $214.39 \pm 54.09$\ 19.25 & $145.11 \pm 51.32$ & $182.13 \pm 59.06$\ 19.75 & $270.40 \pm 62.50$ &\ 20.25 & $253.25 \pm 71.37$ &\ [cccccc]{} @depropris98 & Coma & 1 & H & 0.023 & $-0.78 \pm 0.3$\ @andreon01 & AC 118 & 1 & K$s$ & 0.3 & $-1.18 \pm 0.15$\ @lin04 & & 93 & K$s$ & 0.043 & $-0.84 \pm 0.02$\ @jenkins07 & Coma & 1 & 3.6$\mu$m & 0.023 & $-1.25 \pm 0.05$\ @muzzin07 & & 15 & K & 0.296 & $-0.84 \pm 0.08$\ @skelton09 & Norma & 1 & K$s$ & 0.016 & $-1.26 \pm 0.1$\ @depropris09 & & 10 & K & 0.07 & $-0.98 \pm 0.2$\ @mancone10 & & 35 & 3.6$\mu$m & 0.37 & $-0.60 \pm 0.2$\ This Work & & 7 & 3.6$\mu$m & 1.35 & ${-0.97}\pm {0.14}$\ This Work & & 7 & 4.5$\mu$m & 1.35 & ${-0.91}\pm {0.28}$\ @mobasher03 & Coma & 1 & R & 0.023 & $-1.18^{+0.04}_{-0.02}$\ @depropris03 & & 60 & Bj & $<$0.11 & $-1.28 \pm 0.03$\ @chiboucas06 & Centaurus & 1 & V & 0.0114 & $-1.4^{+0.1}_{-0.18}$\ @strazzullo10 & XMMU J2235-2557 & 1 & H & 1.39 & $-1.2^{+0.2}_{-0.15}$\ Results {#sec:results} ------- Figure \[fig:lfs\] shows the control-subtracted cluster LF and the Schechter fit to the cluster member LF for 3.6$\mu$m (left) and 4.5$\mu$m (right). Maximum likelihood fitting gives a fit to the LF without binning, but for plotting purposes we show the binned and control-subtracted cluster LF in Figure \[fig:lfs\], which is the binned difference between the composite cluster LF and the composite control LF. In each panel the solid curve shows the best fit while the dashed vertical line illustrates the magnitude limit used for the fit. Figure \[fig:contours\] shows the 1, 2, and 3$\sigma$ contours in M$^$ vs. $\alpha$ space derived from our measured likelihoods. We also report the binned LF values in Table \[tbl:lfs\], although we note that our fit was to the unbinned data. We also measure uncertainties using bootstrap resampling, as such an error estimate is more sensitive to systematic uncertainties caused by cluster-to-cluster variations. We generate realizations of the LF by randomly selecting seven clusters from our sample and repeating our LF fitting procedure with the new cluster sample. The cluster selection is done with replacement, which means that an individual cluster can be selected more than once when generating a new cluster sample. This allows us to probe any systematic uncertainty caused by cluster-to-cluster variation, as this process effectively applies random weights to the clusters while fitting. We perform 100 realizations of the cluster LF and take the standard deviation of the fitted $M^$ and $\alpha$ parameters as our random uncertainties. Our best fitting Schechter parameters and bootstrap errors are listed in Table \[tbl:params\]. We note that our measured bootstrap uncertainties agree well with the contours in Figure \[fig:contours\]. The relatively small sizes of the GOODS fields means that cosmic variance in our control fields could be an additional source of systematic uncertainty. To verify that cosmic variance is not strongly biasing our results we redo our fit but use just one of the GOODS fields for our control sample, and then redo it again using the other. In each case the best fitting value of $\alpha$ changes by $\sim$0.1 in both filters, which is smaller than our measured errors. Therefore, cosmic variance is unlikely to be a dominant source of uncertainty in our results. Discussion {#sec:discussion} ========== High-Redshift Comparison ------------------------ For a basic consistency check we compare to our results from @mancone10. In @mancone10 we measured $M^_{3.6\mu\text{m}}$ and $M^_{4.5\mu\text{m}}$ out to $z=1.8$ using the slightly shallower SDWFS data and a statistical background subtraction. As such the methodology is very similar, the filters are the same, and the seven clusters studied herein were also included in @mancone10. Due to our shallower data in @mancone10 we fixed $\alpha$ and reported fitted $M^$ values for $\alpha = -0.6$, $-0.8$, and $-1.0$ in redshift bins from $z = 0.3$ to $z = 2.0$. The median redshift of the clusters in this study ($z=1.35$) fall directly between the $z=1.24$ and $z=1.46$ bins from @mancone10. Therefore we compare our fitted $M^$ values to the average of the $M^$ values for the $z=1.24$ and 1.46 bins with $\alpha = -1.0$, which gives $M^_{3.6\mu\text{m}} = 17.42 \pm 0.1$ and $M^_{4.5\mu\text{m}} = 16.92 \pm 0.1$, in good agreement with the values of $M^$ measured herein. We note that our random errors for $M^$ in this paper are larger than the quoted errors in @mancone10. This is a simple result of number statistics. While we only have seven clusters in this study, we had 25 (22) clusters in our $z=1.24$ (1.46) bins in @mancone10. This leads to a larger random uncertainty in $M^$ for this current work, although we have lower systematic uncertainty for $M^$ in this paper because the requirement of fixing $\alpha$ in @mancone10 introduced a systematic uncertainty of $\sim0.2$ mags into $M^$. We also compare to @strazzullo10 who measure the $H$-band LF of a $z=1.39$ galaxy cluster to $M^ + 4$. They find $\alpha_H = -1.2^{+0.2}_{-0.15}$, also consistent with our results. While there is a difference in passband between our studies, both trace rest-frame wavelengths redward of the 4000Å break so we expect the difference in passband to have a minimal affect on our fitted values of $\alpha$. Recent studies have found a deficit of low-luminosity red-sequence galaxies in high-redshift clusters [@delucia04; @rudnick09; @lemaux12]. At face value this seems in contradiction with the flat $\alpha$ values found in this study as well as @strazzullo10. However, neither our results nor the results from @strazzullo10 are limited to red-sequence galaxies, and therefore this apparent difference can simply be a sign that low-luminosity galaxies are in place in the cluster environment at these redshifts but have not yet finished transitioning onto the red sequence, as was suggested in @lemaux12. Low-Redshift Comparison ----------------------- We compare our results to low-redshift cluster LFs to assess the evolution of $\alpha$ over a substantial fraction ($\sim$9 Gyr) of cosmic history. To accomplish this we have compiled a list of $\alpha$ measurements from the literature for a variety of clusters or cluster samples at different redshifts, which we summarize in Table \[tbl:alpha\_refs\]. We note that @depropris98, @andreon01, and @depropris09 did not present a formal error for $\alpha$ but did plot confidence regions for their fit, so we estimated the error on $\alpha$ from the plots of their confidence regions. Specifically, we derived the error from the full range of values covered by their 1$\sigma$ confidence contours. We split Table \[tbl:alpha\_refs\] up into two sections: studies which trace the rest-frame optical and studies which trace the rest-frame NIR (such as this work). Star formation can be an important contributor to the rest-frame optical LF, and as such $\alpha$ is not necessarily directly comparable between the two sets of studies. The results in Table \[tbl:alpha\_refs\] are presented graphically in Figure \[fig:z\_alpha\]. In this Figure the fitted $\alpha$ values and errors are plotted for all the rest-frame NIR results in Table \[tbl:alpha\_refs\]. There is substantial study-to-study scatter at low redshift, and large error bars at high redshift, but Figure \[fig:z\_alpha\] shows no obvious evidence for evolution in $\alpha$ from $z = 0$ to $z \sim 1.4$, representing nearly 70% of cosmic history. Past work on cluster LFs have primarily characterized the evolution of $M^$, and shallow imaging has required assuming a value for $\alpha$ and fixing it as a function of redshift (see, e.g., @muzzin08 and @mancone10). Fixing $\alpha$ has been a potential source of systematic uncertainty, as the strong coupling between $M^$ and $\alpha$ means that if $\alpha$ is improperly held fixed then the fitted values of $M^$ will also be wrong. This can be particularly important for studies of the evolution of $M^$ because if $\alpha$ is evolving but assumed to be fixed then this false assumption can create spurious evolution in $M^$. This potential source of systematic uncertainty was discussed in detail in @mancone10 because we found that for $z\gtrsim1.3$ the fitted values of $M^$ to the cluster LF deviated strongly from passive evolution. We concluded that while evolution in $\alpha$ could contribute to the measured deviations from passive evolution, it was unlikely to be the underlying cause because the direction of the deviation would require $\alpha$ to become steeper at higher redshift. Having now measured $\alpha$ at high redshift we can conclude that evolution in $\alpha$ was not the cause of our observed deviations from passive evolution as $\alpha$ does not evolve significantly with redshift out to $z \sim 1.4$. Comparison to the Field LF {#sec:field_comparison} -------------------------- We find that at high redshift the faint-end slope of the cluster LF matches that of the field. @saracoo06 measured $\alpha$ in the rest-frame $J$-band for field galaxies, finding $\alpha = -0.94^{+0.16}_{-0.15}$ in a redshift bin centered at $z \sim 1.2$. @cirasuolo07 found $\alpha = -0.92 \pm 0.18$ for galaxies with $1.25 < z < 1.5$ in the rest-frame $K$-band, again in good agreement with the faint-end slope of the cluster LF found herein. Moreover, work in the field [@saracoo06; @cirasuolo07; @stefanon12] has also found that the faint end of the LF for field galaxies is consistent with being constant out to the highest redshifts studied. For example, @kochanek01 find $\alpha = -1.09 \pm 0.06$ for 2MASS galaxies at low redshift, while @stefanon12 find that the faint-end slope of their rest-frame $J$-band LF is consistent with flat from $1.5 < z < 3.5$. They compare their results to lower redshift studies, finding no evidence for evolution in $\alpha$ out to $z = 3.5$ with a mean value of $\alpha = -1.05 \pm 0.03$. This is all consistent with our finding that $\alpha$ does not substantially evolve but is consistent with flat from $z = 0$ to $z \sim 1.4$ for cluster galaxies. Implications for Galaxy Formation --------------------------------- As discussed above, we find no statistically significant evidence for evolution in the faint-end slope of the NIR luminosity function out to $z \sim 1.4$. A lack of evolution in $\alpha$ combined with a lack of evolution in $M^$ [@mancone10] means that both faint and bright galaxies are largely in place at high redshift. This places a strong constraint on the luminosity evolution of cluster galaxies at $z \lesssim 1.4$. Either little evolution is happening at lower redshifts, or the processes responsible for LF evolution have no net impact on the cluster population. To understand the implications of this for galaxy evolution, we must understand how the various process that cause cluster galaxy evolution would affect the luminosity evolution of the cluster galaxy population. The fact that the NIR traces old stellar populations means that our results are most sensitive to processes which would cause evolution in the stellar mass of cluster galaxies. Any process which affects low and high stellar mass cluster galaxies equally will lead to evolution in $M^$. Conversely, any process which leads to differential evolution between galaxies with low and high stellar masses will lead to evolution in $\alpha$. The two primary processes by which galaxies can grow their stellar masses over time are star formation and mergers. The downsizing paradigm (see @fontanot09 and references therein) suggests that star formation will be preferentially found in lower mass galaxies at low redshift. Such a mass dependence for galaxy star formation histories will necessarily imply evolution in the faint-end slope of the LF. The amplitude of this effect however depends upon the total amount of ongoing star formation, which will vary between clusters and may depend upon the total mass of the host cluster halo. There is evidence that substantial star formation is still ongoing in our cluster sample [@snyder12], as well as other clusters at similar redshifts [@hilton10; @tran10; @fassbender11]. In contrast, @muzzin12 find that star formation has already been strongly quenched in their cluster at $z=1.2$. Mergers can also build up the stellar mass of cluster galaxies, although mergers are expected to be suppressed in the cluster environment due to the high relative velocities of cluster galaxies [@alonso12]. Recent theoretical studies [@murante07; @conroy07; @puchwein10] suggest that for massive galaxies growth by mergers becomes very inefficient (but see @rudnick12), and it is possible that mergers can yield a steepening of the faint-end slope if this efficiency is strongly mass-dependent. In contrast gravitational interactions such as galaxy-galaxy interactions, interactions of a galaxy with the cluster potential, galaxy harassment, and the dissolution of cluster galaxies, can all strip mass away from low-mass galaxies in particular [@moore96; @boselli06; @murante07] and therefore cause $\alpha$ to grow flatter or turn over with time. Another process for consideration is the infall of new galaxies into the cluster, the effect of which depends on the shape of the LF for the infalling galaxy population. Since we find that the cluster and field galaxy populations have a similar faint end slope (Section \[sec:field\_comparison\]), we expect that the infall of new galaxies into the cluster will primarily act to mitigate any potential evolution of the cluster LF by driving the cluster LF back towards a flat faint-end slope. Instead, the infall of new galaxies will lead to an increase in $\Phi^$, the normalization of the LF, which we have not constrained Finally, a mass-dependent galactic initial mass function (IMF) can cause the shape of the LF to change relative to the underlying stellar mass function. This is because the rate of luminosity evolution for a stellar population depends sensitively on the IMF [@conroy09]. Recent work [@cappellari12] has suggested that the IMF does indeed depend on galaxy mass, such that lower mass galaxies have a flatter IMF (i.e., a higher fraction of high-mass stars). A flatter IMF leads to a faster fading of the underlying stellar population [@conroy09] and therefore, if true, the results of @cappellari12 suggest that low-mass galaxies should fade faster than high mass galaxies when the stellar masses of both remains fixed. This will cause $\alpha$ to grow flatter or turn over with time, effectively acting against processes which build up the stellar mass of galaxies but without impacting the underlying mass function. Clearly, there are many processes which could potentially cause evolution of the NIR cluster LF at $z < 1.5$. Therefore, the lack of evolution in $\alpha$ observed herein, combined with the lack of evolution in $M^$ observed over a similar redshift range [@mancone10], places an important constraint on these processes. In net, they cannot cause any large evolution in the shape of the NIR cluster LF. This could be because both low mass and high mass cluster galaxies are largely assembled at high redshift, or because the differing effects of these processes causes little evolution in net. The latter might imply an uncomfortable degree of fine tuning. In general though, the ability of infalling galaxies to dilute any evolution in the cluster LF allows for more flexibility in the strength of other processes. Conclusions {#sec:conclusions} =========== We measure the 3.6 and 4.5$\mu$m luminosity functions of seven galaxy clusters at $1 < z < 1.5$, specifically investigating the shape of the LF for faint galaxies. We find the LFs to be well-fit by Schechter distributions with faint-end slopes of $\alpha_{3.6\mu\text{m}} = {-0.97}\pm {0.14}$ and $\alpha_{4.5\mu\text{m}} = {-0.91}\pm {0.28}$, both consistent with having flat faint-end slopes within 1$\sigma$. Our primary conclusions are summarized here: 1. We compare to studies of the NIR LF of low-redshift clusters and find no statistically significant evidence for evolution of the faint-end slope of the cluster LF. Therefore we conclude that the faint end of the cluster LF has not evolved significantly over 70% of cosmic history. 2. Having measured a non-evolving faint-end slope we have removed one source of systematic uncertainty from studies of the evolution of $M^$ as a function of redshift. This is particularly relevant for our recent detection of deviations from passive evolution at high redshift [@mancone10]. Shallow imaging in @mancone10 necessitated fixing $\alpha$, which could have lead to spurious evolution in $M^$ if $\alpha$ was evolving. 3. We compare to the faint end slope for field galaxies at similar redshifts and find good agreement. Field studies [@saracoo06; @cirasuolo07; @stefanon12] find a faint-end slope consistent with flat at high redshift, and the most recent results [@stefanon12] find no evidence for evolution out to $z=3.5$. 4. Given recent studies [@muzzin08; @mancone10] which have found that the evolution of the bright end of the cluster LF is consistent with passive evolution out to $z \sim 1.3$, we conclude that the shape of the cluster LF has been in place and evolved little since $z \sim 1.3$. This could suggest that low-mass galaxies are largely assembled at high redshift. Conversely, it could simply mean that the many processes which cause evolution of the cluster galaxy population have no net impact on the mass and luminosity function of cluster galaxies. This work is based on observations made with the Spitzer Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. The observations are associated with programs P78 and P30950. This work is based on observations made with the NASA/ESA Hubble Space Telescope, obtained from the Data Archive at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. The observations are associated with program GO-10496. CLM and AHG acknowledge support for this work from the National Science Foundation under grant AST-00708490. AHG, MB, and SAS also acknowledge support from NASA through grant HST-GO-10496. [^1]: http://stev.oapd.inaf.it/cmd	{ "pile_set_name": "ArXiv" }
--- abstract: 'It was proved by Linares and Ortega in [@LO] that the [linearized]{} Benjamin-Ono equation posed on a periodic domain ${\mathbb T}$ with a distributed control supported on an arbitrary subdomain is exactly controllable and exponentially stabilizable. The aim of this paper is to extend those results to the [full]{} Benjamin-Ono equation. A feedback law in the form of a localized damping is incorporated in the equation. A smoothing effect established with the aid of a propagation of regularity property is used to prove the semi-global stabilization in $L^2({\mathbb T})$ of weak solutions obtained by the method of vanishing viscosity. The local well-posedness and the local exponential stability in $H^s({\mathbb T})$ are also established for $s>1/2$ by using the contraction mapping theorem. Finally, the local exact controllability is derived in $H^s({\mathbb T})$ for $s>1/2$ by combining the above feedback law with some open loop control.' address: - 'Instituto de Matematica Pura e Aplicada, Estrada Dona Castorina 110, Rio de Janeiro 22460-320, Brazil' - 'Institut Elie Cartan, UMR 7502 UHP/CNRS/INRIA, B.P. 70239, 54506 Vandœuvre-lès-Nancy Cedex, France' author: - Felipe Linares - Lionel Rosier title: 'Control and Stabilization of the Benjamin-Ono Equation on a Periodic Domain' --- Introduction ============ The Benjamin-Ono (BO) equation can we written as $$u_t + {\mathcal H} u_{xx} + uu_x=0,$$ where $u=u(x,t)$ denotes a real-valued function of the variables $x\in{\mathbb R}$ and $t\in {\mathbb R}$, and ${\mathcal H}$ denotes the Hilbert transform defined as $$\widehat {{\mathcal H } u} (\xi )= -i \, \text{sgn} (\xi )\, \hat u(\xi ).$$ This integro-differential equation models the propagation of internal waves in stratified fluids of great depth (see [@benjamin; @ono]) and turns out to be important in other physical situations as well (see [@DR; @ishimori; @MK]). Among noticeable properties of this equation we can mention that: (i) it defines a Hamiltonian system; (ii) it admits infinitely many conserved quantities (see [@case]); (iii) it can be solved by an analogue of the inverse scattering method (see [@AF]); (iv) it admits (multi)soliton solutions (see [@case]). In this paper, we consider the BO equation posed on the periodic domain ${\mathbb T}={\mathbb R}/(2\pi{\mathbb Z})$: $$\label{BOT} u_t + {\mathcal H} u_{xx} + uu_x=0,\quad x\in{\mathbb T},\ t\in{\mathbb R},$$ where the Hilbert transform $\mathcal H$ is defined now by $$(\widehat{{\mathcal H}u})_k=-i\, \text{sgn}(k)\hat u_k.$$ The two first conserved quantities are $$I_1(t)=\int_{{\mathbb T}} u(x,t)dx$$ and $$I_2(t)= \int_{{\mathbb T}} u^2(x,t)dx.$$ From the historical origins [@benjamin; @ono] of the BO equation, involving the behavior of stratified fluids, it is natural to think $I_1$ and $I_2$ as expressing conservation of volume (or mass) and energy, respectively. The Cauchy problem for the equation in the real line has been intensively studied for many years ([@Sa; @Io1; @ABFS; @Po; @MoSaTz; @KoTz1; @KeKo; @tao; @BP; @IoKe; @MP; @FoPo; @GFFLGP]). In the periodic case, there have been several recent developments. (See for instance [@molinet; @MR; @MP] and the references therein.) The best known result so far [@molinet; @MP] is that the Cauchy problem is well-posed in the space $$H^s_0({\mathbb T}) =\{u\in H^s({\mathbb T}); \ \hat u_0:= \frac{1}{2\pi}\int_{{\mathbb T}}u(x)\, dx=0\}$$ for $s\ge 0$. Moreover, the corresponding solution map ($u_0\to u$) is real analytic from the space $H^0_0({\mathbb T})$ to the space $C([0,T],H^0_0({\mathbb T}))$. In this paper we will study the equation from a control point of view with a forcing term $f=f(x,t)$ added to the equation as a control input: $$\label{BOc} u_t + {\mathcal H} u_{xx} + uu_x=f(x,t),\quad x\in{\mathbb T},\ t\in{\mathbb R},$$ where $f$ is assumed to be supported in a given open set $\omega \subset {\mathbb T}$. The following exact control problem and stabilization problem are fundamental in control theory. [Exact Control Problem:]{} Given an initial state $u_0$ and a terminal state $u_1$ in a certain space, can one find an appropriate control input $f$ so that the equation admits a solution $u$ which satisfies $u(\cdot,0)=u_0$ and $u(\cdot,T)=u_1$? [Stabilization Problem:]{} Can one find a feedback law $f=Ku$ so that the resulting closed-loop system $$u_t + {\mathcal H} u_{xx} + uu_x=Ku,\quad x\in{\mathbb T},\ t\in{\mathbb R}^+$$ is asymptotically stable as $t\to +\infty$? Those questions were first investigated by Russell and Zhang in [@RZ] for the Korteweg-de Vries equation, which serves as a model for propagation of surface waves along a channel: $$u_t + u_{xxx} + uu_x=f,\quad x\in{\mathbb T}, \ t\in{\mathbb R}. \label{BOf}$$ In their work, in order to keep the [mass]{} $I_1(t)$ conserved, the control input is chosen to be of the form $$f(x,t)=(Gh)(x,t):=a(x)\left( h(x,t)-\int_{{\mathbb T}} a(y)h(y,t)\, dy \right)$$ where $h$ is considered as a new control input, and $a(x)$ is a given nonnegative smooth function such that $\{x\in {\mathbb T}; \ a(x) >0\}=\omega$ and $$2\pi [a]=\int_{{\mathbb T}} a(x)\, dx=1.$$ For the chosen $a$, it is easy to see that $$\frac{d}{dt}\int_{{\mathbb T}} u(x,t)\, dx=\int_{{\mathbb T}} f(x,t)dx=0\quad \forall t\in{\mathbb R}$$ for any solution $u=u(x,t)$ of the system $$\label{BOG} u_t + u_{xxx} + uu_x=Gh,\quad x\in{\mathbb T},\ t\in{\mathbb R}.$$ Thus the [mass]{} of the system is indeed conserved. The control of dispersive nonlinear waves equations on a periodic domain has been extensively studied in the last decade: see e.g. [@RZ; @RZ2009bis; @LRZ] for the Korteweg-de Vries equation, [@MORZ] for the Boussinesq system, [@RZ2012] for the BBM equation, and [@DGL; @RZ2007b; @laurent; @RZ2009; @laurent2] for the nonlinear Schrödinger equation. By contrast, the control theory of the BO equation is at its early stage. The following results are due to Linares and Ortega [@LO].\ [Theorem A.]{} [@LO] [Let $s\ge 0$ and $T>0$ be given. Then for any $u_0,u_1\in H^s({\mathbb T})$ with $[u_0]=[u_1]$ one can find a control input $h\in L^2(0,T,H^s({\mathbb T}))$ such that the solution of the system $$\label{BOlin} u_t + {\mathcal H} u_{xx} =Gh, \qquad u(x,0)=u_0(x)$$ satisfies $u(x,T)=u_1(x)$.]{} In order to stabilize , Linares and Ortega employed a simple control law $$h(x,t)=-G^* u(x,t).$$ The resulting closed-loop system reads $$u_t+{\mathcal H} u_{xx}=-GG^u.$$ [Theorem B.]{} [@LO] [Let $s\ge 0$ be given. Then there exist some constants $C>0$ and $\lambda >0$ such that for any $u_0\in H^s({\mathbb T})$, the solution of $$u_t+{\mathcal H} u_{xx}=-GG^u,\qquad u(x,0)=u_0(x)$$ satisfies $$\\|u(\cdot,t) - [u_0]\\| _{H^s({\mathbb T}) } \le Ce^{-\lambda t} \\|u_0-[u_0]\\| _{H^s({\mathbb T}) } \qquad \forall t\ge 0.$$]{} The extension of those results to the full BO equation turns out to be a very hard task. Indeed, it is by now well known that the contraction principle cannot be used to establish the local well-posedness of BO in $H^s_0({\mathbb T})$ for $s\ge 0$. The method of proof in [@molinet; @MP] used strongly Tao’s gauge transform, and it is not clear whether this approach can be followed when an additional control term is present in the equation. For the sake of simplicity, we shall assume from now on that $[u_0]=0$, so that $u(t)$ has a zero mean value for all times. To stabilize the BO equation, we consider the following feedback law $$f=-G(D(G u))$$ where $\widehat{D u}_k =\|k\| \hat u_k$. Scaling in by $u$ gives (at least formally) $$\label{id} \frac{1}{2} \\|u(T)\\|_{L^2({\mathbb T})}^2 + \int_0^T \\|D^{\frac{1}{2}}(Gu)\\|_{L^2({\mathbb T})}^2 dt =\frac{1}{2} \\|u_0\\|_{L^2({\mathbb T})}^2.$$ This suggests that the energy is dissipated over time. On the other hand, reveals a smoothing effect, at least in the region $\{ a>0\}$. Using a [propagation of regularity property]{} in the same vein as in [@DGL; @laurent; @laurent2; @LRZ], we shall prove that the smoothing effect holds everywhere, i.e. $$\label{sm} \\|u\\|_{L^2(0,T;H^{\frac{1}{2}}({\mathbb T}))}\le C(T,\\|u_0\\|).$$ Using this smoothing effect and the classical compactness/uniqueness argument, we shall first prove that the corresponding closed-loop equation is semi-globally exponentially stable. \[main1\] Let $R>0$ be given. Then there exist some constants $C=C(R)$ and $\lambda =\lambda (R)$ such that for any $u_0\in H^0_0({\mathbb T})$ with $\\|u_0\\|\le R$, the weak solutions in the sense of vanishing viscosity of $$\label{BOfinal} u_t + {\mathcal H} u_{xx} + uu_x =-GDGu,\qquad u(x,0)=u_0(x)$$ satisfy $$\\|u(t)\\|\le Ce^{-\lambda t}\\|u_0\\| \qquad \forall t\ge 0.$$ A weak solution of in the sense of vanishing viscosity is a distributional solution of $u\in C_w({\mathbb R}^+ , H^0_0({\mathbb T})) \cap L^2_{loc} ( {\mathbb R}^+, H_0^{\frac{1}{2}} ({\mathbb T}))$ that may be obtained as a weak limit in a certain space of solutions of the BO equation with viscosity $$\label{BOviscosity} u_t + ({\mathcal H} -\varepsilon )u_{xx} + uu_x =-GDGu,\qquad u(x,0)=u_0(x)$$ as $\varepsilon\to 0^+$ (see below Definition \[defiweak\] for a precise definition). The issue of the [uniqueness]{} of the weak solutions in the sense of vanishing viscosity seems challenging. Using again the smoothing effect , one can extend (at least locally) the exponential stability from $H^0_0({\mathbb T})$ to $H^s_0({\mathbb T})$ for $s>1/2$. \[main2\] Let $s\in (\frac{1}{2}, 2]$. Then there exists $\rho >0$ such that for any $u_0\in H^s_0({\mathbb T})$ with $\\|u_0\\|_{H^s({\mathbb T})}< \rho$, there exists for all $T>0$ a unique solution $u(t)$ of in the class $C([0,T],H^s_0({\mathbb T}))\cap L^2(0,T, H^{s+\frac{1}{2}}_0({\mathbb T}))$. Furthermore, there exist some constants $C>0$ and $\lambda >0$ such that $$\\|u(t)\\| _s \le Ce^{-\lambda t} \\|u_0\\|_s\qquad \forall t\ge 0.$$ Finally, incorporating the same feedback law $ f = -G(D(Gu))$ in the control input to obtain a smoothing effect, one can derive an exact controllability result for the full equation as well. \[main3\] Let $s\in (\frac{1}{2}, 2] $ and $T>0$ be given. Then there exists $\delta>0$ such that for any $u_0,u_1\in H^s_0({\mathbb T})$ satisfying $$\\|u_0\\|_{H^s({\mathbb T})} \le \delta,\quad \\|u_1\\|_{H^s({\mathbb T})}\le \delta$$ one can find a control input $h\in L^2(0,T,H^{s-\frac{1}{2}} ({\mathbb T}))$ such that the system admits a solution $u\in C([0,T],H^s_0({\mathbb T}))\cap L^2(0,T,H^{s+\frac{1}{2}}_0({\mathbb T}) )$ satisfying $$u(x,0)=u_0(x),\quad u(x,T)=u_1(x).$$ Note that it would be desirable to have a control input $h$ in the class $L^2(0,T,H^s({\mathbb T}))$, but this will require to adapt the analysis in [@molinet; @MP]. Note also that a global controllability result in $H^0_0({\mathbb T})$ would follow from Theorems \[main1\] and \[main3\] if Theorem \[main3\] were also true for $s=0$. The paper is organized as follows. Section 2 is concerned with the local well-posedness and the stability properties of . We first prove the global well-posedness of in the energy space $H^0_0({\mathbb T})$, by using classical energy estimates (Theorem \[GWP\]). Next, we establish several technical properties, namely a commutator estimate (Lemma \[commutator\]), a propagation of regularity property (Propositions \[smoothing\] and \[prop10\]), and a unique continuation property (Proposition \[unique\_continuation\]) that are used to derive the exponential stability of with a decay rate [independent of $\varepsilon$]{} (Theorem \[thmstab1\]). This leads to the proofs of Theorems \[main1\] and \[main2\]. Finally, the control properties of are investigated in Section 3. Stabilization of BO with a localized damping ============================================ Semi-global exponential stabilization in $L^2({\mathbb T})$ ----------------------------------------------------------- \ Pick any function $$\label{defa} a\in C^\infty ({\mathbb T},{\mathbb R}^+) \ \text{ with }\ \int_{{\mathbb T}}a(x)dx=1$$ decomposed as $a(x)=\sum_{k\in{\mathbb Z}}\hat a_ke^{ikx}$. We are interested in the stability properties of the BO equation with localized damping $$\label{BOs} u_t+ {\mathcal H}u_{xx}+(\frac{u^2}{2})_x=-G(D(Gu)), \qquad u(0)=u_0,$$ where $$\label{defG} \widehat{{\mathcal H}u}_k=-i\, \text{sgn}(k)\hat u_k,\quad \widehat{D^s u}_k =\|k\|^s \hat u_k,\quad (G u)(x)=a(x)(u(x)-\int_{{\mathbb T}}a(y)u(y)dy).$$ We shall assume that $u_0\in H^0_0({\mathbb T})$, where for any $s\in{\mathbb R}$, $$H^s_0({\mathbb T})=\{u=\sum_{k\in {\mathbb Z}} {\hat u}_ke^{ikx}\in H^s({\mathbb T});\ {\hat u}_0=0\}.$$ Let $(u,v)=\int_{{\mathbb T}} u(x)v(x)dx$ denote the usual scalar product in $L^2({\mathbb T})$ with $\\|u\\|=\\|u\\|_{L^2({\mathbb T})}$ as associated norm, and for any $s\in {\mathbb R}$, let $(u,v)_s=((1-\partial _x^2)^{\frac{s}{2}} u,(1-\partial _x^2)^{\frac{s}{2}} v)$ denote the scalar product in $H^s({\mathbb T})$ with corresponding norm $\\|u\\|_s=(u,u)_s^{\frac{1}{2} } $. Let $\langle x\rangle := (1+\|x\|^2) ^{\frac{1}{2}}$ for any $x\in{\mathbb R}$. Note that for $s<0$ and $u\in H^s({\mathbb T})$, $Gu$ has to be understood as $$Gu= a\left( u - \langle u,a\rangle_{H^{s}({\mathbb T}),H^{-s}({\mathbb T})}\right) . $$ Assuming that $u_0\in H^0_0({\mathbb T})$, we obtain (formally) by scaling in by $u$ that $$\label{identity} \frac{1}{2} \\|u(T)\\|^2 + \int_0^T \\|D^{\frac{1}{2}}(Gu)\\|^2 dt =\frac{1}{2} \\|u_0\\|^2.$$ This suggests that the energy is dissipated over time. On the other hand, reveals a smoothing effect, at least in the region $\{ a>0\}$. Using a [propagation of regularity property]{} in the same vein as in [@DGL; @laurent; @laurent2; @LRZ], we shall prove that the smoothing effect holds everywhere, i.e. $$\label{SE1} u\in L^2(0,T;H^{\frac{1}{2}}({\mathbb T})).$$ Of course, a rigorous derivation of requires enough regularity for $u$, e.g. $$\label{SE2} u\in L^2(0,T,H^1({\mathbb T}))\cap C([0,T],H^0_0({\mathbb T})).$$ As there is a gap between and , we are let to put some artificial viscosity in (parabolic regularization method) to derive in a rigorous way the energy identity for the $\varepsilon -$BO equation $$u_t+{\mathcal H} u_{xx} +uu_x= \varepsilon u_{xx} -G(D(Gu)), \qquad u(0)=u_0. \label{BOe}$$ We shall prove the global well-posedness (GWP) of in $H^0_0$, together with the semi-global exponential stability in $H^0_0$ with a decay rate [uniform]{} in $\varepsilon >0$. Letting $\varepsilon\to 0$, this will give the semi-global exponential stability in $H^0_0$ of the weak solutions $u\in C_w([0,+\infty ),H^0_0({\mathbb T}))$ of obtained as limits of the (strong) solutions of . The (difficult) issue of the uniqueness of a weak solution to will not be addressed here. We first establish the GWP of . \[GWP\] Let $\varepsilon >0$ and $u_0\in H^0_0({\mathbb T})$. Then for any $T>0$ there exists a unique solution $u\in C([0,T], H^0_0({\mathbb T}))\cap L^2(0,T;H^1({\mathbb T}))$ of . Moreover $$\label{parabolic} u\in C((0,T],H^2({\mathbb T}))\cap C^1((0,T],H^1({\mathbb T})),$$ and for any $t\ge 0$ $$\frac{1}{2} \\|u(t)\\|^2 + \varepsilon \int_0^t \\|u_x(\tau )\\|^2 d\tau + \int_0^t \\|D^{\frac{1}{2}}(G u ) (\tau ) \\|^2 d\tau = \frac{1}{2} \\|u _0\\|^2. \label{identitybis}$$ [Proof:]{} The proof of Theorem \[GWP\] is divided into five parts. Note that the weak smoothing effect will be established later, as it is not needed here.\ [Step 1. Linear Theory]{}\ We consider the linear system $$u_t + ({\mathcal H}-\varepsilon) u_{xx} + G(D(Gu))=0,\qquad u(0)=u_0.$$ Let $ A u =({\mathcal H} -\varepsilon) u_{xx}$ with domain ${\mathcal D}( A ) = H^2_0({\mathbb T})\subset H^0_0({\mathbb T})$, and $ B u = G(D (Gu))$. Clearly $G\in {\mathcal L}(H^r({\mathbb T}),H^r_0({\mathbb T}))$ for all $r\in {\mathbb R}$, hence $ B \in {\mathcal L} ( H^1_0({\mathbb T}), H^0_0({\mathbb T}))$. Let $\theta _0\in (\arctan \varepsilon ^{-1}, \pi/2)$. Then, for $\theta _0 < \|\text{arg }\lambda \| \le \pi$, we have $$\\|( A -\lambda )^{-1}\\| \le \sup_{k\ne 0} \| (\varepsilon + i \, \text{sgn } k ) k^2 -\lambda \|^{-1} \le \frac{C}{\|\lambda \|}\cdot$$ It follows that $A$ is a sectorial operator (see [@henry Definition 1.3.1]) in $H^0_0({\mathbb T})$. Note that $\sigma ( A )= \{ (\varepsilon + i\, \text{sgn } k)k^2;\ k\in {\mathbb Z}^* \}$. Therefore, $\text{Re } \sigma ( A ) \ge \varepsilon$ and $A^{-\alpha}$ is meaningful for all $\alpha >0$. Since for all $s>0$ $$\\|A^{-\frac{s}{2}} u\\|^2_{H^s({\mathbb T})} \le C\sum_{k\ne 0} \|\varepsilon + i\, \text{sgn }k\|^{-s} \|{\hat u}_k\|^2 \le C \\|u\\|^2_{L^2({\mathbb T})}$$ we infer that $ B A^{-\frac{1}{2}} \in {\mathcal L} (H^0_0({\mathbb T}))$. It follows from [@henry Corollary 1.4.5] that the operator ${\mathcal A} := A + B$ is also sectorial, so that $-{\mathcal A}$ generates an analytic semigroup $\big( {\mathcal S}(t) \big) _{t\ge 0} = (e^{-t {\mathcal A}} )_{t\ge 0}$ on $H^0_0({\mathbb T})$ according to [@henry Theorem 1.3.4]. Note that, by [@henry Theorem 1.4.8], $D((A+B+\lambda )^\alpha)=D(A^\alpha )=H_0^{2\alpha} ({\mathbb T}) $ for all $\alpha\ge 0$ and $\lambda >0$ large enough, hence $${\mathcal S}(t)H_0^{s}({\mathbb T}) \subset H_0^{s} ({\mathbb T}),\qquad \forall t>0,\ \forall s \ge 0.$$ Let us derive estimates for the solutions of the Cauchy problem $$\label{AB} u_t + {\mathcal A} u =f,\qquad u(0)=u_0.$$ For any $T>0$ and any $s\in {\mathbb N}$, let $$Y_{s,T}=C([0,T];H^s_0( {\mathbb T})) \cap L^2(0,T; H^{s+1}_0({\mathbb T}))$$ be endowed with the norm $$\begin{aligned} \\|u\\|_{Y_{s,T}} = \\|u\\|_{L^\infty (0,T;H^s({\mathbb T}) )} +\\|u\\|_{L^2(0,T;H^{s+1}({\mathbb T}) )} .\end{aligned}$$ \[linearestim\] We have for some constant $C_0=C_0(\varepsilon, s,T)$ $$\\|u\\|_{Y_{s,T}} \le C_0 \left( \\|u_0\\|_{s} + \\| f \\|_{L^1(0,T,H^s ({\mathbb T}) ) } \right),$$ $u$ denoting the mild solution of associated with $(u_0,f)\in H^s_0({\mathbb T})\times L^1(0,T,H^s_0({\mathbb T}))$. [Proof of Lemma \[linearestim\].]{} It is well known from classical semigroup theory that $$\\|u\\|_{L^\infty (0,T,H^s({\mathbb T}) ) } \le C\left( \\|u_0\\|_s + \\| f \\|_{L^1(0,T,H^s({\mathbb T}) ) } \right) \cdot$$ Next we estimate $\\|u\\|_{L^2(0,T,H^{s+1}({\mathbb T}))}$. We first assume $u_0\in H^{s+2}_0({\mathbb T})$ and $f\in C([0,T]; H^{s+2}_0({\mathbb T}))$, so that $u\in C([0,T];H^{s+2}_0({\mathbb T}))\cap C^1([0,T];H^s_0({\mathbb T}))$. Taking the scalar product of each term of by $u$ in $H^s({\mathbb T})$ results in $$\label{energyeps} \frac{1}{2} \\|u(t)\\|_{s}^2 +\varepsilon \int_0^t \\|u_x\\|_s^2\, d\tau +\int_0^t (G(D(Gu)), u)_s\, d\tau = \frac{1}{2}\\|u_0\\|_s^2 + \int_0^t (f,u)_s \, d\tau.$$ The identity is also true for $u_0\in H^s_0({\mathbb T})$ and $f\in L^1(0,T,H^s_0({\mathbb T}))$, by density. The following claim is needed.\ [Claim 1.]{} For any $s\in {\mathbb R}$, there exists a constant $C=C(s)>0$ such that $$- \big( G(D(Gu) ) , u\big) _s \le C \\|u\\|^2_{s} -\\|D^{\frac{1}{2}} (Gu)\\|^2_s\qquad \forall u\in H_0^{s+1}({\mathbb T}).$$ [Proof of Claim 1.]{} We have $$\begin{aligned} \big( G(D(Gu) ) , u \big)_{s} &=& \big( (1-\partial^2 _x)^{\frac{s}{2}} G(D(Gu) ) , (1-\partial _x^2 )^{\frac{s}{2}} u \big) \\ &=& \big( [(1-\partial ^2_x)^{\frac{s}{2}},G] D(Gu), (1-\partial _x^2)^{\frac{s}{2}} u ) \\ &&\qquad +(G(1-\partial _x^2)^{\frac{s}{2}} D(Gu), (1-\partial _x ^2)^{\frac{s}{2}} u ) \\ &=:& I_1+ I_2.\end{aligned}$$ Since $a\in C^\infty ({\mathbb T})$, we easily obtain that $$\\| [ (1-\partial_x^2)^{\frac{s}{2}},G] u \\| \le C \\|u\\| _{s-1}.$$ It follows that $$\| I_1 \| \le C \\|u\\|^2_{s}.$$ On the other hand $$\begin{aligned} I_2 &=& \big( (1-\partial _x^2) ^{\frac{s}{2}} D(Gu) ,G(1-\partial _x^2)^{\frac{s}{2}} u \big) \\ &=& \\|(1-\partial _x^2 )^{\frac{s}{2}} D^{\frac{1}{2}}(Gu)\\| ^2 + ((1-\partial _x^2)^{\frac{s}{2}}(Gu), D[G,(1-\partial _x^2)^{\frac{s}{2}}]u),\end{aligned}$$ hence $$-I_2 \le C \\|u\\|^2_{s} -\\|D^{\frac{1}{2}} (Gu)\\|^2_s \cdot$$ The claim is proved.\ Combining Claim 1 with , we obtain that for $t=T$ $$\begin{aligned} &&\frac{1}{2} \\|u(T)\\|_s^2 + \varepsilon \int_0^T \\|u_x(\tau )\\|_s^2 d\tau + \int_0^T \\|D^{\frac{1}{2}}(G u) \\|^2_s d\tau \\ &&\qquad \le \frac{1}{2} \\|u_0\\|^2_s + C\\|u\\|^2_{L^2(0,T,H^s({\mathbb T}) )} + \frac{1}{2}\\|u\\|^2_{L^\infty (0,T ,H^s({\mathbb T}) )} + \frac{1}{2}\\| f \\|^2_{L^1(0,T,H^s( {\mathbb T})) } \\ &&\qquad \le C\big( \\|u_0\\|^2_s + \\| f \\|^2_{L^1(0,T,H^s({\mathbb T}))} \big) \cdot\end{aligned}$$ The proof of Lemma \[linearestim\] is achieved. \[rem\] We observe that when $u_0\equiv 0$ in then $$\\|\int_0^t {\mathcal S}(t-\tau )f(\tau )\,d\tau \\|_{Y_{s,T}}\le C(\epsilon,s,T)\,\\|f\\|_{L^1(0,T,H^s(\mathbb T))},$$ and when $f\equiv 0$ in $$\\|{\mathcal S}(t)u_0\\|_{Y_{s,T}}\le C(\epsilon,s,T)\\|u_0\\|_{H^s(\mathbb T)}.$$ [Step 2. Local Well-posedness in $H^s_0({\mathbb T}), \ s\ge 0$]{}\ We prove the following \[prop1\] Let $s\ge 0$. For any $u_0\in H^s_0({\mathbb T})$, there exists some $T>0$ such that the problem admits a unique solution $u\in Y_{s,T}$. [Proof. ]{} Write in its integral form $$u(t)={\mathcal S} (t) u_0-\int_0^t {\mathcal S}(t-\tau ) (uu_x )(\tau ) d\tau$$ where the spatial variable is suppressed throughout. For given $u_0\in H^s_0({\mathbb T}) $, let $r>0$ and $T>0$ be constants to be determined. Define a map $\Gamma$ on the closed ball $$B=\left\{ v\in Y_{s,T};\ \\|v\\|_{Y_{s,T}} \le r \right\}$$ of $Y_{s,T}$ by $$\Gamma (v) (t) ={\mathcal S} (t)u_0 - \int_0^t {\mathcal S}(t-\tau )(vv_x)(\tau )\, d\tau .$$ We aim to prove that $\Gamma$ contracts in $B$ for $T$ small enough and $r$ conveniently chosen. To that end, we shall prove the following estimates $$\begin{aligned} \\|\Gamma (v)\\|_{Y_{s,T}} &\le& C_0\\|u_0\\|_s + C_1T^{\frac{1}{4}} \\|v\\|^2_{Y_{s,T}} ,\quad \forall v\in B,\label{est1}\\ \\|\Gamma (v^1)-\Gamma (v^2)\\|_{Y_{s,T} } &\le& C_1 T^{\frac{1}{4}} (\\|v^1\\|_{Y_{s,T}} + \\|v^2\\|_{Y_{s,T}}) \\|v^1-v^2\\|_{Y_{s,T}} \quad \forall v^1,v^2\in B.\qquad \label{est2}\end{aligned}$$ From Lemma \[linearestim\] and Remark \[rem\], it is adduced that $$\begin{aligned} \\|\Gamma (v^1) -\Gamma (v^2)\\|_{Y_{s,T}} &\le& C\\| v^1 v^1_x - v^2 v^2_x \\|_{L^1(0,T,H^s({\mathbb T})) } \\ &\le& C\int_0^T \big( \\|v^1-v^2\\|_{L^\infty}\\|v^1+v^2\\|_{s+1} + \\|v^1 + v^2\\|_{L^\infty} \\|v^1-v^2\\|_{s+1}\big) d\tau \\ &\le&CT^{\frac{1}{4}} \\|v^1-v^2\\|_{Y_{s,T}} \big( \\|v^1\\|_{Y_{s,T}} + \\|v^2\\|_{Y_{s,T}} \big)\end{aligned}$$ where we used the fact that $$\int_0^T \\|v\\|_{L^\infty}^2 dt \le C\int_0^T \\|v\\|_1 \\|v\\| dt \le C\sqrt{T} \\|v\\|_{L^\infty (0,T,L^2({\mathbb T}))} \\|v\\|_{L^2(0,T,H^1({\mathbb T})) }.$$ This yields . follows from Lemma \[linearestim\], Remark \[rem\] and . Choosing $r>0$ and $T>0$ so that $$\left\{ \begin{array}{l} r=2C_0\\|u_0\\|_s,\\ 2rC_1T^{\frac{1}{4}}\le \frac{1}{2}, \end{array} \right.$$ we obtain that $$\\|\Gamma (v^1)\\| _{Y_{s,T}} \le r, \qquad \\|\Gamma (v^1) - \Gamma (v^2)\\|_{Y_{s,T}} \le \frac{1}{2} \\|v^1-v^2\\|_{Y_{s,T}}$$ for any $v^1,v^2\in B$. Thus, with this choice of $r$ and $T$, $\Gamma$ is a contraction in $B$. Its fixed-point is the unique solution of in $B$.\ [Step 3. Global Well-Posedness in $H^0_0({\mathbb T})$.]{}\ Assume that $u_0\in H^0_0({\mathbb T})$. We first establish for $0\le t\le T$. Since $u\in Y_{0,t}$, we have that $$\begin{aligned} \int_0^t \\|uu_x\\|^2_{-1} d\tau &\le& C \int_0^t \\|u^2\\|^2 d\tau \\ &\le& C \int_0^t \\|u\\|^3\\|u_x\\|\, d\tau \\ &\le& C \sqrt{t} \\|u\\|^4_{Y_{0,t}}.\end{aligned}$$ Thus each term in belongs to $L^2(0,t,H^{-1}({\mathbb T}))$. Scaling in by $u$ yields $$\int_0^t\langle u_t+({\mathcal H}-\varepsilon) u_{xx} +uu_x +G(D(Gu)), u\rangle _{H^{-1}({\mathbb T}),H^1({\mathbb T}) } d\tau =0.$$ We have that for a.e. $\tau\in (0,t)$ $$\begin{aligned} &&\langle ({\mathcal H}-\varepsilon)u_{xx},u\rangle _{H^{-1}({\mathbb T}),H^1({\mathbb T}) }= - ( ({\mathcal H } - \varepsilon )u_x,u_x)=\varepsilon \\|u_x\\|^2,\\ &&\langle uu_x,u\rangle _{H^{-1}({\mathbb T}), H^1({\mathbb T})} = (uu_x,u)=0,\\ &&\langle G(D(Gu)),u\rangle _{H^{-1}({\mathbb T}), H^1({\mathbb T})} =(G(D(Gu)),u) = \\|D^{\frac{1}{2}}(Gu)\\|^2.\end{aligned}$$ follows at once, and we infer that $\\|u(t)\\| \le \\|u_0\\|$. Using the standard extension argument, one sees that $u$ is defined on ${\mathbb R}^+$ with $u\in Y_{0,T}$ for all $T>0$. Furthermore, with the constants $C_0$ and $C_1$ given in Step 2 for $s=0$ and $T=(8C_0C_1\\|u_0\\|)^{-4}$, we obtain $$\\|u(nT+\cdot )\\|_{Y_{0,T}}\le 2C_0\\|u(nT)\\|\le 2C_0\\|u_0\\|.$$ [Step 4. Global Well-Posedness in $H^2_0({\mathbb T})$.]{}\ Pick any $u_0\in H^2_0({\mathbb T})$. By Proposition \[prop1\] and Step 3, admits a unique solution $u\in Y_{0,T}$ for each $T>0$, which belongs to $Y_{2,T_0}$ for some $T_0>0$. We just need to show that $T_0$ may be taken as large as desired. Let $v=u_t$. If $u\in Y_{2,T}$, then $v\in Y_{0,T}$ and it satisfies $$\label{eqv} v_t + ( {\mathcal H} - \varepsilon )v_{xx} + (uv)_x =-G(D(Gv)),\qquad v(0)=v_0$$ where $$v_0:=-\left\{ ({\mathcal H}-\varepsilon)u_{0,xx} +u_0u_{0,x} + G(D(G u_0)) \right\} \in H^0_0({\mathbb T}).$$ We may write in its integral form $$v(t) = {\mathcal S } (t) v_0 - \int_0^t {\mathcal S} (t-s) (uv)_x(s) ds.$$ Let $\Gamma (w) (t)={\mathcal S } (t) v_0 - \int_0^t {\mathcal S} (t-s) (uw)_x(s) ds$ for $w\in Y_{0,T}$. Computations similar to those in Step 2 lead to $$\begin{aligned} \\|\Gamma w\\|_{Y_{0,T}} &\le & C_0\\|v_0\\| + C_1 T^{\frac{1}{4}} \\|u\\|_{Y_{0,T}}\\|w\\|_{Y_{0,T}},\\ \\|\Gamma (w^1) -\Gamma (w^2) \\|_{Y_{0,T}} &\le& C_1 T^{\frac{1}{4}} \\|u\\|_{Y_{0,T}} \\|w^1-w^2\\|_{Y_{0,T}}\end{aligned}$$ where the constants $C_0$ and $C_1$ depend only on $\varepsilon$ for $T<1$. Therefore $\Gamma$ contracts in $B= \{w\in Y_{0,\theta};\ \\|w\\|_{Y_{0,\theta }} \le r:= 2C_0\\|v_0\\|\}$, provided that $$C_1\theta ^{\frac{1}{4}}\\|u\\|_{Y_{0,\theta}} \le \frac{1}{2} \cdot$$ Its fixed point gives the unique solution of the integral equation in $B$. Pick $\theta$ fulfilling $$\theta<\min \{ (8C_0C_1\\|u_0\\|)^{-4}, 1\} .$$ Then, from Step 2, we have that $$\\|u(n\theta +\cdot )\\|_{Y_{0,\theta}}\le 2C_0 \\|u_0\\|$$ for all $n\in {\mathbb N}$ and that $w$ may be extended to $[n\theta, (n+1)\theta ]$ inductively by using the contraction mapping theorem (replacing $v_0$ by $w(\theta )$, $w(2\theta )$, etc.). Therefore, $w$ is defined on ${\mathbb R}^+$ and it holds $$\label{eqw} \\| w ( n\theta + \cdot ) \\|_{Y_{0,\theta }} \le 2C_0 \\|w(n\theta )\\|\le (2C_0)^{n+1} \\|v_0\\|.$$ By uniqueness of the solution of the integral equation, we have that $v(t)=w(t)$ as long as $0<t<T$ and $v\in Y_{0,T}$. shows that $\\|v(t)\\|=\\|w(t)\\|$ is uniformly bounded on compact sets of ${\mathbb R}^+$, namely $$\\|v\\|_{Y_{0,T}} \le C(T,\\|u_0\\|) \\|v_0\\| .$$ The same is true for $\\|u(t)\\|_2$, by . Indeed, since $$\\|uu_x\\| \le \\|u\\|_{L^\infty ({\mathbb T})} \\|u_x\\|\le \\|u\\|^{\frac{5}{4}}\\|u_{xx}\\|^{\frac{3}{4}} \le C_\delta \\|u\\|^5 + \delta \\|u_{xx}\\| ,$$ we infer from that $$\\| ( {\mathcal H}-\varepsilon ) u_{xx} (t)\\| \le C(T,\\|u_0\\|) \\|u_0\\|_2 + C(\\|u\\| + \\|u\\|^5) + \delta \\|u_{xx}\\|$$ hence $$\\|u(t)\\|_2 \le C(T, \\|u_0\\|) \\|u_0\\|_2.$$ Using the standard extension argument, one sees that $u(t)\in H^2_0({\mathbb T})$ for all $t\ge 0$ with $u\in Y_{2,T}$ for all $T > 0$.\ [Step 5. Smoothing effect from $H^0_0({\mathbb T})$ to $H^2_0({\mathbb T})$.]{}\ Pick any $u_0\in H^0_0({\mathbb T})$. Then the solution $u$ to belongs to $Y_{0,1}$. Therefore, for a.e. $t_0\in (0,1)$, $u(t_0)\in H^1_0({\mathbb T})$. The solution of in $Y_{1,T}$ issued from $u(t_0)$ at $t=0$ must coincide with $u(t_0+t)$ in $[0,T]$, by uniqueness of the solution of in $Y_{0,T}$. In particular, $u(t_1)\in H^2_0({\mathbb T})$ for a.e. $t_1>t_0$. Again by uniqueness we conclude that $u\in C([t_1,+\infty ),H^2_0({\mathbb T}))$ for a.e. $t_1>0$, so that $$u\in C((0,+\infty ), H^2_0({\mathbb T}))\cap C^1((0,+\infty ),H^0_0({\mathbb T})).$$ The proof of Theorem \[GWP\] is complete. The following commutator lemma, used several times in the proof of the property of propagation of regularity, is a periodic version of a result from [@DMP]. \[commutator\] Let ${\mathcal N}\subset {\mathbb Z}$ be a set such that for some constant $C>0$ $${\langle}n {\rangle}+ {\langle}k {\rangle}\le C{\langle}n - k {\rangle},\qquad \forall n\not\in{\mathcal N},\ \forall k\in {\mathcal N}. \label{separation}$$ Let $P$ be the projector on the closure of $\text{Span}\{ e^{ikx};\ k\in \mathcal N \}$ in $L^2({\mathbb T})$, namely $$P (\sum_{k\in {\mathbb Z}}\hat u_k e^{ikx})=\sum_{k\in \mathcal N}\hat u_ke^{ikx}.$$ Let $a\in C^\infty ({\mathbb T})$ and let $p\in {\mathbb N},\ q\in {\mathbb N}$. Then there exists some constant $C=C(a,p,q)>0$ such that for all $v\in L^2({\mathbb T})$ $$\label{estim_commutator} \\|\partial _x^p [a,P]\partial _x ^q v\\| \le C \\|v\\|.$$ \[rmk2.6\] Note that condition is fulfilled in the following cases: (i) ${\mathcal N}={\mathbb N}^$; (ii) ${\mathcal N}$ is a finite set, or the complement of a finite set in ${\mathbb Z}$. It follows that is true with $P={\mathcal H}=(-i)(P_{{\mathbb N}^} - P_{-{\mathbb N}^})$. Note, however, that condition and are not true when $\mathcal N=1+2{\mathbb Z}$ (pick e.g. $a(x)=e^{ix}$). [Proof of Lemma \[commutator\].]{} Let ${\mathcal N},a,p$ and $q$ be as in the statement of the lemma, and pick any $v\in C^\infty ({\mathbb T})$. Decompose $a$ and $v$ in using Fourier series $$v(x)=\sum_{n\in {\mathbb Z}}{\hat v}_n e^{inx},\qquad a=\sum_{n\in {\mathbb Z}} {\hat a}_n e^{inx},$$ and denote by $1_{\mathcal N}$ the characteristic function of $\mathcal N$, defined by $1_{\mathcal N}(n)=1$ if $n\in \mathcal N$, and $0$ otherwise. Then $$\begin{aligned} [a,P]v &=& a(Pv) -P(av)\\ &=& a (\sum_{n} 1_{\mathcal N}(n) {\hat v}_n e^{inx}) - P(\sum_n(\sum_k {\hat a}_{n-k}{\hat v}_k )e^{inx}) \\ &=& \sum_{n} \left( \sum_k {\hat a}_{n-k}{\hat v}_k (1_{\mathcal N}(k) -1_{\mathcal N} (n)) \right) e^{inx}.\end{aligned}$$ Taking derivatives, one obtains $$\partial_x^p [a,P]\partial _x^q v = \sum_n \left( \sum_k {\hat a}_{n-k}(ik)^q{\hat v}_k (1_{\mathcal N}(k) -1_{\mathcal N}(n))\right)(in)^p e^{inx} =:\Sigma _1 - \Sigma _2$$ where $\Sigma _1$ (resp. $\Sigma _2$) is the sum over the $(n,k)$ with $n\not\in\mathcal N$ and $k\in\mathcal N$ (resp. with $n\in\mathcal N$ and $k \not\in\mathcal N$). Let us estimate $\Sigma _1$ only, the estimate for $\Sigma _2$ being similar. Since $a\in C^\infty ({\mathbb T})$, for any $s\in {\mathbb N}$ there exists some constant $C_s>0$ such that $$\label{smooth} \|{\hat a}_l\|\le C_s \langle l\rangle ^{-s}\qquad \forall l\in {\mathbb Z}.$$ Then, for $s>\sup\{ 2p+1,2q+1\}$, $$\begin{aligned} \\|\Sigma _1 \\|^2_{L^2({\mathbb T})} &=& \\| \sum_{n \not\in\mathcal N}(\sum_{k\in\mathcal N} {\hat a}_{n-k}{\hat v}_k (ik)^q (in)^p )e^{inx} \\|^2_{L^2({\mathbb T})} \\ &=& C \sum_{n\not\in\mathcal N} \left\vert \sum_{k\in\mathcal N}{\hat a}_{n-k} {\hat v}_k(ik)^q \right\vert ^2 \|n\|^{2p}\\ &\le& C \\|v\\|^2\sum_{n\not\in\mathcal N}\sum_{k\in\mathcal N}\langle n-k\rangle ^{-2s} \|n\|^{2p}\|k\|^{2q}\\ &\le& C \\|v\\|^2\sum_{n\not\in\mathcal N}\sum_{k\in\mathcal N} (\langle n \rangle + \langle k\rangle ) ^{-2s} \|n\|^{2p}\|k\|^{2q}\\ &\le& C\\|v\\|^2\end{aligned}$$ where we used the Cauchy-Schwarz inequality, and . Since $C^\infty ({\mathbb T})$ is dense in $L^2({\mathbb T})$, the proof is complete.\ The propagation of regularity property we need is as follows. \[smoothing\] Let $a\in C^\infty ({\mathbb T}, {\mathbb R}^+)$, $\varepsilon >0$, $\alpha \in{\mathbb R}$, $T>0$, and $R>0$ be given. Pick any $v_0\in H^0_0({\mathbb T})$ with $\\|v_0\\| \le R$ and let $v\in C([0,T];H^0_0({\mathbb T}))\cap L^2(0,T,H^1({\mathbb T}))\cap C((0,T],H^2({\mathbb T}))$ be such that $$\begin{aligned} && v_t + ({\mathcal H} -\varepsilon ) v_{xx} + \alpha v v_x = - G (D(G \,v)),\qquad x\in {\mathbb T},\ t\in (0,T)\\ &&v(0)=v_0.\end{aligned}$$ Then there exists some constant $C=C(T)>0$ (independent of $\varepsilon$, $\alpha$ and $R$) such that $$\label{H12} \int_0^T \\|D^{\frac{1}{2}}v\\|^2dt \le C(R^2 + \alpha ^4 R^6).$$ [Proof of Proposition \[smoothing\]]{}. Pick any $t_0\in (0,T)$. Let $(f,g)_{L^2_{t,x}}:=\int_{t_0}^T \int_{{\mathbb T}} f(x,t)g(x,t)\, dxdt$ denote the scalar product in $L^2(t_0,T,L^2({\mathbb T}))$. $C$ will denote a constant which may vary from line to line, and which may depend on $T$, but not on $t_0$, $\varepsilon$, $\alpha$ and $R$. Setting $Lv:=v_t+{\mathcal H}v_{xx}$, $f:=\varepsilon v_{xx} -G(D(Gv))$ and $g:=-\alpha vv_x$, we have that $$Lv=f+g.$$ Pick any $\varphi\in C^\infty ({\mathbb T})$, and set $Av=\varphi (x)v$. Noticing that $L$ is formally skew-adjoint, we have that $$\begin{aligned} ([L,A]v,v)_{L^2_{t,x}} &=&(L(\varphi v) -\varphi (Lv),v)_{L^2_{t,x}}\\ &=&(\varphi v, L^v)_{L^2_{t,x}} + [(\varphi v,v)]_{t_0}^T -(Lv,\varphi v)_{L^2_{t,x}}\end{aligned}$$ so that $$\| ([L,A]v,v)_{L^2_{t,x}} \| \le 2\|(f + g, \varphi v)_{L^2_{t,x}}\| + 2 \\|\varphi\\|_{L^\infty ({\mathbb T})} R^2 .$$ We first notice that $$\begin{aligned} \|(f,\varphi v)_{L^2_{t,x}}\| &\le& \|(v_x,\varepsilon (\varphi v)_x)_{L^2_{t,x}}\| + \|(D(Gv),G(\varphi v))_{L^2_{t,x}}\|\\ &\le& C \varepsilon \int_0^T \int_{{\mathbb T}} (\|v\|^2 + \|v_x\|^2)dt + C\int_0^T \\|D^{\frac{1}{2}}(Gv)\\|^2 dt \\ &&\qquad +\int_0^T \{ \|(D(Gv),[G,\varphi ] v)\| +\|(D^{\frac{1}{2}}(Gv),[D^{\frac{1}{2}},\varphi](Gv))\| \} d\tau \\ &\le& CR^2\end{aligned}$$ where we used and classical commutator estimates. (Note that Theorem \[GWP\] is still true when $\alpha =1$ is replaced by any value $\alpha \in {\mathbb R}$.) On the other hand $$\|(g,\varphi v)_{L^2_{t,x}}\|=\|(\alpha vv_x,\varphi v)_{L^2_{t,x}}\|=\frac{\| \alpha \|}{3} \|(v^3,\varphi _x)_{L^2_{t,x}}\|.$$ From Sobolev embedding and the fact that the $L^2-$norm is nonincreasing $$\\|v\\|_{L^3} \le \\|v\\|^{\frac{1}{2}}\\|v\\|^{\frac{1}{2}}_{L^6} \le C R^{\frac{1}{2}} \\|v\\|^{\frac{1}{2}}_{\frac{1}{2}} \cdot$$ Therefore, $$\begin{aligned} \|(g, \varphi v)_{L^2_{t,x}} \| &\le& C\| \alpha \| \int_{t_0}^T\\|v\\|^3_{L^3} dt \\ &\le& C\|\alpha \| R^{\frac{3}{2}}T^{\frac{1}{4}} \left( \int_{t_0}^T \\|v\\|^2_{\frac{1}{2}}dt\right)^{\frac{3}{4}}\\ &\le& C\delta ^{-3}\alpha ^4 R^{6} T + \delta \int_{t_0}^T\\| D^{\frac{1}{2}} v\\|^2 dt\end{aligned}$$ where $\delta >0$ will be chosen later on. On the other hand $$\begin{aligned} [L,A]v &=& [{\mathcal H}\partial _x^2,\varphi]v \nonumber \\ &=& {\mathcal H}\big( (\partial _x^2\varphi )v +2 (\partial _x \varphi) (\partial _x v) +\varphi\partial _x^2v \big) -\varphi {\mathcal H}\partial _x^2 v \nonumber \\ &=& [{\mathcal H},\varphi ] \partial _x ^2 v + {\mathcal H}\big( (\partial^2 _x \varphi )v\big) +2[{\mathcal H},\partial _x \varphi ]\partial _x v + 2(\partial _x\varphi) {\mathcal H}\partial _x v. \label{K1}\end{aligned}$$ It follows from Lemma \[commutator\] and Remark \[rmk2.6\] that $$\begin{aligned} &&\| ( [{\mathcal H},\varphi ] \partial _x^2 v , v )_{L^2_{t,x}} \| + \|\big( {\mathcal H}( (\partial _x ^2 \varphi ) v), v \big)_{L^2_{t,x}} \| + \|\big( [{\mathcal H},\partial _x\varphi ] \partial _x v , v \big) _{L^2_{t,x}} \| \nonumber\\ &&\qquad \le C\\|v\\|^2_{L^2(0,T;L^2({\mathbb T}))}\nonumber \\ &&\qquad \le C R ^2. \label{K2}\end{aligned}$$ Therefore $$\|(\partial _x \varphi {\mathcal H}\partial _x v, v)_{L^2_{t,x}}\| \le C( R^2 +\delta ^{-3}\alpha ^4 R^6) + \delta \int_{t_0}^T \\|D^\frac{1}{2} v\\|^2 dt.$$ Let $b\in C_0^\infty(\omega)$, where $\omega = \{ x\in {\mathbb T}; \ a(x)>0 \}$. Then $b=a{\tilde b }$ with ${\tilde b}\in C_0^\infty (\omega)$ and $$\begin{aligned} \int_{t_0}^T \\|D^{\frac{1}{2}} (bv)\\|^2 dt &\le& 2\int_{t_0}^T \big( \\| [D^{\frac{1}{2}},\tilde b](av)\\|^2 + \\|\tilde b D^{\frac{1}{2}}(av)\\|^2\big) dt\nonumber \\ &\le& C\int_{t_0}^T (\\|v\\|^2 + \\|D^{\frac{1}{2}} (av)\\|^2) dt \nonumber \\ &\le& C\int_0^T \big( \\|v\\|^2 + \\|D^{\frac{1}{2}}(Gv)\\|^2 +\\|D^{\frac{1}{2}}a\\|^2 \|\int_{{\mathbb T}}a(y) v(y,t)\, dy\|^2 \big)dt \nonumber\\ &\le& CR^2. \label{Z1}\end{aligned}$$ Pick any $x_0\in {\mathbb T}$. Then $b^2(x)-b^2(x-x_0)=\partial _x\varphi $ for some $\varphi\in C^\infty ({\mathbb T})$. Noticing that ${\mathcal H}\partial _x =D$, we have that $$\begin{aligned} \|(b^2(x){\mathcal H}\partial _x v,v)_{L^2_{t,x}}\| &=& \|(b D v, b v)_{L^2_{t,x}}\| \\ &\le& \| ([b,D]v,bv)_{L^2_{t,x}} \| + \|( D(bv),bv)_{L^2_{t,x}}\|\\ &\le& C\\|v\\|^2_{L^2(0,T;L^2({\mathbb T}))} + \int_{t_0}^T \\| D^{\frac{1}{2}} (bv)\\|^2dt\\ &\le& CR^2\end{aligned}$$ by . It follows that $$\|(b^2(x-x_0)Dv,v)_{L^2_{t,x}}\|\le C(R^2 +\delta ^{-3}\alpha ^4 R^6) + \delta \int_{t_0}^T \\|D^\frac{1}{2} v\\|^2dt.$$ Using a partition of unity and choosing $\delta >0$ small enough, we infer that $$\|(Dv,v)_{L^2_{t,x}}\| \le C(R^2 +\alpha ^4 R^6) +\frac{1}{2} \int_{t_0}^T \\|D^\frac{1}{2} v\\|^2 dt.$$ This gives $$\int_{t_0}^T \\|D^\frac{1}{2} v\\|^2dt \le C(R^2 +\alpha ^4 R^6),$$ where $C=C(T)$. Letting $t_0\to 0$ yields the result.\ A unique continuation property is also required. \[unique\_continuation\] Let $\alpha\in {\mathbb R}$, $\varepsilon \ge 0$, $c\in L^2(0,T)$, and $u\in L^2(0,T;H^0_0({\mathbb T}) )$ be such that $$\begin{aligned} u_t+ ({\mathcal H} -\varepsilon )u_{xx} +\alpha uu_x=0 &&\text{ in }\ {\mathbb T}\times (0,T),\label{UCP1}\\ u(x,t)=c(t) &&\text{ for a.e. }\ (x,t)\in (a,b)\times (0,T)\label{UCP2}\end{aligned}$$ for some numbers $T>0$ and $0\le a<b\le 2\pi$. Then $u(x,t)=0$ for a.e. $(x,t)\in {\mathbb T}\times (0,T)$. [Proof.]{} From , we obtain that $u_{xx}(x,t)=(uu_x)(x,t)=0$ for a.e. $(x,t)\in (a,b)\times (0,T)$. Thus, by using , $${\mathcal H} u_{xx} =-u_t= -c_t \qquad \text{ in } (a,b)\times (0,T).$$ Therefore, for almost every $t\in (0,T)$, it holds $$\begin{aligned} && u_{xxx}(\cdot,t)\in H^{-3}({\mathbb T}),\\ && u_{xxx}(\cdot,t)=0 \ \text{ in }\ (a,b),\\ && {\mathcal H}u_{xxx}(\cdot,t)=0 \ \text{ in }\ (a,b).\end{aligned}$$ Pick a time $t$ as above, and set $v=u_{xxx}(\cdot,t)$. Decompose $v$ as $$v(x)=\sum_{k\in {\mathbb Z}} \hat v_k e^{ikx},$$ the convergence of the Fourier series being in $H^{-3}({\mathbb T})$. Then in $(a,b)$ $$0=iv - {\mathcal H}v = 2i\sum_{k > 0}\hat v_k e^{ikx}.$$ Since $v$ is real-valued, we also have that $\hat v_{-k}=\overline{\hat v_k}$ for all $k$. The following lemma for Fourier series is needed. \[Fourier\] Let $s\in {\mathbb R}$ and let $v(x) =\sum_{k\ge 0}\hat v_k e^{ikx}$ be such that $v\in H^s({\mathbb T})$ and $v=0$ in $(a,b)$. Then $v\equiv 0$. [Proof of Lemma \[Fourier\].]{} It is clearly sufficient to prove the property for $s=-p$, where $p\in {\mathbb N}$. Let us proceed by induction on $p$. Assume first that $p=0$. Then $$\label{L2} \sum_{k\ge 0}\|\hat v_k\|^2 <\infty.$$ Introduce the set $U=\{ z\in {\mathbb C};\ \|z\|<1 \}$ and the Hardy space (see e.g. [@rudin]) $${\mathbf H}^2 (U) =\{ f : U\to {\mathbb C};\ f \text{ is holomorphic in $U$ and } \limsup_{r\to 1^-} \int_{-\pi}^\pi \|f(re^{i\theta})\|^2 d\theta <\infty \}$$ Let $f(z)=\sum_{k\ge 0}\hat v_k z^k$. Then, by [@rudin Thm 17.10] and , we have that $f\in {\mathbf H}^2(U)$. On the other hand, by [@rudin Thm 17.10 and Thm 17.18], it holds that $$\begin{aligned} &&f^* ( e^{i\theta} ): =\lim_{r\to 1^-} f( re^{i\theta} ) \text{ exists for a.e. } \theta \in (0,2\pi); \label{F1}\\ && f^(e^{i\theta}) =\sum_{k\ge 0} \hat v_k e^{ik\theta } = v(\theta ) \qquad \text{ in } L^2({\mathbb T}); \label{F1bis}\\ &&\text{If } f\not\equiv 0,\ \ \text{then} \ f^(e^{i\theta })\ne 0 \ \ \text{for a.e. } \theta\in (0,2\pi ).\label{F2}\end{aligned}$$ Since $$f^(e^{i\theta })=v(\theta )=0\qquad \text{ for a.e. } \theta \in (a,b),$$ it follows from that $f\equiv 0$. Therefore $\hat v_k=0$ for all $k\ge 0$, hence $v\equiv 0$. This gives the result for $p=0$. Assume now that the result has been proved for $s= - p$ for some $p\in {\mathbb N}$, and pick any $v\in H^{-p-1}({\mathbb T})$, decomposed as $v(x)=\sum_{k\ge 0}\hat v_k e^{ikx}$, and such that $v\equiv 0$ in $(a,b)$. Let $w(x)=\sum_{k>0} \frac{\hat v_{k-1}}{ik} e^{ikx}$. Then $w\in H^{-p}({\mathbb T})$ and $$w_x=\sum_{k>0} \hat v_{k-1}e ^{ikx} = e^{ix}v,$$ so $w_x=0$ on $(a,b)$ and we have, for some constant $C\in{\mathbb C}$, $$\label{constant} w(x)=C\ \ \text{ on } \ (a,b).$$ Introducing the function $\tilde w(x)=w(x)-C$, we infer from and the induction hypothesis that $\tilde w\equiv 0$ on ${\mathbb T}$, which yields $v\equiv 0$ on ${\mathbb T}$. This completes the proof of Lemma \[Fourier\]. With Lemma \[Fourier\] we infer that for a.e. $t\in (0,T)$, $u_{xxx}(.,t)=0$ in ${\mathbb T}$, hence with $u(x,t)=c(t)$ a.e. in ${\mathbb T}\times (0,T)$. From we infer that $c_t=0$, which, combined with the fact that $u\in L^2(0,T;H^0_0({\mathbb T}))$, gives that $u(x,t)=0$ a.e. in ${\mathbb T}\times (0,T)$. The proof of Proposition \[unique\_continuation\] is complete. We are now in a position to state a stabilization result for the $\varepsilon$-BO equation. We stress that the decay rate does not depend on $\varepsilon$. \[thmstab1\] Let $R>0$. There exist some numbers $\lambda >0$ and $C>0$ such that for any $\varepsilon \in (0,1]$ and any $u_0\in H^0_0 ({\mathbb T})$ with $\\|u_0\\|\le R$, the solution $u$ of satisfies $$\\|u(t)\\| \le Ce^{-\lambda t}\\|u_0\\|\qquad \forall t\ge 0.$$ [Proof.]{} Note that $\\|u(t)\\|$ is nonincreasing by , so that the exponential decay is ensured if $\\|u((n+1)T)\\|\le \kappa \\|u(nT)\\|$ for some $\kappa <1$. To prove the theorem, it is thus sufficient (with ) to establish the following observability inequality: for any $T>0$ and any $R>0$ there exists some constant $C(T,R) >0$ such that for any $\varepsilon \in (0,1]$ and any $u_0\in H^0_0({\mathbb T}) $ with $\\|u_0\\|\le R$, it holds $$\label{obs} \\|u_0\\|^2 \le C\left( \varepsilon \!\! \int_0^T \\|u_x(t )\\|^2 dt + \int_0^T \\|D^{\frac{1}{2}}(Gu)\\|^2 dt\right) ,$$ where $u$ denotes the solution of . Fix any $T>0$ and any $R>0$, and assume that fails. Then there exist a sequence $ (u_0^n )$ in $H^0_0({\mathbb T})$ and a sequence $ (\varepsilon ^n )$ in $(0,1]$ such that for each $n$ we have $\\|u_0^n\\|\le R$, and $$\\|u^n_0\\|^2 > n \left( \varepsilon ^n \!\! \int_0^T \\|u^n_x (t)\\|^2 \, dt + \int_0^T \\|D^{\frac{1}{2}}(G \,u^n)\\|^2dt\right) .$$ Let $\alpha ^n=\\|u_0^n\\|\in (0,R]$. Extracting a sequence if needed, we may assume that $\alpha ^n\to \alpha \in [0,R]$ and $\varepsilon ^n\to \varepsilon \in [0,1]$. Let $v^n=u^n/\alpha ^n$. Then $v^n$ solves $$\label{BO2} v^n_t+ ({\mathcal H}-\varepsilon ^n)v^n_{xx} + \alpha ^n v^n v^n_x = - G (D ( G v^n)), \qquad v^n(0)=v^n_0$$ with $v^n_0\in H^0_0({\mathbb T}) $ and $\\|v^n_0\\|=1$. Again, we have that $$\begin{aligned} &&\frac{1}{2} \\|v^n(t)\\|^2 + \varepsilon ^n \int_0^t \\|v^n_x \\|^2 d\tau + \int_0^t \\|D^{\frac{1}{2}}(G v^n)\\|^2 d\tau =\frac{1}{2} \\|v_0^n\\|^2\qquad \forall t>0,\label{X1}\\ &&1=\\|v_0^n\\|^2 > n \left( \varepsilon ^n \int_0^T \\|v^n_x(t)\\|^2 dt + \int_0^T \\|D^{\frac{1}{2}}(G v^n)\\|^2\, dt \right) . \label{X2}\end{aligned}$$ We infer from Proposition \[smoothing\] that $$\label{H12bis} \int_0^T \\|D^\frac{1}{2} v^n\\|^2dt \le C.$$ This yields $$\\|G( D(Gv^n))\\|_{ L^2(0,T; H^{-\frac{1}{2}} ({\mathbb T})) } + \\| ( {\mathcal H}-\varepsilon )v^n_{xx}\\|_{L^2(0,T;H^{-\frac{3}{2}} ({\mathbb T})) } \le C.$$ On the other hand, for any $\delta >0$ $$\\|v^nv^n_x\\|_{H^{-\frac{3}{2}-\delta}({\mathbb T})} \le C \\|(v^n)^2\\|_{ H^{-\frac{1}{2}-\delta } ({\mathbb T})} \le C \\|(v^n)^2\\|_{L^1({\mathbb T})} \le C \\|v^n\\|^2 \le C$$ thus $$\\|\alpha ^n v^n v^n_x\\|_{L^2(0,T;H^{-\frac{3}{2}-\delta}({\mathbb T}))} \le C.$$ It follows that $ (v^n_t )$ is bounded in $L^2(0,T; H^{-\frac{3}{2} -\delta }({\mathbb T}))$. Combined with and Aubin-Lions’ lemma, this gives that for a subsequence still denoted by $ (v^n )$, we have $$\begin{aligned} &&v^n\to v\qquad \text{ in } L^2(0,T;H^\alpha ({\mathbb T})) \qquad\forall \alpha <\frac{1}{2}, \\ &&v^n\to v\qquad \text{ in } L^2(0,T;H^{\frac{1}{2}} ({\mathbb T}))\text{ weak} , \ \\ &&v^n\to v\qquad \text{ in } L^\infty (0,T;L^2 ({\mathbb T})) \text{ weak} \\end{aligned}$$ for some function $ v\in L^2(0,T;H_0^\frac{1}{2} ({\mathbb T})) \cap L^\infty (0,T;L^2({\mathbb T}))$. In particular, $$(v^n)^2\to v^2 \qquad \text{ in } L^1({\mathbb T}\times (0,T)).$$ Letting $n\to \infty$ in , we obtain that $$\int_0^T\\|D^{\frac{1}{2}} (Gv)\\|^2 dt =0,$$ hence $Gv=0$ a.e. on ${\mathbb T}\times (0,T)$. Recall that $\omega=\{ x\in {\mathbb T}; \ a(x)>0 \}$. Then $$v(x,t)=\int_{{\mathbb T}}a(y)v(y,t)\, dy= : c(t)\qquad \text{ for a.e. } (x,t)\in \omega\times (0,T).$$ Note that $c\in L^\infty (0,T)$. Taking the limit in gives $$\left\{ \begin{array}{ll} v_t+({\mathcal H}-\varepsilon ) v_{xx}+\alpha vv_x=0,\qquad &\text{in } {\mathbb T}\times (0,T),\\ v(x,t) =c(t)\qquad &\text{for a.e. } (x,t) \in \omega \times (0,T). \end{array} \right.$$ It follows from Proposition \[unique\_continuation\] that $v\equiv 0$. Thus, extracting a subsequence still denoted by $ (v^n )$, we have that $v^n(\cdot ,t)\to 0$ in $L^2({\mathbb T})$ for a.e. $t\in (0,T)$. Using -, we infer that $v^n_0\to 0$ in $L^2({\mathbb T})$. This contradicts the fact that $\\|v^n_0\\|=1$ for all $n$. We are now in a position to define the weak solutions of obtained by the method of vanishing viscosity, and to state the corresponding exponential stability property. \[defiweak\] For $u_0\in H^0_0({\mathbb T})$, we call a [weak solution of in the sense of vanishing viscosity]{} any function $u\in C_w({\mathbb R}^+, H^0_0({\mathbb T}))$ with $u\in L^2(0,T, H^{\frac{1}{2}}({\mathbb T}))$ for all $T>0$ which solves (in the distributional sense) and such that for some sequence $\varepsilon ^n \searrow 0$ we have for all $T>0$ $$\begin{aligned} u^n &\to& u \qquad \text{ in } L^\infty (0,T,H^0_0 ({\mathbb T}))\ \text{weak}\, ,\\ u^n &\to& u \qquad \text{ in } L^2(0,T,H^{\frac{1}{2}} _0({\mathbb T}))\ \text{weak}\end{aligned}$$ where $u^n$ solves for $\varepsilon =\varepsilon ^n$. The main result in this section is the following \[Thmstab\] For any $u_0\in H^0_0({\mathbb T})$ there exists (at least) one weak solution of in the sense of vanishing viscosity. On the other hand, for all $R>0$ there exist some positive constants $\lambda =\lambda (R)$ and $C=C(R)$ such that for any weak solution $u(t)$ of in the sense of vanishing viscosity, it holds $$\label{decay} \\|u(t)\\| \le Ce^{-\lambda t} \\|u_0\\| \qquad \forall t\ge 0$$ whenever $\\|u_0\\|\le R$. [Proof.]{} Pick $R>0$ and $u_0\in H^0_0({\mathbb T}) $ with $\\|u_0\\|\le R$. Pick any sequence $\varepsilon ^n\searrow 0$ and let $u^n(t)$ denote the solution of $$\label{BOn} u^n_t + ({\mathcal H}-\varepsilon ^n) u^n_{xx} +u^n u^n_x = -G(DG u^n),\quad u^n(0)=u_0.$$ It follows from and that $$\begin{aligned} &&\\|u^n\\|_{L^\infty (0,T,H^0_0({\mathbb T}))} \le R,\\ &&\\|u^n\\|_{L^2(0,T,H_0^{\frac{1}{2}} ( {\mathbb T}))} \le C(T,R).\end{aligned}$$ Using a diagonal process, we obtain that for a subsequence, still denoted by $( u^n )$, we have for all $T>0$ $$\begin{aligned} &&u^n\to u \qquad \text{ in } L^\infty(0,T,H^0_0({\mathbb T})) \ \text{ weak},\label{H99}\\ &&u^n\to u \qquad \text{ in } L^2(0,T,H_0^{\frac{1}{2}} ({\mathbb T})) \ \text{weak} \label{H100}\end{aligned}$$ for some function $u\in L^\infty ({\mathbb R}^+ ,H^0_0({\mathbb T}))\cap L^2_{loc}({\mathbb R}^+,H^{\frac{1}{2}}_0({\mathbb T}))$. The same argument as in the proof of Theorem \[thmstab1\] shows that $\{ u^n_t \}$ is bounded in $L^2(0,T; H^{-\frac{3}{2} -\delta }({\mathbb T}))$ for all $\delta >0$. Combined with - and Aubin-Lions’ lemma, this shows that $$u^n\to u \ \text{ in }\ L^2({\mathbb T}\times (0,T)) \text { and in } C([0,T],H_0^{-\delta }({\mathbb T}))$$ for all $T>0$ and all $\delta >0$. On the other hand, $u\in C([0,T],H^{-\delta}({\mathbb T}))$ for all $T>0$ and all $\delta >0$, which, combined to , yields $u\in C_w({\mathbb R}^+,H^0_0({\mathbb T}))$ (the space of weakly continuous functions from ${\mathbb R}^+$ to $H^0_0({\mathbb T})$). By letting $n\to \infty$ in , we see that $u$ solves . Thus $u$ is a weak solution of in the sense of vanishing viscosity. On the other hand, from Theorem \[thmstab1\] we have that $$\\|u^n(t)\\| \le C e^{-\lambda t}\\|u_0\\| ,\quad \forall t\ge 0, \ \forall n\ge 0.$$ where $C=C(R)$, $\lambda = \lambda (R)$. Letting $n\to \infty$ in the above estimate yields . Note also that $\\|u(t)\\|\le \\|u_0\\|$ for all $t\ge 0$, since the same estimate holds for the $u^n$’s and $u\in C_w ({\mathbb R}^+ ,H^0_0({\mathbb T}))$. Local stabilization in $H^s_0({\mathbb T})$ ------------------------------------------- \ ### Main results Let again $a$ and $G$ be as in and , respectively. For $s\ge 0$ and $T>0$, let $$\label{R2} Z_{s,T}=C([0,T],H^s_0({\mathbb T})) \cap L^2(0,T, H_0^{s+\frac{1}{2}} ({\mathbb T}))$$ be endowed with the norm $$\\| v\\|_{Z_{s,T}} =\\|v\\|_{L^\infty(0,T,H^s({\mathbb T}))} + \\|v\\|_{L^2(0,T,H^{s+\frac{1}{2}}({\mathbb T}))} \cdot$$ We are concerned here with the stability properties of the BO equation with localized damping in the space $H^s_0({\mathbb T})$ for $s>0$. Our first aim is to prove the local well-posedness of in $H^s_0({\mathbb T})$ for $s>1/2$. \[thmA\] Let $s\in (\frac{1}{2},2]$. Then there exists $\rho >0$ such that for any $u_0\in H^s_0({\mathbb T})$ with $\\|u_0\\|_s <\rho $, there exists some time $T>0$ such that admits a unique solution in the space $Z_{s,T}$. The proof of Theorem \[thmA\] rests on the smoothing effect due to the damping term, namely $$\label{R3} \int_0^T \\|e^{-t({\mathcal H} \partial _x ^2 +GDG)} u_0\\|^2_{\frac{1}{2}} dt \le C\\|u_0\\|^2.$$ In [@RZ2006], the semi-global exponential stability of the Korteweg-de Vries on a bounded domain $(0,L)$ with a localized damping was first established in $L^2(0,L)$, and next extended to $\{u\in H^3(0,L);\ \ u(0)=u(L)=u_x(L)=0\}$ by using the Kato smoothing effect in the equation fulfilled by the time derivative of the solution. As the smoothing effect is much weaker, that argument cannot be used. The semi-global exponential stability of in $H^0_0({\mathbb T})$, if true, is thus open. However, a local exponential stability in $H^s_0({\mathbb T})$ for $s>1/2$ can be derived. \[thmB\] Let $s\in(\frac{1}{2},2]$. Then there exist some numbers $\rho >0$, $\lambda >0$ and $C>0$ such that for any $u_0\in H^s_0({\mathbb T})$ with $\\|u_0\\|_s <\rho$, there is a (unique) solution $u:{\mathbb R}^+ \to H^s_0({\mathbb T})$ of with $u\in Z_{s,T}$ for all $T>0$ and such that $$\\|u(t)\\|_s \le C e^{-\lambda t} \\|u_0\\|_s \qquad \forall t\ge 0.$$ The proofs of Theorem \[thmA\] and Theorem \[thmB\] are given in the next sections. ### Linear Theory In this section, we focus on the well-posedness and the smoothing property of the linearized BO equation with localized damping: $$\label{R5} u_t + {\mathcal H} u_{xx} + GDGu =0, \qquad u(0)=u_0.$$ Let $s\in {\mathbb R}$ and let $Au=-({\mathcal H}u_{xx} +GDGu)$ with domain ${\mathcal D}(A)=H^{s+2}_0({\mathbb T})\subset H^s_0({\mathbb T})$. Our first result is the \[lem100\] $A$ generates a continuous semigroup in $H^s_0({\mathbb T})$, denoted by $(S(t))_{t\ge 0}$. [Proof.]{} Let $C=C(s)$ be the constant in Claim 1. Clearly, $A-C$ is a densely defined closed operator in $H^s_0({\mathbb T})$. Furthermore, by Claim 1, $$(Au -Cu,u)_s \le - \\| D^{\frac{1}{2}} (Gu)\\| _s^2 \quad \forall u\in H^{s+2}_0({\mathbb T}),$$ which shows that $A-C$ is dissipative. It is easily verified that $D(A^)=D(A)=H^{s+2}_0({\mathbb T})$. Thus $$(A^u-Cu,u)_s=(u,Au-Cu)_s\le 0\quad \forall u\in H^{s+2}_0({\mathbb T}),$$ so that $A^-C$ is dissipative too. Thus, $A-C$ generates a semigroup of contractions in $H^s_0({\mathbb T})$ by [@pazy Cor. 4.4, p. 15]. Now we turn our attention to the smoothing effect. \[prop10\] Let $s\ge 0$, $v_0\in H^s_0({\mathbb T})$ and $g\in L^2(0,T,H^{s-\frac{1}{2}}_0({\mathbb T}))$. Then the solution $v$ of $$\label{R10} v_t +{\mathcal H} v_{xx} +GDG v=g, \qquad v(0)=v_0$$ satisfies $v\in Z_{s,T}$ with $$\\|v\\|_{Z_{s,T}} \le C(s,T) \left( \|\|v_0\|\|_s + \|\|g\|\|_{L^2(0,T,H^{s-\frac{1}{2}}({\mathbb T}))} \right) , \label{R12}$$ $C(s,T)$ being nondecreasing in $T$. [Proof.]{} Let us assume first that $s=0$. To have enough regularity in the computations, we assume that $v_0\in H^2_0({\mathbb T})$ and that $g\in C([0,T],H^2_0({\mathbb T}))$, so that the solution $v$ of satisfies $v\in C([0,T],H^2_0({\mathbb T}))\cap C^1([0,T],H^0_0({\mathbb T}))$. We now proceed as in the proof of Proposition \[smoothing\]. We set $Lv=v_t + {\mathcal H} v_{xx}$, $f=-GDGv$, so that $L v= f+g$. Pick any $\varphi \in C^\infty ({\mathbb T})$, and let $Av=\varphi (x) v$. Then $$\vert \int_0^T ([L,A]v,v)\, dt \vert \le 2 \vert \int_0^T (f+g,\varphi v) dt \vert +\\|\varphi \\|_{L^\infty} \big( \\|v_0\\|^2 + \\|v(T)\\|^2 \big) \cdot$$ Scaling in by $v$ yields $$\begin{aligned} \frac{1}{2}\\|v(t) \\|^2 + \int_0^t \\|D^{\frac{1}{2}} Gv\\|^2 d\tau &= & \frac{1}{2} \\|v_0\\|^2 +\int_0^t (g,v) d\tau \\ &\le& \frac{1}{2} \\|v_0\\|^2 +\int_0^T \\|g\\|_{-\frac{1}{2}} \\|v\\|_{\frac{1}{2}} dt.\end{aligned}$$ This yields $$\label{R15} \\|v\\|^2_{L^\infty (0,T,H^0({\mathbb T}))} +\int_0^T \\|D^{\frac{1}{2}} (Gv)\\|^2 d\tau \le \frac{3}{2} \\|v_0\\|^2 +3\int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}} dt.$$ Computations similar to those in Proposition \[smoothing\] give that $$\begin{aligned} \vert \int _0^T (f+g,\varphi (x) v)\, d\tau \vert &\le & C\\|\varphi \\|_1 \int_0^T ( \\|D^{\frac{1}{2}}(Gv)\\|^2 + \\|v\\|^2 + \\|g\\|_{-\frac{1}{2}} \\|v\\|_{\frac{1}{2}} ) dt \\ &\le& C(T, \\|\varphi \\|_1) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}} \\|v\\|_{\frac{1}{2}} dt \right) ,\end{aligned}$$ hence $$\vert \int_0^T ([L,A]v,v)dt \vert \le C(T,\\|\varphi \\|_1) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}}dt \right).$$ Combined with -, the last inequality gives $$\vert \int_0^T (\partial_x \varphi Dv,v)\, dt \vert \le C(T,\\|\varphi \\|_1) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}}dt \right). \label{K3}$$ We pick again $b\in C^\infty _0(\omega )$, where $\omega =\{ x\in {\mathbb T};\ a(x)>0\}$ and $x_0\in {\mathbb T}$. Writing again $b^2(x)-b^2(x-x_0)=\partial _x \varphi$, we obtain successively, with and , that $$\begin{aligned} \int_0^T \\|D^{\frac{1}{2}} (bv)\\|^2 dt &\le& C (T) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}}dt \right),\\ \vert \int_0^T (b^2Dv,v) dt \vert &\le& C\int_0^T \big( \\|v\\|^2 + \\|D^{\frac{1}{2}} (bv)\\|^2\big) dt \\ &\le& C(T) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}}dt \right)\end{aligned}$$ and therefore, with , $$\vert \int_0^T (b^2(x-x_0)Dv,v)dt \vert \le C(T) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|_{-\frac{1}{2}}\\|v\\|_{\frac{1}{2}}dt \right) .$$ Using a partition of unity, this yields $$\int_0^T \\| v\\|_{\frac{1}{2}} ^2 dt \le C(T) \left( \\|v_0\\|^2 + \int_0^T \\|g\\|^2 _{-\frac{1}{2}}dt \right) +\frac{1}{2}\int_0^T \\|v\\|^2_{\frac{1}{2}}dt.$$ Combined with , this gives for $s=0$ when $v_0\in H^2_0({\mathbb T})$ and $g\in C([0,T],H^2_0({\mathbb T}))$. This is also true for $v_0\in H^0_0({\mathbb T})$ and $g\in L^1(0,T,H^{-\frac{1}{2}}_0({\mathbb T}))$ by density. Let us now assume that $s\in (0,2]$. Pick again any $v_0\in H^2_0({\mathbb T})$, $g\in C([0,T],H^2_0({\mathbb T}))$, and let $v\in C([0,T],H^2_0({\mathbb T}))\cap C^1([0,T],H^0_0({\mathbb T}))$ denote the solution of . Set $w=D^s v$ and $h=D^s g$. Note that $$D^s(GDGv) = GDG w + Ew$$ with $E=[D^s,G]DGD^{-s} +GD [D^s,G]D^{-s}$. Note that $\\|Ew\\|\le C\\|w\\|$ and that $w$ solves $$w_t + {\mathcal H} w_{xx} +GDG w + Ew=h, \quad w(0)=w_0:=D^s v_0.$$ Since $$\begin{aligned} \vert \int_0^T (\varphi w,Ew)\, dt \vert &\le& C\\|\varphi \\|_1 \\|w\\|^2_{L^2 (0,T,H^0({\mathbb T}) ) } \\ &\le& C(T , \\|\varphi \\|_1) \left( \\|w_0\\|^2 + \int_0^T \\| h \\|_{-\frac{1}{2} } \\| w \\| _{\frac{1}{2}}dt \right),\end{aligned}$$ we obtain in a similar fashion as above that $$\\|w\\|^2_{L^\infty (0,T,H^0({\mathbb T}))} + \int_0^T \\| w\\|_{\frac{1}{2}}^2 dt \le C(T) \left( \\|w_0\\|^2 + \int_0^T \\|h\\|^2_{-\frac{1}{2}} dt \right),$$ i.e. $$\\|v\\|^2_{L^\infty (0,T,H^s({\mathbb T}))} + \\| v \\|^2_{L^2(0,T,H^{s+\frac{1}{2}}({\mathbb T}))} \le C(T) \left( \\|v_0\\|_s^2 + \\| g \\|^2_{L^2(0,T,H^{s-\frac{1}{2}}({\mathbb T}))} \right). \label{R30}$$ Inequality and the fact that $v\in C([0,T],H^s_0({\mathbb T}))$ are also true for $v_0\in H^s_0({\mathbb T})$ and $g\in L^2(0,T, H^{s-\frac{1}{2}}_0({\mathbb T}))$ by density. \[cor10\] Let $s\ge 0$ and $B\in {\mathcal L} (H^s_0({\mathbb T}))$. Then for any $v_0\in H^s_0({\mathbb T})$, the solution $v$ of $$\label{G1} v_t +{\mathcal H} v_{xx} +GDG v= Bv, \qquad v(0)=v_0$$ fulfills $v\in Z_{s,T}$ with $$\|\|u\|\|_{Z_{s,T}} \le C(s,T) \|\|v_0\|\|_s. \label{G2}$$ Since $A$ is the generator of a continuous semigroup on $H^s_0({\mathbb T})$ and $B$ is a bounded operator on $H^s_0 ({\mathbb T})$, $A+B$ is the generator of a continuous semigroup on $H^s_0({\mathbb T})$ (see e.g. [@pazy Thm 1.1 p. 76]). Pick any $v_0\in H^s_0({\mathbb T})$, and let $v$ denote the solution of given by the semigroup generated by $A+B$. Noticing that $g:=Bv\in C([0,T]; H^s_0({\mathbb T}))$, we infer from Proposition \[prop10\] that $v\in Z_{s,T}$ with $$\\|v\\|_{L^\infty (0,T,H^s({\mathbb T}))} +\\|v\\|_{L^2(0,T,H^{s+\frac{1}{2}} ({\mathbb T}))} \le C(s,T) \left( \\|v_0\\|_s +\sqrt{T} \,\\|B\\|_{ {\mathcal L} (H^s_0({\mathbb T}))} \\|v\\|_{L^\infty (0,T;H^s({\mathbb T}))}\right) .$$ Selecting $T_0>0$ such that $c(s,T_0) \sqrt{T_0} \\|B\\|_{ {\mathcal L} (H^s_0({\mathbb T}))} <1/2$ yields $$\\|v\\|_{L^\infty (0,T_0,H^s({\mathbb T}))} +\\|v\\|_{L^2(0,T_0,H^{s+\frac{1}{2}} ({\mathbb T}))} \le 2 C(s,T_0) \\|v_0\\|_s. \label{G3}$$ Successive applications of on the intervals $[0,T_0]$, $[T_0,2T_0]$,... give for any $T>0$. ### Proof of Theorem \[thmA\] Pick any $s\in (\frac{1}{2},2]$ and any $T>0$. Let $u_0\in H^s_0({\mathbb T})$. We write in its integral form $$\label{B32} u(t)=S(t)u_0 -\int_0^t S(t-\tau )(uu_x)(\tau ) \, d\tau.$$ Let $\Gamma (v)(t)=S(t)u_0 -\int_0^t S(t-\tau )(vv_x)(\tau )d\tau$. We have, by Proposition \[prop10\], that $$\\| \Gamma (v) \\|_{Z_{s,T}} \le C \left( \\|u_0\\|_s + \\| ( \frac{v^2}{2} )_x \\|_{L^2(0,T,H^{s-\frac{1}{2}} ({\mathbb T}))} \right)\cdot$$ Clearly, for $u,v\in Z_{s,T}$, $$\begin{aligned} \int_0^T \\|(uv)_x\\|^2_{s-\frac{1}{2}} dt &\le& C\int_0^T \\|uv\\|^2_{s+\frac{1}{2}} dt \\ &\le& C\int_0^T \big( \\|u\\|^2_{L^\infty( {\mathbb T})} \\|v\\|^2_{s+\frac{1}{2}} + \\|u\\|^2_{s+\frac{1}{2}} \\|v\\|^2_{L^\infty ({\mathbb T})} \big) dt \\ &\le& C\int_0^T \big( \\|u\\|^2_s\\|v\\|^2_{s+\frac{1}{2}} + \\|u\\|^2_{s+\frac{1}{2}} \\|v\\|^2_s \big) dt\\ &\le& C\left( \\|u\\|^2_{ L^\infty (0,T,H^s({\mathbb T})) } \\|v\\|^2_{L^2(0,T,H^{s+\frac{1}{2}} ({\mathbb T}))} \right. \\ &&\qquad \left. + \\|v\\|^2_{ L^\infty (0,T,H^s({\mathbb T})) } \\|u\\|^2_{L^2(0,T,H^{s+\frac{1}{2}} ({\mathbb T}))} \right) \\ &\le& C\\|u\\|_{Z_{s,T}}^2 \\|v\\|_{Z_{s,T}}^2,\end{aligned}$$ where we used the Sobolev embedding $H^s_0({\mathbb T})\subset L^\infty ({\mathbb T})$ for $s>1/2$. Thus, there are some constants $C_0>0$ and $C_1>0$ such that $$\begin{aligned} \\| \Gamma (v) \\|_{Z_{s,T}} &\le& C_0 \\|u_0\\|_s + C_1 \\| v \\|^2_{Z_{s,T}} \qquad \forall v\in Z_{s,T} ,\\ \\| \Gamma (v^1) -\Gamma (v^2) \\|_{Z_{s,T}} &\le& C_1\big( \\| v^1 \\|_{Z_{s,T}} + \\| v^2 \\|_{Z_{s,T}} \big) \\| v^1 -v^2 \\|_{Z_{s,T} } \qquad \forall v^1,v^2\in Z_{s,T}.\end{aligned}$$ Let $B=\{ v\in Z_{s,T}; \ \\|v\\|_{Z_{s,T}} \le R \}$. We choose $R$ in such a way that $B$ is left invariant by $\Gamma$ and $\Gamma $ contracts in $B$, i.e. $$C_0\\|u_0\\|_s + C_1 R^2 \le R, \ \text{ and } 2C_1R<1.$$ It is sufficient to take $R=(4C_1)^{-1}$ and $u_0\in H^s_0({\mathbb T})$ with $\\|u_0\\|_s\le \rho := R/(2C_0)$. ### Proof of Theorem \[thmB\] We proceed as in [@PR]. It has been proved that is semi-globally exponentially stable in $H^0_0({\mathbb T})$. Obviously, the same analysis shows that the [linearized]{} BO equation with localized damping is also exponentially stable in $H^0_0({\mathbb T})$, i.e. $$\\|S(t)u_0\\| \le Ce^{-\lambda t} \\|u_0\\| \label{Z100}$$ for all $u_0\in H^0_0({\mathbb T})$ and some constants $C,\lambda >0$. If $u_0\in H^2_0({\mathbb T})$, then $u(t)=S(t)u_0$ solves $$\label{Z2} u_t+{\mathcal H} u_{xx} +GDGu =0, \quad u(0)=u_0.$$ Letting $v=u_t$, $v$ solves also $$\label{Z3} v_t + {\mathcal H} v_{xx} +GDG v = 0,\quad v(0)=v_0:= -({\mathcal H } u_{0,xx} +GDG u_0).$$ yields $$\\|v(t)\\| =\\|S(t)v_0\\| \le C e^{-\lambda t} \\|v_0\\|,$$ and thus $$\\|S(t) u_0\\|_2 \le C' e^{-\lambda t} \\|u_0\\|_2.$$ By interpolation, this shows that for any $s\in [0,2]$, for any $u_0\in H^s_0({\mathbb T})$ and for some constant $C>0$ (independent of $s$, $u_0$, and $t$), it holds $$\label{Z5} \\|S(t) u_0\\|_s \le C e^{-\lambda t} \\|u_0\\|_s.$$ Let $s>1/2$ and $u_0\in H^s_0({\mathbb T})$. For $$u\in Z_{s,T}([n,n+1]):=C([n,n+1],H^s_0({\mathbb T})) \cap L^2 (n,n+1,H_0^{s+\frac{1}{2}}({\mathbb T})),$$ let $${\|\!\|\!\|}u {\|\!\|\!\|}_n =\\|u\\|_{L^\infty (n,n+1,H^s({\mathbb T})) } + \\|u\\|_{L^2(n,n+1,H^{s+\frac{1}{2}}({\mathbb T})) }\cdot$$ Finally, let $$\\|u\\|_E =\sup_{n\ge 0} \big( e^{n\lambda} {\|\!\|\!\|}u{\|\!\|\!\|}_n \big) \le +\infty .$$ Introduce the space $$E=\{ u\in C({\mathbb R}^+ , H^s_0({\mathbb T}))\cap L^2_{loc} ({\mathbb R}^+, H_0^{s+\frac{1}{2}}({\mathbb T})); \ \\|u\\|_E <\infty \}.$$ Endowed with the norm $\\|\cdot \\|_E $, $E$ is a Banach space. We search for a solution of in a closed ball $B=\{u\in E;\ \\|u\\|_E \le R \}$ as a fixed point of the map $\Gamma (v)(t)=S(t)u_0 - \int_0^t S(t-\tau ) (vv_x)(\tau ) d\tau$. By , we have $$\label{Z7} \\|S(n) u_0\\|_s \le Ce^{-n\lambda } \\|u_0\\|_s \quad \forall n\ge 0.$$ Combined with Proposition \[prop10\], this gives for some constant $C_0>0$ $$\label{Z8} {\|\!\|\!\|}S(t)u_0 {\|\!\|\!\|}_n \le C_0 \, e^{-n\lambda} \\|u_0\\|_s,$$ hence $$\label{Z10} \\| S(t)u_0\\|_E \le C_0 \\|u_0\\|_s.$$ On the other hand, for any $u,v\in E$, $${\|\!\|\!\|}\int_0^t S(t-\tau ) [(uv)_x(\tau)] d\tau {\|\!\|\!\|}_n \le I_1+I_2$$ with $$\begin{aligned} I_1 &=& {\|\!\|\!\|}S(t-n) \int_0^n S(n-\tau ) [(uv)_x (\tau )]d\tau {\|\!\|\!\|}_n,\\ I_2 &=& {\|\!\|\!\|}\int_n^t S(t-\tau ) [(uv)_x (\tau )]d\tau {\|\!\|\!\|}_n\end{aligned}$$ By and , $$\begin{aligned} I_1 &\le & C\\| \int_0^n S(n- \tau ) [(uv)_x (\tau )] d\tau \\|_s \\ &\le& C\sum_{k=1}^n \\|S(n-k) \int_{k-1}^k S(k-\tau ) [(uv)_x(\tau )] \\|_s \\ &\le& C\sum_{k=1}^n e^{-(n-k)\lambda} \\| \int_{k-1}^k S(k-\tau ) [(uv)_x(\tau ) ]d\tau \\|_s \\ &\le& C \sum_{k=1}^n e^{-(n-k)\lambda } \\|(uv)_x\\|_{L^2(k-1,k,H^{s-\frac{1}{2}} ({\mathbb T}))} \\ &\le& C\sum_{k=1}^n e^{-(n-k)\lambda } {\|\!\|\!\|}u {\|\!\|\!\|}_{k-1} {\|\!\|\!\|}v {\|\!\|\!\|}_{k-1} \\ &\le& Ce^{-n\lambda } \\| u\\|_E \\|v\\|_E .\end{aligned}$$ On the other hand $$I_2\le C\\|(uv)_x\\|_{L^2(n,n+1,H^{s-\frac{1}{2}} ({\mathbb T}))} \le C e^{-2n\lambda } \\|u\\|_E \\|v\\|_E.$$ We have proved that for some constant $C_1>0$ $${\|\!\|\!\|}\int_0^t S(t-\tau ) [(uv)_x(\tau )] d\tau {\|\!\|\!\|}_n \le 2 C_1 e^{-n\lambda } \\|u\\|_E \\|v\\|_E,$$ hence $$\\|\int_0^t S(t-\tau ) [(uv)_x (\tau )] d\tau \\|_E \le 2 C_1 \\|u\\|_E \\|v\\|_E.$$ Thus $$\begin{aligned} \\|\Gamma (v)\\|_E &\le & C_0\\|u_0\\|_s + C_1 \\|v\\|_E^2, \\ \\|\Gamma (v^1)-\Gamma (v^2)\\|_E &\le& C_1( \\|v^1\\|_E + \\|v^2\\|_E) \\|v^1-v^2\\|_E.\end{aligned}$$ It follows that $\Gamma$ contracts in the ball $B=\{ u\in E;\ \\|u\\|_E \le R \}$ if $$\label{Z20} 2C_1R<1, \ \text{ and } \ C_0\\|u_0\\|_s + C_1 R^2 \le R.$$ Let $R=\gamma\rho$ ($\gamma$ and $\rho$ being determined later), and assume that $\\|u_0\\|_s\le \rho$. The conditions become $$\label{ZZZ} 2C_1\gamma \rho < 1 ,\text{ and } C_0+C_1\gamma ^2 \rho \le \gamma.$$ Pick $\gamma =2C_0$ and $\rho >0$ sufficiently small so that holds. Then $\Gamma$ contracts in $B$. Replacing $\rho$ by $\\|u_0\\|_s$, we see that the fixed point $u=\Gamma (u)$ satisfies $$\\|u\\|_{L^\infty (n, n+1,H^s({\mathbb T}))} \le e^{-n\lambda } \\| u \\|_E \le e^{-n\lambda} \gamma \\|u_0\\|_s.$$ It follows that $$\\|u(t)\\|_s \le Ce^{-\lambda t} \\|u_0\\|_s \quad \forall t\ge 0$$ for some constant $C>0$, provided that $\\|u_0\\|_s\le \rho$. Control of the Benjamin-Ono equation ==================================== \ Let again $a$ and $G$ be as in and , respectively. We now focus on the control properties of the full BO equation. More precisely, we aim to prove the exact controllability of the system $$u_t + {\mathcal H} u_{xx} + uu_x =Gh, \qquad u(0)=u_0, \label{Y1}$$ where $h$ is the control input. If the exact controllability of the linearized system is well known (cf. Theorem A), the exact controllability of is challenging, as the contraction mapping theorem cannot be applied directly to BO. To overcome that difficulty, we incorporate the feedback $f=-DGu$ into the control input $h$ to obtain a strong enough smoothing effect to apply the contraction principle. Setting $$\label{Y3} h(t) = -DG u(t) +D^{\frac{1}{2}} k(t),$$ we are thus led to investigate the controllability of the system $$u_t + {\mathcal H} u_{xx} + GDGu + uu_x =GD^{\frac{1}{2}} k, \qquad u(0)=u_0. \label{Y4}$$ We shall derive the following local exact controllability result. \[controllability\] Let $s\in (\frac{1}{2}, 2]$ and $T>0$. Then there exists $\delta >0$ such that for any $u_0,u_1\in H^s_0({\mathbb T})$ with $$\label{Y5} \\|u_0\\|_s \le \delta , \qquad \\|u_1\\|_s \le \delta ,$$ one may find a control $k\in L^2(0,T,H^s_0({\mathbb T}))$ such that the system admits a (unique) solution $u$ in the class $Z_{s,T}$ for which $u(T)=u_1$. The proof of Theorem \[controllability\] is done in three steps. In the first step, we prove the exact controllability of the linearized system $$u_t + {\mathcal H} u_{xx} + GDGu =GD^{\frac{1}{2}} k, \qquad u(0)=u_0, \label{Y6}$$ in $L^2_0({\mathbb T})$. In the second step, we prove the exact controllability of in $H^s_0({\mathbb T})$ for all $s>0$ by following the same approach as in [@RZ2009]. Finally, in the third part we derive the exact controllability of the full BO equation by using the contraction mapping theorem as e.g. in [@rosier97; @RZ2007b; @RZ2009]. Note that Theorem \[main3\] follows at once from Theorem \[controllability\] by letting $$h = -DG u + D^{\frac{1}{2}} k \in L^2(0,T,H_0^{s-\frac{1}{2}} ({\mathbb T})).$$ [Proof of Theorem \[controllability\].]{}\ [Step1. Exact controllability of in $H^0_0({\mathbb T})$.]{}\ First, the solution of belongs to $Z_{s,T}$ for $u_0\in H^s_0({\mathbb T})$ and $k\in L^2(0,T,H^s_0({\mathbb T}))$, according to Proposition \[prop10\]. The adjoint system reads $$-v_t-{\mathcal H} v_{xx} +GDG v = 0, \qquad v(T)=v_T. \label{Y9}$$ Scaling in by $v$ yields $$\int_{{\mathbb T}} uvdx\big\vert_0^T = \int_0^T\!\!\!\int_{{\mathbb T}} k D^{\frac{1}{2}} (Gv)dxdt. \label{Y11}$$ The computations are fully justified when $u_0,v_T\in H^2_0({\mathbb T})$ and $k\in L^2(0,T,H^{\frac{5}{2}}_0({\mathbb T}))$, and next extended to the case when $u_0,v_T\in H^0_0({\mathbb T})$ and $k\in L^2(0,T,H^0_0({\mathbb T}))$ by density. Following the classical duality approach, we are led to prove the following observability inequality $$\\|v_T\\|^2 \le C \int_0^T\!\!\! \int_{{\mathbb T}} \| D^{\frac{1}{2}} (Gv) \|^2 dxdt. \label{Y12}$$ Once is proved, the exact controllability of follows by noticing that the operator $\Gamma \in {\mathcal L} (H^0_0({\mathbb T}))$ defined by $\Gamma (v_T) = u(T)$, where $u$ denotes the solution of associated with $u_0=0$ and $k=D^{\frac{1}{2}} (Gv)$ and $v$ denotes the solution of , is onto by and Lax-Milgram theorem. Let us prove by contradiction. If is not true, then one can pick a sequence $(v^n_T)$ in $H^0_0({\mathbb T})$ such that $$\label{Y13} 1=\\|v^n_T\\| ^2 > n \int_0^T\!\!\!\int_{{\mathbb T}} \|D^{\frac{1}{2}} (Gv^n)\|^2 dxdt ,$$ where $v^n$ denotes the solution of issued from $v_T=v_T^n$. Multiplying each term in by $tv^n$ and integrating by parts results in $$\label{Y14} \frac{T}{2} \\|v_T^n\\|^2 = \frac{1}{2} \int_0^T\!\!\! \int_{{\mathbb T}} \|v^n\|^2 dxdt + \int_0^T \!\!\! \int_{{\mathbb T}} t \, \|D^{ \frac{1}{2}} (Gv^n) \|^2 dxdt.$$ Computations similar to those in the proof of Proposition \[prop10\] (changing $t$ into $\tau :=T-t$) give $$\label{Y15} \\|v^n\\|_{L^2(0,T,H^{\frac{1}{2}}({\mathbb T}))} \le C \\|v_T^n\\| \cdot$$ Thus, by and , $(v^n)$ is bounded in $L^2(0,T,H^{\frac{1}{2}}_0({\mathbb T})) \cap H^1(0,T,H^{-\frac{3}{2}} ({\mathbb T}) ) $. By Aubin-Lions’ lemma, a subsequence of $(v^n)$, still denoted by $(v^n)$, has a strong limit (say $v$) in $L^2(0,T,H^0_0({\mathbb T}))$. It follows from and that $(v_T^n)$ is a Cauchy sequence in $H^0_0({\mathbb T})$, hence it has a strong limit (say $v_T$) in $H^0_0({\mathbb T})$, with $\\|v_T\\|=1$. By standard semigroup theory, $v^n$ converges in $C([0,T],H^0_0({\mathbb T}))$ to the solution of associated with $v_T$, which therefore agrees with $v$. By , $D^{\frac{1}{2}} (Gv) \equiv 0$, hence $Gv\equiv 0$. We conclude that $v$ satisfies $$\begin{aligned} v_t + {\mathcal H} v_{xx} &=& 0,\\ Gv &=& 0.\end{aligned}$$ It follows from Proposition \[unique\_continuation\] that $v\equiv 0$. In particular $v_T=v(T)=0$, a property which contradicts the fact that $\\|v_T\\|=1$. The proof of is achieved.\ [Step 2. Exact controllability of in $H^s_0({\mathbb T})$.]{}\ Picking any number $s>0$, we aim to prove the exact controllability of in $H^s_0({\mathbb T})$. Notice first that the system is (backward) well-posed in $H^{-s}_0({\mathbb T})$, since the conclusion of Lemma \[lem100\] is still valid when ${\mathcal H}u_{xx}$ is replaced by $-{\mathcal H}u_{xx}$ in . Thus, the following estimate holds $$\\| v\\| _{L^\infty (0,T,H^{-s}({\mathbb T}))} \le C \\| v_T\\|_{-s}.$$ On the other hand, setting $w=(1-\partial _x^2)^{-\frac{s}{2}} v$, we see that $w$ solves $$\begin{aligned} &&-w_t-{\mathcal H} w_{xx} + GDG w = (1-\partial _x ^2 )^{-\frac{s}{2}} [(1-\partial _x ^2) ^{\frac{s}{2}} , GDG ]w =: B w\\ && w(T)= (1-\partial _x ^2) ^{-\frac{s}{2}} v_T =: w_T.\end{aligned}$$ Note that $B\in {\mathcal L} (H^\sigma_0({\mathbb T}))$ for all $\sigma\in{\mathbb R}$ (see e.g. [@laurent]). Using computations similar to those to prove Corollary \[cor10\], we see that $$\label{AAA} \\|w \\|_{L^2(0,T,H^{\frac{1}{2}}({\mathbb T}))} \le C \\|w_T\\|,$$ and hence $$\label{BBB} \\|v \\|_{L^2(0,T,H^{-s+\frac{1}{2}}({\mathbb T}))} \le C \\|v_T\\|_{-s}.$$ Assuming again that $u_0=0$, we first note that may be written $$\langle v_T,u(T) \rangle _{-s,s} = \int_0^T \langle D^{\frac{1}{2}} (Gv) , k \rangle _{-s,s} dt,$$ where $\langle \cdot ,\cdot \rangle _{-s,s} $ denotes the duality pairing $\langle \cdot ,\cdot \rangle _{H^{-s}_0 ({\mathbb T}), H^s_0({\mathbb T}) }$. We aim to prove the observability inequality $$\label{Y20} \\|v_T\\|^2_{-s} \le C \int_0^T \\|Gv \\|^2 _{-s+ \frac{1}{2}} dt.$$ Once is proved, the exact controllability of in $H^s_0({\mathbb T})$ follows easily. Indeed, if $\Gamma _{-s} \in {\mathcal L} (H^{-s}_0({\mathbb T}))$ is defined by $\Gamma _{-s} (v_T) = (1-\partial _x^2 ) ^s u (T) $ where $u$ solves with $k=(1-\partial _x^2)^{-s} D^{\frac{1}{2}} (Gv)$ and $v$ still denotes the solution of , then $$(v_T, \Gamma _{-s} (v_T) )_{-s} = \int_0^T \\|D^{\frac{1}{2}} (Gv) \\|^2 _{-s} dt \ge C\int_0^T \\|Gv\\|^2_{-s+\frac{1}{2}} dt \ge C \\|v_T\\|^2 _{-s},$$ so that $\Gamma _{-s}: H^{-s}_0({\mathbb T})\to H^{-s}_0({\mathbb T}) $ is onto. The same is true for the map $v_T\in H^{-s}_0({\mathbb T}) \to u(T) \in H^s_0({\mathbb T})$. To prove , we establish first a weaker estimate $$\label{Y21} \\|v_T\\|^2_{-s} \le C \left( \int_0^T \\|Gv\\|^2_{-s+\frac{1}{2}} dt + \\|v_T\\|^2_{-s-\frac{1}{2} } \right) .$$ We argue again by contradiction. If is false, then there is a sequence $(v_T^n)$ in $H^{-s}_0({\mathbb T})$ such that $$\label{Y22} 1=\\|v_T^n\\| ^2_{-s} > n \left( \int_0^T \\|Gv^n\\|^2_{-s+\frac{1}{2}} dt + \\|v_T^n\\| ^2 _{-s -\frac{1}{2}} \right) .$$ It follows that $$\begin{aligned} v_T^n &&\to 0 \quad \text{ in } H_0^{-s-\frac{1}{2}} ({\mathbb T}), \label{Y23} \\ v^n &&\to 0 \quad \text{ in } C([0,T],H^{-s-\frac{1}{2}}_0({\mathbb T})). \label{Y24}\end{aligned}$$ Let $w^n=(1-\partial _x^2)^{-\frac{s}{2}} v^n$. Then $w^n$ solves $$\begin{aligned} &&-w_t^n-{\mathcal H} w^n_{xx} + GDG w^n = (1-\partial _x ^2 )^{-\frac{s}{2}} [(1-\partial _x ^2) ^{\frac{s}{2}} , GDG ]w^n = B w^n,\\ && w^n(T)= (1-\partial _x ^2) ^{-\frac{s}{2}} v^n_T =: w_T^n.\end{aligned}$$ Then $\\|w_T^n\\|=1$, $w_T^n \to 0$ in $H_0^{-\frac{1}{2}} ({\mathbb T})$, and $$\begin{aligned} && w^n\to 0\quad \text{ in } C([0,T],H_0^{-\frac{1}{2}} ({\mathbb T})) \label{Y25a}\\ && \int_0^T \\|G w^n\\|^2_{\frac{1}{2}} dt \to 0. \label{Y25b}\end{aligned}$$ For , we notice that $$\begin{aligned} \int_0^T \\|G w^n\\|^2_{\frac{1}{2}} dt &=& \int_0^T \\|G (1-\partial _x ^2)^{-\frac{s}{2}} v^n \\|^2_{\frac{1}{2}} dt \nonumber \\ &\le& \int_0^T \\|G v^n\\|^2_{-s+\frac{1}{2}} dt + \int_0^T \\|[G,(1-\partial _x^2)^{-\frac{s}{2}} ] v^n \\|^2_{\frac{1}{2}} dt . \label{Y25c}\end{aligned}$$ The first term in the right hand side of tends to $0$ by . For the second one, we have that $$\int_0^T \\|[G, (1-\partial _x ^2 ) ^{-\frac{s}{2}} ] v^n \\|^2 _{\frac{1}{2}} dt \le C \int_0^T \\|v^n\\|^2_{-s-\frac{1}{2}} dt \le C\\|v^n\\|^2_{L^\infty (0,T,H^{-s-\frac{1}{2}} ({\mathbb T}) ) } \to 0,$$ by . From , we infer that $$\label{Y26} \\|w^n\\|_{L^2(0,T,H^{\frac{1}{2}}({\mathbb T}))} \le C \\|w^n_T\\|.$$ Arguing as in Step 1 and using , we can derive the following observability inequality $$\\|w_T^n\\|^2 \le C\left( \int_0^T\!\!\! \int_{{\mathbb T}} \|D^{\frac{1}{2}}(Gw^n)\|^2 dxdt + \\|Bw^n\\|^2_{L^2(0,T,H^{-\frac{1}{2}}({\mathbb T}))} \right) \cdot$$ Combined with and , this yields $w_T^n\to 0$ in $H^0_0({\mathbb T})$, contradicting the fact that $\\|w_T^n\\|=1$ for all $n$. The proof of is complete. Finally, we prove by contradiction. If is false, then there is a sequence $(v_T^n)$ in $H^{-s}_0({\mathbb T})$ such that $$1=\\|v_T^n\\|^2_{-s} >n \int_0^T \\|Gv^n\\|^2 _{-s+\frac{1}{2}} dt. \label{Y27}$$ Extracting a subsequence still denoted by $(v_T^n)$, we can assume that $(v_T^n)$ is strongly convergent in $H_0^{-s-\frac{1}{2} } ({\mathbb T})$ by compactness of the embedding $H_0^{-s} ({\mathbb T}) \subset H_0^{-s-\frac{1}{2} } ({\mathbb T})$. Using , we infer from that $(v_T^n)$ is also strongly convergent in $H_0^{-s} ({\mathbb T})$. Its limit $v_T$ satisfies $\\|v_T\\|_{-s} = 1 $, and the solution $v$ of satisfies $Gv=0$ by . Thus for a.e. $t\in (0,T)$ $$v_{xxx}(\cdot ,t ) = {\mathcal H } v_{xxx} (\cdot , t ) = 0 \qquad \text{ on } \omega .$$ We conclude with Lemma \[Fourier\] that $v\equiv 0$, hence $v_T=0$, which contradicts $\\|v_T\\|_{-s}=1$. The proof of is achieved.\ [Step 3. Fixed-point argument in $H^s_0({\mathbb T})$.]{}\ We proceed as in [@rosier97]. Pick any $s\in (\frac{1}{2}, 2]$ and any $T>0$. We still denote by $(S(t))_{t\ge 0}$ the semigroup introduced in Lemma \[lem100\] and by $Z_{s,T}$ the space introduced in . For $v\in Z_{s,T}$, we set $$\omega (v) = \int_0^T S(T-t) (vv_x)(t)\, dt.$$ From Step 2 we know that the linearized system, namely , with initial data $u_0\in H^s_0({\mathbb T}) $ and control function $k\in L^2(0,T , H^s_0({\mathbb T}) )$ is well-posed and exactly controllable in $H^s_0({\mathbb T}) $. By a classical functional analysis argument (see e.g. [@coron-book Lemma 2.48 p. 58]), one can construct a continuous operator $\Lambda : H^s_0({\mathbb T}) \to L^2(0,T, H^s_0({\mathbb T}))$ such that for any $u_1\in H^s_0({\mathbb T}) $ the solution $u$ of associated with $u_0=0$ and $k=\Lambda (u_1)$ satisfies $u(T) = u_1$. Let us denote by $u=W(k)$ the corresponding trajectory. We know from Proposition \[prop10\] that $W$ is continuous from $L^2(0,T,H^s_0({\mathbb T}))$ into $Z_{s,T}$. Let $u_0,u_1\in H^s_0({\mathbb T})$ be given with $$\\|u_0\\|_{H^s_0({\mathbb T}) } < \delta, \qquad \\|u_1\\|_{H^s_0({\mathbb T}) } <\delta ,$$ where $\delta >0$ will be chosen later. Let $v\in Z_{s,T}$. If we choose $k=\Lambda (u_1-S(T)u_0 +\omega (v))$, then $$S(t)u_0 - \int_0^t S(t-\tau ) (vv_x)(\tau ) d\tau + W ( k ) (t) = \left\{ \begin{array}{ll} u_0 \quad &\text{\rm if } t=0;\\ u_1 &\text{\rm if } t=T. \end{array} \right.$$ It suggests to consider the nonlinear map $v\to \Gamma (v)$, where $$\Gamma (v)(t) = S(t)u_0 - \int_0^t S(t-\tau ) (vv_x)(\tau ) d\tau + W (\Lambda ( u_1-S(T)u_0 +\omega (v) ) ) (t).$$ The proof will be complete if we can show that this map has a fixed point in the space $Z_{s,T}$. Using the estimates in the proof of Theorem \[thmA\], we see that $$\\|\omega (v)\\|_s \le C \\| \int_0^t S(t-\tau ) (vv_x)( \tau ) d\tau \\|_{Z_{s,T}} \le C \\|v\\|^2_{Z_{s,T}}$$ and that there are some constants $C_0>0$ and $C_1>0$ such that $$\begin{aligned} \\| \Gamma (v)\\|_{Z_{s,T}} &\le& C_0 (\\|u_0\\|_s + \\|u_1\\|_s) + C_1 \\| v \\|^2_{Z_{s,T}} \qquad \forall v\in Z_{s,T},\\ \\| \Gamma (v^1) -\Gamma (v^2) \\|_{Z_{s,T}} &\le& C_1\big( \\| v^1 \\|_{Z_{s,T}} + \\| v^2 \\|_{Z_{s,T}} \big) \\| v^1 -v^2\\|_{Z_{s,T}} \qquad \forall v^1,v^2\in Z_{s,T}.\end{aligned}$$ Let $B=\{ v\in Z_{s,T}; \ \\|v\\|_{Z_{s,T}} \le R \}$. We choose the radius $R$ in such a way that the ball $B$ is left invariant by $\Gamma$ and $\Gamma $ contracts in $B$, i.e. $$C_0 (\\|u_0\\|_s + \\|u_1\\|_s) + C_1 R^2 \le R, \ \text{ and } 2C_1R<1.$$ It is sufficient to take $R=(4C_1)^{-1}$ and $\delta := R/(4C_0)$. The proof of Theorem \[controllability\] is complete. Acknowledgements {#acknowledgements .unnumbered} ================ The authors wish to thank Institut Henri Poincaré (Paris, France) for providing a very stimulating environment during the “Control of Partial Differential Equations and Applications” program in the Fall 2010. LR was partially supported by the Agence Nationale de la Recherche, Project CISIFS, grant ANR-09-BLAN-0213-02. FL was partially supported by CNPq and FAPERJ/Brazil. [99]{} L. Abdelouhab, J. L. Bona, M. Felland, and J.-C. Saut, Nonlocal models for nonlinear dispersive waves, Physica D. [40]{} (1989) 360–392. M. J. Ablowitz and A. S. Fokas, [The inverse scattering transform for the Benjamin-Ono equation – A pivot to multidimensional problems]{}, Stud. Appl. Math. [68]{} (1983), 1–10. C. Amick and J. Toland, Uniqueness and related analytic properties for the Benjamin-Ono equation – a nonlinear Neumann problem in the plane, Acta Math. [167]{} (1991) 107–126. T. B. Benjamin, [Internal waves of permanent form in fluids of great depth]{}, J. Fluid Mech. [29]{} (1967), 559–592. N. Burq and F. Planchon, [On the well-posedness for the Benjamin-Ono equation]{}, Math. Ann., [340]{} (2008), 497–542. K. M. Case, [Benjamin-Ono related equations and their solutions]{}, Proc. Nat. Acad. Sci. U.S.A. [53]{} (1965), 1092-1099. R. Coifman and M. Wickerhauser, The scattering transform for the Benjamin-Ono equation, Inverse Problems [6]{} (1990) 825–860. J.-M. Coron, Control and Nonlinearity, Mathematical Surveys and Monographs, Vol. 136, American Mathematical Society, Providence, 2007. K. D. Danov and M. S. Ruderman, [Nonlinear waves on shallow water in the presence of a horizontal magnetic field]{}, Fluid Dynam. [18]{} (1983), 751–756. L. Dawson, H. McGahagan, and G. Ponce, [On the decay properties of solutions to a class of Schrödinger equations]{}, Proc. Amer. Math. Soc. [136]{} (2007), 2081-2090. B. Dehman, P. Gérard, and G. Lebeau, [Stabilization and control for the nonlinear Schrödinger equation on a compact surface]{}, Math. Z* [254]{}(2006), 729–749. G. Fonseca and F. Linares, [Benjamin-Ono Equation with Unbounded Data]{}, Journal of Mathematical Analysis and Applications [247]{} (2000), 426–447. G. Fonseca and G. Ponce, The I.V.P for the Benjamin-Ono equation in weighted Sobolev spaces, J. Funct. Anal. [260]{} (2011) 436–459. G. Fonseca, F. Linares, and G. Ponce, The I.V.P for the Benjamin-Ono equation in weighted Sobolev spaces II, J. Funct. Anal. [262]{} (2012) 2031–2049. D. Henry, Geometric Theory of Semilinear Parabolic Equations. Lecture Notes in Mathematics, 840, Springer-Verlag, Berlin Heidelberg New York, 1981. A. D. Ionescu and C. E. Kenig, Global well-posedness of the Benjamin-Ono equation on low-regularity spaces, J. Amer. Math. Soc. [20]{} (2007) 753–798. R. J. Iorio, On the Cauchy problem for the Benjamin-Ono equation, Comm. P. D. E. [11]{} (1986) 1031–1081. Y. Ishimori, [Solitons in a one-dimensional Lennard/Mhy Jones lattice]{}, Prog. Theoret. Phys. [68]{} (1982), no. 2, 402–410. C. E. Kenig and K. D. Koenig, On the local well-posedness of the Benjamin-Ono and modified Benjamin-On equations, Math. Res. Letters [10]{} (2003) 879–895. H. Koch and N. Tzvetkov, On the local well-posedness of the Benjamin-Ono equation on $H^{s}(\mathbb{R})$, Int. Math. Res. Not. [26]{} (2003) 1449–1464. C. Laurent, [Global controllability and stabilization for the nonlinear Schrödinger equation on an interval]{}, ESAIM Control Optim. Calc. Var. [16]{} (2010), no. 2, 356–379. C. Laurent, [Global controllability and stabilization for the nonlinear Schrödinger equation on some compact manifolds of dimension 3]{}, SIAM J. Math. Anal. [42]{} (2010), no. 2, 785–832. C. Laurent, L. Rosier, B.-Y. Zhang, [Control and Stabilization of the Korteweg-de Vries Equation on a Periodic Domain]{}, Commun. Partial Differ. Equations [35]{} (2010), no. 4, 707–744. F. Linares, J. H. Ortega, [On the controllability and stabilization of the linearized Benjamin-Ono equation]{}, ESAIM Control Optim. Calc. Var., [11]{} (2005), 204–218. F. Linares, A. F. Pazoto, [On the exponential decay of the critical generalized Korteweg-de Vries equation with localized damping]{}, Proc. Amer. Math. Soc. [135]{} (2007), no. 5, 1515–1522. Y. Matsuno and D. J. Kaup, [Initial value problem of the linearized Benjamin-Ono equation and its applications]{}, J. Math. Phys. [38]{} (1997), no. 10, 5198–5224. S. Micu, J. H. Ortega, L. Rosier, and B.-Y. Zhang, [Control and stabilization of a family of Boussinesq systems]{}, Discrete and Continuous Dynamical Systems, [24]{} (2009), no. 2, 273–313. L. Molinet, [Global well-posedness in $L^2$ for the periodic Benjamin-Ono equation]{}, Amer. J. Math. [130]{} (2008), no. 3, 635Ð683. L. Molinet and D. Pilod, [The Cauchy problem for the Benjamin-Ono equation in $L^2$ revisited]{}, preprint (09/07/2010), arXiv:1007.1545. L. Molinet and F. Ribaud, [Well-posedness in $H^1$ for generalized Benjamin-Ono equations on the circle]{}, Discrete Contin. Dyn. Syst. 23 (2009), no. 4, 1295–1311. L. Molinet, J.-C. Saut, and N. Tzvetkov, Ill-posedness issues for the Benjamin-Ono and related equations, SIAM J. Math. Anal. [33]{} (2001) 982–988. G. Ponce, On the global well-posedness of the Benjamin-Ono equation, Diff. Int. Eqs. [4]{} (1991) 527–542. H. Ono, [Algebraic solitary waves in stratified fluids]{}, J. Phys. Soc. Japan [39]{} (1975), 1082–1091. A. F. Pazoto, L. Rosier, [Stabilization of a Boussinesq system of KdV-KdV type]{}, System $\&$ Control Letters [57]{} (2008), 595–601. A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations, Applied Mathematical Sciences, vol. 44, Springer-Verlag, New York Inc. L. Rosier, [Exact boundary controllability for the Korteweg-de Vries equation]{}, ESAIM Control Optim. Cal. Var. 2 (1997), pp. 33-55. L. Rosier and B.-Y. Zhang, [Global stabilization of the generalized Korteweg-de Vries equation posed on a finite domain]{}, SIAM J. Control Optim., [45]{} (2006), 927–956. L. Rosier and B.-Y. Zhang, [Local exact controllability and stabilizability of the nonlinear Schrödinger equation on a bounded interval]{}, SIAM J. Control Optim. [48]{} (2009), no. 2, 972–992. L. Rosier and B.-Y. Zhang, [Exact boundary controllability of the nonlinear Schrödinger equation]{}, J. Differential Equations [246]{} (2009), 4129–4153. L. Rosier and B.-Y. Zhang, [Control and stabilization of the Korteweg-de Vries equation: recent progresses]{}, Jrl Syst $\&$ Complexity (2009) 22: 647–682. L. Rosier and B.-Y. Zhang, [Control and stabilization of the nonlinear Schrödinger equation on rectangles]{}, M3AS: Math. Models Methods Appl. Sci. [20]{} (12) (2010), 2293–2347. L. Rosier and B.-Y. Zhang, [Unique continuation property and control for the Benjamin-Bona-Mahony equation on a periodic domain]{}, J. Differential Equations (2012), doi=10.1016/j.jde.2012.08.014. W. Rudin, Real and Complex Analysis, McGraw-Hill, Inc, 1966. D. L. Russell and B.-Y. Zhang, [Exact controllability and stabilizability of the Korteweg-de Vries equation]{}, Trans. Amer. Math. Soc., [348 ]{}(1996), 3643–3672. J.-C. Saut, Sur quelques généralisations de l’ équations de Korteweg-de Vries, J. Math. Pures Appl. [58]{} (1979) 21–61. T. Tao, [Global well-posedness of the Benjamin-Ono equation in $H^1({\mathbb R})$]{}, J. Hyperbolic Differ. Equ. 1 (2004), 27Ð49.	{ "pile_set_name": "ArXiv" }
--- abstract: \| For fixed $m > 1$, we study the product of $m$ independent $N \times N$ elliptic random matrices as $N$ tends to infinity. Our main result shows that the empirical spectral distribution of the product converges, with probability $1$, to the $m$-th power of the circular law, regardless of the joint distribution of the mirror entries in each matrix. This leads to a new kind of universality phenomenon: the limit law for the product of independent random matrices is independent of the limit laws for the individual matrices themselves. Our result also generalizes earlier results of Götze–Tikhomirov [@GTprod] and O’Rourke–Soshnikov [@OS] concerning the product of independent iid random matrices. address: - 'Department of Mathematics, University of Colorado at Boulder, Boulder, CO 80309 ' - 'Department of Mathematics, UCLA ' - 'Department of Mathematics, University of California, Davis, One Shields Avenue, Davis, CA 95616-8633 ' - 'Department of Mathematics, Yale University, New Haven, CT 06520, USA' author: - 'Sean O’Rourke' - David Renfrew - Alexander Soshnikov - Van Vu title: Products of independent elliptic random matrices --- [^1] [^2] [^3] [^4] Introduction ============ We begin by recalling that the eigenvalues of a $N \times N$ matrix ${\mathbf}{M}$ are the roots in $\mathbb{C}$ of the characteristic polynomial $\det( {\mathbf}{M} - z {\mathbf}{I})$, where ${\mathbf}{I}$ is the identity matrix. We let $\lambda_1({\mathbf}{M}), \ldots, \lambda_N({\mathbf}{M})$ denote the eigenvalues of ${\mathbf}{M}$. In this case, the empirical spectral measure $\mu_{{\mathbf}{M}}$ is given by $$\mu_{\mathbf}{M} := \frac{1}{N} \sum_{i=1}^N \delta_{\lambda_i({\mathbf}{M})}.$$ The corresponding empirical spectral distribution (ESD) is given by $$F^{{\mathbf}{M}}(x,y) := \frac{1}{N} \# \left\{1 \leq i \leq N: {\operatorname{Re}}(\lambda_i({\mathbf}{M})) \leq x, {\operatorname{Im}}(\lambda_i({\mathbf}{M})) \leq y \right\}.$$ Here $\# E$ denotes the cardinality of the set $E$. If the matrix ${\mathbf}{M}$ is Hermitian, then the eigenvalues $\lambda_1({\mathbf}{M}), \ldots, \lambda_N({\mathbf}{M})$ are real. In this case the ESD is given by $$F^{{\mathbf}{M}}(x) := \frac{1}{N} \# \left\{ 1 \leq i \leq N : \lambda_i({\mathbf}{M}) \leq x \right\}.$$ One of the simplest random matrix ensembles is the class of random matrices with independent and identically distributed (iid) entries. Let $\xi$ be a complex random variable. We say ${\mathbf}{Y}_N$ is an $N \times N$ iid random matrix with atom variable $\xi$ if the entries of ${\mathbf}{Y}_N$ are iid copies of $\xi$. When $\xi$ is a standard complex Gaussian random variable, ${\mathbf}{Y}_N$ can be viewed as a random matrix drawn from the probability distribution $${\mathbb{P}}(d {\mathbf}{M}) = \frac{1}{\pi^{N^2}} e^{- \operatorname{tr}({\mathbf}{M} {\mathbf}{M}^\ast)} d {\mathbf}{M}$$ on the set of complex $N \times N$ matrices. Here $d {\mathbf}{M}$ denotes the Lebesgue measure on the $2N^2$ real entries $$\{ {\operatorname{Re}}(m_{ij}) : 1 \leq i, j \leq N\} \cup \{ {\operatorname{Im}}(m_{ij}) : 1 \leq i, j \leq N\}$$ of ${\mathbf}{M}=(m_{ij})_{i,j=1}^N$. The measure ${\mathbb{P}}(d {\mathbf}{M})$ is known as the [complex Ginibre ensemble]{}. The [real Ginibre ensemble]{} is defined analogously. Following Ginibre [@Gi], one may compute the joint density of the eigenvalues of a random matrix ${\mathbf}{Y}_N$ drawn from the complex Ginibre ensemble. Indeed, $(\lambda_1({\mathbf}{Y}_N), \ldots, \lambda_N({\mathbf}{Y}_N))$ has density $$\label{eq:ginibre} p_N(z_1, \ldots, z_N) := \frac{1}{\pi^N \prod_{i=1}^N k! } \exp \left( - \sum_{k=1}^N \|z_k\|^2 \right) \prod_{1 \leq i < j \leq N} \|z_i - z_j \|^2.$$ Mehta [@M; @M:B] used the joint density function to compute the limiting spectral measure of the complex Ginibre ensemble. In particular, he showed that if ${\mathbf}{Y}_N$ is drawn from the complex Ginibre ensemble, then the ESD of $\frac{1}{\sqrt{N}} {\mathbf}{Y}_N$ converges to the [circular law]{} $F_{\mathrm{circ}}$ as $N \to \infty$, where $$F_{\mathrm{circ}}(x,y) := \mu_{\mathrm{circ}} \left( \left\{ z \in \mathbb{C} : {\operatorname{Re}}(z) \leq x, {\operatorname{Im}}(z) \leq y \right\} \right)$$ and $\mu_{\mathrm{circ}}$ is the uniform probability measure on the unit disk in the complex plane. Edelman [@Ed-cir] verified the same limiting distribution for the real Ginibre ensemble. For the general (non-Gaussian) case, there is no formula for the joint distribution of the eigenvalues and the problem appears much more difficult. The universality phenomenon in random matrix theory asserts that the spectral behavior of an iid random matrix does not depend on the distribution of the atom variable $\xi$ in the limit $N \rightarrow \infty$. In other words, one expects that the circular law describes the limiting ESD of a large class of random matrices (not just Gaussian matrices). An important result was obtained by Girko [@G1; @G2] who related the empirical spectral measure of a non-Hermitian matrix to that of a family of Hermitian matrices. Using this Hermitization technique, Bai [@Bcirc; @BSbook] gave the first rigorous proof of the circular law for general (non-Gaussian) distributions. He proved the result under a number of moment and smoothness assumptions on the atom variable $\xi$, and a series of recent improvements were obtained by Götze and Tikhomirov [@GTcirc], Pan and Zhou [@PZ] and Tao and Vu [@TVcirc; @TVesd]. In particular, Tao and Vu [@TVbull; @TVesd] established the law with the minimum assumption that $\xi$ has finite variance. \[thm:tvcirc\] Let $\xi$ be a complex random variable with mean zero and unit variance. For each $N \geq 1$, let ${\mathbf}{Y}_N$ be a $N \times N$ iid random matrix with atom variable $\xi$. Then the ESD of $\frac{1}{\sqrt{N}} {\mathbf}{Y}_N$ converges almost surely to the circular law $F_{\mathrm{circ}}$ as $N \rightarrow \infty$. More recently, Götze and Tikhomirov [@GTprod] consider the ESD of the product of $m$ independent iid random matrices. They show that, as the sizes of the matrices tend to infinity, the limiting distribution is given by $F_m$, where $F_m$ is supported on the unit circle in the complex plane and has density $f_m$ given by $$\label{plotnost} f_m(z) := \left\{ \begin{array}{ll} \frac{1}{m \pi} \|z\|^{\frac{2}{m}-2}, & \text{ for } \|z\|\leq 1 \\ 0, & \text{ for } \|z\| > 1 \end{array} \right.$$ in the complex plane. It can be verified directly, that if $\psi$ is a random variable distributed uniformly on the unit disk in the complex plane, then $\psi^m$ has distribution $F_m$. \[thm:GTprod\] Let $m \geq 1$ be an interger, and assume $\xi_1, \ldots, \xi_m$ are complex random variables with mean zero and unit variance. For each $N \geq 1$ and $1 \leq k \leq m$, let ${\mathbf}{Y}_{N,k}$ be an $N \times N$ iid random matrix with atom variable $\xi_k$, and assume ${\mathbf}{Y}_{N,1}, \ldots, {\mathbf}{Y}_{N,m}$ are independent. Define the product $${\mathbf}{P}_N := N^{-m/2} {\mathbf}{Y}_{N,1} \cdots {\mathbf}{Y}_{N,m}.$$ Then ${\mathbb{E}}F^{{\mathbf}{P}_N}$ converges to $F_m$ as $N \to \infty$. The convergence of $F^{{\mathbf}{P}_N}$ to $F_m$ in Theorem \[thm:GTprod\] was strengthened to almost sure convergence in [@B; @OS]. The Gaussian case was originally considered by Burda, Janik, and Waclaw [@BJW]; see also [@Bsurv]. We refer the reader to [@AB; @ABK; @AIK; @AIK2; @AKW; @AS; @BJLNS; @F; @F2] and references therein for many other interesting results concerning products of Gaussian random matrices. New results =========== In this paper, we generalize Theorem \[thm:GTprod\] by considering products of independent real elliptic random matrices. Elliptic random matrices were originally introduced by Girko [@Gorig; @Gten] in the 1980s. \[def:elliptic\] Let $(\xi_1, \xi_2)$ be a random vector in $\mathbb{R}^2$, and let $\zeta$ be a real random variable. We say ${\mathbf}{Y}_N = (y_{ij})_{i,j=1}^N$ is a $N \times N$ real elliptic random matrix with atom variables $(\xi_1,\xi_2), \zeta$ if the following conditions hold. - (independence) $\{ y_{ii} : 1 \leq i \leq N\} \cup \{ (y_{ij}, y_{ji}) : 1 \leq i < j \leq N\}$ is a collection of independent random elements. - (off-diagonal entries) $\{ (y_{ij}, y_{ji}) : 1 \leq i < j \leq N\}$ is a collection of iid copies of $(\xi_1,\xi_2)$. - (diagonal entries) $\{y_{ii} : 1 \leq i \leq N\}$ is a collection of iid copies of $\zeta$. Real elliptic random matrices generalize iid random matrices. Indeed, if $\xi_1, \xi_2, \zeta$ are iid, then ${\mathbf}{Y}_N$ is just an iid random matrix. On the other hand, if $\xi_1 = \xi_2$ almost surely, then ${\mathbf}{Y}_N$ is a real symmetric matrix. In this case, the eigenvalues of ${\mathbf}{Y}_N$ are real and ${\mathbf}{Y}_N$ is known as a real symmetric Wigner matrix [@W]. Suppose $\xi_1, \xi_2$ have mean zero and unit variance. Set $\rho := {\mathbb{E}}[\xi_1 \xi_2]$. When $\|\rho\| < 1$ and $\zeta$ has mean zero and finite variance, it was shown in [@NgO] that the ESD of $\frac{1}{\sqrt{N}} {\mathbf}{Y}_N$ converges almost surely to the elliptic law $F_\rho$ as $N \to \infty$, where $$F_\rho(x,y) = \mu_{\rho}( \{z \in \mathbb{C} : {\operatorname{Re}}(z) \leq x, {\operatorname{Im}}(z) \leq y\})$$ and $\mu_\rho$ is the uniform probability measure on the ellipsoid $$\mathcal{E}_{\rho} = \left\{ z \in {\mathbb{C}}: \frac{{\operatorname{Re}}(z)^2}{(1+\rho)^2} +\frac{{\operatorname{Im}}(z)^2}{(1-\rho)^2} < 1 \right\} .$$ This is a natural generalization of the circular law (Theorem \[thm:tvcirc\]). Figure \[fig:elliptic\] displays a numerical simulation of the eigenvalues of a real elliptic random matrix. In this note, we consider the product ${\mathbf}{Y}_{N,1} \cdots {\mathbf}{Y}_{N,m}$ of $m$ independent real elliptic random matrices. In particular, we assume each real elliptic random matrix ${\mathbf}{Y}_{N,k}$ has atom variables $(\xi_{k,1}, \xi_{k,2}), \zeta_k$ which satisfy the following conditions. \[ass:atom\] There exists $\tau > 0$ such that the following conditions hold. 1. $\xi_{k,1}, \xi_{k,2}$ both have mean zero and unit variance. 2. ${\mathbb{E}}\| \xi_{k,1}\|^{2 + \tau} + {\mathbb{E}}\|\xi_{k,2}\|^{2 + \tau} < \infty$. \[item:tau\] 3. $\rho_k := {\mathbb{E}}[\xi_{k,1} \xi_{k,2}]$ satisfies $\|\rho_k\| < 1$. \[item:rhoass\] 4. $\zeta_k$ has mean zero and finite variance. In our main result below, we show that the limiting distribution $F_m$ (with density given by ) from Theorem \[thm:GTprod\] for the product of independent iid random matrices is also the limiting distribution for the product of independent elliptic random matrices. In other words, the limit law for the product of independent random matrices is independent of the limit laws for the individual matrices themselves. This type of universality was first considered by Burda, Janik, and Waclaw in [@BJW] for matrices with Gaussian entries; see also [@Bsurv]. Figure \[fig:plots\] displays several numerical simulations which illustrate this phenomenon. \[thm:simple\] Let $m > 1$ be an integer. For each $1 \leq k \leq m$, let $(\xi_{k,1}, \xi_{k,2}), \zeta_k$ be real random elements that satisfy Assumption \[ass:atom\]. For each $N \geq 1$ and $1 \leq k \leq m$, let ${\mathbf}{Y}_{N,k}$ be an $N \times N$ real elliptic random matrix with atom variables $(\xi_{k,1}, \xi_{k,2}), \zeta_k$, and assume ${\mathbf}{Y}_{N,1}, \ldots, {\mathbf}{Y}_{N,m}$ are independent. Then the ESD of the product $${\mathbf}{P}_N := N^{-m/2} {\mathbf}{Y}_{N,1} \cdots {\mathbf}{Y}_{N,m}$$ converges almost surely to $F_m$ (with density given by ) as $N \to \infty$. More generally, we establish a version of Theorem \[thm:simple\] where each elliptic random matrix ${\mathbf}{Y}_{N,k}$ is perturbed by a deterministic, low rank matrix ${\mathbf}{A}_{N,k}$ with small Hilbert-Schmidt norm. In fact, Theorem \[thm:simple\] will follow from Theorem \[thm:main\] below. We recall that, for any $m \times n$ matrix ${\mathbf}{M}$, the Hilbert-Schmidt norm $\\|{\mathbf}{M}\\|_2$ is given by the formula $$\label{eq:def:hs} \\|{\mathbf}{M}\\|_2 := \sqrt{ \operatorname{tr}({\mathbf}{M} {\mathbf}{M}^\ast) } = \sqrt{ \operatorname{tr}({\mathbf}{M}^\ast {\mathbf}{M})}.$$ \[thm:main\] Let $m > 1$ be an integer. For each $1 \leq k \leq m$, let $(\xi_{k,1}, \xi_{k,2}), \zeta_k$ be real random elements that satisfy Assumption \[ass:atom\]. For each $N \geq 1$ and $1 \leq k \leq m$, let ${\mathbf}{Y}_{N,k}$ be an $N \times N$ real elliptic random matrix with atom variables $(\xi_{k,1}, \xi_{k,2}), \zeta_k$, and assume ${\mathbf}{Y}_{N,1}, \ldots, {\mathbf}{Y}_{N,m}$ are independent. For each $1 \leq k \leq m$, let ${\mathbf}{A}_{N,k}$ be a $N \times N$ deterministic matrix, and assume $$\label{eq:Aassump} \max_{1 \leq k \leq m} \operatorname{rank}({\mathbf}{A}_{N,k}) = O(N^{1-{\varepsilon}}) \quad \text{and} \quad \sup_{N \geq 1} \max_{1 \leq k \leq m} \frac{1}{N^2} \\| {\mathbf}{A}_{N,k} \\|_2 < \infty,$$ for some ${\varepsilon}> 0$. Then the ESD of the product $$\label{def:PN} {\mathbf}{P}_N := N^{-m/2} \prod_{k=1}^m ( {\mathbf}{Y}_{N,k} + {\mathbf}{A}_{N,k})$$ converges almost surely to $F_m$ (with density given by ) as $N \to \infty$. We conjecture that items and from Assumption \[ass:atom\] are not required for Theorem \[thm:main\] to hold. Indeed, in view of Theorem \[thm:tvcirc\] and [@NgO], it is natural to conjecture that $\xi_1, \xi_2$ need only have two finite moments. Also, our proof of Theorem \[thm:main\] can almost be completed under the assumption that $-1 < \rho_k \leq 1$. We only require that $\rho_k \neq 1$ in Section \[sec:singularvalue\] in order to control the least singular value of matrices of the form ${\mathbf}{Y}_{N,k} + {\mathbf}{F}_{N,k}$, where ${\mathbf}{F}_{N,k}$ is a deterministic matrix whose entries are bounded by $N^{\alpha}$, for some $\alpha > 0$. See Remark \[rem:lsv\] and Theorem \[thm:prodWigner\] below for further details. Among other things, the perturbation by ${\mathbf}{A}_{N,k}$ in Theorem \[thm:main\] allows one to consider elliptic random matrices with nonzero mean. Indeed, let $\mu_k$ be a real number, and assume each entry of ${\mathbf}{A}_{N,k}$ takes the value $\mu_k$. Then ${\mathbf}{Y}_{N,k} + {\mathbf}{A}_{N,k}$ is an elliptic random matrix whose atom variables have mean $\mu_k$. In [@GNT], a result similar to Theorem \[thm:simple\] is proved under a different set of assumptions. As noted above, when $\rho_k = 1$, the matrix ${\mathbf}{Y}_{N,k}$ is known as a real symmetric Wigner matrix. Theorem \[thm:main\] requires that $\|\rho_k\| < 1$, but in the special case when $m=2$, we are able to extend our proof to show that the same result holds for the product of two independent real symmetric Wigner matrices. \[thm:prodWigner\] Let $\xi_{1,1}, \xi_{2,1}$ be real random variables with mean zero and unit variance, and which satisfy $${\mathbb{E}}\|\xi_{1,1}\|^{2+\tau} + {\mathbb{E}}\|\xi_{2,1}\|^{2+\tau} < \infty$$ for some $\tau > 0$. For each $N \geq 1$ and $k=1,2$, let ${\mathbf}{Y}_{N,k}$ be an $N \times N$ real symmetric matrix whose diagonal entries and upper diagonal entries are iid copies of $\xi_{k,1}$, and assume ${\mathbf}{Y}_{N,1}$ and ${\mathbf}{Y}_{N,2}$ are independent. Then the ESD of the product $${\mathbf}{P}_N := N^{-1} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2}$$ converges almost surely to $F_2$ (with density given by when $m=2$) as $N \to \infty$. Overview and outline -------------------- We begin by outlining the proof of Theorem \[thm:main\]. Instead of directly considering $ {\mathbf}{P}_N := N^{-m/2} \prod_{k=1}^m ( {\mathbf}{Y}_{N,k} + {\mathbf}{A}_{N,k})$, we introduce a linearized random matrix, ${\mathbf}{Z}_N:= \frac{1}{\sqrt{N}} ({\mathbf}{Y}_N + {\mathbf}{A}_N)$, where ${\mathbf}{Y}_N$ and ${\mathbf}{A}_N$ are $mN \times mN$ block matrices of the form $$\label{def:YN} {\mathbf}{Y}_N := \begin{bmatrix} 0 & {\mathbf}{Y}_{N,1} & & & 0 \\ 0 & 0 & {\mathbf}{Y}_{N,2} & & 0 \\ & & \ddots & \ddots & \\ 0 & & & 0 & {\mathbf}{Y}_{N,m-1} \\ {\mathbf}{Y}_{N,m} & & & & 0 \end{bmatrix}$$ and $$\label{def:AN} {\mathbf}{A}_N := \begin{bmatrix} 0 & {\mathbf}{A}_{N,1} & & & 0 \\ 0 & 0 & {\mathbf}{A}_{N,2} & & 0 \\ & & \ddots & \ddots & \\ 0 & & & 0 & {\mathbf}{A}_{N,m-1} \\ {\mathbf}{A}_{N,m} & & & & 0 \end{bmatrix}.$$ The following theorem gives the limiting distribution of ${\mathbf}Z_N$, from which we will deduce our main theorem as a corollary. \[thm:circular\] Under the assumptions of Theorem \[thm:main\], the ESD of ${\mathbf}{Z}_N := \frac{1}{\sqrt{N}} ({\mathbf}{Y}_N + {\mathbf}{A}_N)$ converges almost surely to the circular law $F_{\mathrm{circ}}$ as $N \to \infty$. In Section \[sec:linear\], we show that Theorem \[thm:main\] is a short corollary of Theorem \[thm:circular\]. This same linearization trick was used in [@OS] to study products of non-Hermitian matrices with iid entries. Similar techniques were also used in [@A; @HT] to study general self-adjoint polynomials of self-adjoint random matrices. Sections \[sec:resolvent\], \[sec:singularvalue\], and \[sec:complete\] are dedicated to proving Theorem \[thm:circular\]. Following the ideas of Girko [@G1; @G2], we compute the limiting spectral measure of a non-Hermitian random matrix ${\mathbf}{M}$, by employing the method of Hermitizaition. Given an $N \times N$ matrix ${\mathbf}{M}$, we recall that the empirical spectral measure of ${\mathbf}{M}$ is given by $$\mu_{\mathbf}{M} := \frac{1}{N} \sum_{i=1}^N \delta_{\lambda_i({\mathbf}{M})},$$ where $\lambda_1({\mathbf}{M}), \ldots, \lambda_N({\mathbf}{M})$ are the eigenvalues of ${\mathbf}{M}$. We let $\nu_{{\mathbf}{M}}$ denote the symmetric empirical measure built from the singular values of ${\mathbf}{M}$. That is, $$\nu_{{\mathbf}{M}} := \frac{1}{2N} \sum_{i =1}^N \left( \delta_{\sigma_i({\mathbf}{M})} + \delta_{-\sigma_i({\mathbf}{M})} \right),$$ where $\sigma_1({\mathbf}{M}) \geq \cdots \geq \sigma_N({\mathbf}{M}) \geq 0$ are the singular values of ${\mathbf}{M}$. In particular, $$\sigma_1({\mathbf}{M}) := \sup_{\\|x\\| = 1} \\| {\mathbf}{M} x \\|$$ is the largest singular value of ${\mathbf}{M}$ and $$\sigma_N({\mathbf}{M}) := \inf_{\\|x\\| = 1} \\| {\mathbf}{M} x \\|$$ is the smallest singular value, both of which will play a key role in our analysis below. The key observation of Girko [@G1; @G2] relates the empirical spectral measure of a non-Hermitian matrix to that of a Hermitian matrix. To illustrate the connection, consider the Cauchy–Stieltjes transform $s_N$ of the measure $\mu_{{\mathbf}{M}}$, where ${\mathbf}{M}$ is an $N \times N$ matrix, given by $$s_N(z) := \frac{1}{N} \sum_{i=1}^N \frac{1}{ \lambda_i({\mathbf}{M}) - z} = \int_{\mathbb{C}} \frac{1}{x -z } \mu_{{\mathbf}{M}} (dx),$$ for $z \in \mathbb{C}$. Since $s_N$ is analytic everywhere except at the poles (which are exactly the eigenvalues of ${\mathbf}{M}$), the real part of $s_N$ determines the eigenvalues. Let $\sqrt{-1}$ denote the imaginary unit, and set $z = s + \sqrt{-1} t$. Then we can write the real part of $s_N(z)$ as $$\begin{aligned} {\operatorname{Re}}( s_N(z)) &= \frac{1}{N} \sum_{i=1}^N \frac{ {\operatorname{Re}}(\lambda_i({\mathbf}{M})) - s }{ \left\| \lambda_i({\mathbf}{M}) - z \right\|^2} \\ &= \frac{-1}{2N} \sum_{i=1}^N \frac{\partial}{\partial s} \log \left\| \lambda_i({\mathbf}{M}) - z \right\|^2 \\ &= \frac{-1}{2N} \frac{\partial}{\partial s} \log \det \left( {\mathbf}{M} - z {\mathbf}{I} \right) \left( {\mathbf}{M} - z{\mathbf}{I} \right)^{\ast} \\ &= -\frac{\partial}{\partial s} \int_{\mathbb{C}} \log x^2 \nu_{{\mathbf}{M} - z {\mathbf}{I}}(dx), \end{aligned}$$ where ${\mathbf}{I}$ denotes the identity matrix. In other words, the task of studying $\mu_{{\mathbf}{M}}$ reduces to studying the measures $\{\nu_{{\mathbf}{M} - z {\mathbf}{I}}\}_{z \in \mathbb{C}}$. The difficulty now is that the $\log$ function has two poles, one at infinity and one at zero. The largest singular value can easily be bounded by a polynomial in $N$. The main difficulty is controlling the least singular value. In order to study $\nu_{{\mathbf}M}$ it is useful to note that it is also the empirical spectral measure for the Hermitization of ${\mathbf}M$. The Hermitization of ${\mathbf}M$ is defined to be $${\mathbf}H:=\begin{bmatrix} 0 & {\mathbf}M\\ {\mathbf}M^& 0 \end{bmatrix}.$$ For an $N \times N$ matrix, the Stieltjes transform of $\nu_{{\mathbf}{M} - z {\mathbf}{I}}$ is also the trace of the Hermitized resolvent. That is, for $\eta \in \mathbb{C}^{+} := \{ w \in \mathbb{C} : {\operatorname{Im}}(w) > 0\}$, we have $$\int \frac{ 1 }{x-\eta}\nu_{{\mathbf}{M} - z {\mathbf}{I}}(dx) = \frac{1}{2N} \operatorname{tr}({\mathbf}R( {\mathbf}q)),$$ where $${\mathbf}{R}({\mathbf}q) := ( {\mathbf}{H} - {\mathbf}{q} \otimes {\mathbf}{I}_N)^{-1}, \quad {\mathbf}{q} := \begin{bmatrix} \eta & z \\ \bar{z} & \eta \end{bmatrix}.$$ Here ${\mathbf}{q} \otimes {\mathbf}{I}_N$ denotes the Kronecker product of the matrix ${\mathbf}{q}$ and the identity matrix ${\mathbf}{I}_N$. Typically, in order to estimate the measures $\nu_{{\mathbf}{M} -z {\mathbf}{I}}$, one shows that the Stieltjes transform approximately satisfies a fixed point equation. Then one can show that this Stieltjes transform is close to the Stieltjes transform that exactly solves the fixed point equation. Because of the dependencies between entries in the matrix ${\mathbf}Z_N$, directly computing the trace of the resolvent of the Hermitization of ${\mathbf}Z_N$ is troublesome. To circumvent this issue, in Section \[sec:resolvent\], instead of taking the trace of the resolvent, we instead take the partial trace and consider a $2m \times 2m$ matrix-valued Stieltjes transform. Then we show this partial trace approximately satisfies a matrix-valued fixed point equation. In Section \[sec:singularvalue\], we deduce a bound for the least singular value of the matrix $\frac{1}{\sqrt{N}}({\mathbf}{Y}_N + {\mathbf}{A}_N) - z {\mathbf}{I}$ from the known bounds on the least singular values of the individual matrices ${\mathbf}Y_{N,k} +{\mathbf}A_{N,k}$. We finally complete the proof of Theorem \[thm:circular\] in Section \[sec:complete\]. The proof of Theorem \[thm:prodWigner\] is very similar to the proof of Theorem \[thm:main\]. In fact, there are only a few places in the proof of Theorem \[thm:main\] where the condition $\|\rho_k\| < 1$ is required. We prove Theorem \[thm:prodWigner\] in Section \[sec:Wigner\]. A remark from free probability ------------------------------ Notation -------- We use asymptotic notation (such as $O,o,\Omega$) under the assumption that $N \rightarrow \infty$. We use $X \ll Y, Y \gg X, Y=\Omega(X)$, or $X = O(Y)$ to denote the bound $X \leq CY$ for all sufficiently large $N$ and for some constant $C$. Notations such as $X \ll_k Y$ and $X=O_k(Y)$ mean that the hidden constant $C$ depends on another constant $k$. We always allow the implicit constants in our asymptotic notation to depend on the integer $m$ from Theorem \[thm:main\]; we will not denote this dependence with a subscript. $X=o(Y)$ or $Y=\omega(X)$ means that $X/Y \rightarrow 0$ as $N \rightarrow \infty$. $\\|{\mathbf}{M}\\|$ is the spectral norm of the matrix ${\mathbf}{M}$. $\\|{\mathbf}{M}\\|_2$ denotes the Hilbert-Schmidt norm of ${\mathbf}{M}$ (defined in ). We let ${\mathbf}{I}_N$ denote the $N \times N$ identity matrix. Often we will just write ${\mathbf}{I}$ for the identity matrix when the size can be deduced from the context. We write a.s., a.a., and a.e. for almost surely, Lebesgue almost all, and Lebesgue almost everywhere respectively. We use $\sqrt{-1}$ to denote the imaginary unit and reserve $i$ as an index. We let ${\ensuremath{\mathbf{1}_{{E}}}}$ denote the indicator function of the event $E$. We let $C$ and $K$ denote constants that are non-random and may take on different values from one appearance to the next. The notation $K_p$ means that the constant $K$ depends on another parameter $p$. We always allow the constants $C$ and $K$ to depend on the integer $m$ from Theorem \[thm:main\]; we will not denote this dependence with a subscript. In view of Theorem \[thm:main\] and Assumption \[ass:atom\], we define the correlations $\rho_k := {\mathbb{E}}[ \xi_{k,1} \xi_{k,2} ]$ for the atom variables $\xi_{k,1}, \xi_{k,2}$. In addition, we let $\tau > 0$ be such that $$\sum_{k=1}^m \left( {\mathbb{E}}\|\xi_{k,1}\|^{2+\tau} + {\mathbb{E}}\|\xi_{k,2}\|^{2+\tau} \right) < \infty.$$ Proof of Theorem \[thm:main\] {#sec:linear} ============================== A matrix-valued Stieltjes transform {#sec:resolvent} =================================== Least singular value bound {#sec:singularvalue} ========================== Completing the argument {#sec:complete} ======================= Proof of Theorem \[thm:prodWigner\] {#sec:Wigner} =================================== This section is devoted to the proof of Theorem \[thm:prodWigner\]. As noted above, the proof of Theorem \[thm:prodWigner\] is very similar to the proof of Theorem \[thm:main\]. We define the linearized matrix ${\mathbf}{Z}_N := \frac{1}{\sqrt{N}} {\mathbf}{Y}_N$, where $${\mathbf}{Y}_N := \begin{bmatrix} 0 & {\mathbf}{Y}_{N,1} \\ {\mathbf}{Y}_{N,2} & 0 \end{bmatrix}.$$ As before, the proof of Theorem \[thm:prodWigner\] reduces to showing that the ESD of ${\mathbf}{Z}_N$ converges to the circular law as $N$ tends to infinity. \[thm:circularWigner\] Under the assumptions of Theorem \[thm:prodWigner\], the ESD of ${\mathbf}{Z}_N := \frac{1}{\sqrt{N}} {\mathbf}{Y}_N$ converges almost surely to the circular law $F_{\mathrm{circ}}$ as $N \to \infty$. The proof that Theorem \[thm:prodWigner\] follows from Theorem \[thm:circularWigner\] is identical to the proof given in Section \[sec:linear\]. The proof of Theorem \[thm:circularWigner\] follows the same arguments outlined in Section \[sec:complete\] (taking ${\mathbf}{A}_N = 0$). Indeed, the proof in Section \[sec:complete\] requires three key inputs: Lemma \[lemma:concentrate\], Theorem \[approxthm\], and Theorem \[thm:least-sing-value\]. Thus, in order to complete the proof, we will need versions of these results for the matrix ${\mathbf}{Z}_N$ defined above. We first observe that both Lemma \[lemma:concentrate\] and Theorem \[approxthm\] hold for the matrix ${\mathbf}{Z}_N$ defined above. Indeed, the matrices ${\mathbf}{Y}_{N,1}$ and ${\mathbf}{Y}_{N,2}$ (as well as their truncated counterparts) are elliptic random matrices. In fact, the proofs of Lemma \[lemma:concentrate\] and Theorem \[approxthm\] do not require any conditions on the correlations $\rho_k$. Therefore, it only remains to prove the following analogue of Theorem \[thm:least-sing-value\]. \[thm:least-sing-valueWigner\] Under the assumptions of Theorem \[thm:prodWigner\], there exists $A >0$ such that for almost every $z \in \mathbb{C}$, almost surely $$\lim_{N \to \infty} {\ensuremath{\mathbf{1}_{\{\sigma_{2N}( {\mathbf}{Z}_N - z{\mathbf}{I}) \leq N^{-A}\}}}} = 0.$$ In order to prove Theorem \[thm:least-sing-valueWigner\], we will need the following result due to Nguyen [@NgLSV]. The version stated in [@NgLSV] requires that the deterministic matrices ${\mathbf}{F}_{N,k}$ be real symmetric. However, the proof given in [@NgLSV] only requires that ${\mathbf}{F}_{N,k}$ be complex symmetric [@NgPC]. In particular, the proof only uses that the $(i,j)$-entry of ${\mathbf}{F}_{N,k}$ be equal to the $(j,i)$-entry. We present the most general version below. \[thm:NgLSV\] For each $k=1,2$, let ${\mathbf}{F}_{N,k}$ be a $N \times N$ complex symmetric matrix whose entries are bounded in magnitude by $N^\alpha$, for some $\alpha > 0$. Then, under the assumptions of Theorem \[thm:prodWigner\], for any $B > 0$, there exists $A>0$ (depending on $\alpha, B$) such that $${\mathbb{P}}\left( \min_{k=1,2} \sigma_{N}( {\mathbf}{Y}_{N,k} + {\mathbf}{F}_{N,k}) \leq N^{-A} \right) = O(N^{-B}).$$ We now prove Theorem \[thm:least-sing-valueWigner\] Let $A$ be a large positive constant to be chosen later. By the Borel-Cantelli lemma, it suffices to show that, for almost every $z \in \mathbb{C}$, $${\mathbb{P}}\left( \left\\| ({\mathbf}{Z}_N - z {\mathbf}{I})^{-1} \right\\| \geq N^{A} \right) = O(N^{-2}).$$ By , we observe that, for $z \neq 0$, $({\mathbf}{Z}_N - z {\mathbf}{I})^{-1}$ has the form $$\begin{bmatrix} z \left(\frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} & \left(\frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} \\ z^2 \left(\frac{1}{N} {\mathbf}{Y}_{N,2} {\mathbf}{Y}_{N,1} - z^2 {\mathbf}{I} \right)^{-1} \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} & z \left( \frac{1}{N} {\mathbf}{Y}_{N,2} {\mathbf}{Y}_{N,1} - z^2 {\mathbf}{I} \right)^{-1} \end{bmatrix}$$ provided the relevant inverses exist. Thus, it suffices to show that the spectral norm of each block above is $O(N^{A})$ with probability $1-O(N^{-2})$. The treatment of each block is similar; as an illustration, we will show that $${\mathbb{P}}\left( \left\\| \left(\frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} \right\\| \geq N^{A} \right) = O(N^{-2}).$$ Indeed, since $${\mathbb{P}}\left( \left\\| \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} \right\\| \geq N^{100} \right) \leq N^{-200} {\mathbb{E}}\left\\| \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} \right\\|^2_2 = O(N^{-2}),$$ it suffices to show that there exists $A>200$ such that $$\label{eq:show2prodbnd} {\mathbb{P}}\left( \left\\| \left( \frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} \right\\| \geq N^{A} \right) = O(N^{-2}).$$ By Theorem \[thm:NgLSV\], there exists $A'>0$ such that the event $$\Omega_N := \left\{ \left\\| \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} \right)^{-1} \right\\| \leq N^{A'} \right\}$$ holds with probability $1-O(N^{-2})$. On this event, we observe that $$\left( \frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} = \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} \right)^{-1} \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} - z^2 \left(\frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} \right)^{-1} \right)^{-1}.$$ Therefore, we conclude that $$\begin{aligned} {\mathbb{P}}&\left( \left\\| \left( \frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} \right\\| \geq N^{A} \right) \\ &\leq {\mathbb{P}}\left( \left\\| \left( \frac{1}{N} {\mathbf}{Y}_{N,1} {\mathbf}{Y}_{N,2} - z^2 {\mathbf}{I} \right)^{-1} \right\\| \geq N^{A} \bigg\| \Omega_N \right) {\mathbb{P}}(\Omega_N) + {\mathbb{P}}(\Omega_N^C) \\ &\leq O(N^{-2}) + {\mathbb{P}}\left( \left\\| \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} - z^2 \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2}\right)^{-1} \right)^{-1} \right\\| \geq N^{A/2} \bigg\| \Omega_N \right)\end{aligned}$$ for $A$ sufficiently large. We now recall that ${\mathbf}{Y}_{N,1}$ and ${\mathbf}{Y}_{N,2}$ are independent, and, on the event $\Omega_N$, the entries of $\left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} \right)^{-1}$ are bounded in magnitude by $N^{A'}$. In addition, we observe that $z^2 \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2} \right)^{-1}$ is a complex symmetric matrix since ${\mathbf}{Y}_{N,2}$ is a real symmetric matrix. Therefore, by Theorem \[thm:NgLSV\], there exists $A>0$ such that $${\mathbb{P}}\left( \left\\| \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,1} - z^2 \left( \frac{1}{\sqrt{N}} {\mathbf}{Y}_{N,2}\right)^{-1} \right)^{-1} \right\\| \geq N^{A/2} \bigg\| \Omega_N \right) = O(N^{-2}).$$ This verifies , and hence the proof of Theorem \[thm:least-sing-valueWigner\] is complete. Truncation {#sec:truncationproof} ========== [99]{} G. Akemann, Z. Burda, Universal microscopic correlation functions for products of independent Ginibre matrices, J. Phys. A: Math. Theor. 45 (2012). G. Akemann, Z. Burda, M. Kieburg, Universal distribution of Lyapunov exponents for products of Ginibre matrices, available at [arXiv:1406.0803]{}. G. Akemann, J. R. Ipsen, M. Kieburg, Products of Rectangular Random Matrices: Singular Values and Progressive Scattering, Phys. Rev. E 88, (2013). G. Akemann, J. R. Ipsen, E. Strahov, Permanental processes from products of complex and quaternionic induced Ginibre ensembles, available at [arXiv:1404.4583]{}. G. Akemann, M. Kieburg, L. Wei, Singular value correlation functions for products of Wishart random matrices, J. Phys. A: Math. Theor. 46 (2013). G. Akemann, E. Strahov, Hole probabilities and overcrowding estimates for products of complex Gaussian matrices, J. Stat. Phys. (2013), Volume 151, Issue 6, pp 987–1003. G. Anderson, Convergence of the largest singular value of a polynomial in independent Wigner matrices, Ann. Probab. Volume 41, Number 3B (2013), 2103–2181. Z. D. Bai, [Circular law]{}, Ann. Probab. 25* (1997), 494–529. Z. D. Bai, J. Silverstein, [Spectral analysis of large dimensional random matrices]{}, Mathematics Monograph Series 2, Science Press, Beijing 2006. C. Bordenave, [On the spectrum of sum and product of non-hermitian random matrices]{}, Elect. Comm. in Probab. 16 (2011), 104–113. C. Bordenave, D. Chafaï, [Around the circular law.]{} Probability Surveys 9 (2012). 1–89 P. Biane, F. Lehner Computation of some examples of Brown’s spectral measure in free probability. Colloq. Math. 90 (2001), no. 2, 181–211. Z. Burda, R. A. Janik,and B. Waclaw, Spectrum of the product of independent random Gaussian matrices, Phys. Rev. E 81 (2010). Z. Burda, A. Jarosz, G. Livan, M. A. Nowak, A. Swiech, Eigenvalues and Singular Values of Products of Rectangular Gaussian Random Matrices, Phys. Rev. E 82 (2010). Z. Burda, Free products of large random matrices - a short review of recent developments, available at [arXiv:1309.2568]{}. D. L. Burkholder, Distribution function inequalities for martingales, Ann. Probab. 1 19–42 (1973). A. Edelman, [The Probability that a random real Gaussian matrix has $k$ real eigenvalues, related distributions, and the circular Law]{}, J. Multivariate Anal. 60, 203–232 (1997). P. J. Forrester, Lyapunov exponents for products of complex Gaussian random matrices, available at [arXiv:1206.2001]{}. P. J. Forrester, Probability of all eigenvalues real for products of standard Gaussian matrices, available at [arXiv:1309.7736]{}. J. Ginibre, [Statistical ensembles of complex, quaternion and real matrices]{}, J. Math. Phys. 6 (1965), 440–449. V. L. Girko, [Circular law]{}, Theory Probab. Appl. (1984), 694–706. V. L. Girko, [The strong circular law, twenty years later]{}, II. Random Oper. Stochastic Equations 12 (2004), no. 3, 255–312. V. L. Girko, Elliptic law, Theory of Probability and Its Applications, Vol. 30, No. 4 (1985). V. L. Girko, The elliptic law: ten years later I, Random Oper. and Stoch. Equ., Vol. 3, No. 3, pp. 257–302 (1995). I. Goldsheid, B. A. Khoruzhenko, [The Thouless formula for random non-Hermitian Jacobi matrices]{}, Israel J. Math., 148 (2005), 331–346. F. Götze, A. Naumov, T. Tikhomirov, On one generalization of the elliptic law for random matrices, available at [arXiv:1404.7013]{}. F. Götze, T. Tikhomirov, The circular law for random matrices, Ann. Probab. Volume 38, Number 4 1444–1491 (2010). F. Götze, T. Tikhomirov, On the Asymptotic Spectrum of Products of Independent Random Matrices, available at [arXiv:1012.2710]{} U. Haagerup, F. Larsen Brown’s Spectral Distribution Measure for R-Diagonal Elements in Finite von Neumann Algebras, Journal of Functional Analysis, Volume 176, Issue 2, 331–367 (2000). U. Haagerup, S. Thorbjørnsen, A new application of random matrices: $Ext(C_{red}^(F_2))$ is not a group.* Ann. of Math. (2) 162 711–775 (2005). J. Helton, R. Far, R. Speicher, Operator-valued semicircular elements: solving a quadratic matrix equation with positivity constraints, Int. Math. Res. Not. (2007). F. Hiai and D. Petz, The Semicircle Law, Free Random Variables and Entropy, Mathematical Surveys and Monographs, 77, American Mathematical Society, Providence, RI, 2000. R. A. Horn, C. R. Johnson, Matrix Analysis, Cambridge Univ. Press (1991). F. Larsen, Powers of R-diagonal Elements, Journ. Operator Theory 47, 197–212 (2002). M. L. Mehta, [Random matrices and the statistical theory of energy levels]{}, Acad. Press (1967). M. L. Mehta, [Random Matrices]{}, third edition. Elsevier/Academic Press, Amsterdam (2004). A. Naumov, Elliptic law for real random matrices, available at [arXiv:1201.1639 \[math.PR\]]{}. H. Nguyen, On the least singular value of random symmetric matrices, Electronic Journal of Probability 17 (2012), no. 53, 1–19. H. Nguyen, personal communications, September 2014. H. Nguyen, S. O’Rourke, The elliptic law, available at [arXiv:1208.5883 \[math.PR\]]{}. H. Nguyen, S. O’Rourke, On the concentration of random multilinear forms and the universality of random block matrices, available at [arXiv:1309.4815]{}. Nica A., Speicher R. R-diagonal pairs - a common approach to Haar unitaries and circular elements Fields Institute Communications, Vol. 12 (D. Voiculescu, ed.), AMS, 1997, pp 149-188. S. O’Rourke, D. Renfrew, Low rank perturbations of large elliptic random matrices, submitted, available at [arXiv:1309.5326]{}. S. O’Rourke, A. Soshnikov, Products of Independent Non-Hermitian Random Matrices, Electronic Journal of Probability, Vol. 16, Art. 81, 2219–2245, (2011). G. Pan, W. Zhou, [Circular law, extreme singular values and potential theory]{}, Journal of Multivariate Analysis, 101 645–656 (2010). T. Tao, V. Vu, Random matrices: The Circular Law, Communication in Contemporary Mathematics 10 (2008), 261–307. T. Tao, V. Vu, [From the Littlewood-Offord problem to the circular law: universality of the spectral distribution of random matrices]{}, [Bull. Amer. Math. Soc.]{} (N.S.) 46 (2009), no. 3, 377–396. T. Tao, V. Vu, Random matrices: Universality of ESDs and the circular law, Ann. Probab. Volume 38, Number 5 (2010), 2023–2065. T. Tao, V. Vu, Random matrices: universality of local eigenvalue statistics, Acta Math 206 (2011), 127–204. E. P. Wigner, [On the distributions of the roots of certain symmetric matrices]{}, Ann. Math. 67, (1958) 325–327. [^1]: S. O’Rourke has been supported by grant AFOSAR-FA-9550-12-1-0083 [^2]: D. Renfrew is partly supported by NSF grant DMS-0838680 [^3]: A. Soshnikov has been supported in part by NSF grant DMS-1007558 [^4]: V. Vu is supported by research grants DMS-0901216, DMS-1307797, and AFOSAR-FA-9550-12-1-0083.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Public policy-makers use cost-effectiveness analyses (CEA) to decide which health and social care interventions to provide. Appropriate methods have not been developed for handling missing data in complex settings, such as for CEA that use data from cluster randomised trials (CRTs). We present a multilevel multiple imputation (MI) approach that recognises when missing data have a hierarchical structure, and is compatible with the bivariate multilevel models used to report cost-effectiveness. We contrast the multilevel MI approach with single-level MI and complete case analysis, in a CEA alongside a CRT. The paper highlights the importance of adopting a principled approach to handling missing values in settings with complex data structures.' address: - 'Centre for Primary Care & Public Health, Queen Mary University of London, 58 Turner Street, London E1 2AB, UK. ' - 'London School of Hygiene and Tropical Medicine, Keppel Street, London, WC1E 7HT, UK.' - 'London School of Hygiene and Tropical Medicine, Keppel Street, London, WC1E 7HT, UK.' author: - 'Karla Díaz-Ordaz' - 'Michael G. Kenward' - Richard Grieve title: 'Handling missing values in cost-effectiveness analyses that use data from cluster randomised trials.' --- Introduction ============ Public policy-makers use cost-effectiveness analyses (CEA) in deciding which health and social care interventions to prioritise [@NICE; @CADTH2006; @PBCA2008; @IQWIG2009]. CEA exploit evidence from randomised studies, and if they adopt appropriate statistical methods, can provide accurate assessments of which interventions are most worthwhile [@Gold1996; @OHagan2001c; @Willan2006; @Glick2007; @Gray2010]. CEA raise major challenges for the analytical approach as the data tend to have complex structures, with correlated cost and effectiveness endpoints [@Willan2003; @Willan2006a], hierarchical data [@Manca2005b; @Pinto2005], and costs with right-skewed distributions [@Manning2006; @Jones2011a]. Most CEA that use individual-level data have observations with incomplete information [@Noble2011]. Statistical methods have not been developed that can simultaneously address all these issues. Hence studies may fail to provide the unbiased, precise cost-effectiveness estimates that decision-makers require. This paper is motivated by CEA that use data from cluster randomised trials (CRTs), but the approach we propose addresses three issues of general relevance. The first issue is raised by the bivariate nature of the outcomes, which implies the need for joint modelling. Here, one endpoint is highly skewed, but inferences about means are still required on the original scales of measurement. The second issue is that randomisation is at the cluster level, which implies that the data are hierarchical. The third issue, and the focus of this paper, is the presence of missing data. Approaches have been proposed for jointly modelling costs and health outcomes while acknowledging that individual costs tend to have right-skewed distributions [@Thompson2005]. There is a large literature on methods for handling clustered data, see for example [@Hayes2009], [@Eldridgebook], [@Aerts2002], [@Goldstein]. Methods for CEA alongside CRT include: a ‘two-stage’ non-parametric bootstrap procedure [@Flynn2005b; @Bachmann2007]; bivariate Generalised Estimating Equations with robust standard errors [@Gomes2011], and bivariate multilevel models (MLMs). Amongst the MLMs proposed are bivariate Normal models estimated by maximum likelihood [@Gomes2011], or with Bayesian Markov-Chain Monte Carlo (MCMC) methods [@Grieve2010; @Bachmann2007]. An outstanding issue is handling missing data in CEA with complex structures. For example, in a CRT, the prevalence of missing endpoint data may differ according to individual and cluster-level characteristics (e.g cluster size). CEA methods guidance recommends multiple imputation (MI) [@Blough2009; @Briggs2003; @Ramsey2005], but most published CEA still use complete case analysis [@Noble2011]. While MI approaches have been proposed for handling missing data with a clustered structure [@CarpenterGold; @Schafer], no previous study has developed methods for handling missing hierarchical data in complex settings, such as those seen in CEA that use cluster trials. The aim of this paper is to develop and illustrate an overall approach to analysing studies which have: bivariate outcomes with one highly skewed endpoint, a clustered structure and missing data. We do this using MI within a frequentist paradigm. At the same time, we explore the implications of failing to acknowledge relevant features of the setup in the handling of the missing data, in particular the potential consequences of ignoring clustering in the imputation step, and departures from Normality. We also compare the results we obtain with those from an analysis restricted to those individuals with complete data. In Section \[sec:ponder\], we introduce our case study which is a typical CEA that uses CRT data. In Section \[sec:substantive\], we develop a simple modelling framework for a clustered bivariate pair of outcomes, one of which has a potentially non-Normal distribution. Section \[sec:missing\] considers in some detail the handling of missing data in this setup, and explores the use of multilevel MI for this problem. In Section \[sec:results\], we compare the results obtained from a range of alternative strategies. We close with a discussion in Section \[sec:discussion\]. Motivating example: The PoNDER study {#sec:ponder} ==================================== The PoNDER study (psychological interventions for post-natal depression trial and economic evaluation) was a CRT evaluating an intervention for preventing postnatal depression, [@Morrell2009]. It included 2659 patients who attended 101 primary care providers in the UK (general practices). Clusters were randomly allocated to provide either usual care (control) or an intervention delivered by a health visitor (treatment). The intervention comprised health visitor training to identify and manage patients with postnatal depression. As is common, the PoNDER CRT had an unbalanced design; the number of patients per cluster varied widely (from 1 to 101 in the control group and from 1 to 81 in the treatment group). Patients were followed up for 18 months with costs ($\pounds$ sterling) and health-related quality of life (HRQoL) recorded at six monthly intervals. This paper considers costs and HRQoL reported at six months. These HRQoL data were used to adjust life years and present quality-adjusted life years (QALYs) over six months. Intra-cluster correlation coefficients (ICCs) were moderate for QALYs (ICCq=0.04), but high for costs (ICCc=0.17). While QALYs were approximately Normally distributed, costs were moderately skewed. Baseline measurements were collected from mothers at six weeks post-natally, for variables anticipated to be prognostic for either cost or effectiveness endpoints. Table \[Tabmiss\] reports the percentage of observations with missing data, by treatment group. For each baseline variable, less than $2.5\%$ of participants had missing data, but a relatively high proportion of individuals had missing data for the cost endpoint; 31 clusters were without any observed cost data (15 in the control arm, including one cluster that withdrew from the study). [lllllll]{}&&& &\ &&& &\ &type&symbol&&$\%$&& $\%$\ \ Cost&continuous &$c\ij$ &402 &41.1&460&26.6\ QALY&continuous &$q\ij$ &39&4.3&59&3.4\ \ & && & &\ \ \ &continuous &$\epd_{ij}$ &0 &0&0&0\ &binary & $\eth_{ij}$&0&0&0&0\ &binary&$\eco_{ij}$ & 0 & 0&0&0\ Age&continuous &$\mbox{age}_{ij}$ & 1&0.1&0&0\ &binary&$\mbox{eng}\ij$& 0&0&0&0\ &binary&$\mbox{al}\ij$ &9&1.0&7&0.4\ &ordinal&$\mbox{pe}\ij$ & 7& 0.8&10 &0.6\ &binary& $\mbox{b}\ij$&19&2.1&38&2.2\ &binary&$\mbox{d}\ij$&5&0.5&6&0.3\ &binary&$\mbox{l}\ij$&8&0.9&9&0.5\ &ordinal & $\mbox{rb}\ij$ &12&1.3&20&1.2\ The CEA presents incremental QALYs and costs as the differences in means, between the treatment and control groups (@MorrelHTA). Cost-effectiveness is then reported as incremental net monetary benefits (or INB, see equation for definition). To simplify the exposition, we restrict our analyses to those individuals with positive costs, by excluding 18 observations with zero costs (15 in the treatment group). See Section \[sec:discussion\] for further discussion. Substantive model {#sec:substantive} ================= Let $C_{ij}$ and $Q_{ij}$ be the cost and QALY outcomes respectively from the $j$th patient in cluster $i$ of a two-armed CEA alongside a CRT. We assume that the observations from different clusters are independent. We are principally concerned with estimating the linear additive effect of treatment on mean costs and health outcomes, with no additional covariates. Because of the simplicity of our setup, we are able to model the data from the two treatment groups entirely separately, and then make the comparison. So, in the following, we show the development for one treatment group; exactly the same arguments apply to the other. First, we introduce bivariate Normal latent variables $\{u_{i},w_{i}\}$ to represent possible cluster effects for cost and QALYs respectively, with $$\label{clustereffectsrho} \left(\begin{array}{c} u_{i}\\ w_{i}\end{array}\right) \sim \mbox{N}\left[\left(\begin{array}{c} 0\\ 0\end{array}\right), \left(\begin{array}{c c}\s_{u} & \rho\sigma_u\sigma_w \\ \rho\sigma_u\sigma_w & \s_{w} \end{array}\right) \right],$$ where $\s_u, \s_w,$ and $\rho$ are the variances and correlation of the two latent variables respectively. We now build the bivariate substantive model on the expectations of the two outcomes, $C_{ij}$ and $Q_{ij}$, defined conditionally on the two cluster effects, first for cost: $$\label{eq:cost} \mu_C=\mbox{E}[C_{ij} \mid u_i, w_i] = \b_1+ u_i$$ with $\b_1$ the mean appropriate for the first treatment group, and then for QALYs, conditional on the costs and cluster effects $$\label{eq:qualy} \mu_Q=\mbox{E}[Q_{ij}\mid c_{ij}, u_i, w_i]= \g_1+\alpha c_{ij}+w_i,$$ with $\g_1$ the intercept for $Q_{ij}$ for the first treatment group, and $\alpha$ the corresponding regression coefficient for the costs. We now introduce distributions for $C_{ij}$ and $Q_{ij}$, conditional on the cluster effects. It is assumed that the conditional distribution of $Q_{ij}$ given $C_{ij}$ is Normal, with variance $\sigma^2_q$. We consider three possible distributions for $C_{ij}$: Normal, Lognormal and Gamma. Other distributions could of course also be considered if thought appropriate. The choice of the Normal is straightforward, the mean is given by (\[eq:cost\]), with some variance $\sigma_c^2$ say. The Gamma alternative is introduced with a parameterisation that implies that the coefficient of variation, $\sqrt{\eta}$ say, is constant across clusters; in contrast to the Normal which implies constant variance. For $\mu_C$, the conditional mean as given in (\[eq:cost\]), the chosen Gamma density can then be written $$\label{eq:gammapdf} f_C(c) = \frac{1}{\Gamma(\eta)}\left(\frac{\eta}{\mu_C}\right)^{\eta}x^{\eta-1}\exp(-\eta x/\mu_C).$$ To maintain comparability with the Gamma distribution, we introduce the Lognormal with a somewhat unusual parameterisation, in which the coefficient of variation is again constant across clusters. This gives $$\label{eq:lgnormpdf} f_C(c)=\frac{1}{c\sqrt{2\pi \log(1+\eta)}}\exp{\left\{- \frac{(\log c-\mu_C)^2}{2\log(1+\eta)} \right\}.}$$ We assume that, conditional on the cluster effects, $(C_{ij},Q_{ij})$ is independent of $(C_{ij'},Q_{ij'})$ for $j\neq j'$, and so the required joint density, still conditional on the cluster effects, can be obtained from the product of the densities for $[ C_{ij} \mid u_i, w_i ]$ and $[Q_{ij} \mid C_{ij}, u_i, w_i]$. Finally, to obtain the marginal likelihood for the data for one treatment group, it is then necessary to combine this joint density over all relevant patients, and then integrate over the distribution of the cluster effects. This needs to be done numerically. There are several approaches for this, here we have used adaptive Gaussian quadrature as implemented in SAS PROC NLMIXED. We provide sample code for this in Appendix \[appendix:sas\]. Using conventional likelihood procedures we can then obtain estimated means for cost and QALYs ($\hat{\mu}_{C,k}$ and $\hat{\mu}_{Q,k}$ say) for treatment groups $k=1,2$ respectively, together with their estimated variances and covariances. Note that the separate modelling steps for the two treatment groups implies that estimates are independent between groups. The increments between the two groups are then estimated as $\hat{\delta}_C = \hat{\mu}_{C,2} - \hat{\mu}_{C,1} $ and $\hat{\delta}_Q = \hat{\mu}_{Q,2} - \hat{\mu}_{Q,1}$. The relative cost-effectiveness of treatment 2 against treatment 1 can be summarised by the INB defined as $$\label{inbdef} \mbox{INB}(\lambda) = \lambda \delta_Q-\delta_C$$ for $\lambda$, a given threshold willingness to pay for a unit of health gain. Its standard error can be calculated from the estimated variances and covariances of $\hat{\delta}_C$ and $\hat{\delta}_Q$ in the usual way. Missing Data {#sec:missing} ============ Handling the missing data ------------------------- It is well known that missing data can be the source of selection bias, and we are rarely able to construct analyses in which we can be confident that such bias has been eliminated. Rather, we use what information is available both in the data and the substantive setting in an attempt to reduce potential bias. Following this, carefully targeted sensitivity analysis can play an valuable role. There are many ways in which analyses can attempt to deal with missing data, and in which sensitivity analysis can be constructed, see for example [@lr02] and [@MKbook]. One important source of information that can be used to potentially reduce bias is contained in observed variables that are associated both with the outcome and with the missing value process itself. If these variables are not part of the substantive model, they are termed [auxiliary]{} variables in the missing value context. There are several potential auxiliary variables in the current setting, and we will use an approach which can incorporate them. To explain the intended role of these variables we need to introduce some definitions due to [@rubin76]. We use these in a fairly loose way here, more formal expositions can be found in the books mentioned above, and in the references given there. One important distinction here from Rubin’s original definitions, is our use of these terms in a frequentist framework, which implies rather stronger conditions than Rubin’s likelihood based definitions. Let $\bY_{ij} = \{Y_{ij1},Y_{ij2}\} = \{C_{ij},Q_{ij}\}$ be the pair of observations from subject $(i,j)$ and define the random variable $R_{ijl}$ to take the value 1 if $Y_{ijl}$ is observed and 0 if missing. We say that the missing data are [Missing Completely at Random]{} (MCAR) if $R_{ijl}$ and $Y_{ijl}$ are independent. By contrast, the data are [Missing at Random]{} (MAR) if there are observed variables, contained in $\bV$ say, such that $R_{ijl}$ and $Y_{ijl}$ are conditionally independent given $\bV$. It can be seen that MCAR implies MAR. We can reject the MCAR assumption in favour of MAR if we see associations between observed variables and $R_{ijl}$, which is of course completely observed. If neither MCAR nor MAR hold, we say that the missing data are [Missing Not at Random]{} (MNAR). It usually impossible to rule out MNAR in practice from the data at hand, because this depends critically on the existence of associations between [unobserved]{} variables and the $R_{ijl}$, which the observed data cannot exclude. It is this dependence between $R_{ijl}$ and $Y_{ijl}$ that is the potential source of bias. It is therefore usually sensible to try at least to reduce this dependence by identifying potential auxiliary variables from among those observed, and this will form the first step in handling the missing data. This will be done separately for the two outcomes, because it is entirely plausible that very different missing value mechanisms will operate with the two outcomes. We make the simplifying assumption that our auxiliary variables are completely observed. This is not strictly necessary, and in principle the approach used here can be extended to the situation when they are not, but for our present purposes the restriction to complete variables permits a simpler exposition. Multiple Imputation ------------------- Having identified potential auxiliary variables, it is necessary to incorporate them into the analysis. If these variables were part of the substantive model, we could simply include them and so condition on them, and in this way reduce or remove the unwanted dependence between $R_{ijl}$ and $Y_{ijl}$. But as auxiliary variables, they are not in the substantive model, so this route is not available to us. Alternatively, we could construct an overall joint model in which these auxiliary variables are included as additional outcome variables. In the current setting this is awkward, although not infeasible, because of the clustered structure. We instead choose to use multiple imputation [@rubin78; @Kenward2007]. This has the advantage of retaining the original substantive model, adding to this an [imputation model]{}. This is essentially determined by the conditional distribution of the missing data given the observed data, which we allow to differ between outcomes and treatment arms. In the present setting, in which we are only considering missing data in the outcome, the conditional model follows from the substantive model. We note the role of the clustering in this: the observations from one cluster are mutually dependent, and so the conditional distribution of a missing value involves all the other observed values in the same cluster. To this basic model (or models) we add the auxiliary variables, again acknowledging the clustered structure. Given the substantive and imputation models, conventional MI procedures can be followed. These are set out in detail in many references, including [@lr02] and [@MKbook]. The overall MI procedure is as follows.\ (1) The imputation model is fitted to the observed data and Bayesian draws are taken from the posterior of the model parameters.\ (2) The missing data are imputed from the imputation model, using the parameters drawn in step (1).\ (3) The substantive model is fitted (here using maximum likelihood) to the data set that has been [completed]{} using the imputations from step (2), producing parameter estimates and their estimated covariance matrix.\ (4) Steps (1)-(3) are repeated a fixed number, $K$ say, of times.\ (5) The $K$ sets of parameter and covariance estimates from step (3) are then combined using Rubin’s formulae [@rubin87] to produce a single MI estimate of the substantive model parameters and associated covariance matrix. Under the MAR assumption, this will produce consistent estimators and in the absence of auxiliary variables, is asymptotically (as $K$ increases) equivalent to maximum likelihood. For the current analysis, the whole MI procedure has been done separately for the two treatment groups. To carry this out, we do need the facilities to make the required Bayesian draws from the imputation model, which is bivariate and includes a cluster random effect. One route for this is the MLwiN procedures developed by [@CarpenterGold]. This is restricted however to the multivariate Normal distribution, and the imputation step has been undertaken on the log of the costs, transforming back to the original scale in step (3) above. The approximation implied under the Lognormal and Gamma substantive models is unlikely to be critical. The more flexible multilevel imputation procedure of [@Carpenter2011] might be considered for future work, or a bespoke sampler developed for the specific bivariate models used here. Specifying the imputation model for the PoNDER case study {#PonderIM} --------------------------------------------------------- To investigate the associations of observed variables with the $R_{ijl}$, it is natural to use logistic regression, in this example with and without random cluster effects. This has been done separately here for cost ($l=1$) and QALYs ($l=2$) and also for each treatment group ($k=1,2$). In addition to the patient-level covariates described, we added the cluster-level variable [cluster size, $n_i$]{}, defined as the number of participants randomised in each cluster. Previous studies suggest that cluster size may be associated with costs or health outcomes [@Campbell2000; @Omar2000; @Neuhaus2011]. We also consider that the number of participants recruited in each cluster may be associated with missingness. In PoNDER, because clinical protocols were less restrictive in the control than treatment group, it was anticipated that any relationship between the cluster size and the endpoints would be stronger in the control group. In the control group, before allowing for clustering, ethnicity, economic status and cluster size were associated with missing costs, but after including a random effect to allow for clustering no covariates were associated with missing costs. Cluster size and $\epd$ seem to be associated with unobserved QALYs. In addition, $\epd$ and cluster size ($n_i$) are associated with costs and economic status, ethnicity and $\epd$ are associated with QALYs. In the treatment group, cluster size was seen to be associated with missing cost at the individual level, while adjusting for clustering resulted in economic status being predictive of missing costs. Only age is predictive of QALYs missingness, both ignoring and accounting for clustering. In addition, $\epd$ is associated with the value of both cost and QALYs. The MI algorithm implemented here assumes all variables included in the model are multivariate Normally distributed. We exploit this by choosing the same imputation models for both outcomes, adding all auxiliary variables which seem associated with either endpoint and their missingness, and modelling the two outcomes simultaneously. The imputation models chosen are summarised in Table \[impmodels\]. Type Model Control Group Intervention Group -------------- ------- ------------------------------------- ------------------------------------- Single level SL $\epd_{ij}+\eco_{ij}+\eth_{ij}$ $\epd_{ij}+\eco_{ij}+\age_{ij}$ SL-C $\epd_{ij}+\eco_{ij}+\eth_{ij}+n_i$ $\epd_{ij}+\eco_{ij}+\age_{ij}+n_i$ Multilevel ML $\epd_{ij}+\eco_{ij}+\eth_{ij}$ $\epd_{ij}+\eco_{ij}+\age_{ij}$ ML-C $\epd_{ij}+\eco_{ij}+\eth_{ij}+n_i$ $\epd_{ij}+\eco_{ij}+\age_{ij}+n_i$ : \[impmodels\] Single-level (ignoring clustering) and multilevel (accounting for clustering) imputation models used for the cost and QALY endpoints in the PoNDER case study. Models that included a cluster-level auxiliary variable are indicated by -C. Multiple imputation estimates for the example data set {#sec:results} ====================================================== The imputation models that ignore clustering (SL, SL-C) were implemented with the [ice]{} command in STATA (by chained equations), while the multilevel imputation models (ML, ML-C) used multivariate Normal MCMC algorithms implemented in MLwiN [mi]{} macros. For each imputation model in Table \[impmodels\], we obtained five imputed datasets. Figure \[Fig2\] highlights the impact that accounting for clustering in the MI model can have on the distributions of “imputed” values. It shows imputed cost data for the six clusters in the control arm with the highest number of observations with missing cost data. The Figure contrasts data imputed after applying the single-level imputation models versus the multilevel imputation which included cluster size as an auxiliary variable. The cost distribution appears somewhat less clustered after the single-level imputation than after multilevel imputation. ![\[Fig2\]Difference in the “spread” of imputed data, depending on whether the imputation model ignored or accounted for clustering (Model SL versus Model ML-C). The Figure shows the distribution of costs in the six control clusters with the highest number of missing values in the original dataset.](Graph1.eps) The five multiply imputed datasets were each analysed with the three substantive models defined in Section \[sec:substantive\], i.e. random cluster effects models with bivariate Normal (N-N), Lognormal-Normal (L-N) and Gamma-Normal (G-N) distributions. Table \[miest\] reports the MI estimates for mean cost and QALYs by treatment arm, and for comparison also includes estimates from complete cases (CC). Table \[miest\] shows that, as anticipated, the standard errors for both endpoints are larger for the control than for the treatment group. It is also clear that ignoring the hierarchical structure of the data in the imputation model results in different point estimates for mean cost, especially in the control arm. For all approaches, the estimated correlations between cost and QALYs are small. We use the estimates from Table \[miest\] to obtain incremental costs, QALYs and INBs for a willingness to pay, $\lambda$, of $\pounds 20000$ per QALY. These are reported in Table \[MItable\]. ------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ Model N-N L-N G-N N-N L-N G-N CC $273.3$ $286.6$ $277.2$ $256.8$ $258.5$ $254.6$ (SE) ($25.3$) ($23.0$) ($24.8$) ($11.7$) ($11.6$) ($10.7$) $0.027$ $0.027$ $0.027$ $0.030$ $0.030$ $0.030$ (SE) (0.0014) (0.0014) (0.0014) (0.0008) (0.0008) (0.0008) $-0.03$ $ -0.03$ $-0.04$ $-0.04$ $-0.04$ $-0.04$ SL 295.0 299.1 295.00 251.4 253.1 251.3 (SE) (16.8) (16.9) (19.1) (9.9) (9.5) (9.1) $0.027$ $0.027$ $0.027$ $0.030$ $0.030$ $0.030$ (SE) ($0.0011$) ($0.0011$) ($0.0011$) ($0.0007$) ($0.0007$) ($0.0007$) $0.00$ $-0.09$ $-0.06$ $-0.09$ $ -0.08$ $-0.06$ SL-C Mean cost $268.1$ $275.5$ $270.4$ $257.2$ $257.5$ $255.4$ (SE) ($18.8$) ($16.9$) ($18.1$) ($9.1$) ($9.2$) ($8.8$) Mean QALYs $0.026$ $0.026$ $0.026$ $0.030$ $0.030$ $0.030$ (SE) ($0.0011$) ($0.0011$) ($0.0011$) ($0.0007$) ($0.0007$) ($0.0007$) corr(c,q) $0.01$ $0.05$ 0.001 $-0.06$ $-0.08$ $-0.04$ ML Mean cost $264.6$ $265.0$ $262.5$ $256.9$ $257.9$ $255.1$ (SE) ($31.6$) ($25.8$) ($29.4$) ($12.8$) ($12.9$) ($12.1$) Mean QALYs $0.026$ $0.026$ $0.026$ $0.030$ $0.030$ $0.030$ (SE) ($0.0011$) ($0.0011$) ($0.0011$) ($0.0007$) ($0.0007$) ($0.0007$) corr(c,q) $0.06$ $0.03$ $0.04$ $-0.07$ $ -0.07$ $-0.06$ ML-C Mean cost $270.0$ $280.6$ $275.0$ $262.1$ $262.5$ $259.5$ (SE) ($21.0$) ($23.4$) ($23.5$) ($14.0$) ($12.2$) ($12.1$) Mean QALYs $0.026$ $0.026$ $0.026$ $0.030$ $0.030$ $0.030$ (SE) ($0.0011$) ($0.0011$) ($0.0011$) ($0.0007$) ($0.0007$) ($0.0007$) corr(c,q) $-0.09$ $ -0.09$ $-0.08$ 0.01 $0.00$ 0.04 ------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ : \[miest\]Mean (SE) costs (in $\pounds$ sterling) and QALYs, and estimated correlations between the two endpoints. Estimates are according to choice of approach for handling missing data, and for alternative bivariate substantive models. Estimates (SE) Model N-N L-N G-N ------------------- ------- ------------------- ------------------- ------------------- Incremental cost CC $-16.5$ ($27.9$) $-28.1$ ($25.8$) $-22.5$ ($27.00$) $\delta c$ SL $-43.6$ ($19.5$) $-46.0$ ($19.4$) $-43.7$ ($21.2$) SL-C $-10.9$ ($20.9$) $-18.0$ ($19.2$) $-15.0$ ($20.1$) ML $-7.7$ ($34.1$) $-7.1$ ($28.8$) $-7.4$ ($31.8$) ML-C $-7.9$ ($25.2$) $-18.1$ ($26.4$) $-15.5$ ($26.4$) Incremental QALYs CC $0.003$ ($0.002$) $0.003$ ($0.002$) $0.003$ ($0.002$) $\delta q$ SL $0.004$ ($0.001$) $0.004$ ($0.001$) $0.004$ ($0.001$) SL-C $0.004$ ($0.001$) $0.004$ ($0.001$) $0.004$ ($0.001$) ML $0.004$ ($0.001$) $0.004$ ($0.001$) $0.004$ ($0.001$) ML-C $0.004$ ($0.001$) $0.004$ ($0.001$) $0.004$ ($0.001$) INB CC $76.5$ ($43.7$) $81.3$ ($41.6$) $75.3$ ($43.5$) SL $117.0$ ($34.0$) $117.4$ ($34.9$) $117.9$ ($34.0$) SL-C $82.6$ ($33.8$) $94.0$ ($32.4$) $90.8$ ($33.3$) ML $82.7$ ($42.8$) $82.5$ ($38.6$) $82.6$ ($41.2$) ML-C $84.5$ ($38.0$) $96.0$ ($38.2$) $93.0$ ($38.5$) : \[MItable\]Estimated incremental cost (in $\pounds$ sterling), QALYS and INBs at a threshold of $\pounds 20000$ per QALY. Estimates are according to choice of approach for handling missing data for alternative bivariate substantive models Table \[MItable\] shows that the estimates of incremental cost, incremental QALYs and INB are relatively insensitive to the choice of cost distribution. In fact, for the incremental QALYs, where there is little missing data and the ICCs are low, the estimates and their standard errors are virtually identical following each missing data approach. However, inferences about the estimated incremental costs and the INBs differ depending on the approach for handling missing data. Firstly, complete case estimates are likely to be biased, as the missing mechanism is probably not MCAR and the substantive model is not adjusting for any covariates. Single-level MI approaches produce smaller standard errors than those obtained with multilevel MI and CC. This is because cost has a large ICC and we are looking at a between-cluster estimator. As a consequence, there is an increased risk of type I error, regardless of the choice of cost distribution used for the substantive model. Moreover, ignoring informative cluster size in the multilevel imputation model increases the magnitude of the estimated standard errors. This cluster-level covariate (cluster size) is associated with cost and with cost missingness, and so excluding it from the imputation model reduces the precision of the estimate, as information is lost. By contrast, including cluster size in the single-level imputation model results in point estimates for the incremental cost which are similar to those following multilevel MI, although estimates for the standard errors are still smaller than the corresponding multilevel MI estimates. ![\[Fig1GN\] MI Estimates of mean ($95\%$) incremental net benefit (INB) from model ML-C (multilevel multiple imputation with informative cluster size) versus model SL (ignoring clustering) using bivariate Gamma-Normal substantive model](MVCEAFigure1GammaN_2.eps) Figure \[Fig1GN\] shows the INB (with $95\%$ CI) at alternative thresholds of willingness to pay for a QALY gained. With the single-level MI, the INB and $95\%$ CI are positive throughout, indicating that the treatment is cost-effective. For the multilevel MI that acknowledges informative cluster size, the $95\%$ CIs around the INB are wider and include zero at realistic thresholds for a QALY gained. While for both approaches, the INB remains positive throughout, the single level MI approach appears to overstate both the absolute level of the INB, and the precision surrounding the estimate. Hence, in the PoNDER case study, once a more appropriate approach is taken to handling the missing data, it is less certain that the intervention is cost-effective. Discussion {#sec:discussion} ========== This paper provides a principled approach to handling missing data with complex structures, exemplified by CEA that use data from CRTs. The proposed approach follows the general principle that the imputation model should reflect the structure of the analytical model. In the context of cluster trials, just as the substantive model can account for clustering with a MLM, so must the imputation model. Moreover, because the analytical models typically used in CEA estimate the linear additive effects of treatment on mean costs and health outcomes without covariate adjustment, MI has particular appeal in this setting. By separating the imputation and substantive models, information on those auxiliary variables, such as baseline patient characteristics, associated with missingness and the endpoints of interest can and should be used, without the analyst having to modify the substantive model. Our study highlights that a single-level imputation model can underestimate the uncertainty surrounding the estimates of interest. More generally, [@Taljaard2008] showed that MI approaches that ignore clustering can increase Type I errors. Another common approach to handling missing data in CRTs is to include cluster as a fixed effect in a single-level imputation model [@White2011; @Graham2009], but this does not produce an imputation model that properly captures the conditional distribution of the missing data given the observed. Indeed, including cluster as a fixed effect represents the limiting case where the ICC tends to one, and does not reflect the variability of the imputed values. The simulation study by [@Andridge2011] found that including cluster as a fixed effect in the imputation model, can overestimate the variance of the estimates, especially when ICCs are low, and there are few clusters. Both features are common in our setting, a recent review found that out of 63 published CEA alongside cluster trials, $40\%$ had fewer than 15 clusters per treatment arm, and one third reported ICCs of 0.01 or less for health outcomes [@Gomes2012]. The case study presented here suggests that if there are cluster-level covariates associated with the missingness patterns and the value of the outcomes to be imputed, then including them even in a single-level imputation model may potentially provide more accurate point estimates. This is because including covariates that predict dependency on cluster reduces the ICC. However, unless such covariates fully explain the between-cluster variance, such single-level MI approaches would still overstate precision. Hence, we propose imputation models with random effects for clusters. A general challenge in CEA is choosing appropriate statistical models for costs, which tend to have right-skewed distributions. The bivariate models developed here use marginal log-likelihoods for one outcome and conditional for the other, by expressing the relationship between the two responses as a linear regression (see sample code provided in Appendix \[appendix:sas\]). In principle, these models are generalisable to allow mixed distribution log-likelihoods, provided the conditional likelihood of the dependent outcome is known explicitly and can be optimised. The advantage of this approach is that, by parameterising the density according to the coefficient of variation, and maximising the log-likelihood obtained, we avoid log-transforming and re-transforming costs in the presence of heteroscedasticity [@Manning1998; @Duan1983; @Mullahy1998; @Manning2001]. We consider three cost models that make alternative distributional assumptions, but keep an essentially Lognormal imputation model throughout, and use standard optimisation routines to obtain maximum likelihood estimates. Our findings suggest that assuming a different distribution for the imputation versus analytical model appears to have little impact, whereas the choice of whether or not the imputation model accounts for clustering can be important. A previous barrier to adopting principled MI approaches for hierarchical data was the lack of available software, but this is no longer the case. There are now three options for performing multilevel MI based on multivariate Normal MCMC algorithms: PAN [@Schafer] which is available as an [R]{} package [@R], the [mi]{} macro [@CarpenterGold] which operates within MLwiN [@mlwin] and can handle up to four hierarchical levels and binary variables, and REALCOM-impute macros [@Carpenter2011], which can also handle categorical variables and cluster-level variables with missing data. The approach presented in this paper has some limitations. For simplicity, we assume the missing data mechanism is MAR throughout. However, MI provides a flexible and convenient route for investigating sensitivity to alternative MNAR mechanisms [e.g. @Carpenter2007], and in principle standard procedures should apply without much modification. A further advantage of MI, which this case study could not exploit, is that the imputation model may include post-randomisation variables associated with missingness and endpoints, which should not be included in the substantive model. A further concern is that the imputation and analytical models may make incorrect distributional assumptions. Simulation studies by [@SchaferBook] have shown that MI can be fairly robust to model misspecification, but their simulation settings did not include multilevel structures; [@Yucel2010] recently investigated the impact of misspecifing the multilevel imputation model but focused on violations of the distributional assumptions for the random-effects. They find that when the imputation model has sufficient auxiliary variables, inferences are insensitive to non-Normal random-effects, unless the rates of missingness are very high or the sample size is small. They obtained similar results when the assumption that level-1 residuals were Normally distributed was violated. While we propose a general approach to handling missing data in cluster trials, it is illustrated through a single case study which cannot represent all the circumstances faced by CEA that use CRTs. There may be circumstances when the data display quite different structures to those considered here, for example where there are a high proportion with zero costs [@Mullahy1998], QALYs with highly irregular distributions [@Basu2012], or there are many auxiliary variables available. This paper suggests further extensions. Here, we combine multilevel MI with a MLM estimated by maximum likelihood, but there may be circumstances where it would be advantageous to combine multilevel MI with MLM estimated by Bayesian MCMC [@Lambert2005], for example when synthesising evidence across multiple sources [@Welton2008]; or indeed adopt a fully Bayesian approach to handling the missingness and specifying the analytical models [@Mason2012]. Future simulation studies could be useful in contrasting the relative performance of the alternative approach across a broad range of settings including those where there are a high proportion of observations with zero costs, health outcomes with irregular distributions, and few clusters. Clearly, ignoring clustering in the imputation model will have less impact as the ICC decreases. One way of reducing the outstanding variation at the cluster-level within a single-level imputation model is to introduce more cluster-level covariates. Further work is needed to assess under what circumstances this simple MI approach would provide reliable inferences. Finally, it would also be useful to extend the approaches to handling missing data to other settings with hierarchical data. These could include trials with repeated measures over time, studies with a high proportion of zero costs, censored costs, or non-randomised studies where covariate adjustment is required. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to Jane Morrell (PI) and Simon Dixon for permission to use, and for providing access to, the PoNDER data. We thank James Carpenter, Simon Thompson, Richard Nixon, John Cairns, Manuel Gomes and Edmond Ng for helpful discussions. KDO was supported by a NIHR Research Methods Fellowship, and RG was partly funded by the UK Medical Research Council. Implementation in SAS {#appendix:sas} ===================== We have developed a method that allows us to exploit the optimisation of general likelihood functions available in SAS procedure NLMIXED. Briefly, we duplicated the data and created an indicator variable for the first copy of the data, $\mbox{flag}=1$. We then used an $\mbox{if}$ statement indicating we wished to estimate cost parameters if $\mbox{flag}=1$. With this method, we were able to use marginal expressions of corresponding log-likelihood to estimate parameters for costs using in turn either a Normal, Lognormal or Gamma log-likelihood; the last two parameterised by the coefficient of variation. We use Gauss-Hermite quadrature, with 70 quadrature points, and the Newton-Raphson maximisation technique to estimate the maximum-likelihood parameters. As likelihood maximisation is sensitive to the initial parameters chosen in the NLMIXED model, we ran this twice, using different initial values, to ensure optimization had achieved convergence. Sample SAS code can be found below. proc nlmixed data=ponder2 method=Gauss qpoints=70 cov corr; title "Control group bivariate Gamma-Normal with 2 Cluster Effects"; where group=0; x=cost; y=qalygain; parms b0=268 c0=0.27 a=3 lsyx=-7 lnsc=8 lnse=2 r=0.01; mux= b0+u1; varyx = exp(lsyx); muyx= c0+u2+(varx(a/mux2))x; if (flag=1) then ll=-xa/mux+(a-1)log(x)-alog(mux)+alog(a)-log(Gamma(a)); else ll=-(1/2)log(2constant('pi'))-log(varyx)-((y-muyx)*2)/(2varyx); if (flag=1) then z = x; else z = y; model z ~ general(ll); random u1 u2 ~ normal([0,0],[exp(lnsc),r, exp(lnse)]) subject=cluster; estimate "my" c0+(varyx(a/b02))b0; run; erts, M., [M]{}olenberghs, G., [R]{}yan, L. M., and [G]{}eys, H. (2002). . . ndridge, R. R. (2011). [53]{}(1), 57–74. achmann, M. O., [F]{}airall, L., [C]{}lark, A. and [M]{}ugford, M. (2007). [5]{}(Icc), 12. asu, A. and [M]{}anca, A. (2012). egression estimators for generic health-related quality of life and quality-adjusted life years. [32]{}, 56–69. lough, D. K., [R]{}amsey, S. D., [S]{}ullivan, S. D. and [Y]{}usen, Y., for the [NETT]{} Research group (2009). he impact of using different imputation methods for missing quality of life scores on the estimation of the cost-effectiveness of lung volume-reduction surgery. [18]{}, 91–101. riggs, A., [C]{}lark, T., [W]{}olstenholme, J. and [C]{}larke, P. (2003). [12]{}(5), 377–92. (2006). . Canadian Agency for Drugs and Technologies in Health. Ottawa, Canada. ampbell, M. K., [M]{}ollison, J. [S]{}teen, N. [G]{}rimshaw, J. M. and [E]{}ccles, M. (2000). nalysis of cluster randomized trials in primary care: a practical approach. [17]{}(2), 192–196. arpenter, J. R. and [G]{}oldstein, H. (2004). ultiple imputation in [ML]{}wi[N]{}. [16]{}, 9–18. arpenter, J. R., [G]{}oldstein, H. and [K]{}enward, M. G. (2011). . [45]{}(5), 1–14. arpenter, J. R., [K]{}enward, M. G. and [W]{}hite, I. R. (2007). ensitivity analysis after multiple imputation under missing at random: a weighting approach. [16]{} (3), 259–275. evelopment [C]{}ore [T]{}eam. (2011). he [R]{} project for statistical computing. URL http://www.R-project.org/. uan, N. (1983). . [ 78]{}(383), 605. ldridge, [S]{}. and [K]{}erry, S. (2012). . . lynn, T. N. and [P]{}eters, T. J. (2005). luster randomized trials: [A]{}nother problem for cost-effectiveness ratios. [21]{}(3), 403–409. lick, H. A., [D]{}oshi, J. [S]{}onnad, S. and [P]{}olsky, D. (2007). . Oxford, UK: Oxford University Press. Gold, M., Siegel, J., Russell, L., and Weinstein, M. (Eds.) (1996). . Oxford University Press. oldstein, H. (2003). . . omes, M., [G]{}rieve, R., [E]{}dmunds, J. and [N]{}ixon, R. (2012). tatistical methods for cost-effectiveness analyses that use data from cluster randomised trials: a systematic review and checklist for critical appraisal. [32]{} (1), 209–220. omes, M., [N]{}g,E. S., [G]{}rieve, R., [N]{}ixon, R., [C]{}arpenter, J. R. and [T]{}hompson, S. G. (2012). eveloping appropriate analytical methods for cost-effectiveness analyses that use cluster randomized trials. [32]{} (2), 350–361. Graham, J. W. (2009). issing [D]{}ata [A]{}nalysis: [M]{}aking [I]{}t [W]{}ork in the [R]{}eal [W]{}orld. [60]{}, 549–576. ray, A., [C]{}larke, P., [W]{}olstenholme, J. and [W]{}ordsworth, S. (2010). . Oxford University Press. rieve, R., [N]{}ixon, R., and [T]{}hompson, S. G. (2010). [30]{} (2), 163–175. ayes, R. and [M]{}oulton, L. (2009). . . (2009). Institute for Quality and Efficiency in Health Care., Cologne, Germany. ones, A. M., [L]{}omas, J. and [R]{}ice, N. (2011). pplying [B]{}eta-type [S]{}ize [D]{}istributions to [H]{}ealthcare [C]{}ost [R]{}egressions. Health, econometrics and data group working papers, [HEDG]{}, c/o Department of Economics, University of York. enward, M. G. and [C]{}arpenter, J. R. (2007). ultiple imputation: current perspectives. [16]{} (3), 199–218. ambert, P., [S]{}utton, A., [B]{}urton, P., [A]{}brams, K., and [J]{}ones, D. (2005). ow vague is vague? [A]{} simulation study of the impact of the use of vague prior distributions in [MCMC]{} using [W]{}in[BUGS]{}. [24]{}, 2401–-2428. ittle, R. J. A. and [R]{}ubin, D. B. (2002). . Chichester: Wiley. anca, A., [R]{}ice,N., [S]{}culpher, M. J., and [B]{}riggs, A. (2005). ssessing generalisability by location in trial-based cost-effectiveness analysis: the use of multilevel models. [14 ]{}(5), 471–485. anning, W. (2006). , Chapter Dealing with skewed data on costs and expenditures. Cheltenham, UK: Edward Elgar. anning, W. G. (1998). [17]{}(3), 283–295. anning, W. G. and [M]{}ullahy, J. (2001). [20]{}(4), 461–494. ason, A., [R]{}ichardson, S., [P]{}lewis, I., and [B]{}est, N. (2012). trategy for modelling non-random missing data mechanisms in observational studies using [B]{}ayesian methods. [(in press)]{}. olenberghs, G. and [K]{}enward, M. G. (2007). Wiley, Chichester. orrell, C. J., [S]{}lade,P., [W]{}arner, R., [D]{}ixon, S., [N]{}icholl, J., [W]{}alters, S. J., [B]{}rugha, T., [B]{}arkham, M. and [P]{}arry, G. (2009). linical effectiveness of health visitor training in psychologically informed approaches for depression in postnatal women: pragmatic cluster randomised trial in primary care. [338]{}, 3045. orrell, C. J., [W]{}arner, R., [S]{}lade, P., [D]{}ixon, S., [W]{}alters, S. J., [P]{}aley, G., and [B]{}rugha, T. (2009)b. sychological interventions for postnatal depression: cluster randomised trial and economic evaluation. [T]{}he [P]{}o[NDER]{} trial. Technical report, Health Technology Assessment, 13, iii-iv, xi-xiii, 1-153. ullahy, J. (1998). [17]{}(3), 247–81. euhaus, J. M. and [M]{}c[C]{}ulloch, C. E. (2011). . [98]{}(1), 147–162. (2008). National Institute for Health and Clinical Excellence, London, UK. ixon, R. M. and [T]{}hompson, S. G. (2005). [14]{}(12), 1217–29. oble, S., [H]{}ollingworth, W. and [T]{}illing, K. (2012). issing data in trial-based cost-effectiveness analysis: the current state of play. [21]{}(2), 187–200. ’[H]{}agan, A., [S]{}tevens, J. W., and [M]{}ontmartin, J. (2001). ayesian cost-effectiveness analysis from clinical trial data. [20]{}(5), 733–753. mar, R. Z. and [T]{}hompson, S. G. (2000). nalysis of a cluster randomized trial with binary outcome data using a multi-level model. [19]{}(19), 2675–2688. (2008). Australian Government - Department of Health and Ageing., Camberra, Australia. into, E. M., [W]{}illan, A. R., and [O]{}’[B]{}rien, B. J. (2005). ost-effectiveness analysis for multinational clinical trials. [24]{}(13), 1965–1982. amsey, S., [W]{}illke, R., and [B]{}riggs, A. (2005). est practices for cost-effectiveness analysis alongside clinical trials: an [ISPOR]{} [RCT]{}-[CEA]{} task force report. [8]{}, 521–33. asbash, J., [C]{}harlton, C., [B]{}rowne, W., [H]{}ealy, M., and [C]{}ameron, B. (2011). . URL http://www.bristol.ac.uk/cmm/. ubin, D. (1976). [63]{}, 581–592. ubin, D. (1978). . , 20–34. ubin, D. (1987). . chafer, J. (1997). Chapman and Hall. London. chafer, J. and [Y]{}ucel, R. (2002). omputational strategies for multivariate linear mixed-effects models with missing values. [11]{}, 421–442. aljaard, M., [D]{}onner, A., and [K]{}lar, N. (2008). [50]{}(3), 329–45. hompson, S. G. and [N]{}ixon, R. M. (2005). [25]{}(4), 416–423. elton, N. J., [A]{}des, A. E., [C]{}aldwell, D. M., and [P]{}eters, T. J. (2008). esearch prioritization based on expected value of partial perfect information: a case-study on interventions to increase uptake of breast cancer screening. [171]{}, 807–841. hite, I. R., [R]{}oyston, P., and [W]{}ood, A. M. (2011). [30]{}(4), 377–99. illan, A. (2006). tatistical [A]{}nalysis of cost-effectiveness data from randomised clinical trials. [6]{}, 337–346. illan, A. and [B]{}riggs, A. (2006). John Wiley & Sons Ltd. illan, A. R., [C]{}hen, E., [C]{}ook, R., and [L]{}in, D. (2003). ncremental net benefit in randomized clinical trials with qualify-adjusted survival. [22]{}, 353–362. ucel, R. and [D]{}ermitas, H. (2010). mpact of non-normal random effects on inference by multiple imputation: [A]{} simulation assessment. [54]{}(3), 790–801.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The grand potential of a system of interacting electrons is considered as a stationary point of a self-energy functional. It is shown that a rigorous evaluation of the functional is possible for self-energies that are representable within a certain reference system. The variational scheme allows to construct new non-perturbative and thermodynamically consistent approximations. Numerical results illustrate the practicability of the method.' author: - 'Michael Potthoff [@ad]' title: 'Self-energy-functional approach to systems of correlated electrons' --- Introduction {#sec:intro} ============ Systems of strongly correlated electrons continue to represent a central subject of current research. Different interesting correlation phenomena, such as high-temperature superconductivity, [@OM00] Mott metal-insulator transitions [@Geb97] or itinerant ferromagnetism, [@BDN01] are far from being finally clarified. Progress in this field crucially depends on the development of new theoretical methods as even highly idealized model systems pose notoriously difficult problems. There are only a few general approaches which are able to access the equilibrium thermodynamics as well as excitation properties of an extended system of correlated electrons. General methods can be based on the Green’s-function formalism of Luttinger and Ward [@LW60] and Baym and Kadanoff: [@BK61] Here the grand potential $\Omega$ is expressed in terms of the time- or frequency-dependent one-electron Green’s function ${\bf G}$. The functional $\Omega[{\bf G}]$ can be shown to be stationary at the physical ${\bf G}$. In principle, this is an exact variational approach which provides information not only on static equilibrium but also on dynamic excitation properties. The functional dependence $\Omega[{\bf G}]$, however, is generally not known explicitly as it must be constructed by summation of an infinite series of renormalized skeleton diagrams. In the standard approximation the exact but unknown functional is replaced by an explicitly known but approximate one which is based on an incomplete summation of the diagram series. This leads to the well-known perturbational (“conserving”) theories. [@BK61] Higher-order approaches as the fluctuation-exchange approximation [@BSW89] are mainly applied to discrete lattice models while for continuum systems, e.g. for the inhomogeneous electron gas, one has to be content with lowest-order theories as the GW method. [@SMH80; @SMH82; @AG98] A second type of general methods is based on density-functional (DF) approaches. [@HK64; @KS65] Normally these aim at the inhomogeneous electron gas but can also be applied to Hubbard-type lattice models. [@SGN95] Compared with the Green’s-function formalism, there is a conceptually similar situation for DF approaches: In the latter the ground-state energy $E$ (or the grand potential $\Omega$) [@Mer65] is given as a functional of the (static) density ${\bf n}$. The variational principle associated with the functional $E[{\bf n}]$ is rigorous but cannot be evaluated as $E[{\bf n}]$ is generally unknown. In the standard local-density approximation (LDA) the (unknown part of the) functional is replaced by an explicitly known but approximate functional which is taken from the homogeneous system. For systems with weakly varying density the LDA should be justified. Information on excitation properties is contained in dynamic response functions which are in principle accessible via time-dependent DF theory [@RG84] where the action $A$ is considered as a functional of the time-dependent density ${\bf n}$. Again, the exact but unknown functional $A[{\bf n}]$ is approximated to make it explicit and the variational principle is exploited afterwards. The method proposed here rests on a variational principle which uses the electron self-energy ${\bf \Sigma}$ as the basic dynamic variable. A new functional $\Omega[{\bf \Sigma}]$ is constructed which can be shown to be stationary at the physical self-energy. The main result is that the variational principle can be exploited without any approximation of the functional dependence. Namely, a [rigorous]{} evaluation of the functional $\Omega[{\bf \Sigma}]$ is possible on a certain subspace of trial self-energies. Trial self-energies must be representable within an exactly solvable reference system sharing the same interaction with the original system. This result has important consequences as it opens a route for constructing a novel class of approximations. Although the self-energy essentially contains the same information as the Green’s function or the (time-dependent) density, the new approach is conceptually contrary to the Green’s-function approach and to the DF approach as there is no approximation to be tolerated for the central functional. Instead of approximating the functional itself, it is considered on a restricted domain. The self-energy-functional approach is completely general and yields approximations which are non-perturbative, thermodynamically consistent and systematic. Opposed to numerical techniques directly applied to systems of finite size, the self-energy-functional approach provides a variational or self-consistent embedding of finite systems and thus yields results in the thermodynamical limit. Such techniques are needed to construct phase diagrams from standard correlated lattice models. A potentially fruitful field of application are systems with competing types of order resulting from spin, charge or orbital correlations as it is typical e.g. for numerous transition-metal oxides. [@OM00; @Geb97; @BDN01] In the present paper the approach is introduced and a number of general aspects are discussed in detail (Sec. \[sec:sft\]). To demonstrate its usefulness, two applications will be considered for the single-band Hubbard model: In Sec. \[sec:dmft\] it is shown that the dynamical mean-field theory (DMFT) [@GKKR96] can be recovered within the self-energy-functional approach, namely by choosing a decoupled set of impurity Anderson models as a reference system. The DMFT generally requires the treatment of a quantum-impurity problem with an infinite number of degrees of freedom ($n_{\rm s} = \infty$). In Sec. \[sec:ed\] a new approximation is discussed which is based on an impurity model with a [finite]{} number of degrees of freedom only and which approaches the DMFT for $n_{\rm s} = \infty$. The method is closely related to the exact-diagonalization approach (ED). [@CK94; @SRKR94] Opposed to the ED, however, thermodynamical consistency is guaranteed at any stage of the approximation. New approaches beyond the mean-field level will be discussed elsewhere. The conclusions and an outlook are given in Sec. \[sec:con\]. Self-energy-functional approach {#sec:sft} =============================== Consider a general Hamiltonian $H = H_0({\bf t}) + H_1 ({\bf U})$ with one-particle (“hopping”) parameters ${\bf t}$ and two-particle interaction parameters ${\bf U}$: $$H = \sum_{\alpha\beta} t_{\alpha\beta} c_{\alpha}^\dagger c_{\beta} + \frac{1}{2} \sum_{\alpha\beta\gamma\delta} U_{\alpha\beta\delta\gamma} c_{\alpha}^\dagger c^\dagger_{\beta} c_{\gamma} c_{\delta} \: .$$ Here $\alpha,\beta,...$ refer to an orthonormal and complete set of one-particle basis states. We are interested in the equilibrium thermodynamics and in elementary one-particle excitations of the system for temperature $T$ and chemical potential $\mu$. This is described by the one-particle Green’s function $G_{\alpha\beta}(i\omega)=\langle \langle c_\alpha ; c_\beta^\dagger \rangle \rangle$ of the imaginary fermionic Matsubara frequencies $i\omega = i(2n+1) \pi T$ with integer $n$. [@AGD64] The Green’s function can be calculated from the self-energy $\Sigma_{\alpha\beta}(i\omega)$ via the Dyson equation. Using a matrix notation, this reads as ${\bf G} = {\bf G}_0 + {\bf G}_0 {\bf \Sigma} {\bf G}$ where ${\bf G}_0 = 1/(i\omega + \mu - {\bf t})$ is the “free” Green’s function. The self-energy is given by ${\bf \Sigma} = {\bf \Sigma}[{\bf G}] = T^{-1} \delta \Phi[\bf G] / \delta {\bf G}$, where $\Phi[{\bf G}]$ is the so-called Luttinger-Ward functional. [@LW60; @BK61] This allows to derive the Green’s function from a variational principle: One has $\delta {\Omega} [{\bf G}] / \delta {\bf G} = 0$ where ${\Omega} [{\bf G}] = \Phi[{\bf G}] + \mbox{Tr} \ln (- {\bf G}) - \mbox{Tr} (({\bf G}_0^{-1} - {\bf G}^{-1}) {\bf G})$ and using the notation $\mbox{Tr} \, {\bf A} = T \sum_{\omega,\alpha} A_{\alpha\alpha}(i\omega)$. In general, however, the functional $\Phi[{\bf G}]$ is not known explicitly which prevents an evaluation of ${\Omega} [{\bf G}]$ for a given ${\bf G}$. So-called conserving approximations [@BK61] provide an explicit but approximate functional $\Phi_{\rm pert.}[{\bf G}] \approx \Phi[{\bf G}]$. However, these are weak-coupling approaches where a certain subclass of $\Phi$ diagrams is summed up. Here a different but still rigorous variational principle is proposed which is based on a functional ${\bf G} = {\bf G}[{\bf \Sigma}]$ defined as the inverse of ${\bf \Sigma} = {\bf \Sigma}[{\bf G}]$. We can assume the latter to be invertible (locally) provided that the system is not at a critical point for a phase transition (see Appendix \[sec:gofs\]). Consider then: $$\Omega_{\bf t}[{\bf \Sigma}] \equiv {\rm Tr} \ln (- ({\bf G}_0^{-1} - {\bf \Sigma})^{-1}) + F[{\bf \Sigma}] \label{eq:var}$$ where $F[{\bf \Sigma}] \equiv \Phi[{\bf G}[{\bf \Sigma}]] - \mbox{Tr} ({\bf \Sigma} \, {\bf G}[{\bf \Sigma}])$ is the Legendre transform of $\Phi[{\bf G}]$. The subscript ${\bf t}$ indicates the explicit ${\bf t}$ dependence of $\Omega$ due to the free Green’s function ${\bf G}_0$. Using $T^{-1} \delta F[{\bf \Sigma}] / \delta {\bf \Sigma} = {\bf G}[{\bf \Sigma}]$, one finds that $$\delta \Omega_{\bf t} [{\bf \Sigma}] / \delta {\bf \Sigma} = 0 \; \Leftrightarrow \; {\bf G}[{\bf \Sigma}] = ({\bf G}_0^{-1} - {\bf \Sigma})^{-1} \: . \label{eq:stat}$$ Thus $\Omega_{\bf t}[{\bf \Sigma}]$ is stationary at the exact (physical) self-energy and its value is the exact grand potential of the system. Again, the problem is that the functional $\Omega_{\bf t}[{\bf \Sigma}]$ is in general not known explicitly. As the domain of the self-energy functional $\Omega_{\bf t}[{\bf \Sigma}]$ we define the class of all ${\bf t}'$ representable self-energies. ${\bf \Sigma}$ is termed ${\bf t}'$ representable, if there is a set of hopping parameters ${\bf t}'$ such that ${\bf \Sigma}$ is the exact self-energy of the model $H_0({\bf t}') + H_1 ({\bf U})$. This implies that any self-energy in the domain of $\Omega_{\bf t}[{\bf \Sigma}]$ can be parameterized as ${\bf \Sigma} = {\bf \Sigma}({\bf t}')$. The interaction parameters ${\bf U}$ are taken to be fixed. Suppose we are interested in the model $H = H_0({\bf t}) + H_1 ({\bf U})$. Then the function $\Omega_{\bf t}({\bf t}') \equiv \Omega_{\bf t}[{\bf \Sigma}({\bf t}')]$ is stationary at ${\bf t}' = {\bf t}$. Thus $\partial \Omega_{\bf t}({\bf t}') / \partial {\bf t}' = 0$. It is important to note that $F[{\bf \Sigma}]$ is universal: The functional dependence is the same for any ${\bf t}$, i.e. it remains unchanged for an arbitrary reference system $H'$ with the same interaction but modified hopping parameters: $H' = H_0({\bf t}') + H_1 ({\bf U})$. $F[{\bf \Sigma}]$ is universal as it is the Legendre tranform of $\Phi[{\bf G}]$ which in turn is universal because it can be constructed formally as the sum of all closed, irreducible, and renormalized skeleton diagrams which, apart from ${\bf G}$, include the vertices ${\bf U}$ only. Consequently, one has: $$\Omega_{{\bf t}'}[{\bf \Sigma}] = {\rm Tr} \ln (- ({{\bf G}'_0}^{-1} - {\bf \Sigma})^{-1}) + F[{\bf \Sigma}] \: , \label{eq:varp}$$ for the reference system $H'$ with ${{\bf G}'_0}^{-1} = i\omega + \mu - {\bf t}'$. Combining Eqs. (\[eq:var\]) and (\[eq:varp\]), $F[{\bf \Sigma}]$ can be eliminated: $$\begin{aligned} \Omega_{\bf t}[{\bf \Sigma}] = \Omega_{{\bf t}'}[{\bf \Sigma}] &+& {\rm Tr} \ln (- ({\bf G}_0^{-1} - {\bf \Sigma})^{-1}) \nonumber \\ &-& {\rm Tr} \ln (- ({{\bf G}'_0}^{-1} - {\bf \Sigma})^{-1}) \: . \label{eq:felim}\end{aligned}$$ Evaluating the functional $\Omega_{\bf t}[{\bf \Sigma}]$ for self-energies parameterized as ${\bf \Sigma} = {\bf \Sigma}({\bf t}')$, one obtains: $$\Omega_{\bf t}[{\bf \Sigma}({\bf t}')] = \Omega' + {\rm Tr} \ln (- ({\bf G}_0^{-1} - {\bf \Sigma}({\bf t}'))^{-1}) - {\rm Tr} \ln (-{\bf G}') \: . \label{eq:vvv}$$ Here it has been used that $\Omega_{{\bf t}'}[{\bf \Sigma}({\bf t}')] = \Omega'$, the exact grand potential of the reference system $H'$, and $({{\bf G}'_0}^{-1} - {\bf \Sigma}({\bf t}'))^{-1} = {\bf G}'$, the exact Green’s function of $H'$. Suppose that the reference system $H'$ is much simpler than the original system $H$ so that it can be solved exactly for any ${\bf t}'$ belonging to a certain subspace of the entire space of hopping parameters. The resulting Eq. (\[eq:vvv\]) is remarkable, as it shows that the functional $\Omega_{{\bf t}}[{\bf \Sigma}]$ can be evaluated rigorously for trial self-energies ${\bf \Sigma}={\bf \Sigma}({\bf t}')$ taken from the reference system $H'$. This is the main result. Contrary to previous approaches (e.g. conserving theories, LDA), there is no need to approximate the functional dependence in a fundamental variational principle. Approximations are constructed by searching for a stationary point of $\Omega_{{\bf t}}[{\bf \Sigma}]$ on a [restricted]{} set of trial self-energies ${\bf \Sigma}({\bf t}')$. The stationary point is determined by the Euler equation: $\partial \Omega_{\bf t}[{\bf \Sigma}({\bf t}')] / \partial {\bf t}'=0$. Calculating the derivative, $$T \sum_{\omega} \sum_{\alpha\beta} \left( \frac{1}{{\bf G}_0^{-1} - {\bf \Sigma}({\bf t}')} - {\bf G}' \right)_{\beta \alpha} \frac{\partial \Sigma_{\alpha\beta}({\bf t}')} {\partial {{\bf t}'}} = 0 \: . \label{eq:euler}$$ Note that the equation involves, apart from ${\bf G}_0$, quantities of the reference system $H'$ only. The linear response of the self-energy of $H'$ due to a change of the hopping ${\bf t}'$ can be calculated along the lines of Ref. . It turns out that $\partial {\bf \Sigma}({\bf t}') / \partial {\bf t}'$ is given by a two-particle Green’s function of $H'$. Since ${\bf G}' = {\bf G}[{\bf \Sigma}({\bf t}')]$, the exact self-energy of the system $H$ is determined by the condition that the bracket in (\[eq:euler\]) be zero. Hence, one can consider Eq. (\[eq:euler\]) to be obtained from the [exact]{} equation that determines the “vector” ${\bf \Sigma}$ in the self-energy space through [projection]{} onto the hypersurface of ${\bf t}'$ representable trial self-energies ${\bf \Sigma}({\bf t}')$ by taking the scalar product with vectors $\partial {\bf \Sigma}({\bf t}') / \partial {\bf t}'$ tangential to the hypersurface. An analysis of the second derivative $\partial^2 \Omega_{\bf t} [{\bf \Sigma}({\bf t}')] / \partial t'_{\alpha\beta} \partial t'_{\gamma\delta}$ shows that a stationary point is not an extremum point in general. This feature is shared with the time-dependent DF approach, [@RG84] the Green’s-function approach [@BK61] and also with a recently considered variant. [@CK01] Only in the static DF theory there is a convex (density) functional. [@HK64; @KS65; @Mer65] Nevertheless, the proposed self-energy-functional approach is systematic: For any sequence of reference systems $H'$ including more and more degrees of freedom and converging to the original system $H$ there is, from the variational principle, a corresponding sequence of grand potentials which must converge to the exact $\Omega = \Omega_{\bf t} [{\bf \Sigma}({\bf t})]$ as the subspace of trial self-energies increases and eventually includes the exact self-energy ${\bf \Sigma}({\bf t})$. Relation to the DMFT {#sec:dmft} ==================== Given an original model $H$, what could a suitable reference system $H'$ look like? Consider, for example, $H$ to be the Hubbard model [@hub] which is shown in Fig. 1a schematically: A filled dot represents a correlated site $i$ with on-site Hubbard interaction $U$, and a line connecting two sites $i$ and $j$ represents the nearest-neighbor hopping $t_{i,j}$. The number of sites is $L \mapsto \infty$. Fig. 1c shows a conceivable reference system $H'$. $H'$ is obtained from $H$ (Fig. 1a) by (i) adding to each correlated site $i$ a number of $n_{\rm s}-1$ uncorrelated (“bath”) sites $k=2,...,n_{\rm s}$ (open dots) which are disconnected from the rest of the system, by (ii) switching off the hopping $t_{i,j}$ between the correlated sites and (iii) switching on a hopping $V_{i,k}$ to the bath sites. After step (i) the Hamiltonian Fig. 1b (in the figure $n_{\rm s}=5$) has an enlarged Hilbert space but the same self-energy. It is important to note that steps (i) - (iii) leave the interaction part unchanged and thus preserve the functional dependence $F[{\bf \Sigma}]$. Actually, the system $H'$ is a set of $L$ decoupled single-impurity Anderson models (SIAM) [@And61] with $n_{\rm s}$ sites each. Compared to $H$, the problem posed by $H'$ is strongly simplified. This is achieved at the cost of restricting the set of trial self-energies. In particular, as the correlated sites are decoupled in $H'$, the trial self-energies are local: $\Sigma_{ij}(i\omega,{\bf t}') \propto \delta_{ij}$. One has to consider $H'$ for arbitrary one-particle parameters, namely the on-site energies of the correlated (“c”) and of the bath sites (“a”), $\epsilon^{(c)}_{i}$ and $\epsilon^{(a)}_{i,k}$, respectively, and the hopping (“hybridization”) $V_{i,k}$ between them and take these as variational parameters in the principle $\delta \Omega_{\bf t}[{\bf \Sigma}({\bf t}')] = 0$. ![ Schematic representation of the Hubbard model $H$ (a), an equivalent model (b), and a possible reference system $H'$ (c). See text for discussion. []{data-label="fig:hhp"}](fig1.eps){width="90mm"} Let us discuss the case $n_{\rm s} \mapsto \infty$. For a homogeneous phase of the (translationally invariant) original system, $\Omega_{\bf t}[{\bf \Sigma}({\bf t}')]$ will be stationary at a homogeneous set of variational parameters: ${\bf t}'=\{ \epsilon^{(c)}_{i}, \epsilon^{(a)}_{i,k}, V_{i,k} \}= \{ \epsilon^{(c)}, \epsilon^{(a)}_{k}, V_{k}\}$. Consequently, it is sufficient to consider one SIAM only. As the different equivalent SIAM’s are spatially decoupled, not only the self-energy but also its linear response is local: $\partial \Sigma_{ij}({\bf t}') / \partial {\bf t}' \propto \delta_{ij}$. To solve the Euler equation (\[eq:euler\]), it is thus sufficient to fulfill the “locally projected” equation $$\left( \frac{1}{{\bf G}_0^{-1}(i\omega) - {\bf \Sigma}(i\omega)} \right)_{ii} = G'_{ii}(i\omega) \: . \label{eq:dmft}$$ This is just the self-consistency equation of the DMFT: [@GKKR96] The SIAM parameters have to be found such that the on-site (“impurity”) Green’s function at a correlated site $i$ coincides with the on-site Green’s function of the Hubbard model which is calculated from ${\bf G}_0$ and the (“impurity”) self-energy of $H'$ by means of the Dyson equation. Therefore, one can state that the DMFT (as an approximation for any finite-dimensional system or as the exact theory in infinite dimensions) is recovered as a stationary point of $\Omega_{\bf t}[{\bf \Sigma}]$ when restricting the search to local self-energies representable by a SIAM. Within the DMFT the computation of the self-energy requires an iterative procedure: ${\bf \Sigma}_{\rm old} \mapsto {\bf \Sigma}_{\rm new}$. Here it turns out that this corresponds to a certain (discrete) path on the hypersurface of SIAM trial self-energies. Convergence of the iteration (${\bf \Sigma}_{\rm old} = {\bf \Sigma}_{\rm new}$), however, is by no means guaranteed physically but depends on the contracting properties of the map ${\bf \Sigma}_{\rm old} \mapsto {\bf \Sigma}_{\rm new}$. The self-energy-functional approach offers an alternative as instead of solving Eq. (\[eq:dmft\]) one may calculate $\Omega_{\bf t}[{\bf \Sigma}({\bf t}')]$ by Eq. (\[eq:vvv\]) and determine the stationary point. Hence, the DMFT can also be obtained by a more [direct]{} computation avoiding any iterations – similar (in this respect) to the random-dispersion approximation. [@NG99] Note that in case of more than a single stationary point there is also an equally direct access to metastable phases. For any inhomogeneous situation, Eq. (\[eq:dmft\]) represents a system of self-consistency equations to fix the parameters of non-equivalent impurity models labeled by the site index $i$. The models can be solved independently but are coupled indirectly due to the matrix inversion in (\[eq:dmft\]). This exactly recovers the DMFT generalized to systems with reduced translational symmetry. [@PN99c; @PN99d] A consistent ED method {#sec:ed} ====================== A brief discussion of two limiting cases of the Hubbard model may be instructive. Consider the band limit with $U=0$ first. Here $H = \sum_{ij\sigma} t_{ij} c_{i\sigma}^\dagger c_{j\sigma}$ describes a system of non-interacting electrons. This case is exceptional as obviously the functional $F[{\bf \Sigma}] \equiv 0$ and therefore $\Omega_{\bf t}[{\bf \Sigma}] = {\rm Tr} \ln (- ({\bf G}_0^{-1} - {\bf \Sigma})^{-1})$. Any valid reference system $H'$ must have the same (i.e. a vanishing) interaction part as $H$, and thus ${\bf \Sigma}({\bf t}') \equiv 0$ and $\Omega_{\bf t}[{\bf \Sigma}({\bf t}')] = {\rm Tr} \ln (-{\bf G}_0^{-1})$, the exact grand potential for non-interacting electrons. The atomic limit, $H = \sum_{i\sigma} ( t_{0} c_{i\sigma}^\dagger c_{i\sigma} + (U/2) n_{i\sigma} n_{i-\sigma})$ is more interesting as $\Phi[{\bf G}]$ and $F[{\bf \Sigma}]$ cannot be constructed explicitly. Within the self-energy-functional approach one has to compute ${\bf \Sigma({\bf t}')}$, ${\bf G}'$, and $\Omega'$ for a suitable reference system $H'$ and to insert into Eq. (\[eq:vvv\]) for optimization. The only meaningful choice for the reference system is $H'=H$ in this case. Obviously, this yields the exact solution. Generally, whenever the original model $H$ is exactly solvable, the choice $H'=H$ will do. For a non-trivial model $H$, the choice $H'=H$ is useless for a practical computation. Any simplified reference system, however, yields a consistent approximation. The case of the DMFT discussed in Sec. \[sec:dmft\] is an illustrative example. On the other hand, in the context of the DMFT actually both, $H$ and $H'$, are highly non-trivial models, and further approximations or large-scale numerics are needed to treat the reference system $H'$. More simple approximations for the Hubbard model which are still constistent are generated by considering reference systems with a [finite]{} number of degrees of freedom. The reference system of Fig. 1c with $n_{\rm s} < \infty$ is an interesting example which shall be discussed in the following. For small $n_{\rm s}$ one can easily obtain numerical results as a complete diagonalization of $H'$ is feasible. ![ Grand potential $\Omega$ (per lattice site) and the different contributions (per lattice site) according to Eq. (\[eq:vvv\]): $\Omega'$, ${\rm Tr} \ln (- ({\bf G}_0^{-1} - {\bf \Sigma}({\bf t}'))^{-1})$, and ${\rm Tr} \ln (-{\bf G}')$ for $U=4$, $T=0$, and $\mu=U/2$ (half-filling) as functions of $V$. Stationary points (arrows) at $V=\pm 0.519$ (metal) and at $V=0$ (insulator). $\epsilon_c=0$ and $\epsilon_a=2$. []{data-label="fig:e1"}](fig2.eps){width="65mm"} Consider the Hubbard model $$H = \sum_{ij\sigma} t_{ij} c_{i\sigma}^\dagger c_{j\sigma} + \frac{U}{2} \sum_{i\sigma} n_{i\sigma} n_{i-\sigma}$$ at temperature $T=0$ and chemical potential $\mu = U/2$. For symmetric conditions this implies half-filling. The Hamiltonian of the reference system is given by $H'= \sum_i H'(i)$ with $$\begin{aligned} H'(i) &=& \sum_{\sigma} \epsilon^{(c)}_{i} c_{i\sigma}^\dagger c_{i\sigma} + \frac{U}{2} \sum_{\sigma} n_{i\sigma} n_{i-\sigma} \nonumber \\ &+& \sum_{\sigma,k=2}^{n_{\rm s}} \epsilon^{(a)}_{i,k} a^\dagger_{ik\sigma} a_{ik\sigma} + \sum_{\sigma,k} \left( V_{i,k} c_{i\sigma}^\dagger a_{ik\sigma} + \mbox{h.c.} \right) \: . \nonumber \\ \label{eq:siam2}\end{aligned}$$ For the sake of simplicity we consider a homogeneous paramagnetic phase and the most simple case $n_{\rm s}=2$ where one is left with three independent variational parameters only, namely $V \equiv V_{i,k=2}$, $\epsilon_a \equiv \epsilon^{(a)}_{i,k=2}$, and $\epsilon_c \equiv \epsilon^{(c)}_{i}$. The computation of the different contributions to the grand potential, Eq. (\[eq:vvv\]), is straightforward: Diagonalization of $H'$ yields the ground-state energy $E_0'$ and $\Omega' = E_0' -\mu \langle N' \rangle$ as well as the excitation energies, the ground state and the excited states. The Green’s function ${\bf G}'$ and the free Green’s function ${\bf G}'_0$ can be computed from their respective Lehmann representations. The self-energy of the reference system is obtained as ${\bf \Sigma}({\bf t}') = {{\bf G}'_0}^{-1} - {{\bf G}'}^{-1}$. Since the self-energy is local, as in the DMFT, the lattice structure enters via the free ($U=0$) density of states only. Therefore, the ${\bf k}$-sum which appears in the first trace in Eq. (\[eq:vvv\]) can be performed conveniently by a one-dimensional density-of-states integration. A semi-elliptical free density of states with the band width $W=4$ is used for the calculations. This sets the energy scale for the results discussed below. Fig. \[fig:e1\] shows the grand potential $\Omega$ and the three different contributions as functions of $V$. The interaction is kept fixed at $U=W=4$. The remaining variational parameters are set to $\epsilon_c=0$ and $\epsilon_a=U/2=2$, as required by particle-hole symmetry. Each of the three contributions strongly depends on $V$ and none of them has a stationary point at a finite $V\ne 0$. Two of them show a singular behavior at $V=0$. Contrary, the resulting $\Omega$ is regular for any $V$ and shows a much weaker $V$ dependence. There are three stationary points which are indicated by the arrows. The maximum at $V=0$ corresponds to an insulating phase since ${\Sigma}(\omega)$ for $n_{\rm s}=2$ and $V=0$ is the Hubbard-I self-energy which implies a vanishing spectral density $-(1/\pi)\mbox{Im}{\bf G}(\omega+i0^+)$ at $\omega=0$. The minima at $V=\pm 0.519$ correspond to a metallic phase. $\Omega$ as well as the different contributions are symmetric functions of $V$. As ${\Sigma}(V,\omega)={\Sigma}(-V,\omega)$, however, this symmetry is trivial and does not yield an additional physical phase (see also Appendix \[sec:gofs\]). Due to the lower $\Omega$ the metallic phase is stable as compared to the insulating one. ![ $\Omega$ as a function of $\epsilon_{\rm c}$ for $V=0.519$ (metal) and $\epsilon_{\rm a}=2$. $U=4$. Inset: impurity- and bath-orbital filling, $n_{\rm c}$ and $n_{\rm a}$, as functions of $\epsilon_{\rm c}$. []{data-label="fig:e2"}](fig3.eps){width="78mm"} ![ $\Omega$ and the impurity- and bath-orbital filling, $n_{\rm c}$ and $n_{\rm a}$, as functions of $\epsilon_{\rm a}$ for $V=0.519$ and $\epsilon_{\rm c}=0$. $U=4$. []{data-label="fig:e3"}](fig4.eps){width="78mm"} The minimum at $V=0.519$ is actually a saddle point if the entire space of variational parameters is considered. This is demonstrated by Fig. \[fig:e2\] which shows $\Omega$ as a function of $\epsilon_c$ for fixed $V=0.519$ and $\epsilon_a=2$. While $\Omega(V)$ is at a minimum for $V=0.519$, $\Omega(\epsilon_c)$ is at a [maximum]{} for $\epsilon_c=0$. In the $(V,\epsilon_c)$ space one therefore encounters a saddle point. As already noted in Sec. \[sec:sft\], there is no reason to expect an extremum in general. It is worth mentioning that stationarity at $\epsilon_c = 0$ is consistent with the requirements of particle-hole symmetry. For any $\epsilon_c \ne 0$ the impurity model is asymmetric. This can be seen from the inset where the average occupations of the impurity and of the bath site are plotted as functions of $\epsilon_c$. The total particle number $\langle N' \rangle = \sum_\sigma (\langle c_{\sigma}^\dagger c_{\sigma} \rangle + \langle a^\dagger_{\sigma} a_{\sigma}\rangle) = n_{\rm c} + n_{\rm a}$ ($i$ and $k=2$ fixed) is constant: $\langle N' \rangle = 2$. With respect to the third variational parameter $\epsilon_a$, the grand potential $\Omega$ is at a maximum for $\epsilon_a = 2 = \mu$, see Fig. \[fig:e3\]. Again, this value is required by particle-hole symmetry. If $\epsilon_a$ exceeds a certain critical value (away from the stationary point), the ground-state of the reference system $H'$ no longer lies within $N'=2$ subspace but is found in the $N'=1$ or $N'=3$ subspace, respectively (see inset of Fig. \[fig:e3\]). While $\Omega'$ is continuous at the level crossing, the symmetry of the ground state changes. Consequently, there is a discontinuous change of the trial self-energy which implies a discontinuous change of $\Omega$. Consider now the [original]{} model at slightly modified parameters, e.g. $U$, $\mu$, or $T$. Clearly, the stationary point of $\Omega$ will be expected then at slightly different values of the variational parameters $V$, $\epsilon_c$, $\epsilon_a$. This implies that all physical quantities which derive from the thermodynamical potential $\Omega$ will be continuous functions of the (original) model parameters in general – irrespective of the fact that the reference system includes a finite number of degrees of freedom only: It is a typical feature of any mean-field approach that results are directly provided for the thermodynamical limit. A discontinuous jump of $\Omega$ due to a symmetry change of the ground state of the reference system (see Fig. \[fig:e3\]) usually occurs away from stationarity and is thus irrelevant physically. It is conceivable, however, that the stationary point moves to a point of discontinuity as a function a parameter of the original model. In this case the approach would generate an artifact which is a reminiscence of the finiteness of $H'$. ![ $U$ dependence of the quasi-particle weight $z = 1 / (1-d \Sigma(\omega = 0)/d\omega)$ within different approximations. BR: Brinkman-Rice (Gutzwiller) approach. [@BR70] 2-site DMFT: a non-variational two-site approach. [@Pot01] DIA, $n_{\rm s}=2$: self-energy-functional approach with a reference system $H'$ consisting of decoupled two-site impurity models. DIA, $n_{\rm s}=4$: dynamical-impurity approximation with $n_{\rm s}=4$ sites. DIA, $n_{\rm s}=\infty$: DMFT limit (circles: NRG, [@Bul99] line: ED using 8 sites). []{data-label="fig:zofu"}](fig5.eps){width="80mm"} So far we discussed the case $U=4=W$ only. As a function of $U$ the half-filled paramagnetic Hubbard model at $T=0$ is expected to undergo a transition from a metal to a Mott-Hubbard insulator. [@Geb97] This is marked by a divergence of the effective mass or, equivalently, by a vanishing quasi-particle weight $z=1/(1-d \Sigma(\omega)/d\omega\|_{\omega=0})$ as $U$ approaches a critical value $U_{\rm c}$ from below. [@GKKR96; @BR70] The result for $z(U)$ as obtained by the use of the two-site reference system (“dynamical impurity approximation”, DIA with $n_{\rm s}=2$) is shown in Fig. \[fig:zofu\]. As there are less degrees of freedom contained in $H'$, the approximation should be considered to be [inferior]{} as compared to the full DMFT the results of which are in Fig. \[fig:zofu\], too. It is remarkable that the simple $n_{\rm s}=2$-DIA (which requires the diagonalization of a dimer model only) yields $z(U)$ in an almost quantitative agreement with the full DMFT. The results of the $n_{\rm s}=2$-DIA may also be compared with those of the recently developed “linearized” or “two-site” DMFT [@BP00; @Pot01] where the Hubbard model is mapped onto the two-site SIAM (\[eq:siam2\]) by means a strongly simplified self-consistency condition. As compared to the two-site DMFT, the present self-energy-functional approach not only represents a clear conceptual improvement but also improves the actual results for $z(U)$ and for $U_{\rm c}$ (see Fig. \[fig:zofu\] and note that $U_{\rm c}=6$ within the linearized DMFT, $U_{\rm c}=5.8450$ within the DIA for $n_{\rm s}=2$, and $U_{\rm c} = 5.84$ and $U_{\rm c} = 5.88$ from numerical evaluations [@MSK+95; @Bul99] of the full DMFT). The self-energy-functional approach with reference system of Fig. 1c and small $n_{\rm s}$ actually represents a new variant of the DMFT-exact-diagonalization method (ED). [@CK94; @SRKR94] As compared to previous formulations of the ED, the convergence with respect to $n_{\rm s}$ appears to be faster: Compare the results for $n_{\rm s}=2$, $n_{\rm s}=4$ and $n_{\rm s}=\infty$ (full DMFT) in Fig. \[fig:zofu\] with those of Ref. . More important, however, there is no need for a fit procedure in the present approach; any arbitrariness in the method to find the SIAM parameters is avoided completely. Furthermore, consistent results will be obtained for any finite $n_{\rm s}$ while in the usual ED this can be expected in the DMFT limit $n_{\rm s} \mapsto \infty$ only. Conclusions and outlook {#sec:con} ======================= Concluding, the proposed self-energy-functional method is a systematic scheme for the construction of new non-perturbative and consistent approximations for extended systems of interacting fermions. For Hubbard-type lattice models with on-site interaction several relations to and generalizations of existing approaches are obtained immediately. The numerical results obtained by considering a rather simple reference system clearly demonstrate the practicability of the theory. Its generality promises that the approach may successfully be applied also in different contexts: For a Hubbard-type system including $M>1$ orbitals per site, a consistent DMFT can only be defined when using $M$ baths. There is no such necessity within the self-energy-functional approach. While clearly the optimal local approximation requires $M$ baths, any $M'<M$ will nevertheless lead to a fully consistent approximation. This represents an interesting option for numerical studies of multi-band systems. Non-local trial self-energies can be constructed by grouping the sites into identical clusters of finite size $N_{\rm s}$, switching off the inter-cluster hopping and treating the intra-cluster hopping as variational parameters. Each of the $N_{\rm s}$ sites in a cluster can be coupled to $n_{\rm s}-1$ additional bath sites. The relation of such an approach to cluster extensions of the DMFT [@HTZ+98; @KSPB01] and to the cluster perturbation theory [@GV93; @SPPL00; @SPP02] will be the subject of a forthcoming paper. Extensions and applications of the method to continuous models (inhomogeneous electron gas) and to Bose systems deserve further investigations. The work is supported by the Deutsche Forschungsgemeinschaft (Sonderforschungsbereich 290). The functional ${\bf G}[{\bf \Sigma}]$ {#sec:gofs} ====================================== For the definition of ${\bf G}[{\bf \Sigma}]$, invertibility of the functional ${\bf \Sigma}[{\bf G}]$ is required. The [local]{} invertibility of ${\bf \Sigma}[{\bf G}]$ is controlled by the Jacobian $\Gamma_{\alpha\beta'\alpha'\beta}(i\omega,i\omega') = \delta{\Sigma_{\alpha\beta}}(i\omega) / \delta{G_{\alpha'\beta'}}(i\omega')$. The two-particle self-energy [@BK61] $\bf \Gamma = \delta{\bf \Sigma} / \delta {\bf G}$ can be assumed to be non-singular in general. For a further analysis we need the following [lemma:]{} Consider the interaction ${\bf U}$ to be fixed. Then two different Green’s functions ${\bf G}_1$ and ${\bf G}_2$ must result from two different sets of one-particle parameters ${\bf t}'_1$ and ${\bf t}'_2$, respectively. The [proof]{} is straightforward: Consider the high-frequency expansion of the Green’s function $G_{\alpha\beta}(\omega)=\sum_{n=1}^\infty M_{\alpha\beta}^{(n)} \omega^{-n}$. The coefficients are given by the moments $M_{\alpha\beta}^{(n)} = \int d\omega \: \omega^n (-1/\pi) \, \mbox{Im} \, G_{\alpha\beta}(\omega+i0^+) = \langle [ {\cal L}^n c_\alpha , c_\beta ]_+\rangle$ with ${\cal L O} \equiv [{\cal O} , H]_-$. Using the symmetry $U_{\alpha\beta\gamma\delta}=U_{\beta\alpha\delta\gamma}$ one has: $$\begin{aligned} G_{\alpha\beta}(\omega) &=& \delta_{\alpha\beta} \frac{1}{\omega} \nonumber \\ &+& \Big( t'_{\alpha\beta} + \sum_{\gamma\delta} \left( U_{\alpha\gamma\beta\delta} - U_{\alpha\gamma\delta\beta} \right) \langle c^\dagger_\gamma c_\delta \rangle \Big) \frac{1}{\omega^2} \nonumber \\ &+& {\cal O}(\omega^{-3})\end{aligned}$$ Now ${\bf t}'_1 \ne {\bf t}'_2$ implies the $\omega^{-2}$ coefficients to be different because $$\langle c^\dagger_\alpha c_\beta \rangle = - \frac{1}{\pi} \int_{-\infty}^\infty d\omega \: \frac{1}{e^{\omega/T}+1} \, \mbox{Im} \, G_{\beta\alpha}(\omega+i0^+) \: .$$ Consequently, we must have ${\bf G}_1 \ne {\bf G}_2$. The lemma shows that the relation ${\bf t}' \leftrightarrow {\bf G}$ is one-to-one. Consequently, we can write ${\bf \Sigma}[{\bf G}] = {\bf \Sigma}[{\bf t}'[{\bf G}]]$ and $\bf \Gamma = \delta{\bf \Sigma} / \delta {\bf G} = \delta{\bf \Sigma} / \delta {\bf t}' \cdot \delta{\bf t}' / \delta {\bf G}$ with a non-singular Jacobian $\delta{\bf t}' / \delta {\bf G}$. Hence, a singular $\bf \Gamma = \delta{\bf \Sigma} / \delta {\bf G}$ implies a singular $\delta{\bf \Sigma} / \delta {\bf t}'$ and vice versa. However, $\delta{\bf \Sigma} / \delta {\bf t}'$ is just the “projector” in the Euler equation (\[eq:euler\]). We conclude that local non-invertibility of the functional ${\bf \Sigma}[{\bf G}]$ at ${\bf G}={\bf G}({\bf t}')$ is indicated by $\partial {\bf \Sigma}[{\bf t}'] / \partial t'_{\bf n} = 0$ with $t'_{\bf n} = {\bf t}' \cdot {\bf n}$ for a certain “direction” ${\bf n}$ in the space of hopping paramters. For such a direction, the Euler equation (\[eq:euler\]) would be satisfied [trivially]{}. Referring to the present numerical results, one can state that generally the projector $\delta{\bf \Sigma} / \delta {\bf t}'$ is non-singular in fact, as has been expected. There is one exception, however, namely points in the hopping-parameter space where one or more bath sites are [decoupled]{} from the rest of the system (vanishing hybridization $V$). Here the one-particle energy of a decoupled bath site can be varied without changing the trial self-energy. Even for this case, however, there are no formal difficulties with the inverse functional ${\bf G}[{\bf \Sigma}]$: To ensure the local invertibility of ${\bf \Sigma}[{\bf G}]$, one simply has to restrict the space of variational parameters ${\bf t}'$ by excluding the one-particle energies of the decoupled bath sites, i.e. one has to focus on the physically relevant parameters. This implies a respective restriction of the space of ${\bf t}'$-representable Green’s functions ${\bf G}({\bf t}')$ and ensures the local invertibility of ${\bf \Sigma}[{\bf G}]$ on the restricted domain. Similarly, a restriction of the ${\bf t}'$ space becomes necessary to ensure the local invertibility of ${\bf \Sigma}[{\bf G}]$ in case of a system where the self-energy is trivial (as e.g. for a model of spinless fermions with nearest-neighbor Coulomb interaction in the limit of infinite spatial dimensions where the self-energy is given by the Hartree term, cf. Ref. ). Finally, it should be mentioned that generally the functional ${\bf \Sigma}[{\bf G}]$ cannot be inverted [globally]{}. Consider, for example, the Hubbard model on the infinite-dimensional hypercubic lattice with nearest-neighbor hopping $t$ at half-filling. Due to manifest particle-hole symmetry, a sign change of the hopping $t \mapsto -t$ leaves the (local) self-energy invariant but transforms (the non-local elements of) the Green’s function ${\bf G}$. We conclude that ${\bf G}[{\bf \Sigma}]$ in general cannot be defined uniquely. Due to this non-uniqueness, and also due to non-linearity, there may be more than a single solution of Eq. (\[eq:stat\]). However, this does not cause any problem since for any $\bf \Sigma$ satisfying (\[eq:stat\]) we have: $$\begin{aligned} {\bf G}[{\bf \Sigma}] &=& ({\bf G}_0^{-1} - {\bf \Sigma})^{-1} \nonumber \\ \Rightarrow \qquad {\bf \Sigma} &=& {\bf \Sigma}({\bf G}_0^{-1} - {\bf \Sigma})^{-1} \nonumber \\ \Rightarrow \qquad {\bf \Sigma} &=& {\bf \Sigma}[{\bf G}] \quad \mbox{and} \quad {\bf G} = ({\bf G}_0^{-1} - {\bf \Sigma})^{-1} \: .\end{aligned}$$ This means that ${\bf \Sigma}$ is given by the (formal) sum of all skeleton diagrams built up by a propagator ${\bf G}$ which is constructed via the Dyson equation from the same ${\bf \Sigma}$ in turn. Hence, any stationary point should be regarded as a physically meaningful solution. Among different physical solutions (corresponding e.g. to different phases) the minimum grand potential selects the stable one. [10]{} bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} Present address: Institut für Theoretische Physik und Astrophysik, Universität Würzburg, Am Hubland, 97074 Würzburg, Germany , **, (). , (, , ). , , , eds., (, , ). , , (). , , (). , , , , (). , , , , (). , , , , (). , , (). , , (). , , (). , , , , (). , , (). , , (). , , , , , (). , , (). , , , , , (). , , , (, , ). , **, (). J. Hubbard, Proc. R. Soc. London A 276, 238 (1963); M. C. Gutzwiller, Phys. Rev. Lett. 10, 159 (1963); J. Kanamori, Prog. Theor. Phys. (Kyoto) 30, 275 (1963). , , (). , , (). , , (). , , (). , , (). , , (). , , (). , , (). , , , , , , (). , , , , , , (). , , , , , (). , , (). , , , , (). , , , , (). , **, ().	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper studies $n$-person simultaneous-move games with linear best response function, where individuals interact within a given network structure. This class of games have been used to model various settings, such as, public goods, belief formation, peer effects, and oligopoly. The purpose of this paper is to study the effect of the network structure on Nash equilibrium outcomes of this class of games. Bramoullé et al. derived conditions for uniqueness and stability of a Nash equilibrium in terms of the smallest eigenvalue of the adjacency matrix representing the network of interactions. Motivated by this result, we study how local structural properties of the network of interactions affect this eigenvalue, influencing game equilibria. In particular, we use algebraic graph theory and convex optimization to derive new bounds on the smallest eigenvalue in terms of the distribution of degrees, cycles, and other relevant substructures. We illustrate our results with numerical simulations involving online social networks.' author: - 'Victor M. Preciado, Jaelynn Oh and Ali Jadbabaie [^1] [^2]' title: '[Analysis of Equilibria and Strategic Interaction in Complex Networks]{}' --- Introduction ============ In most social and economic settings, individuals do not interact uniformly with the rest of a society. Instead, they influence each other according to a structured network of interactions. This network can represent friendships in a social network, transactions among firms in a market, or communication links in a process of belief formation. In this context, an interesting question is to study how the network structure affects the outcome of the interactions of the agents. With this purpose, one can model strategic interactions in a networked society as a multi-player simultaneous-move game. In particular, we focus our attention on the broad class of games with linear best response function [@FT91]. This class of games have been used to model various settings such as belief formation [@AP07], peer effects [@BCZ06], and public goods [@BK07]. In order to analyze the influence of the network structure on the game outcome, we use two recent results by Bramoullé et al. [@BKA10] and Ballester et al. [@BC07] relating the Nash equilibria of the game with the largest and smallest eigenvalues of the adjacency matrix of interactions. For example, in [@BKA10], the authors show that uniqueness and stability of a Nash equilibrium in games with linear best responses can be determined by the smallest eigenvalue of the network. In [@BKA10], it was also illustrated how the smallest eigenvalue of the adjacency matrix determines the capacity of the network to absorb perturbations on the actions of the agents. Motivated by the results of Bramoullé et al. [@BKA10], we study how local structural properties of the network affect the smallest eigenvalue of the adjacency matrix of interactions, affecting game equilibria. Therefore, our results build a bridge between structural properties of a network of interactions and the outcome of games with linear best responses. In particular, we use algebraic graph theory and convex optimization to derive bounds on the smallest eigenvalue of the adjacency matrix in terms of the distribution of degrees, cycles, and other important substructures. As we illustrate with numerical simulations in online social networks, these bounds can be used to estimate the effect of structural perturbations on the smallest eigenvalue. The paper is organized as follows. In the next subsection, we review graph-theoretical terminology needed in our derivations. In Section [Strategic Section]{}, we review the relationship between the equilibria of games with linear best responses in a network and the smallest eigenvalue of the adjacency matrix of interactions. In Section \[Spectral Section\], we use algebraic graph theory to derive closed-form expressions for the so-called spectral moments of a network in terms of local structural features. In Section \[Optimal Bounds\], we use convex optimization to derive optimal bounds on the smallest (and largest) eigenvalue of the interaction network from these moments. Our bounds help us to understand how structural properties of a network impact the stability properties of the Nash equilibria in the game. We illustrate our results with numerical simulations in real online social networks in Section \[Simulations\]. Notation -------- Let $\mathcal{G}=\left( \mathcal{V},\mathcal{E}\right) $ denote an undirected graph with $n$ nodes, $e$ edges, and no self-loops[^3]. We denote by $\mathcal{V}\left( \mathcal{G}\right) =\left\{ v_{1},\dots ,v_{n}\right\} $ the set of nodes and by $\mathcal{E}\left( \mathcal{G}% \right) \subseteq \mathcal{V}\left( \mathcal{G}\right) \times \mathcal{V}% \left( \mathcal{G}\right) $ the set of undirected edges of $\mathcal{G}$. If $\left\{ v_{i},v_{j}\right\} \in \mathcal{E}\left( \mathcal{G}\right) $ we call nodes $v_{i}$ and $v_{j}$ adjacent (or neighbors), which we denote by $v_{i}\sim v_{j}$ and define the set of neighbors of $v_{i}$ as $% \mathcal{N}_{i}=\{w\in \mathcal{V}\left( \mathcal{G}\right) :\left\{ v_{i},w\right\} \in \mathcal{E}\left( \mathcal{G}\right) \}$. The number of neighbors of $v_{i}$ is called the degree of the node, denoted by $% d_{i}$. We define a walk of length $k$ from $v_{0}$ to $v_{k}$ to be an ordered sequence of nodes $\left( v_{0},v_{1},...,v_{k}\right) $ such that $v_{i}\sim v_{i+1}$ for $i=0,1,...,k-1$. If $v_{0}=v_{k}$, then the walk is closed. A closed walk with no repeated nodes (with the exception of the first and last nodes) is called a cycle. For example, triangles, quadrangles and pentagons are cycles of length three, four, and five, respectively. Graphs can be algebraically represented via matrices. The adjacency matrix of an undirected graph $\mathcal{G}$, denoted by $A_{\mathcal{G}}=[a_{ij}]$, is an $n\times n$ symmetric matrix defined entry-wise as $a_{ij}=1$ if nodes $v_{i}$ and $v_{j}$ are adjacent, and $a_{ij}=0$ otherwise[^4]. The eigenvalues of $A_{\mathcal{% G}}$, denoted by $\lambda _{1}\geq \lambda _{2}\geq \ldots \geq \lambda _{n}$, play a key role in our paper. The spectral radius of $A_{\mathcal{G}}$, denoted by $\rho \left( A_{\mathcal{G}}\right) $, is the maximum among the magnitudes of its eigenvalues. Since $A_{\mathcal{G}}$ is a symmetric matrix with nonnegative entries, all its eigenvalues are real and the spectral radius is equal to the largest eigenvalue, $\lambda _{1}$. We define the $k$-th spectral moment of the adjacency matrix $A_{\mathcal{G}}$ as:$$m_{k}\left( A_{\mathcal{G}}\right) =\frac{1}{n}\sum_{i=1}^{n}\lambda _{i}^{k}. \label{Spectral Moment}$$As we shall show in Section \[Spectral Section\], there is a direct connection between the spectral moments and the presence of certain substructures in the graph, such as cycles of length $k$. Strategic Interactions in Networks\[Strategic Section\] ======================================================= In this section we present the game-theoretical model of strategic interactions considered in this paper and present interesting connections between the Nash equilibria and the eigenvalues of the adjacency matrix of the network. The Model --------- We represent the network of influences using a simple graph $\mathcal{G}$. Let $\mathcal{N}=\{1,\cdots ,n\}$ denote the set of $n$ players located at each node of the graph $\mathcal{G}$. We denote by $x_{i}\in \lbrack 0,\infty )$ the action chosen by agent $i$, and by $\mathbf{x}$ the vector that represents the joint actions for all agents. We denote by $\mathbf{x}% _{-i}$ the vector of actions of all players excluding player $i$. As mentioned before, players interact according to a network of influences that we describe using its adjacency matrix $A_{\mathcal{G}}$. The interactions are assumed to be symmetric, $a_{ij}=a_{ji}$, and we do not allow self-loops, $a_{ii}=0$. The payoff function for agent $i$ is given by:$$U_{i}(x_{i},\mathbf{x}_{-i};\delta ,A_{\mathcal{G}})$$where $\delta \in \mathbb{R}$ is a parameter that can be tuned to change the influence of neighboring nodes on each player’s action. \[Game Description\]Games with Linear Best Response Functions ------------------------------------------------------------- We study a class of games whose best response functions take a linear form. One well known example of this class of games is the differentiated-product Cournot oligopoly with linear inverse demand and constant marginal cost with payoff function defined as [@BKA10]:$$U_{i}(x_{i},\mathbf{x}_{-i};\delta ,A_{\mathcal{G}})=x_{i}\left( a-b(x_{i}+2\delta \sum_{j=1}^{n}a_{ij}x_{j})\right) -dx_{i}, \label{Game Payoffs}$$where $d$ is the constant marginal cost, and $x_{i}$ represents the amount produced by agent $i$ in the oligopoly. Here, the inverse demand for agent $i $ is given by $P_{i}(x_{i},\mathbf{x}_{-i};\delta ,A_{\mathcal{G}% })=a-b(x_{i}+2\delta \sum_{j=1}^{n}a_{ij}x_{j})$. One can prove that the best response function for this type of games yield the form [@BKA10]:$$f_{i}(\mathbf{x},\delta ,A_{\mathcal{G}})=\max (0,\bar{x}_{i}-\delta \sum_{j=1}^{n}a_{ij}x_{j}), \label{Best Reply Function}$$where $\bar{x}_{i}$ is the action that agent $i$ would take in isolation, i.e., $\bar{x}_{i}\in \arg \max_{x_{i}}U_{i}(x_{i},\mathbf{x}_{-i};\delta ,A_{\mathcal{G}})$ with $a_{ij}=0$, for all $j$. Without loss of generality, one can normalize $\bar{x}_{i}\equiv 1$ for all $i$, so that $% f_{i}(\mathbf{x})\in \lbrack 0,1]$. Then, a Nash equilibrium for this game is a vector $\mathbf{x}\in \lbrack 0,1]^{n}$ that satisfies $x_{i}=f_{i}(% \mathbf{x},\delta ,A_{\mathcal{G}})$, for all agents $i\in \mathcal{N}$, simultaneously. In what follows, we briefly describe a strategy to compute the complete set of Nash equilibria for $\delta \in \lbrack 0,1]$. Complete Set of Nash Equilibria ------------------------------- Using the best response function in (\[Best Reply Function\]), one can determine the entire set of equilibria by simultaneously solving for the best response of each player. In [@BKA10], an algorithm that finds the full set of Nash equilibria in exponential time is proposed. For a vector $% \mathbf{x}$, let $S$ denote the set of active agents, i.e., $S=\{i:x_{i}>0\}$. Let $\mathbf{x}_{S}$ denote the vector of actions of the agents in $S$. The set of active players induce a subgraph $\mathcal{G}_{S}\subseteq \mathcal{G}$, with node-set $S\subseteq \mathcal{V}\left( \mathcal{G}\right) $ and a set of edges $\mathcal{E}_{S}\subseteq \mathcal{E}\left( \mathcal{G}% \right) $ connecting active agents. We denote by $\mathcal{G}_{\mathcal{N}% \backslash S,S}$ the subgraph of $\mathcal{G}$ whose edges connect active agents in $S$ to inactive agents in $\mathcal{N}\backslash S$. The adjacency matrices of $\mathcal{G}_{S}$ and $\mathcal{G}_{\mathcal{N}\backslash S,S}$ are denoted by $A_{S}\in \mathbb{R}^{\left\vert S\right\vert \times \left\vert S\right\vert }$ and $A_{\mathcal{N}\backslash S,S}\in \mathbb{R}% ^{n-\left\vert S\right\vert \times \left\vert S\right\vert }$, respectively. Then, one can show the following [@BKA10]: \[Nash Conditions\]A profile $\mathbf{x}$ with active agents $S$ is a Nash equilibrium if and only if: $$\begin{aligned} (I_{\left\vert S\right\vert }+\delta A_{S})\mathbf{x}_{S} &=&\mathbf{1}% _{\left\vert S\right\vert }\mathbf{,}\text{ and} \\ \delta A_{\mathcal{N}\backslash S,S}~\mathbf{x}_{S} &\geq &\mathbf{1}% _{n-\left\vert S\right\vert }\mathbf{,}\end{aligned}$$where $I_{p}$ is the $p\times p$ identity matrix and $\mathbf{1}_{q}$ is the $q$-dimensional vector of ones. Thus, in order to determine the complete set of all Nash equilibria, one can check the conditions in Proposition \[Nash Conditions\] for each one of the $2^{n}$ possibilities of $S$. For each possible $S$, these conditions can be checked by computing $\mathbf{x}_{S}=(I_{\left\vert S\right\vert }+\delta A_{S})^{-1}\mathbf{1}_{\left\vert S\right\vert }$, and checking whether $\delta A_{\mathcal{N}\backslash S,S}~\mathbf{x}% _{S}\geq \mathbf{1}_{n-\left\vert S\right\vert }$. If the last inequality holds, then $\mathbf{x}_{S}$ is an equilibrium outcome. Note that using this approach to compute the set of equilibria runs in exponential time. However, we can analyze some properties of the Nash equilibria, such as uniqueness and stability, by looking into the eigenvalues of the adjacency matrix. The Shape of Nash Equilibria ---------------------------- In order to relate the equilibrium outcomes of the game to the network structure, the authors in [@BKA10] defined the following potential function:$$\varphi (\mathbf{x};\delta ,A_{\mathcal{G}})\triangleq \sum_{i=1}^{n}\left( x_{i}-\frac{1}{2}x_{i}^{2}\right) -\frac{1}{2}\delta \sum_{i=1}^{n}\sum_{j=1}^{n}a_{ij}x_{i}x_{j}.$$Then, they proved, using Kuhn-Tucker conditions, that the set of Nash equilibria coincides with the critical points of the following optimization problem:$$\begin{array}{lll} (\text{P}) & \max_{\mathbf{x}} & \varphi (\mathbf{x};\delta ,A_{\mathcal{G}}) \\ & s.t. & x_{i}\geq 0,\text{ for all }i,% \end{array}%$$for a given network structure $A_{\mathcal{G}}$ and a parameter $\delta $. Eigenvalues and Nash Equilibria ------------------------------- We can find several results in the literature providing sufficient conditions for the existence of a unique Nash equilibrium in games with linear best response functions in terms of the eigenvalues of the network of influences. We enumerate below some sufficient conditions that are related with our work: \[Kranton condition\]Consider the class of games with linear best response functions described in Section \[Game Description\]. For these games, we have the following sufficient conditions for the existence of a unique Nash equilibrium: (i) : $\delta <-1/\lambda _{n}(A_{\mathcal{G}})$, (Bramoullé et al., [@BKA10]). (ii) : $\delta <1/\rho \left( A_{\mathcal{G}}\right) $, (Ballester et al., [@BC07]). We can compare the set of spectral conditions in Proposition \[Kranton condition\] using the following inequalities [@BKA10]: For any simple graph $\mathcal{G}$, we have that $-1/\lambda _{n}(A_{% \mathcal{G}})\geq 1/\rho \left( A_{\mathcal{G}}\right) $, where this inequality is strict when no component of $\mathcal{G}$ is bipartite. \[Active Remark\]Hence, Condition (i) in Proposition \[Kranton condition\] provides the best sufficient condition for the uniqueness of Nash equilibrium in these games. Furthermore, one can also prove that under Condition (i) or (ii), all players are active at the equilibrium point, i.e., $S=\mathcal{N}$ [@BKA10]. In Section \[Optimal Bounds\], we shall derive upper bounds on $\lambda _{n}\left( A_{\mathcal{G}}\right) $ in terms of structural properties of the network. These bounds, in combination with Condition (i) in Proposition \[Kranton condition\], will allow us to derive sufficient conditions for the existence of a Nash equilibrium in terms of structural properties of the network. The Stability of Nash Equilibria -------------------------------- We present conditions for stability of a Nash equilibrium in terms of $% \lambda _{n}(A_{S})$. A Nash equilibrium $\mathbf{x}$ is asymptotically stable when the system of differential equations:$$\begin{array}{c} \dot{x}_{1}=h_{1}(\mathbf{x})=f_{1}(\mathbf{x};\delta ,A_{\mathcal{G}% })-x_{1}, \\ \vdots \\ \dot{x}_{n}=h_{n}(\mathbf{x})=f_{n}(\mathbf{x};\delta ,A_{\mathcal{G}% })-x_{n},% \end{array}%$$is locally asymptotically stable around $\mathbf{x}$. One can prove the following necessary and sufficient condition for an equilibrium $\mathbf{x}$ to be asymptotically stable [@BKA10]: An equilibrium profile $\mathbf{x}$ is asymptotically stable if and only if $% \delta <-1/\lambda _{n}(A_{S})$ and $\delta \sum_{j=1}^{n}a_{ij}x_{j}>1,$ for all inactive agents $i\in \mathcal{N}\backslash S$. From the above lemma and Remark \[Active Remark\], we conclude that if $% \delta <-1/\lambda _{n}(A_{\mathcal{G}})$, there is a unique Nash equilibrium and it is asymptotically stable. These results show the close connection between the smallest and the largest eigenvalues of the adjacency matrix of interactions and the outcome of games with linear best response functions in a network. On the other hand, the results in this section are applicable if we are able to compute the eigenvalues of $A_{\mathcal{G}}$. For many large-scale complex networks, the structure of the network can be very intricate [@S01]-[@Pre08] —in many cases not even known exactly— and an explicit eigenvalue decomposition can be very challenging to compute, if not impossible. In many practical settings, instead of having access to the complete network topology, we have access to local neighborhoods of the network structure around a set of nodes. In this context, it is important to understand the impact of local structural information on the eigenvalues of the adjacency matrix. In the rest of the paper, we propose a novel methodology to compute optimal bounds on relevant eigenvalues of $A_{\mathcal{G}}$ from local information regarding the network structure. Our results allow us to study the role of local structural information in the outcome of games with linear best response functions. Spectral Analysis of the Interaction Matrix\[Spectral Section\] =============================================================== We study the relationship between a network’s local structural properties and the smallest eigenvalue of its adjacency matrix. Algebraic graph theory provides us with tools to relate the eigenvalues of a network with its structural properties. Particularly useful is the following well-known result relating the $k$-th spectral moment of $A_{\mathcal{G}}$ with the number of closed walks of length $k$ in $\mathcal{G}$ [@Big93]: \[Moments from Walks\]Let $\mathcal{G}$ be a simple graph. The $k$-th spectral moment of the adjacency matrix of $\mathcal{G}$ can be written as$$m_{k}(A_{\mathcal{G}})=\frac{1}{n}\sum_{i=1}^{n}\lambda_{i}^{k}=\frac{1}{n}% \left\vert \Psi_{\mathcal{G}}^{\left( k\right) }\right\vert , \label{Moments as Walks in Graph}$$ where $\Psi_{\mathcal{G}}^{\left( k\right) }$ is the set of all closed walks of length $k$ in $\mathcal{G}$. [^5] From (\[Moments as Walks in Graph\]), we can easily compute the first three spectral moments of $A_{\mathcal{G}}$ in terms of the number of nodes, edges and triangles as follows [@Big93]: \[One corollary\]Let $\mathcal{G}$ be a simple graph with adjacency matrix $A_{\mathcal{G}}$. Denote by $n$, $e$ and $\Delta $ the number of nodes, edges and triangles in $\mathcal{G}$, respectively. Then,$$\begin{aligned} m_{1}(A_{\mathcal{G}})& =0, \label{Moments as Averages} \\ m_{2}(A_{\mathcal{G}})& =2e/n, \notag \\ m_{3}(A_{\mathcal{G}})& =6\,\Delta /n. \notag\end{aligned}$$ Notice that the coefficients 2 (resp. 6) in the above expressions corresponds to the number of closed walks of length 2 (resp. 3) enabled by the presence of an edge (resp. triangle). Similar expressions can be derived for higher-order spectral moments, although a more elaborated combinatorial analysis in required. In our case, we are interested in the following expressions, derived in [PJ10]{}, for the first five spectral moments of $\mathcal{G}$: \[Lemma 4th metrics\]Let $\mathcal{G}$ be a simple graph. Denote by $e$, $\Delta $, $Q$ and $\Pi $ the total number of edges, triangles, quadrangles and pentagons in $\mathcal{G}$, respectively. Define $W_{2}=% \sum_{i=1}^{n}d_{i}^{2}$ and $\mathcal{C}_{dt}=\sum_{i}d_{i}t_{i}$, where $% t_{i}$ is the number of triangles touching node $i$. Then,$$\begin{aligned} m_{4}\left( A_{\mathcal{G}}\right) &=&\frac{1}{n}\left[ 8Q+2W_{2}-e\right] , \label{Fourth Moment Subgraphs} \\ m_{5}\left( A_{\mathcal{G}}\right) &=&\frac{1}{n}\left[ 10\Pi +10\mathcal{C}% _{dt}-30\Delta \right] . \label{Fifth Moment Subgraphs}\end{aligned}$$ Observe how, as we increase the order of the moments, more complicated structural features appear in the expressions. In the following example, we illustrate how to use our expressions to compute the spectral moments of an online social network from empirical structural data. ![image](Fig7C.jpg){width="100.00000%"} In this example, we study a subgraph of Facebook obtained by exploring the online social network around a particular node, as follows. From a particular starting node, we crawl the network (using a breadth-first search) until we discover the set of all nodes that are within a radius $2$ from the starting node. Using this set of nodes, and their interconnections (friendships), we construct a social subgraph $\mathcal{F}$ that has $2,404$ nodes and $22,786$ edges. Using this real dataset, we compute the degrees $% d_{i}$, the number of triangles $t_{i}$, quadrangles $q_{i}$, and pentagons $% p_{i}$ touching each node $i\in \mathcal{V}\left( \mathcal{F}\right) $. In Fig. 1, we plot the distributions of degrees and triangles, as well as a scatter plot of $t_{i}$ versus $d_{i}$ (where each point has coordinates $% \left( d_{i},t_{i}\right) $, in log-log scale, for all $i\in \mathcal{V}% \left( \mathcal{F}\right) $). We can aggregate those quantities that are relevant to compute the spectral moments to obtain the following numerical values:$$\begin{array}{ll} e/n=\sum d_{i}/2n=9.478, & \Delta /n=\sum t_{i}/3n=28.15, \\ Q/n=\sum q_{i}/4n=825.3, & \Pi /n=\sum p_{i}/5n=31,794, \\ W_{2}/n=\sum d_{i}^{2}/n=1,318, & \mathcal{C}_{dt}/n=\sum d_{i}t_{i}/n=8,520.% \end{array}%$$Hence, using Corollary \[One corollary\] and Lemma \[Lemma 4th metrics\], we obtain the following values for the spectral moments: $m_{1}\left( A_{% \mathcal{F}}\right) =0,$ $m_{2}\left( A_{\mathcal{F}}\right) =18.95,$ $% m_{3}\left( A_{\mathcal{F}}\right) =168.9,$ $m_{4}\left( A_{\mathcal{F}% }\right) =9,230,$ and $m_{5}\left( A_{\mathcal{F}}\right) =402,310.$ In this section, we have derived expressions to compute the first five spectral moment of $A_{\mathcal{G}}$ from network structural properties. In the next section, we use semidefinite programming to extract information regarding eigenvalues of interest from a sequence of spectral moments. \[Optimal Bounds\]Optimal Spectral Bounds from Spectral Moments =============================================================== Here, we introduce an approach to derive an upper bound on the smallest eigenvalues of $A_{\mathcal{G}}$ from its sequence of spectral moments[^6]. Since we have expressions for the spectral moments in terms of local structural properties, our bounds relate the eigenvalues of a network with these properties. There is a large literature studying the relationship between structural and spectral properties of graphs (see [CDS80]{},[@DK04], and references therein, for an extensive list of spectral results). For many real-world networks, there is a particular set of structural properties that play a key role in the network’s functionality. For example, it is well-known that social networks contain a large number of triangles (and other cycles). Hence, it would be useful to have spectral bounds where these structural features are jointly represented. In this section, we derive new upper bounds on the smallest eigenvalue of the adjacency matrix in terms of the structural properties involved in (\[Moments as Averages\]), (\[Fourth Moment Subgraphs\]) and (\[Fifth Moment Subgraphs\]). Our results can be easily extended to derive lower bounds on the spectral radius of the adjacency matrix, although this bound is not of relevance in our analysis of games with linear best responses. Now, we derive bounds on the smallest eigenvalue of the adjacency matrix in terms of relevant structural properties by adapting the optimization framework proposed in [@Las11]. We first need to introduce a probabilistic interpretation of a network eigenvalue spectrum and its spectral moments. For a simple graph $\mathcal{G}$, we define its spectral density as:$$\mu _{\mathcal{G}}\left( x\right) =\frac{1}{n}\sum_{i=1}^{n}\delta \left( x-\lambda _{i}\right) , \label{Spectral Measure}$$where $\delta \left( \bullet \right) $ is the Dirac delta function and $% \left\{ \lambda _{i}\right\} _{i=1}^{n}$ is the set of (real) eigenvalues of the (symmetric) adjacency matrix $A_{\mathcal{G}}$. Consider a random variable $X$ with probability density $\mu _{\mathcal{G}}$. The moments of $% X\sim \mu _{\mathcal{G}}$ are equal to the spectral moments of $A_{\mathcal{G% }}$, i.e.,$$\begin{aligned} \mathbb{E}_{\mu _{\mathcal{G}}}\left( X^{k}\right) & =\int_{\mathbb{R}% }x^{k}\mu _{\mathcal{G}}\left( x\right) dx \\ & =\frac{1}{n}\sum_{i=1}^{n}\int_{\mathbb{R}}x^{k}\delta \left( x-\lambda _{i}\right) dx \\ & =\frac{1}{n}\sum_{i=1}^{n}\lambda _{i}^{k}=m_{k}\left( A_{\mathcal{G}% }\right) ,\end{aligned}$$for all $k\geq 0$. In [@Las11], Lasserre proposed a technique to compute the smallest interval $\left[ a,b\right] $ containing the support[^7] of a positive Borel measure $\mu $ from its complete sequence of moments $\left( m_{r}\right) _{r\geq 0}$. In our spectral problem, the positive Borel measure under consideration is the spectral density $\mu _{\mathcal{G}}\left( x\right) $, defined in ([Spectral Measure]{}). Hence, in the context of our problem, the sequence of moments $\left( m_{r}\right) _{r\geq 0}$ is equal to $\left( m_{r}\left( A_{% \mathcal{G}}\right) \right) _{r\geq 0}$, and the smallest interval $\left[ a,b\right] $ containing the support of $\mu _{\mathcal{G}}\left( x\right) $ is equal to $\left[ \lambda _{n},\lambda _{1}\right] $, by the definition in (\[Spectral Measure\]). Lasserre also proposed in [@Las11] a numerical scheme to compute tight bounds on the values of $a$ and $b$ when a truncated sequence of moments $\left( m_{r}\right) _{0\leq r\leq k}$ is known. This numerical scheme involves a series of semidefinite programs (SDP) in one variable. As we show below, at step $s$ of this series of SDP’s, we are given a sequence of moments $\left( m_{1},...,m_{2s+1}\right) $ and solve two SDP’s whose solution provides an inner approximation $\left[ \alpha _{s},\beta _{s}% \right] \subseteq \left[ a,b\right] $. In our case, since we have expressions for the first five spectral moments, $\left( m_{1}\left( A_{% \mathcal{G}}\right) ,...,m_{5}\left( A_{\mathcal{G}}\right) \right) $, we can solve the first two steps of this series of SDP’s to find inner approximations $\left[ \alpha _{s},\beta _{s}\right] \subseteq \left[ \lambda _{n},\lambda _{1}\right] $. In other words, the solution to the SDP’s provide us with the bounds $\alpha _{s}\geq \lambda _{n}$ and $\beta _{s}\leq \lambda _{1}$. In order to formulate the series of SDP’s proposed in [@Las11], we need to define the so-called localizing matrix of our problem [@LasBOOK]. Given a sequence of moments, $\mathbf{m}^{\left( 2s+1\right) }=\left( m_{1},...,m_{2s+1}\right) $, our localizing matrix is a Hankel matrix defined as:$$H_{s}\left( c\right) \triangleq R_{2s+1}-c~R_{2s}\text{,} \label{Localizing matrix}$$where $R_{2s}$ and $R_{2s+1}$ are the Hankel matrices of moments defined as$$\begin{aligned} R_{2s}& =\left[ \begin{array}{cccc} 1 & m_{1} & \cdots & m_{s} \\ m_{1} & m_{2} & \cdots & m_{s+1} \\ \vdots & \vdots & \ddots & \vdots \\ m_{s} & m_{s+1} & \cdots & m_{2s}% \end{array}% \right] , \label{Hankel matrices} \\ R_{2s+1}& =\left[ \begin{array}{cccc} m_{1} & m_{2} & \cdots & m_{s+1} \\ m_{2} & m_{3} & \cdots & m_{s+2} \\ \vdots & \vdots & \ddots & \vdots \\ m_{s+1} & m_{s+2} & \cdots & m_{2s+1}% \end{array}% \right] . \notag\end{aligned}$$Hence, for a given sequence of moments, the entries of $H_{s}\left( c\right) $ depend affinely on the variable $c$. We can compute $\alpha _{s}$ and $% \beta _{s}$ as the solution to the following semidefinite programs [Las11]{}: \[Lasserre bound\]Let $\mathbf{m}^{\left( 2s+1\right) }=\left( m_{1},...,m_{2s+1}\right) $ be the truncated sequence of moments of a positive Borel measure $\mu $. Then,$$\begin{aligned} a& \leq \alpha _{s}\triangleq \max_{\alpha }\left\{ \alpha :H_{s}\left( \alpha \right) \succcurlyeq 0\right\} , \label{Bound min eigenval} \\ b& \geq \beta _{s}\triangleq \min_{\beta }\left\{ \beta :-H_{s}\left( \beta \right) \succcurlyeq 0\right\} , \label{Bound max eigenval}\end{aligned}$$for $\left[ a,b\right] $ being the smallest interval containing $supp(\mu )$. Observe that $\alpha _{s}$ and $\beta _{s}$ are the solutions to two SDP’s in one variable, since the constraint $H_{s}\left( \alpha \right) \succcurlyeq 0$ (resp. $-H_{s}\left( \beta \right) \succcurlyeq 0$) indicates that the matrix $H_{s}\left( \alpha \right) $ (resp. $-H_{s}\left( \beta \right) $) is positive semidefinite and this matrix has affine entries with respect to $\alpha $ (resp. $\beta $). Hence, they can be efficiently computed using standard optimization software (for example, CVX [@CVX]). As we increase $s$ in Proposition \[Lasserre bound\], more moments are involved in the SDP’s, and the resulting bounds become tighter, i.e., $% \alpha _{s+1}\leq \alpha _{s}$ and $\beta _{s+1}\geq \beta _{s}$. In the context of our spectral analysis, the Borel measure in Proposition \[Lasserre bound\] corresponds to the spectral density of a graph $% \mathcal{G}$, and the smallest interval $\left[ a,b\right] $ corresponds to $% \left[ \lambda _{n}\left( A_{\mathcal{G}}\right) ,\rho \left( A_{\mathcal{G}% }\right) \right] $. Thus, Proposition \[Lasserre bound\] provides an efficient numerical scheme to compute the bounds $\alpha _{s}\geq \lambda _{n}\left( A_{\mathcal{G}}\right) $ and $\beta _{s}\leq \rho \left( A_{% \mathcal{G}}\right) $. When we are given a sequence of five spectral moments, we can solve the SDP’s in (\[Bound min eigenval\]) and ([Bound max eigenval]{}) analytically for $s=2$. In this case, the localizing matrix is:$$H_{2}\left( c\right) =\left[ \begin{array}{ccc} m_{1}-c & m_{2}-cm_{1} & m_{3}-cm_{2} \\ m_{2}-cm_{1} & m_{3}-cm_{2} & m_{4}-cm_{3} \\ m_{3}-cm_{2} & m_{4}-cm_{3} & m_{5}-cm_{4}% \end{array}% \right] . \label{H2c}$$As we proved in Section \[Spectral Section\], the spectral moments in the localizing matrix depend on the number of nodes, edges, cycles of length 3 to 5, the sum-of-squares of degrees $W_{2}$, and the degree-triangle correlation $\mathcal{C}_{dt}$. Furthermore, for $s=2$, the optimal values $\alpha _{2}$ and $\beta _{2}$ can be analytically computed, as follows. First, note that $% H_{2}(c)\succcurlyeq 0$ (resp. $-H_{2}(c)\succcurlyeq 0$) if and only if all the eigenvalues of $H_{2}$ are nonnegative (resp. nonpositive). For a given sequence of five moments, the characteristic polynomial of $H_{2}\left( c\right) $ can be written as$$\phi _{2}\left( \lambda \right) \triangleq \det \left( \lambda I-H_{2}(c)\right) =\lambda ^{3}+p_{1}\left( c\right) \lambda ^{2}+p_{2}\left( c\right) \lambda +p_{3}\left( c\right) ,$$where $p_{j}\left( c\right) $ is a polynomial of degree $j$ in the variable $% c$ (with coefficients depending on the moments). Thus, by Descartes’ rule, all the eigenvalues of $H_{2}\left( c\right) $ are nonpositive if and only if $p_{j}\left( c\right) \geq 0$, for $j=1,2,$ and $3$. Similarly, all the eigenvalues are nonnegative if and only if $p_{2}\left( c\right) \geq 0$ and $p_{1}\left( c\right) ,p_{3}\left( c\right) \leq 0$. In fact, one can prove that the optimal value of $\alpha _{2}$ and $\beta _{2}$ in (\[Bound max eigenval\]) can be computed as the smallest and the largest roots of $\det H_{2}\left( c\right) =0$, which yields a third degree polynomial in the variable $c$ [@Las11]. There are closed-form expressions for the roots of this polynomial (for example, Cardano’s formula [@AS65]), although the resulting expressions for the roots are rather complicated. In this subsection, we have presented a convex optimization framework to compute optimal bounds on the maximum and the minimum eigenvalues of a graph $\mathcal{G}$ from a truncated sequence of its spectral moments. Since we have expressions for spectral moments in terms of local structural properties, these bounds relate the eigenvalues of a graph with its structural properties. \[Simulations\]Numerical Simulations ==================================== As we illustrated in Section \[Strategic Section\], there is a close connection between the largest and the smallest eigenvalues of a network and the outcome of a game with linear best response functions. In this section, we use our bounds on the support of the eigenvalue spectrum to study the role of structural properties in the existence and the stability of a Nash equilibrium. For this purpose, we analyze real data from a regional network of Facebook that spans $63,731$ users (nodes) connected by $817,090$ friendships (edges) [@VMCG09]. In order to corroborate our results in different network topologies, we extract multiple medium-size social subgraphs from the Facebook graph by running a Breath-First Search (BFS) around different starting nodes. Each BFS induces a social subgraph spanning all nodes 2 hops away from a starting node, as well as the edges connecting them. We use this approach to generate a set $\mathbf{G}=\{G_{i}\}_{i\leq 100}$ of 100 different social subgraphs centered around 100 randomly chosen nodes.[^8] ![Scatter plot of the spectral radius, $\protect\rho \left( G_{i}\right) $, versus the lower bound $\protect\beta _{2}\left( G_{i}\right) $ , where each point is associated with one of the $100$ social subgraphs considered in our experiments.](BoundMax.png){width="43.00000%"} From Corollary \[One corollary\] and Lemma \[Lemma 4th metrics\] we can compute the first five spectral moments of a graph $G_{i}$ from the following structural properties: number of nodes ($n_{i}$), edges ($e_{i}$), triangles ($\Delta _{i}$), quadrangles ($Q_{i}$), pentagons ($\Pi _{i}$), as well as the sum-of-squares of the degrees ($W_{i}$), and the degree-triangle correlation ($\mathcal{C}_{i}$). For convenience, we define $S\left( G_{i}\right) \triangleq \left\{ n_{i},e_{i},\Delta _{i},Q_{i},\Pi _{i},W_{i},% \mathcal{C}_{i}\right\} $ as a set of relevant structural properties of $% G_{i}$. In our numerical experiment, we first measure the set of relevant properties $S_{i}$ for each social subgraph $G_{i}\in \mathbf{G}$, and then compute the first five spectral moments of its adjacency matrix. From these moments, we then compute the bounds $\alpha _{2}\left( G_{i}\right) $ and $% \beta _{2}\left( G_{i}\right) $ using Proposition \[Lasserre bound\]. As we mentioned before, these bounds can be computed as the maximum and minimum roots of a third order polynomial, for which closed form expressions are known. We illustrate the quality of our bounds in the following figures. Fig. 2 is a scatter plot where each circle has coordinates $\left( \rho \left( G_{i}\right) ,\beta _{2}\left( G_{i}\right) \right) $, for all $G_{i}\in \mathbf{G}$. Observe how the spectral radii $\rho \left( G_{i}\right) $ of these social subgraphs are remarkably close to the theoretical lower bound $\beta _{2}\left( G_{i}\right) $. Therefore, we can use $\beta _{2}\left( G_{i}\right) $ as an estimate of $\rho \left( G_{i}\right) $ for social subgraphs. In Fig. 3 we include a scatter plot where each circle has coordinates $\left( -\lambda _{n}\left( G_{i}\right) ,-\alpha _{2}\left( G_{i}\right) \right) $, for all $G_{i}\in \mathbf{G}$. Although $\alpha _{2}\left( G_{i}\right) $ is a looser bound than $\beta _{2}\left( G_{i}\right) $, we observe how there is a strong correlation between the value of $\lambda _{n}\left( G_{i}\right) $ and $\alpha _{2}\left( G_{i}\right) $. In these numerical experiments, we have first showed that $\alpha _{2}\left( G_{i}\right) $ and $\beta _{2}\left( G_{i}\right) $ bound the smallest and the largest eigenvalues of the adjacency matrix, and that these bounds are tight, specially $\beta _{2}\left( G_{i}\right) $. Since these bounds can be written as explicit functions of the structural properties in $S\left( G_{i}\right) $, we can estimate the impact of structural perturbations on the spectral radius and the smallest eigenvalue by studying $\left. \partial \beta _{2}\right/ \partial p_{i}$ and $\left. \partial \alpha _{2}\right/ \partial p_{i}$ for $p_{i}\in S\left( G_{i}\right) $. (Details of this perturbation analysis are left for future work due to space limitations.) ![Scatter plot of $-\protect\lambda _{\min }\left( G_{i}\right) $ versus the bound $-\protect\alpha _{2}\left( G_{i}\right) $ , where each point is associated with one of the $100$ social subgraphs considered in our experiments.](BoundMin.png){width="43.00000%"} Conclusions =========== In this paper, we have studied games with linear best response functions in a networked context. We have focused on analyzing the role of the network structure on the game outcome. In particular, the existence and the stability of a unique Nash equilibrium in this class of games are closely related with the smallest eigenvalue of the adjacency matrix of the network. We take this spectral result as the foundation to our work, and use algebraic graph theory and convex optimization to study how local structural properties of the network affect this eigenvalue. In particular, we have derived expressions for the first five spectral moments of the adjacency matrix in terms of local structural properties. These structural properties are: the number of nodes and edges, the number of cycles of length up to 5, the sum-of-squares of the degrees, and the degree-triangle correlation. From this sequence of five spectral moments, we propose a novel methodology to compute optimal bounds on the smallest and the largest eigenvalues of the adjacency matrix by solving two semidefinite programs. In our case, we are able to find analytical solutions to these optimal bounds by computing the roots of a cubic polynomial, for which closed-form expressions are available. Finally, we have verified the quality of our bounds by running numerical simulations in a set of 100 online social subgraphs. For future work, we shall use the results herein presented to study the effect of structural perturbations in the relevant eigenvalues of the adjacency matrix, and in properties of the Nash equilibrium. [99]{} D. Fudenberg and J. Tirole, Game Theory, MIT Press, 1991. G.-M. Angeletos and A. Pavan, Efficient Use of Information and Social Value of Information, Econometrica, vol. 75, pp. 1103-1142, 2007. C. Ballester, A. Calvó-Armengol, and Y. Zenou, Who's Who in Networks. Wanted: The Key Player, Econometrica, vol 74, pp. 1403-1417, 2006. Y. Bramoullé and R. Kranton, Public Goods in Networks, Journal of Economic Theory, vol. 135, pp. 478-494, 2007. Y. Bramoullé, R. Kranton, and M. D’Amours, Strategic Interaction and Networks, CIRPEE Working Paper 10-18, 2010. C. Ballester and A. Calvó-Armengol, Moderate Interactions in Games with Induced Complementaries, mimeo, Universidad Autónoma de Barcelona, 2007. N. Biggs, Algebraic Graph Theory, Cambridge University Press, 2$^{nd}$ Edition, 1993. V.M. Preciado and A. Jadbabaie, From Local Measurements to Network Spectral Properties: Beyond Degree Distributions, Proc. IEEE Conference on Decision and Control, 2010. S.H. Strogatz, Exploring Complex Networks, Nature, vol. 410, pp. 268-276, 2001. S. Boccaletti S., V. Latora, Y. Moreno, M. Chavez, and D.-H. Hwang, Complex Networks: Structure and Dynamics, Physics Reports, vol. 424, no. 4-5, pp. 175-308, 2006. V.M. Preciado, Spectral Analysis for Stochastic Models of Large-Scale Complex Dynamical Networks, Ph.D. dissertation, Dept. Elect. Eng. Comput. Sci., MIT, Cambridge, MA, 2008. D. Cvetković, M. Doob and H. Sachs, Spectra of Graphs, Wiley-VCH, $3^{rd}$ Edition, 1998. K.C. Das and P. Kumar, Some New Bounds on the Spectral Radius of Graphs, Discrete Mathematics, vol. 281, pp. 149-161, 2004. J.B. Lasserre, Bounding the Support of a Measure from its Marginal Moments, Proc. AMS, in press. J.B. Lasserre, Moments, Positive Polynomials and Their Applications, Imperial College Press, London, 2009. $<$ http://cvxr.com/cvx/$>$ M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, Dover, 1965. B. Viswanath, A. Mislove, M. Cha, and K.P. Gummadi, On the Evolution of User Interaction in Facebook, Proc. ACM SIGCOMM Workshop on Social Networks, 2009. M. Stumpf, C. Wiuf, and R. May, Subnets of Scale-Free Networks are not Scale-Free: Sampling Properties of Networks, Proceedings of the National Academy of Sciences, vol. 102, pp. 4221-4224, 2005. [^1]: This work is supported by ONR MURI N000140810747. [^2]: V.M. Preciado and A. Jadbabaie are with the GRASP Laboratory, School of Engineering and Applied Science, University of Pennsylvania, Philadelphia, PA 19104, USA `{preciado,jadbabai}@seas.upenn.edu`. J. Oh is with the Department of Operations and Information Management, Wharton School, University of Pennsylvania, Philadelphia, PA 19104, USA `[email protected]` [^3]: An undirected graph with no self-loops is also called a simple graph. [^4]: For simple graphs, $a_{ii}=0$ for all $i$. [^5]: We denote by $\left\vert Z\right\vert $ the cardinality of a set $Z$. [^6]: As a by-product of our analysis, we also derive lower bounds on the spectral radius of $A_{\mathcal{G}}$, although these bounds are not essential in our analysis. [^7]: Recall that the support of a finite Borel measure $\mu $ on $R$, denoted by $% supp\left( \mu \right) $, is the smallest closed set $B$ such that $\mu \left( R\backslash B\right) =0$. [^8]: Although this procedure is common in studying large social network , it introduces biases that must be considered carefully [@SWM05].	{ "pile_set_name": "ArXiv" }
--- abstract: \| In this paper we will think of certain abelian categories with favorable properties as non-commutative surfaces. We show that under certain conditions a point on a non-commutative surface can be blown up. This yields a new non-commutative surface which is in a certain sense birational to the original one. This construction is analogous to blowing up a Poisson surface in a point of the zero-divisor of the Poisson bracket. By blowing up $\le 8$ points in the elliptic quantum plane one obtains global non-commutative deformations of Del-Pezzo surfaces. For example blowing up six points yields a non-commutative cubic surface. Under a number of extra hypotheses we obtain a formula for the number of non-trivial simple objects on such non-commutative surfaces. address: \| Departement WNI\ Limburgs Universitair Centrum\ Universitaire Campus\ Building D\ 3590 Diepenbeek\ Belgium author: - Michel Van den Bergh title: 'Blowing up of non-commutative smooth surfaces' --- Introduction {#ref:1a} ============ Throughout this paper $k$ will be an algebraically closed field. Motivation ---------- Let $X$ be a smooth connected surface over $k$ and let $q$ be a Poisson bracket on $X$. Since we are in the dimension two, $q$ corresponds to a section of $\omega_X^\ast$. Let $p\in X$ and let $\alpha:\tilde{X}\r X$ be the blowup of $X$ in $p$. From the fact that $\tilde{X}$ and $X$ share the same function field it is easily seen that $q$ extends to $\tilde{X}$ if and only if $q$ vanishes in $p$. Denote the extended Poisson bracket by $q'$ and let $Y$ resp. $T$ be the zero divisors of $q$ and $q'$. One verifies that as divisors : $T=\alpha^{-1}(Y)-L$, where $L=\alpha^{-1}(p)$ is the exceptional curve. In particular $T$ contains the strict transform $\tilde{Y}$ of $Y$, and if $p\in Y$ is simple then actually $T=\tilde{Y}$. Our aim in this paper is to show that there exists a non-commutative version of this situation. That is we show that it is possible to view the blowup of a Poisson surface as the quasi-classical analogue of a blowup of a non-commutative surface. Our motivation for doing this is to provide a step in the ongoing project of classifying graded domains of low Gelfand-Kirillov dimension. Since the case of dimension two was completely solved in [@Staf5] the next interesting case will very likely be dimension three (leaving aside rings with fractional dimension which seem to be quite exotic). One may view three dimensional graded rings as non-commutative projective surfaces. Motivated by some heuristic evidence Mike Artin conjectures in [@Ar2] that, up to birational equivalence, there will be only a few classes, the largest one consisting of those algebras that are birational to a quantum $\PP^2$ (see below). Once a birational classification exists, one might hope that there would be some version of Zariski’s theorem saying that if two (non-commutative) surfaces are birationally equivalent then they are related through a sequence of blowing ups and downs. With the current level of understanding it seems rather unlikely that Artin’s conjecture or a non-commutative version of Zariski’s theorem will be proved soon, but this paper provides at least one piece of the puzzle. This being said, it is perhaps the right moment to point out that in this paper we won’t really define the notion of a non-commutative surface. Instead we first introduce non-commutative schemes (or quasi-schemes, to follow the terminology of [@rosenberg]). These will simply be abelian categories having sufficiently nice homological properties. Then we will impose a few convenient additional hypotheses which would hold for a commutative smooth surface (see §\[ref:5.1a\]). To fix ideas we will first discuss two particular cases of quasi-schemes. If $R$ is a ring then $\Spec R$ is the category of right $R$-modules (the “affine case”). If $A=A_0\oplus A_1\oplus\cdots$ is a graded ring then $\Proj A$ is (roughly) the category of graded right $A$-modules, modulo the modules with right bounded grading (the “projective case”). Let us first consider the affine case. Assume that $R$ is a finitely generated $k$-algebra and let $C$ the commutator ideal. $C$ is the natural analog of the zero divisor of a Poisson bracket. Now $\Spec R/C$ is a commutative affine scheme and a $k$-point in $\Spec R/C$ corresponds to a maximal ideal $m$ in $R$ with $R/m=k$. Hence a natural idea is to define the blowup of $\Spec R$ in $p$ as $\Proj D$ where $D$ is the Rees algebra associated to $m$. $$D=R\oplus m\oplus m^2\oplus m^3\oplus\cdots$$ It is easily seen however that this definition is faulty. Consider the following example [@Ar2] $R=k\langle x,y\rangle/(yx-xy-y)$, $m=(x,y)$. Then $m^n=(x^n,y)$. Hence the analog of the exceptional curve $$D/mD=R/m\oplus m/m^2\oplus m^2/m^3\oplus$$ is isomorphic to $k[x]$. Thus $\Proj D/m$ is a point, whereas intuitively we would expect it to be one-dimensional in some sense. It turns out however that in this example one can use a certain twisting of the Rees algebra which yields a reasonably behaved blowing up. This is based on the observation that the commutator ideal $C=(y)$ is an invertible $R$-bimodule. Let $J$ be its inverse. Then we define $I^n$ as the image $I^{\otimes n}$ in $J^{\otimes n}$ and we define the modified Rees algebra $D$ as $$D=R\oplus I\oplus I^2\oplus I^3\oplus\cdots$$ The blowup of $\Spec R$ in $p$ is now defined as $\Proj D$ for this new $D$. We refer the reader to [@Ar2] for a detailed workout of this example. However we will indicate how one finds the analog of the exceptional curve. Let $\tau$ be the automorphism of $R$ given by $a\mapsto y^{-1}a y$. Then $J=R_\tau$ and hence $I^n=m\tau(m)\cdots \tau^{n-1}(m)_{\tau^n}$. Put $L=D/\tau^{-1}(m) D$. One now verifies that $\dim L_u=u+1$. So $L$ plays the role of the exceptional curve. Note however that $L$ is a right $D$-module but not a left module. In retrospect this was to be expected since, as we have said in the first paragraph, if we blow up a Poisson surface, then the extended Poisson bracket will in general not vanish on the exceptional curve. This example indicates the way to go for rings whose commutator ideal is invertible. The latter hypotheses is not unreasonable since if we look at the case of a Poisson surface then we see that we expect a non-commutative smooth surface to contain a commutative curve. Additional motivation comes from considering the local rings $k\langle\langle x,y\rangle\rangle /(\phi)$ with $\phi$ a (non-commutative) formal power series whose lowest degree term is a non-degenerate quadratic tensor in $x,y$. These local rings are the non-commutative analogs of complete two dimensional regular local rings and one verifies that their commutator ideal is indeed invertible (see e.g. [@VdBVG]). There is one important hitch however. The commutator ideal is not invariant under Morita equivalence! This indicates that it is important to develop the theory in a more category-theoretical frame work. This will make it possible to talk about non-commutative schemes containing a commutative curve, without refering to rings or ideals at all. To stress this point even more let us consider the case of graded rings. In [@AS] Artin and Schelter introduced so-called regular rings. These are basically graded rings which have the Hilbert series of three dimensional polynomial rings, together with a few other reasonable properties. They were classified in [@AS; @ATV1; @ATV2] and also, with different methods, in [@Bondal]. Let $A$ be such a regular ring. We view $X= \Proj A$ as a quantum $\PP^2$. Since on $\PP^2$ the anti-canonical sheaf has degree three, the zero divisor of a Poisson bracket will be a (possibly singular and non-reduced) elliptic curve. Therefore we would also expect $X=\Proj A$ to contain an elliptic curve in some reasonable sense. It turns out that this is indeed true! It was shown in [@AS; @AVdB; @ATV1] that $A$ contains a normal element $g$ in degree three such that $\Proj A/gA$ is equivalent with the category of quasi-coherent sheaves over an elliptic curve $Y$. Thus if we actually identify $Y$ with $\Proj A/gA$ then $Y\hookrightarrow X$. Now let $p\in Y$. The previous discussion suggests that it should be possible to blow up $p$. However it is not clear how to proceed. Under the inclusion $Y\hookrightarrow X$, $p$ corresponds to a so-called point module [@ATV2] over $A$. This is by definition a graded right module, which is generated in degree zero and which is one-dimensional in every degree. However such a point module is only a right module and hence it cannot be used to construct a Rees algebra. Our solution is to construct the Rees algebra directly over $\Proj A$. To do this we have to invoke the theory of monads [@ML]. Since we only consider monads satisfying a lot of additional hypotheses we prefer to call our monads algebras. This is at variance with the use of “algebra” in the theory of categories [@ML] but in our context it seems reasonable. For us an algebra over an abelian category $\Cscr$ is in principle an algebra object in the monoidal category of right exact functors from $\Cscr$ to itself. There are however some technical problems with this so we end up using a less intuitive definition (see below). The importance of monads in non-commutative algebraic geometry was noticed by various people, in particular by Rosenberg. See for example [@rosenberg; @RL]. In the last chapter of his book Rosenberg actually defines a blow up of an arbitrary “closed” subcategory of an abelian category. While this definition is also in terms of monads, it is as far as I can see, somewhat different from ours. To see this let us again consider the affine case. Then Rosenberg’s construction is in terms of the functor $M\mapsto Mm$, which is not right exact. If we replace this functor by $M\mapsto M\otimes_R m$ then one would get the $\Proj$ of the ordinary Rees algebra of $R$, which (depending on what one wants to achieve) might not be the right answer (as we have shown above). Construction ------------ Following [@rosenberg] we introduce the notion of a quasi-scheme. For us this will be a Grothendieck category (that is : an abelian category with a generator and exact direct limits). However we tend to think of quasi-schemes as geometric objects, so we denote them by roman capitals $X$, $Y$, …. If we really refer to the category represented by a quasi-scheme $X$ then we write $\Qch(X)$. Note that in fact $X=\Qch(X)$, but it is very useful to nevertheless make this notational distinction since it allows us to introduce other notations in a consistent way. For example we will denote the noetherian objects in $\Mod(X)$ by $\mod(X)$. Furthermore we can absorb additional structure into the symbol $X$ (such as a map to a base quasi-scheme) which is not related to $\Mod(X)$. This would be awkward without the two different notations $X$ and $\Mod(X)$. A map $\alpha:X\r Y$ between quasi-schemes will be a right exact functor $\alpha^\ast:\Mod(Y)\r \Mod(X)$ possessing a right adjoint (denoted by $\alpha_\ast$). In this way the quasi-schemes form a category (more precisely a two-category, see Appendix \[ref:Aa\]). If $X$ is a quasi-compact quasi-separated commutative scheme then $\Qch(X)$ will be the category of quasi-coherent sheaves on $X$. It is proved in [@thomasson] that this is a Grothendieck category. Rosenberg in [@rosenberg1] has proved a reconstruction theorem which allows one to recover $X$ from $\Qch(X)$ (generalizing work of Gabriel in the noetherian case). He has also announced that the functor which assigns to a commutative scheme its associated quasi-scheme is fully faithful if we work over $\Spec\ZZ$. Let $X$ be a quasi-scheme. We think of objects in $\Qch(X)$ as sheaves of right modules on $X$. However to define algebras on $X$, it is clear that we need bimodules on $X$ (see [@VdB11] for the case where $X$ is commutative). Let us for the moment define a bimodule on $X$ as a right exact functor from $\Qch(X)$ to itself commuting with direct limits. Then the corresponding category is monoidal (the tensor product being given by composition) and hence we can define algebra objects. Let $\Ascr$ be such an algebra object. It is routine to define an abelian category $\Mod(\Ascr)$ of $\Ascr$-modules. So this seems like a reasonable starting point for the theory. However a difficulty emerges if one wants to define Rees algebras. As we have seen, the main point is to take the sum of the $I^n$ for some subbimodule $I$ of an invertible bimodule $\Lscr$. $I^n$ was defined as the image of $I^{\otimes n}\r \Lscr^{\otimes n}$. Unfortunately to take an image one needs an abelian category, and I don’t see how to prove that the above definition of a bimodule yields an abelian category, even if we drop the requirement that bimodules should commute with direct limits. In this paper we sidestep this difficulty by defining the category of bimodules on $X$ as the opposite category of the category of left exact functors from $\Qch(X)$ to itself. Since left exact functors are determined by their values on injectives, they trivially form an abelian category. In this way one can define Rees algebras in reasonable generality (see Definition \[ref:3.5.13a\]). We will say that a quasi-scheme $X$ is noetherian if $\Mod(X)$ is locally noetherian. That is, if $\Mod(X)$ is generated by $\mod(X)$. As already has been indicated above, in this paper we will study a noetherian quasi-scheme $X$ which contains a commutative curve $Y$ as a divisor. To make this more precise we denote the identity functor on $\Qch(X)$ by $o_X$. This is an algebra on $X$ such that $\Mod(o_X)=\Qch(X)$. We will assume that $o_X$ contains an invertible subbimodule $o_X(-Y)$ such that $\Mod(o_X/o_X(-Y))$ is equivalent with $\Qch(Y)$. We also need some sort of smoothness condition on $X$. Since it is obviously sufficient to impose this in a neighborhood of $Y$, we assume that every object in $\Qch(Y)$ has finite injective dimension in $\Qch(X)$. This is the same setting as in [@VdBVG], albeit cast in a somewhat different language. Now $p\in Y$ defines a subbimodule $m_p$ of $o_X$ which is the analog of the maximal ideal corresponding to $p$. We put $I=m_po_X(Y)$. Define $$\Dscr=o_X\oplus I\oplus I^2\oplus \cdots$$ This is the Rees algebra associated to $I$. We define the blowup $\tilde{X}$ of $X$ in $p$ as $\Proj \Dscr$. General properties ------------------ A large part of this paper is devoted to proving that $\tilde{X}$ satisfies similar properties as $X$ and furthermore that we have obtained an analogue of the blowup of a commutative surface. However before we start we need to have a better understanding of the formal neighborhood of a point $p\in Y$. This was in fact already done in [@VdBVG]. The answer is given in terms of certain topological rings (see Theorem \[ref:5.1.4a\] for precise results). It turns out that the formal local structure of $p$ depends heavily on a certain automorphism $\tau$ of $Y$ which was also introduced in [@VdBVG]. To be more precise we define the normal bundle of $Y$ in $X$ as $o_X(Y)/o_X$. This bimodule defines an autoequivalence of $\Mod(Y)$ and by a result in [@AZ] such an autoequivalence must necessarily be of the form $\tau_\ast(-\otimes_{\Oscr_Y}\Nscr)$ where $\Nscr$ is a line bundle on $Y$ and $\tau$ is an automorphism of $Y$. In particular if $\tau$ has infinite order (and hence $p\in Y$ is smooth) then the formal local structure of $p$ is given by the ring of doubly infinite lower triangular matrices over the ring $\hat{\Oscr}_{Y,p}$. In particular this is independent of $p$ and $Y$ (as is the case for the completion at a smooth point on a commutative scheme). In this paper we complete the results in [@VdBVG] by showing that various completion functors are exact. For this we refer the reader to §\[ref:5a\]. An interesting application of completion is given in section §\[ref:5.7b\]. In this section we define (roughly) the multiplicity in $p$ and the points infinitely near to $p$ of an object in $\mod(X)$ in the case that the $\tau$-orbit of $p$ is infinite. As a starting point for the study of $\tilde{X}$ we construct a commutative diagram of quasi-schemes. $$\begin{CD} \tilde{Y} @>\beta>> Y\\ @ViVV @VVjV\\ \tilde{X} @>>\alpha> X \end{CD}$$ where the vertical arrows are inclusions. $\tilde{X}$ is again a noetherian quasi-scheme (Theorem \[ref:6.3.1a\]). $\tilde{Y}$ is a commutative curve which plays the role of the strict transform of $Y$. $\tilde{Y}$ is again a divisor in $X$ and every object in $\Qch(\tilde{Y})$ has finite injective dimension in $\Qch(\tilde{X})$ (see Theorem \[ref:6.6.3a\]). Associated to the point $p\in Y$ there is a simple object $\Oscr_p$. We define $\Oscr_L$ as $\alpha^\ast(\Oscr_p)$ and we consider $\Oscr_L$ as the structure sheaf on the exceptional curve in $\tilde{X}$. In fact following a recipe given in [@SmithZhang] we can define a category $\Mod(L)$. Roughly $\Mod(L)$ is generated by subquotients of direct sums of twists of $\Oscr_L$. In this way we can speak of the quasi-scheme $L$. It follows from Proposition \[ref:6.5.2a\] (together with Corollary \[ref:6.7.4a\]) that if we view $\Mod(X)$ modulo the objects supported on the $\tau$-orbit of $p$, and $\Mod(\tilde{X})$ modulo the objects supported on $L$ then we obtain equivalent categories. This is the obvious analogue of the situation in the commutive case where $X-p$ and $\tilde{X}-L$ are isomorphic (and hence in particular $X$ and $\tilde{X}$ are birational). In section §\[ref:6.7b\] we give a precise description of $\Mod(L)$ (using results of [@SmithZhang] in the case that $\tau$ has infinite order). It will follow that $\Mod(L)$ is very closely related to the category of quasi-coherent sheaves on $\PP^1$, illustrating again the analogy with the commutative case. In section §\[ref:7a\] we compute the derived category of $\tilde{X}$. Our main result is that this derived category has a semi-orthogonal decomposition given by the derived category of $X$ and the derived category of $k$. This generalizes a result by Orlov [@Orlov]. Non-commutative Del-Pezzo surfaces ---------------------------------- The results starting from Section \[ref:9a\] are inspired by the construction in the commutative case of (most) Del-Pezzo surfaces by blowing up a collection of points in $\PP^2$. Let $(Y,\sigma,\Lscr)$ be an elliptic triple $Y\subset \PP^2$ as in [@ATV1]. We assume that $Y$ is smooth and that $\sigma$ is a translation of infinite order. Let $A$ be the regular algebra associated to this triple [@ATV1] and let $X_0=\Proj A$. As above we consider $X_0$ as a quantum version of $\PP^2$. The curve $Y$ is contained as a divisor in $X_0$ and $\tau$ is equal in this case to $\sigma^3$. We choose points $p_1,\ldots, p_n\in Y$ ($n\le 8$) and we construct a commutative diagram of quasi-schemes $$\xymatrix{ &\tilde{X}_1\ar[dl]_{\alpha_1}\ar[dr]^{\delta_0} &&\tilde{X}_2\ar[dl]_{\alpha_2}\ar[dr]^{\delta_1} &&&&\tilde{X}_n\ar[dl]_{\alpha_n}\ar[dr]^{\delta_n} &\\ X_1 &&X_2 &&X_3 &\cdots&X_n &&X_{n+1} \\ &&&&Y\ar[ullll]\ar[ull]\ar[u]\ar[urr]\ar[urrrr] &&&& }$$ Here the map $\alpha_i$ is the blowup of $X_i$ in $p_i$. Morally $X_{i+1}$ is constructed from $\tilde{X}_i$ by putting $X_{i+1}=\Proj \left(\bigoplus_n H^0(X_i,o_{X_i}(nY))\right)$ (actually for simplicity $X_{i+1}$ is constructed using a slightly different method (see §\[ref:9.2b\]), which is easily seen to be equivalent). The point is that in the commutative case $\delta_i$ would be an isomorphism if the points $p_1,\ldots,p_n$ are in general position (this follows from the Nakai criterion for ampleness, see [@H]). In the non-commutative $\delta_i$ will not in general be an isomorphism. However we show in Theorem \[ref:11.1.3a\] that $\delta_i$ yields a derived equivalence, under suitable general position hypotheses. Exceptional simple objects -------------------------- One of the aims of these notes is to classify (or rather count) the simple objects in $\Mod(X_i)$ which are not of the form $\Oscr_q$ for some $q\in Y$. We call such simple objects “exceptional” because, firstly, they do not always exist, and secondly when they exist they are not easy to construct or to count. Using some results on geometry in the projective quantum plane (§\[ref:10a\]) together with the above results on derived categories we obtain in Theorem \[ref:11.2.1a\] a formula for the number of exceptional simple objects in $\Mod(X_{i+1})$ (if $n\le 6$, $\tau$ has infinite order and suitable general position hypotheses hold). It turns out that the number of exceptional simple objects depends in a very sensitive way upon the position of the points $p_1,\ldots,p_n$. For example if $n=6$ (and the other hypotheses are satisfied) then our formula yields that there may be between $0$ and $6$ exceptional simple objects in $\Mod(X_7)$. Non-commutative cubic surfaces ------------------------------ Being near the end of this introduction we now indicate our original motivation for starting this project. It concerns a problem which is not quite completely solved but which at least has become more tractable. As above let $(Y,\sigma,\Lscr)$ be an elliptic triple with $Y\in \PP^2$ and let $A$ be the associated three dimensional Artin-Schelter regular algebra. It is easy to show that the representation theory of $A$ is fairly trivial. At a certain point Lieven Le Bruyn (see for example [@L]) suggested that one could obtain more interesting representation theories by considering filtered rings $C$ such that $\gr C=A$. This was motivated by the example of enveloping algebras of Lie algebras which are related in a similar way to polynomial algebras. Instead of studying the filtered rings $C$, it is easier to study their Rees rings $D=\oplus_n C_n$. These are characterized as the $\NN$-graded rings containing a regular central element $t$ in degree one such that $D/tD=A$. Such graded rings were studied in [@LSV] but not much progress was made towards their representation theory. As usual a first step in the study of the representation theory of $D$ is the construction of a “Casimir” element. Indeed $A$ contains a canonical central element $g$ in degree $3$ and by a computer computation one can show that $g$ lifts to a central element $G$ in $D$. Then instead of studying $D$ on may study the central quotients $D_\mu=D/(G+\mu t^3)$, $\mu \in k$. The $D_\mu$ may be considered as quantum versions of cubic surfaces in $\PP^3$. A well known result in commutative algebraic geometry states that a cubic surface in $\PP^3$ is obtained by blowing up six points in $\PP^2$, so one may ask whether this is also true in the noncommutative case. I have not been able to show this in general but a converse result is obtained in Section \[ref:12a\]. We show that if we blow up six points in the elliptic quantum plane then the resulting quasi-scheme is contained as a cubic divisor in a quantum $\PP^3$. Very recently Mike Artin has explained to me that one can probably obtain the complete analogue of the commutative result by deformation theory. Indeed $Z=\Proj D_\mu$ can be obtained as a deformation of a cubic surface $Z_0$ in $\PP^3$. In $Z_0$ we can choose six mutually skew exceptional curves [@H] and since the structure sheaves of these exceptional curves have no higher $\Ext$’s they deform well. So we should find on $Z$ six corresponding exceptional curves. Then we can blow these down, for example using the procedure exhibited in [@VdB24]. To carry out this program there are quite a few technical details that remain to be filled in. There are some recent notes by Mike Artin on specializing birational equivalences in the non-commutative case. Acknowledgement --------------- The author wishes to thank Mike Artin and Paul Smith for stimulating discussions about the material in earlier versions of this manuscript and about non-commutative geometry in general. I also wish to thank Paul Smith for showing me the interesting preprint [@SmithZhang] which contains related results. Preliminaries on category theory ================================ Our main references for categories are [@Gabriel; @Groth1; @ML; @stenstrom]. If nothing is specified then categories will have small homsets. This will not always be the case for some secondary categories such as categories of bimodules (see below). However limits and colimits are always taken over small sets. In particular complete and cocomplete will refer to the existence of small limits and colimits. When we speak of a direct limit, we mean a colimit over a directed set. A similar convention applies to inverse limits. We will often implicitly use the fact that in an abelian category $$\label{ref:2.1a} \begin{split} \oplus_{i\in I} Y_i&=\dirlim_{\substack{j\in J\\ J\subset I\text{ finite}}} Y_j\\ \prod_{i\in I} Y_i&=\invlim_{\substack{j\in J\\ J\subset I\text{ finite}}} Y_j \end{split}$$ Thus an additive functor commuting with direct limits commutes with coproducts. A similar statement applies to products. We use the following specialized version of the standard adjoint functor theorems [@ML]. \[ref:2.1b\] Let $\Cscr$, $\Dscr$ be abelian categories and let $F:\Dscr\r\Cscr$, $G:\Cscr\r\Dscr$ be respectively a right and a left exact functor. Then 1. If $\Cscr$ is complete and has a cogenerator then $G$ has a left adjoint if and only if $G$ commutes with products. 2. (Dual version) If $\Dscr$ is cocomplete and has a generator then $F$ has a right adjoint if and only if $F$ commutes with coproducts. Most of the abelian categories we use will be Grothendieck categories. These are abelian categories which have a generator and exact direct limits. We use the following results which are well-known [@Groth1; @stenstrom]. Let $\Cscr$ be a Grothendieck category. Then 1. $\Cscr$ has products (not necessarily exact). 2. $\Cscr$ has enough injectives. 3. The product of the injective hulls of the quotients of a fixed generator is an injective cogenerator. A Grothendieck category which is generated by noetherian objects is called locally noetherian. Such categories were studied by Gabriel in [@Gabriel]. One important property they have is the following. In a locally noetherian category the direct limit of injective objects is injective. A noetherian category is an abelian category in which every object is noetherian. One has [@Gabriel III, Th 1] Every noetherian category $\Cscr$ is equivalent with the category of noetherian objects in a locally noetherian category $\tilde{\Cscr}$. $\tilde{\Cscr}$ is characterized up to equivalence by this property. A Serre subcategory $\Sscr$ of an abelian category $\Cscr$ is by definition a full subcategory which is closed under subquotients and extensions. In that case there exists a quotient category $\Cscr/\Sscr$ which is characterized by an appropriate universal property. If $\Cscr$ is a Grothendieck category then we say that $\Sscr$ is localizing if $\Sscr$ is a Serre subcategory which is closed under direct limits. In that case $\Cscr/\Sscr$ is also a Grothendieck category. Furthermore the quotient functor $\pi:\Cscr\r \Cscr/\Sscr$ has a right adjoint which we will usually denote by $\omega$. The composition $\omega\pi$ is called the localization functor and will be denoted by $\tilde{(-)}$. In this paper we basically work with the category of Grothendieck categories, the morphisms being certain functors. This is an example of a two-category. Although we usually only implicitly use this concept, we explain some basic notions in Appendix \[ref:Aa\]. Non-commutative geometry {#ref:3a} ======================== In this paper we will sometimes work with categories which do not come from categories of modules over (graded) rings. Therefore in this section we introduce some rudiments of a formalism which may be used to imitate some of the more elementary features of commutative algebraic geometry. This section is closely related to [@rosenberg][@VdB11]. Bimodules {#ref:3.1a} --------- Below $\Cscr$, $\Dscr$, $\Escr$, …will be abelian categories. We define the following categories. $$\begin{aligned} \Lscr(\Dscr,\Cscr)&=\{\text{left exact functors $\Dscr\r \Cscr$}\} \\ \BIMOD(\Cscr-\Dscr)&=\Lscr(\Dscr,\Cscr)^{\text{opp}}\\ \Bimod(\Cscr-\Dscr)&=\{\Mscr\in\BIMOD(\Cscr-\Dscr)\mid \Mscr\text{ has a left adjoint }\}\end{aligned}$$ Objects in $\BIMOD(\Cscr-\Dscr)$ will be called weak $\Cscr$-$\Dscr$ bimodules, whereas objects in $\Bimod(\Cscr-\Dscr)$ will simply be called $\Cscr$-$\Dscr$ bimodules. \[ref:3.1.1a\] 1. $\BIMOD(\Cscr-\Dscr)$, $\Bimod(\Cscr-\Dscr)$ have finite colimits, and the inclusion $\Bimod(\Cscr-\Dscr)\r \BIMOD(\Cscr-\Dscr)$ preserves those colimits. 2. If $\Cscr$ is complete then $\BIMOD(\Cscr-\Dscr)$ is cocomplete. If $\Dscr$ is in addition cocomplete then so is $\Bimod(\Cscr-\Dscr)$ and furthermore the inclusion $\Bimod(\Cscr-\Dscr)\r \BIMOD(\Cscr-\Dscr)$ preserves colimits. 3. If $\Dscr$ has enough injectives then $\BIMOD(\Cscr-\Dscr)$ is an abelian category. 4. If $\Dscr$ is complete and has a cogenerator then an object in $\BIMOD(\Cscr-\Dscr)$ is in $\Bimod(\Cscr-\Dscr)$ iff it commutes with products. 5. If $\Dscr$ is complete and has an injective cogenerator then an object in $\BIMOD(\Cscr-\Dscr)$ is in $\Bimod(\Cscr-\Dscr)$ iff it commutes with products when evaluated on injectives. 6. If $\Dscr$ is complete and has an injective cogenerator and products are exact in $\Cscr$ then $\Bimod(\Cscr-\Dscr)$ is an abelian subcategory of $\BIMOD(\Cscr-\Dscr)$. <!-- --> 1. To show that $\BIMOD(\Cscr-\Dscr)$ has finite colimits we have to show that the opposite category $\Lscr(\Dscr,\Cscr)$ has finite limits. Now one easily verifies that for a functor $G:I\r \Lscr(\Dscr,\Cscr)$ with $I$ a finite category $$G'(M)=\invlim_i G(i)(M)$$ defines the inverse limit of $G$ in $\Lscr(\Dscr,\Cscr)$. Assume now that the $G(i)$ are in $\Bimod(\Cscr-\Dscr)$ so they have left adjoints $F(i)$. Define $F'$ by $$F'(N)=\dirlim_i F(i)(N)$$ One easily verifies that $F'$ is a left adjoint to $G'$ and hence $G'\in \Bimod(\Cscr-\Dscr)$. 2. This is similar to (1). 3. Let $\Inj(\Dscr)$ denote the additive category of injectives in $\Dscr$. Then $ \Lscr(\Dscr,\Cscr) $ is equivalent to $\Funct(\Inj(\Dscr),\Cscr)$, where “$\Funct$” denotes the category of additive functors. It is now clear that $\Lscr(\Dscr,\Cscr)$ inherits the property of being an abelian category from $\Cscr$. 4. This follows from Theorem \[ref:2.1b\]. 5. This follows from (4), using the fact that products are left exact. 6. We have to show that $\Bimod(\Cscr-\Dscr)$ is closed under kernels. Equivalently the subcategory of $\Funct(\Inj(\Dscr),\Cscr)$ of functors commuting with products has to be closed under cokernels. Let $G_1\r G_2$ be a map of functors in $\Funct(\Dscr,\Cscr)$ commuting with products and let $G_3=\coker(G_1\r G_2)$. Then for $E\in \Inj(\Dscr)$ we have $G_3(E)=\coker(G_1(E)\r G_2(E))$ and using the fact that products are exact one easily obtains that $G_3$ commutes with products. We will now introduce some more suggestive notations for dealing with $\BIMOD(\Cscr-\Dscr)$. If $\Mscr\in \BIMOD(\Cscr-\Dscr)$ then we denote the corresponding functor in $\Lscr(\Dscr,\Cscr)$ by $ \HHom_\Dscr(\Mscr,-) $. More or less by definition we have the following facts. 1. $\HHom_\Dscr(-,-)$ is left exact in its two arguments. 2. If $\Cscr$ is complete then $\HHom_\Dscr(-,-)$ transforms colimits in its first argument into limits. 3. If $\Mscr\in \Bimod(\Cscr-\Dscr)$ then $\HHom_\Dscr(\Mscr,-)$ transforms limits in its second argument into limits. 4. If $\Dscr$ has enough injectives and $E\in \Inj(\Dscr)$ then $\HHom_\Dscr(-,E)$ is exact. <!-- --> 1. That $\HHom_\Dscr(-,-)$ is left exact in its second argument is by definition. That it is left exact in its first argument follows from the explicit construction of colimits (and hence of cokernels) in the proof of Prop. \[ref:3.1.1a\](1). 2. This follows as in the proof of Prop. \[ref:3.1.1a\](2). 3. If $\Mscr\in\Bimod(\Cscr-\Dscr)$ then $\HHom_\Dscr(\Mscr,-)$ has by definition a left adjoint. Hence it commutes with limits. 4. This follows from the explicit structure of an abelian category on $\BIMOD(\Cscr-\Dscr)$ given by the proof Prop. \[ref:3.1.1a\](3). We write composition of functors $$\BIMOD(\Cscr-\Dscr)\times \BIMOD(\Dscr-\Escr)\r \BIMOD(\Cscr-\Escr)$$ as $-\otimes_\Dscr-$. In this way we obtain for $\Mscr\in \BIMOD(\Cscr-\Dscr)$, $\Nscr\in \BIMOD(\Dscr-\Escr)$ the satisfying formula $$\HHom_\Escr(\Mscr\otimes_\Dscr\Nscr,-)=\HHom_\Dscr(\Mscr,\HHom_\Escr(\Nscr,-))$$ Again more or less by definition we obtain the following properties. 1. $-\otimes_\Dscr-$ is right exact in its two arguments. 2. If $\Cscr$ is complete then $-\otimes_\Dscr-$ preserves colimits in its first argument. 3. If $\Dscr$ is complete and if $\Mscr\in \Bimod(\Cscr-\Dscr)$ then $\Mscr\otimes_\Dscr-$ preserves colimits in its second argument. 4. $-\otimes_\Dscr-$ sends $\Bimod(\Cscr-\Dscr)\times \Bimod(\Dscr-\Escr)$ to $\Bimod(\Cscr-\Escr)$. Now define $$\MOD(\Cscr)=\BIMOD(\Ab-\Cscr)$$ The functor $M\mapsto \Hom_\Cscr(M,-)$ defines a full faithful embedding of $\Cscr$ in $\MOD(\Cscr)$. Throughout we will identify $\Cscr$ with its essential image under this embedding. We now obtain the following alternative “characterization” of $\Bimod(\Cscr-\Dscr)$ inside $\BIMOD(\Cscr-\Dscr)$. \[ref:3.1.4a\] Let $\Mscr\in\BIMOD(\Cscr-\Dscr)$. Then $\Mscr$ is in $\Bimod(\Cscr-\Dscr)$ if and only if the functor $$-\otimes_\Cscr\Mscr:\MOD(\Cscr)\r \MOD(\Dscr)$$ sends $\Cscr$ to $\Dscr$. A few concepts from the theory of bimodules over rings can be generalized to our setting. We denote the derived functors of $\HHom_\Dscr(-,-)$ in the second argument by $\HExt^i_\Dscr(-,-)$ (if they exist). Assume that $\Escr$ has enough injectives. We say that $\Nscr\in \BIMOD(\Dscr-\Escr)$ is (left) flat if $\HHom_\Escr(\Nscr,-)$ preserves injectives. More generally for $\Mscr\in \BIMOD(\Cscr-\Dscr)$ and $\Nscr\in\BIMOD(\Dscr-\Escr)$ we define $\HTor^\Dscr_i(\Mscr, \Nscr)$ as the object in $\BIMOD(\Cscr-\Dscr)$ satisfying $$\label{ref:3.1b} \HHom_\Escr(\HTor^\Dscr_i(\Mscr,\Nscr),E)\overset{}{=} \HExt^i_\Dscr(\Mscr,\HHom_\Escr(\Nscr,E))$$ for every injective $E$ of $\Escr$. A similar definition holds for $\Mscr\in\Dscr$. \[ref:3.1.5a\] Assume that $\Escr$ has enough injectives. 1. $\HTor^\Dscr_i(-,-)$ is a $\delta$-functor in its two arguments. 2. $\Nscr\in \BIMOD(\Dscr-\Escr)$ is flat if and only if $\HTor^\Dscr_1(\Mscr,\Nscr)=0$ for all $\Mscr\in\Dscr$. 3. If $\Nscr\in \BIMOD(\Dscr-\Escr)$ is flat then $\HTor^\Dscr_i(\Mscr,\Nscr)=0$ for all $\Mscr\in \BIMOD(\Cscr-\Dscr)$. All statements follow directly from the definitions. Assume that $\Cscr$, $\Dscr$ have colimits. Then we will say that $\Mscr\in\BIMOD(\Cscr-\Dscr)$ is coherent if $\HHom_\Dscr(\Mscr,-)$ commutes with direct limits. \[ref:3.1.6a\] Assume that $\Mscr\in\Bimod(\Cscr-\Dscr)$ and consider the following statements. 1. $\Mscr$ is coherent. 2. $-\otimes_{\Cscr}\Mscr$ preserves finitely presented objects in $\Cscr$. Then (1) implies (2). The converse holds if $\Cscr$ is generated by finitely presented objects. - Assume that $T$ is a finitely presented object in $\Cscr$. I.e. $\Hom_\Cscr(T,-)$ commutes with direct limits. We have to show that $\Hom_\Cscr(T\otimes_\Cscr \Mscr,-)$ commutes with direct limits. This follows from the fact that $$\Hom_\Cscr(T\otimes_\Cscr\Mscr,-)=\Hom_\Cscr(T,\HHom_\Dscr(\Mscr,-))$$ - We have to construct a natural isomorphism between $\Hom_\Cscr(U,\dirlim_i \Hom_\Dscr(\Mscr,\Nscr_i))$ and $\Hom_{\Cscr}(U,\Hom_\Dscr(\Mscr,\dirlim \Nscr_i))$ for an arbitrary inverse system $(\Nscr_i)_i$ in $\Dscr$ and $U\in\Cscr$. Since $\Cscr$ is generated by finitely presented objects, it suffices to do this for $U$ finitely presented. But this is clear by adjointness. \[ref:3.1.7a\] Assume $\Dscr$ is a Grothendieck category. Suppose furthermore that $\Cscr$ has exact direct limits. Let $\Mscr\in \BIMOD(\Cscr-\Dscr)$. Then the following are equivalent 1. $\Mscr$ is coherent. 2. $\HHom_\Dscr(\Mscr,-)$ commutes with direct limits of injectives. We only have to prove 2.$\Rightarrow$1. In a Grothendieck category embeddings into injectives can be constructed functorially [@Gr]. Hence if $(T_i)_{i\in I}$ is an inverse system in $\Dscr$ then there is a copresentation $$0\r (T_i)_{i\in I}\r (E_i)_{i\in I}\r (F_i)_{i\in I}$$ with $(E_i)_i$, $(F_i)_i$ injective. The left exactness of $\HHom_\Dscr(\Mscr,-)$ together with the fact that direct limits are exact in $\Cscr$ and $\Dscr$ now shows what we want. \[ref:3.1.8a\] Assume that $\Cscr$ has exact direct limits and that $\Dscr$ is locally noetherian. Then the category of coherent objects in $\BIMOD(\Cscr-\Dscr)$ is an abelian subcategory of $\BIMOD(\Cscr-\Dscr)$, closed under extensions. According to Proposition \[ref:3.1.7a\] the fact whether $\Mscr\in\BIMOD(\Cscr-\Dscr)$ is coherent can be tested on inverse systems of injectives $(E_i)_{i\in I}$. Since $\Dscr$ is locally noetherian we also have that $F=\dirlim\nolimits_i E_i$ is injective. Hence by construction $\HHom_\Dscr(-,E_i)$ and $\HHom_\Dscr(-,F)$ are exact functors on $\BIMOD(\Cscr-\Dscr)$. The corollary is now a simple application of the five-lemma. Sometimes it is convenient to use “virtual” inverse limits of bimodules. These are defined below. \[ref:3.1.9a\] Assume that $\Cscr$ is cocomplete and that $(\Nscr_n,\phi_{m,n})$ is an inverse system in $\BIMOD(\Cscr-\Dscr)$ indexed by $\NN$. We define $``{\invlim_n}{}"\Nscr_n$ by the rule $$\label{ref:3.2a} \HHom_\Dscr(``{\invlim_n}{}"\Nscr_n,\Mscr)=\dirlim_n \HHom_\Dscr(\Nscr_n,\Mscr)$$ for all $\Mscr\in \Dscr$. An inverse system $(\Nscr_n,\phi_{m,n})$ such that $``{\invlim_n}{}"\Nscr_n=0$ is called a torsion inverse system. Assume that $\Cscr$ has exact direct limits and $\Dscr$ has enough injectives. Then $``{\invlim_n}{}"$ is exact. In particular the category of torsion inverse systems is closed under subobjects, quotients and extensions. This is trivial. \[ref:3.1.11a\] Let $(\Nscr_n,\phi_{m,n})$ be an inverse system in $\Dscr$ (viewed as a subcategory of $\MOD(\Dscr)=\BIMOD(\Ab-\Dscr)$). Then the following are equivalent 1. $(\Nscr_n,\phi_{m,n})$ is torsion. 2. For every $m$ there exists $n\ge m$ such that $\phi_{m,n}:\Nscr_n\r\Nscr_m$ is the zero map. We prove $(1)\Rightarrow (2)$ the other direction being obvious. We apply with $\Mscr=\Nscr_m$. Then the identity map $\Id_{\Nscr_m}$ must become zero in some $\Hom_{\Dscr}(\Nscr_n,\Nscr_m)$. This is exactly (2). Note that in the previous lemma $(1)\Rightarrow (2)$ holds more generally for inverse systems in $\BIMOD(\Cscr-\Dscr)$. If $\Cscr=(\Cscr,\otimes, I)$ is a monoidal category [@ML] then an algebra object in $\Cscr$ is a triple $(\Ascr,\eta,\mu)$ where $\Ascr$ is an object in $\Cscr$ equipped with two maps $\eta:I\r \Ascr$ (the unit) and $\mu:\Ascr\otimes\Ascr\r\Ascr$ (the multiplication) satisfying the usual compatibilities. It is clear that $(\BIMOD(\Dscr-\Dscr),\otimes_\Dscr,\Id_\Dscr)$ and $(\Bimod(\Dscr-\Dscr),\otimes_\Dscr,\Id_\Dscr)$ are monoidal categories so we denote the algebra objects in them respectively by $\ALG(\Dscr)$ and $\Alg(\Dscr)$. The objects in $\ALG(\Dscr)$ will be called weak algebras and those of $\Alg(\Dscr)$ will simply be called algebras. Furthermore we define $\Mod(\Ascr)$ as the category consisting of pairs $(\Mscr,h)$ where $\Mscr\in \Dscr$ and $h$ is a morphism $\Mscr\otimes_\Dscr\Ascr\r\Mscr$ in $\MOD(\Dscr)$ satisfying the usual compatibilities. Note that objects of $\ALG(\Dscr)$ are basically monads in the sense of [@ML] with some extra structure. In order to interprete these definitions more concretely we recall that $\BIMOD(\Dscr-\Dscr)=\Lscr(\Dscr,\Dscr)^{\text{opp}}$. Thus $\ALG(\Dscr)$ is equivalent with the category of coalgebra objects in $\Lscr(\Dscr,\Dscr)$. Thus if $\Ascr=(\Ascr,\eta,\mu)\in\ALG(\Dscr)$ then the unit $\eta$ is in fact a natural transformation $$\eta:\HHom_\Dscr(\Ascr,-)\r \Id_\Dscr$$ and the multiplication $\mu$ is a natural transformation $$\label{ref:3.3a} \mu:\HHom_\Dscr(\Ascr,-)\r \HHom_\Dscr(\Ascr,\HHom_\Dscr(\Ascr,-))$$ Likewise if $\Mscr=(\Mscr,h)\in\Mod(\Ascr)$ then $h$ is a natural transformation $$\label{ref:3.4a} h: \Hom_\Dscr(\Mscr,-)\r \Hom_\Dscr(\Mscr,\HHom_\Dscr(\Ascr,-))$$ Then $\bar{h}=h_\Mscr(\Id_\Mscr)$ defines a morphism $$\label{ref:3.5a} \bar{h}:\Mscr\r \HHom_\Dscr(\Ascr,\Mscr)$$ Conversely if one is given a morphism $\bar{h}:\Mscr\r \HHom_\Dscr(\Ascr,\Mscr)$ as in then the composition $$\Hom_\Dscr(\Mscr,-)\r \Hom_\Dscr(\HHom_\Dscr(\Ascr,\Mscr), \HHom_\Dscr(\Ascr,-)) \xrightarrow{\bar{h}^\ast} \Hom_\Dscr(\Mscr,\HHom_\Dscr(\Ascr,-))$$ is a natural transformation as in . Elaborating on this one obtains the following results : Let $\Ascr\in\ALG(\Dscr)$. Then $(\Mscr,h)\mapsto (\Mscr,\bar{h})$ defines an isomorphism between $\Mod(\Ascr)$ and the category of $\Ascr$-comodules where we consider $\Ascr$ as a coalgebra in $\Lscr(\Dscr,\Dscr)$. \[ref:3.1.14a\] Let $\Ascr\in \ALG(\Dscr)$. The forgetful functor $$\label{ref:3.6a} (-)_\Dscr:\Mod(\Ascr)\r\Dscr:(\Mscr,h)\mapsto \Mscr$$ has a right adjoint given by $\HHom_\Dscr(\Ascr,-)$ (with its canonical $\Ascr$-structure given by ) and furthermore if $\Ascr\in\Alg(\Dscr)$ then also has a left adjoint given by $-\otimes_\Dscr\Ascr$. $\Mod(\Ascr)$ inherits most of the good properties of $\Dscr$, as is shown in the following proposition. \[ref:3.1.15a\] Let $\Ascr=(\Ascr,\eta,\mu) \in \ALG(\Dscr)$. 1. $\Mod(\Ascr)$ is an abelian category. 2. $\Mod(\Ascr)$ possesses all colimits which exist in $\Dscr$. 3. The forgetful functor $(-)_\Dscr$ is exact, faithful and commutes with colimits. 4. $\Mod(\Ascr)$ is cogenerated by objects of the form $\HHom_\Dscr(\Ascr,\Mscr)$, $\Mscr\in\Dscr$. 5. If $E\in \Inj(\Dscr)$ then $\HHom_\Dscr(\Ascr,E)\in \Inj(\Mod(\Ascr))$. In particular if $\Dscr$ has enough injectives then so does $\Mod(\Ascr)$. 6. If $\Dscr$ has exact direct limits then so does $\Mod(\Ascr)$. Assume now in addition that $\Ascr\in\Alg(\Dscr)$ 7. $\Mod(\Ascr)$ possesses all limits that exist in $\Dscr$ and $(-)_\Dscr$ commutes with these limits. 8. $\Mod(\Ascr)$ is generated by objects of the form $\Mscr\otimes_\Dscr\Ascr$, $\Mscr\in\Dscr$. Hence if $\Dscr$ has a generator then so does $\Mod(\Ascr)$. In particular combining (6),(8) we find that if $\Ascr\in\Alg(\Dscr)$ and $\Dscr$ is a Grothendieck category then so is $\Mod(\Ascr)$. - Let $I$ be a small category and $\Mscr:I\r \Mod(\Ascr)$ a functor. We write $\Mscr(i)=(\Mscr_i,h_i)$. Then one easily verifies that $\dirlim \Mscr$ is given by the pair $(\dirlim\Mscr_i,h)$ where $\bar{h}$ is given by the composition $$\dirlim\Mscr_i\xrightarrow{\dirlim \bar{h}_i} \dirlim \HHom_\Dscr(\Ascr,\Mscr_i)\xrightarrow{\text{can}} \HHom_\Dscr(\Ascr,\dirlim\Mscr_i)$$ - From (2) it follows that $\Mod(\Ascr)$ has cokernels. So we have to show that $\Mod(\Ascr)$ has kernels. Let $f:(\Mscr,h)\r (\Nscr,j)$ be a morphism in $\Mod(\Ascr)$. Then $\ker f$ is the pair $(\Kscr,s)$ where $\Kscr=\ker(\Mscr\r\Nscr)$ and $\bar{s}$ is as in the following commutative diagram with exact rows. $$\begin{CD} 0 @>>> \Kscr @>>> \Mscr @>f>>\Nscr\\ @. @V\bar{s}VV @V\bar{h}VV @V\bar{j}VV\\ 0 @>>> \HHom_\Dscr(\Ascr,\Kscr) @>>> \HHom_\Dscr(\Ascr,\Mscr) @>f>> \HHom_\Dscr(\Ascr,\Nscr) \end{CD}$$ - This follows from the explicit constructions of kernels and colimits in (1) and (2). - Assume that $(\Mscr,h)\in \Mod(\Ascr)$. The composition $$\Mscr \xrightarrow{\bar{h}} \HHom_\Dscr(\Ascr,\Mscr) \xrightarrow{\eta} \Mscr$$ is the identity so $\bar{h}$ is a monomorphism. We know already that $\HHom_\Dscr(\Ascr,\Mscr)$ has a canonical structure as $\Ascr$-module and it is easy to see that $\bar{h}$ is compatible with it. - This follows from the fact that $\HHom_\Dscr(\Ascr,-)$ has a left adjoint which is exact by (3). - This follows from the explicit construction of kernels, cokernels and colimits in (1)(2). - Let $\Mscr:I\r \Mod(\Ascr)$ be as in (2) Then $\invlim \Mscr$ is the pair $(\invlim\Mscr_i, {h}')$ where $\bar{h}'$ is the composition $$\invlim \Mscr_i \xrightarrow{\invlim \bar{h}_i} \invlim \HHom_\Dscr(\Ascr,\Mscr_i) \xrightarrow {\text{can}^{-1}} \HHom_\Dscr(\Ascr,\invlim\Mscr_i)$$ where we have used the fact that $\Ascr$ preserves products. - If $(\Mscr,h)\in\Mod(\Ascr)$ then $\Mscr\otimes_\Dscr\Ascr\in\Mod(\Ascr)$ by lemma \[ref:3.1.14a\]. Furthermore the composition $$\Mscr\xrightarrow{\eta} \Mscr\otimes_\Dscr \Ascr \xrightarrow{h} \Mscr$$ is the identity so $\Mscr\otimes_\Dscr \Ascr\r \Mscr$ is an epimorphism. Assume now that $f:\Ascr\r\Bscr$ is a morphism in $\ALG(\Dscr)$. Let $\Mscr=(\Mscr,h)\in \Mod(\Ascr)$, $\Nscr=(\Nscr,j)\in \Mod(\Bscr)$. Then we define $\Nscr_\Ascr\in\Mod(\Ascr)$ as the pair $(\Nscr,j')$ were $\bar{j}'$ is the composition $$\Nscr \xrightarrow {\bar{j}} \HHom_\Dscr(\Bscr,\Nscr) \xrightarrow {f^\ast} \HHom_\Dscr(\Ascr,\Nscr)$$ $\HHom_\Ascr(\Bscr,\Mscr)\in\Mod(\Bscr)$ is the pair $(\Uscr,u)$ where $\Uscr$ is obtained as the equalizer of $$\label{ref:3.7a} \begin{CD} \HHom_\Dscr(\Bscr,\Mscr) @>\text{mult}^\ast >> \HHom_\Dscr(\Bscr\otimes_\Dscr\Bscr, \Mscr)\\ @V \bar{h} VV @V (1\otimes f)^\ast VV\\ \HHom_\Dscr(\Bscr,\HHom_\Dscr(\Ascr,\Mscr)) @= \HHom_\Dscr(\Bscr\otimes_\Dscr\Ascr,\Mscr) \end{CD}$$ and $\bar{u}$ is obtained from the fact that all objects in carry a $\Bscr$-structure, and the maps are compatible with it. Finally if $\Ascr,\Bscr\in \Alg(\Dscr)$ then $\Mscr\otimes_\Ascr\Bscr$ is defined as the coequalizer of $$\Mscr\otimes_\Dscr\Ascr\otimes_\Dscr\Bscr \mathop{\Longrightarrow}\limits^{\bar{h}\otimes 1}_{(1\otimes\text{mult}) \circ (1\otimes f\otimes 1)} \Mscr\otimes_\Dscr\Bscr$$ So, summarizing, we have constructed the standard functors $$\begin{aligned} (-)_\Ascr&:\Mod(\Bscr)\r \Mod(\Ascr)\label{ref:3.8a}\\ \HHom_\Ascr(\Bscr,-)&:\Mod(\Ascr)\r \Mod(\Bscr) \label{ref:3.9a}\end{aligned}$$ and if $\Ascr, \Bscr\in \Alg(\Dscr)$ $$\label{ref:3.10a} -\otimes_\Ascr\Bscr :\Mod(\Ascr) \r \Mod(\Bscr)$$ In general is a right adjoint to and if is defined then it is a left adjoint to . From the constructions of limits, colimits, kernels and cokernels in (the proof of) Proposition \[ref:3.1.15a\] one verifies that have the standard exactness properties and satisfy the usual compatibilities with (co)limits. For $\Ascr,\Bscr\in \ALG(\Dscr)$ we define $$\begin{aligned} \BIMOD(\Ascr-\Bscr)&=\BIMOD(\Mod(\Ascr)-\Mod(\Bscr))\\ \Bimod(\Ascr-\Bscr)&=\Bimod(\Mod(\Ascr)-\Mod(\Bscr))\end{aligned}$$ We denote $-\otimes_{\Mod(\Bscr)}-$ by $-\otimes_\Bscr-$. In general we will replace in our notations $\Mod(\Xscr)$ by $\Xscr$ when no confusion can arise. Assume that $\Ascr\r \Ascr'$, $\Bscr\r \Bscr'$ are morphisms in $\ALG(\Dscr)$. We will now define functors $$\begin{aligned} \Ascr'\otimes_\Ascr-\otimes_\Bscr\Bscr':\BIMOD(\Ascr-\Bscr)\r \BIMOD(\Ascr'-\Bscr') \label{ref:3.11a}\\ {}_\Ascr(-)_\Bscr:\BIMOD(\Ascr'-\Bscr')\r \BIMOD (\Ascr-\Bscr) \label{ref:3.12a}\end{aligned}$$ as we did for modules. Let $\Mscr\in\BIMOD(\Ascr-\Bscr)$, $\Nscr\in\BIMOD(\Ascr'-\Bscr')$. Then we define $$\HHom_{\Bscr'}(\Ascr'\otimes_\Ascr\Mscr\otimes_\Bscr\Bscr',-)\overset{ \text{def}}{=} \HHom_\Ascr(\Ascr',\HHom_\Bscr(\Mscr,(-)_\Bscr))$$ $$\label{ref:3.13a} \HHom_\Bscr({}_\Ascr\Nscr_\Bscr,-)\overset{ \text{def}}{=} \HHom_{\Bscr'}(\Nscr,\HHom_\Bscr(\Bscr',-))_\Ascr$$ The functor $-\otimes_{\Ascr'}\Ascr'\otimes_\Ascr\Mscr\otimes_\Bscr\Bscr'$ should be given by $$(-)_\Ascr\otimes_\Ascr\Mscr\otimes_\Bscr\Bscr'$$ and the functor $-\otimes_\Ascr{}_\Ascr\Nscr_\Bscr$ should be given by $$\label{ref:3.14a} ((-\otimes_\Ascr\Ascr')\otimes_{\Ascr'}\Nscr)_\Bscr$$ So we conclude that if $\Mscr\in \Bimod(\Ascr-\Bscr)$ and $\Nscr\in\Bimod(\Ascr'-\Bscr')$ then if $\Bscr'\in \Alg(\Dscr)$ then respects “Bimod” and if $\Ascr'\in \Alg(\Dscr)$ then respects “Bimod”. Our definition of $\BIMOD(\Ascr-\Bscr)$ has the advantage that we can directly apply Proposition \[ref:3.1.1a\] to obtain the properties of this category. However we would also like to have a definition which resembles more closely that of modules. Therefore we state the following proposition. The following categories are equivalent. 1. $\BIMOD(\Ascr-\Bscr)$ 2. The category of triples $(\Mscr,h,h')$ where $\Mscr\in\BIMOD(\Dscr-\Dscr)$ and $h:\Ascr\otimes_\Dscr \Mscr\r \Mscr$, $h':\Mscr\otimes_\Dscr\Bscr\r \Mscr$ are maps in $\BIMOD(\Dscr-\Dscr)$ satisfying the usual compatibilities. If $\Ascr,\Bscr\in \Alg(\Dscr)$ then in the previous statement, “$\BIMOD$” may be replaced by “$\Bimod$”. This is a rather tedious verification which we leave to the reader. Let us simply state how one associates a left exact functor $\HHom_\Bscr(\Mscr,-):\Mod(\Bscr)\r \Mod(\Ascr)$ to a triple $(\Mscr,h,h')$. Let $\Nscr=(\Nscr,p)\in \Mod(\Bscr)$. Then $\HHom_\Bscr(\Mscr,\Nscr)$ will be the equalizer of $$\label{ref:3.15a} \begin{CD} \HHom_\Dscr(\Mscr,\Nscr) @= \HHom_\Dscr(\Mscr,\Nscr)\\ @V h^{\prime\ast}VV @V \bar{p} VV\\ \HHom_\Dscr(\Mscr\otimes_\Dscr\Bscr,\Nscr) @=\HHom_\Dscr(\Mscr, \HHom_\Dscr(\Bscr,\Nscr)) \end{CD}$$ The $\Ascr$-structure on $\HHom_\Bscr(\Mscr,\Nscr)$ is obtained from the fact that all objects in carry a canonical $\Ascr$-structure and the maps are compatible with it. Now let $\Ascr\r \Bscr$ be a map in $\ALG(\Dscr)$. From we obtain a functor $$-\otimes_{\Ascr}\Bscr:\MOD(\Ascr)\r \MOD(\Bscr)$$ Assume that $\Mod(\Ascr)$ and $\Mod(\Bscr)$ have enough injectives. Then we define associated functors $$\HTor_i^\Ascr(-,\Bscr):\MOD(\Ascr)\r\MOD(\Bscr)$$ by $$\Hom_\Bscr(\HTor_i^\Ascr(-,\Bscr),E) = \Ext^i_\Ascr(-,E_\Ascr)$$ for injectives $E\in\Mod(\Bscr)$. One easily verifies that $$\HTor_i^\Ascr(-,\Bscr)_\Ascr=\HTor_i^\Ascr(-,\Bscr_\Ascr)$$ \[ref:3.1.17a\] Assume that $\Dscr$ has colimits and that $\Mod(\Bscr)$ and $\Mod(\Ascr)$ are locally noetherian. Then $\HTor_i^\Ascr(-,\Bscr)$ sends coherent objects in $\MOD(\Ascr)$ to coherent objects in $\MOD(\Bscr)$. This is a formal verification. We will give the proof as an illustration of how the various hypotheses are used. Assume that $\Mscr\in\MOD(\Ascr)$ is coherent and let $(E_i)_i$ be a directed system of injectives in $\Mod(\Bscr)$. By Proposition \[ref:3.1.7a\] we have to show that $$\Hom_\Bscr(\HTor_i^\Ascr(\Mscr,\Bscr),\dirlim_i E_i) = \dirlim_i \Hom_\Bscr(\HTor_i^\Ascr(\Mscr,\Bscr),E_i)$$ (taking into account that $\Ab$ has exact direct limits). From the fact that $\Mod(\Bscr)$ is locally noetherian it follows that $\dirlim_i E_i$ is injective. Hence we have to show that $$\Ext^i_\Ascr(\Mscr,(\dirlim_i E_i)_\Ascr)= \dirlim_i \Ext^i_\Ascr(\Mscr,E_{i\Ascr})$$ The fact that $\Mod(\Ascr)$ is locally noetherian and that $\Mscr$ is coherent implies that the righthand side of this equation is equal to $ \Ext^i_\Ascr(\Mscr,\dirlim_i E_{i\Ascr})$. So it remains to show that $(\dirlim_i E_i)_\Ascr=\dirlim_i E_{i\Ascr}$. This follows easily from the explicit construction of $\dirlim$ in the proof of Proposition \[ref:3.1.15a\].2 and the definition of $(-)_\Ascr$. Graded modules, bimodules and algebras {#ref:3.2b} -------------------------------------- In this section we have stated all our definitions in the ungraded case but we will mainly need them in the graded case. Luckily the generalization is trivial. If $\Dscr$ is an abelian category then we denote by $\tilde{\Dscr}$ the category of $\ZZ$-graded objects over $\Dscr$. Thus by definition an object in $\tilde{\Dscr}$ is a sequence of objects $(\Mscr_n)_n$ in $\Dscr$ and $\Hom_{\tilde{\Dscr}}((\Mscr_n)_n,(\Nscr_n)_n)= \prod_n\Hom_\Dscr(\Mscr_n,\Nscr_n)$. We identify an object $\Mscr\in\Dscr$ with the object of $\tilde{\Dscr}$ which is $\Mscr$ in degree zero and zero elsewhere. As usual $\tilde{\Dscr}$ is equipped with the shiftfunctor $\Mscr\mapsto \Mscr(1)$ where $\Mscr(1)_n=\Mscr_{n+1}$. Of course $\Mscr(m)$, $m\in\ZZ$ is defined similarly. To simplify the notation we will often write $\oplus_n\Mscr_n$ for $(\Mscr_n)_n$. By analogy with the ungraded case we define $$\BIGR(\Cscr-\Dscr)=\{\text{left exact functors $\tilde{\Dscr}\r\tilde{\Cscr}$ commuting with shift}\}^{\text{opp}}$$ $$\Bigr(\Cscr-\Dscr)=\{\Mscr\in\BIGR(\Cscr-\Dscr)\mid\Mscr\text{ has a left adjoint}\}$$ and $$\GR(\Cscr)=\BIGR(\Ab-\Cscr)$$ Objects in $\BIGR(\Cscr-\Dscr)$ will be called graded weak $\Cscr$-$\Dscr$ bimodules and objects in $\Bigr(\Cscr-\Dscr)$ will be called graded bimodules. We will denote the left exact functor corresponding to $\Mscr \in \BIGR(\Cscr-\Dscr)$ with $\underline{\HHom}_\Dscr(\Mscr,-)$ in order to avoid confusion with the ungraded case. Similar conventions will apply to the use of $\underline{\HExt}$, $\underline{\HTor}$. Note that if $\Mscr\in \BIGR(\Cscr-\Dscr)$ (resp. $\Bigr(\Cscr-\Dscr)$) then the composition (for $n\in\ZZ$) $$\Dscr\hookrightarrow \tilde{\Dscr}\xrightarrow{\Mscr} \tilde{\Cscr} \xrightarrow{ \text{degree $-n$}} \Cscr$$ defines objects $\Mscr_{n}$ in $\BIMOD(\Cscr-\Dscr)$ (resp. $\Bimod(\Cscr-\Dscr)$) which determine $\Mscr$. We write $\Mscr=\oplus_n \Mscr_n$. Obviously $\BIGR(\Dscr-\Dscr)$, $\Bigr(\Dscr-\Dscr)$ are monoidal categories and we will denote the algebra objects in them by $\GRALG(\Dscr)$ resp. $\Gralg(\Dscr)$. Consistent with our earlier coventions we speak respectively of weak algebras and algebra. It is easy to see that $\Ascr\in \GRALG(\Dscr)$ is of the form $\Ascr=\oplus_n\Ascr_n$ with multiplication maps $\Ascr_m\otimes_\Dscr\Ascr_n\r\Ascr_{m+n}$ and a unit $\Id_\Dscr\r \Ascr_0$ satisfying the usual compatibilities. If $\Ascr\in \GRALG(\Dscr)$ then $\Gr(\Ascr)$ is defined as $\Mod(\Ascr)$ but with $\Dscr$ replaced by $\tilde{\Dscr}$. Finally if $\Ascr,\Bscr\in\GRALG(\Dscr)$ then $$\begin{aligned} \Bigr(\Ascr-\Bscr)&=\BIGR(\Gr(\Ascr)-\Gr(\Bscr))\\ \Bigr(\Ascr-\Bscr)&=\Bigr(\Gr(\Ascr)-\Gr(\Bscr))\end{aligned}$$ In the sequel we will freely use graded versions of ungraded results stated in this section. Algebras which are strongly graded with respect to a Serre subcategory {#ref:3.3b} ---------------------------------------------------------------------- We put this section here for lack of a better place. The results will be used in §\[ref:3.11b\]. Let $\Dscr$ be an abelian category and let $\Ascr\in\Gralg(\Dscr)$. Put $\Qscr_n=\coker (\Ascr_n\otimes_\Dscr\Ascr_{-n}\r \Id_\Dscr)$. It is clear that $\Qscr_n\in\Bimod(\Dscr-\Dscr)$. We call $\Ascr$ strongly graded if $\Qscr_n=0$ for all $n$. By copying the proof from the ring case [@NVO Thm. I.3.4] one easily shows that the functors $(-)_0$ and $-\otimes_{\Ascr_0}\Ascr$ define inverse equivalences between $\Gr(\Ascr)$ and $\Mod(\Ascr_0)$. Let $\Sscr$ be as Serre subcategory of $\Mod(\Ascr_0)$. Define $\Sscr(\Ascr)$ as the full subcategory of objects $\Tscr$ in $\Gr(\Ascr)$ such that $\Tscr_n$ is in $\Sscr$ for every $n$. It is clear that $\Sscr(\Ascr)$ is a Serre subcategory of $\Gr(\Ascr)$. We say that $\Ascr$ is strongly graded with respect to $\Sscr$ if $-\otimes_{\Ascr_0}\Qscr_n$ sends $\Mod(\Ascr_0)$ to $\Sscr$ and if $-\otimes_{\Ascr_0}\Ascr$ sends $\Sscr$ to $\Sscr(\Ascr)$. Then one has the following: \[ref:3.3.1a\] Assume that $\Ascr$ is strongly graded with respect to $\Sscr$. Then the functors $(-)_0$ and $-\otimes_{\Ascr_0}\Ascr$ factor over $\Sscr(\Ascr)$ and $\Sscr$ and in this way they define inverse equivalences between $\Gr(\Ascr)/\Sscr(\Ascr)$ and $\Mod(\Ascr_0)/\Sscr$. For the convenience of the reader we copy the proof of [@NVO Thm. I.3.4] suitably modified. If $\Mscr\in\Gr(\Ascr)$ and if $\Pscr_n$ is a submodule of $\Mscr_n$ then we define $\Pscr_n\Ascr_m$ as the image of $\Pscr_n\otimes_\Dscr\Ascr_m$ in $\Mscr_{m+n}$. We show that $\Mscr=\Mscr_0\otimes_{\Ascr_0}\Ascr$ modulo $\Sscr(\Ascr)$. Working modulo $\Sscr$ we have the following. $$\label{ref:3.16a} \Mscr_{m+n}=\Mscr_{m+n}\Ascr_0=\Mscr_{m+n}\Ascr_{-m}\Ascr_{m}\subset \Mscr_n\Ascr_m\subset\Mscr_{m+n}$$ The second equality follows from $$\Mscr_{m+n}/\Mscr_{m+n}\Ascr_{-m}\Ascr_{m}=\coker (\Mscr_{m+n}\otimes_\Dscr\Ascr_{-m}\otimes_\Dscr\Ascr_{m}\r \Mscr_{m+n})=\Mscr_{m+n}\otimes_\Dscr\Qscr_{-m}$$ yields $\Mscr_n\Ascr_m=\Mscr_{m+n}$ modulo $\Sscr$. Now we claim that the multiplication map $$\mu:\Mscr_0\otimes_{\Ascr_0}\Ascr\r \Mscr$$ is in fact an isomorphism modulo $\Sscr(\Ascr)$. By what we have done so far $\mu$ is clearly surjective modulo $\Sscr(\Ascr)$. Let $\Kscr$ be the kernel of $\mu$. We have $\Kscr_0=0$ and hence modulo $\Sscr$, $\Kscr_n=\Kscr_0\Ascr_n=0$. Thus $\Kscr\in\Sscr(\Ascr)$ and we are done. We can now finish the proof, provided we can show that the functor $-\otimes_{\Ascr_0}\Ascr$ is well defined on $\Mod(\Ascr_0)/\Sscr$. Inspection shows that we only have to check the following case : assume that $\Mscr\r \Nscr$ is injective in $\Mod(\Ascr_0)$. Then $\Kscr=\ker (\Mscr\otimes_{\Ascr_0}\Ascr\r \Nscr\otimes_{\Ascr_0}\Ascr)$ should be in $\Sscr(\Ascr)$. Since $\Kscr_0=0$ this follows from the previous discussion. Quotients of the identity functor {#ref:3.4b} --------------------------------- This section is related to [@rosenberg]. $\Cscr$, $\Dscr$, $\Escr$ will be abelian categories having enough injectives so that the weak bimodule categories are abelian (cf Prop. \[ref:3.1.1a\]). Let $\Bscr$ be a quotient of $\Id_\Dscr$ in $\ALG(\Dscr)$ (by this we mean that the underlying $\Dscr$-bimodule map is an epimorphism). We will denote the functor $(-)_\Dscr$ by $i_\ast$ and its right adjoint $\HHom_\Dscr(\Bscr,-)$ by $i^!$. If $\Bscr\in\Alg(\Dscr)$ then we denote $-\otimes_\Dscr\Bscr$ by $i^\ast$. This is the left adjoint to $i_\ast$. \[ref:3.4.1a\] 1. $i^!i_\ast=\Id_{\Mod(\Bscr)}$ 2. $i_\ast$ is fully faitful. 3. Let $\Mscr\in\Dscr$. The counitmap $i_\ast i^!\Mscr\r \Mscr$ is injective. Furthermore $\Mscr\in i_\ast\Mod(\Bscr)$ if and only if the counit map is an isomorphism. 4. $i_\ast\Mod(\Bscr)$ is closed under subquotients. 5. If $\Bscr\in\Alg(\Dscr)$ then $i^\ast i_\ast=\Id_{\Mod(\Bscr)}$. <!-- --> 1. Let $\Mscr\in\Mod(\Bscr)$. The canonical map $\Mscr\r\HHom_\Dscr(\Bscr,\Mscr)$ (cfr. ) is the unit for the adjunction $(i_\ast,i^!)$. Thus we have to show that it is an isomorphism. Since $\Bscr$ is a quotient of $\Id_\Bscr$ we can consider the commutative diagram $$\begin{CD} \Mscr @>>> \HHom_\Dscr(\Bscr,\Mscr)\\ @\| @VVV\\ \Mscr @= \HHom_\Dscr(\Id_\Dscr,\Mscr) \end{CD}$$ where the right and upper maps are injective (see Prop.3.1.10.4). This implies that the upper map is an isomorphism. 2. This is formally a consequence of (1). Indeed if $\Mscr,\Nscr\in\Mod(\Bscr)$ then $$\Hom_\Dscr(i_\ast\Mscr,i_\ast\Nscr)=\Hom_\Bscr(\Mscr,i^!i_\ast\Nscr)= \Hom_\Bscr(\Mscr,\Nscr)$$ 3. The counit map is given by the composition $$i_\ast i^!\Mscr=\HHom_\Dscr(\Bscr,\Mscr)\hookrightarrow \HHom_\Dscr(\Id_\Dscr,\Mscr)=\Mscr$$ so it is certainly injective. If it is an isomorphism then clearly $\Mscr=i_\ast(i^!\Mscr)\in i_\ast\Mod(\Bscr)$. Conversely if $\Mscr=i_\ast\Nscr$ then $i_\ast i^! \Mscr =i_\ast i^! i_\ast\Nscr=i_\ast\Nscr=\Mscr$. 4. Since $i_\ast$ is left exact $i_\ast\Mod(\Bscr)$ is closed under kernels. So it suffices to show that $i_\ast\Mod(\Bscr)$ is closed under quotients. Let $\Mscr\in i_\ast\Mod(\Bscr)$ and let $\Nscr$ be a quotient of $\Mscr$ in $\Dscr$. Then we have a commutative diagram $$\begin{CD} i_\ast i^!\Mscr@>\alpha>> \Mscr\\ @VVV @V\beta VV\\ i_\ast i^! \Nscr@>\gamma >>\Nscr \end{CD}$$ where $\alpha$ is an isomorphism, $\beta$ is surjective and $\gamma$ is injective. This implies that $\gamma$ is an isomorphism. 5. This is proved similarly as (1). Let us call a full subcategory of an abelian category closed if it is closed under subquotients and if the inclusion functor has a right adjoint. In most cases this is equivalent to the definition in [@Gabriel IV.4]. Let us call a closed subcategory biclosed if the inclusion functor also has a left adjoint. Proposition \[ref:3.4.1a\] basically tells us that the functor $\Bscr\r i_\ast\Mod(\Bscr)$ associates with every quotient of $\Id_\Dscr$ a closed subcategory of $\Dscr$ and if the quotient is in $\Alg(\Dscr)$ then the resulting category is biclosed. We will now show that the converse to this essentially holds. Assume that $\Escr\subset\Dscr$ is a closed subcategory. Let $i_\ast$ be the inclusion functor and $i^!$ its right adjoint. Let $\Bscr$ be the left exact functor $i_\ast i^!$. Then the counit $i_\ast i^!\r \Id_\Dscr$ and the comultiplication $i_\ast i^!=i_\ast\Id_\Dscr i^!\r i_\ast i^!i_\ast i^!$ make $\Bscr$ into a weak algebra (remember that a “weak algebra” is actually a coalgebra object in the the monoidal category of left exact functors, see the discussion after Prop. \[ref:3.1.5a\]). Furthermore the counit $i_\ast i^!\r \Id_\Dscr$ makes $\Bscr$ into a quotient of $\Id_\Dscr$ (again remember that arrows are reversed if we pass from functors to bimodules). Now by a routine verification (or using an appropriate version of Beck’s theorem [@ML]) one shows that $\Mod(\Bscr)\cong \Escr$. Elaborating on this one may prove the following result. \[ref:3.4.2a\] With notations as above. The functors $\Bscr\mapsto i_\ast \Mod(\Bscr)$ and $\Escr \mapsto i_\ast i^!$ induce inverse bijections between the quotients of $\Id_\Dscr$ in $\ALG(\Dscr)$ (resp. $\Alg(\Dscr)$) and the closed (resp. biclosed) subcategories in $\Dscr$. Below we state a few elementary results concerning closed and biclosed categories which will be useful in the sequel. Recall that a category is said to be well-powered if the set of subobjects of an arbitrary object is small. In an abelian category this holds if there is a generator or a cogenerator. \[ref:3.4.3a\] Let $\Escr$ be a full subcategory of the abelian category $\Dscr$. 1. If $\Escr$ is closed in $\Dscr$ and if $\Dscr$ is cocomplete, then so is $\Escr$. 2. If $\Dscr$ is cocomplete and well-powered and if $\Escr$ is closed under subquotients and direct sums (in $\Dscr$) then $\Escr$ is closed. 3. If $\Escr$ is biclosed and if $\Dscr$ is complete then so is $\Escr$. 4. If $\Dscr$ is complete and well-powered and if $\Escr$ is closed and is closed under products (in $\Dscr$) then $\Escr$ is biclosed. Let $i_\ast:\Escr\r\Dscr$ be the inclusion functor and let $i^!$, $i^\ast$ be the right and left adjoint to $i_\ast$ if they exist. 1. It is sufficient to show that $\Escr$ is closed under direct sums. Let $\oplus_{j\in J} M_j$ be such a direct sum. By construction $i^!(\oplus_{j\in J} M_j)$ is the maximal subobject of $\oplus_{j\in J} M_j$ contained in $\Escr$. Hence for all $j$, $M_j\subset i^!(\oplus_{j\in J} M_j)$. But this implies $i^!(\oplus_{j\in J} M_j)=\oplus_{j\in J} M_j$ and thus $\oplus _{j\in J} M_j\in \Escr$. 2. Since the set of subobjects of an object $M$ is small and since $\Escr$ is closed under (small) direct unions, $M$ has a largest subobject $N$ which lies in $\Escr$. The assignment $M\mapsto N$ is the right adjoint to the inclusion functor. 3. This is proved by an argument dual to (1). 4. This is proved by an argument dual to (2). \[ref:3.4.4a\] If $\Dscr$ is a closed subcategory of $\Cscr$ and $\Escr$ is a closed subcategory of $\Dscr$ then $\Escr$ is closed in $\Cscr$. A similar statement holds if we replace “closed” by “biclosed”. This follows from the fact that adjoint functors are compatible with composition. Let us recall the definition of the Gabriel product. If $\Escr_1,\Escr_2$ are full subcategories of an abelian category $\Dscr$ then $\Escr_1\cdot\Escr_2$ is the full subcategory of $\Dscr$ whose objects are given by middle terms of exact sequences $$0\r M_2\r M\r M_1 \r 0$$ with $M_1\in \Escr_1$, $M_2\in \Escr_2$. It is easy to see that if $\Escr_1$, $\Escr_2$ are closed under subquotients then so is $\Escr_1\cdot \Escr_2$. \[ref:3.4.5a\] [@rosenberg] If $\Escr_1,\Escr_2\subset \Dscr$ are (bi)closed then so is $\Escr_1\cdot\Escr_2$. We give the proof for closedness. Biclosedness is similar. Let $i_{1\ast}$, $i_{2\ast}$, $i_{12\ast}$ be the embeddings $\Escr_1\subset \Dscr$, $\Escr_2\subset \Dscr$, $\Escr_1\cdot\Escr_2\subset \Dscr$. One has to construct the right adjoint to $i_{12\ast}$. Let $q:\Mscr\r \Mscr/i^!_2\Mscr$ be the quotient map. Then one verifies that $$i^!_{12}\Mscr \overset{\text{def}}{=} q^{-1} (i^!_1(\Mscr/i^!_2\Mscr))$$ has the required properties. Ideals in the identity functor {#ref:3.5b} ------------------------------ In this section $\Cscr$, $\Dscr$, $\Escr$ will be abelian categories having enough injectives. If $\Ascr\in \ALG(\Dscr)$ (resp. in $\Alg(\Dscr)$) then a weak ideal (resp. an ideal) in $\Ascr$ is a subobject of ${}_\Ascr \Ascr_\Ascr$ in $\BIMOD(\Ascr-\Ascr)$ (resp. in $\Bimod(\Ascr-\Ascr)$). If $I$, $J$ are weak ideals in $\Ascr\in \ALG(\Dscr)$ then we define the weak ideal $IJ$ as the image of the composition $$I\otimes_\Ascr J\r \Ascr\otimes_\Ascr\Ascr=\Ascr$$ If $f:\Ascr\r\Bscr$ is a morphism in $\ALG(\Dscr)$ then by $\ker f$ we denote $\ker ({}_\Dscr\Ascr_\Dscr\r {}_\Dscr\Bscr_\Dscr)$. Since $\ker f$ is canonically an object in $\BIMOD(\Ascr-\Ascr)$ we find that $\ker f$ is a weak ideal in $\Ascr$. We will say that $f$ is surjective if the underlying bimodule map ${}_\Dscr\Ascr_\Dscr\r {}_\Dscr \Bscr_\Dscr$ is surjective. If this is the case then we call the pair $(\Bscr,f)$ a quotient object of $\Ascr$ in $\ALG(\Dscr)$. If $I$ is a weak ideal in $\Ascr$ then there is a unique weak algebra structure on $\Ascr/I$ which makes the quotient map $q:\Ascr\r \Ascr/I$ into a morphism in $\ALG(\Dscr)$. $I\mapsto (\Ascr/I,q)$ and $(\Bscr,f)\mapsto \ker f$ define inverse bijections between 1. Subobject of ${}_\Ascr\Ascr_\Ascr$ in $\BIMOD(\Ascr-\Ascr)$ 2. Isomorphism classes of quotient object of $\Ascr$. We leave this verification to the reader. The lemma can of course be proved more generally in the setting of monoidal categories $(\Cscr,\otimes,I)$ where $\Cscr$ is abelian and $\otimes$ is right exact. Recall that $\BIMOD(\Ascr-\Ascr)$ was defined as $\BIMOD(\MOD(\Ascr)-\MOD(\Ascr))$ and similarly for “$\Bimod$”. Hence there is a 1-1 correspondence between (weak) ideals in $\Ascr$ and (weak) ideals in $\Id_{\Mod(\Ascr)}$ and furthermore this correspondence is compatible with products. Hence to simplify the notation below we will replace $\Mod(\Ascr)$ by $\Dscr$ and $\Ascr$ by $\Id_\Dscr$. If $M\in \Mod(\Dscr)$ and $I$ is a weak ideal in $\Id_\Dscr$ then we define $M_I$ as the quotient object of $M$ given by the image of $M\r \HHom_\Dscr(I,M)$. One quickly verifies : 1. $M\in \Mod(\Id_\Dscr/I)$ iff $M_I=0$. 2. $(M_J)_I=M_{IJ}$. \[ref:3.5.4a\] Let $I,J$ be weak ideals in $\Id_\Dscr$. Then $$\Mod(\Id_\Dscr/IJ)=\Mod(\Id_\Dscr/I)\cdot \Mod(\Id_\Dscr/J)$$ - Let $M\in \Mod(\Id_\Dscr/IJ)$. Then $M_{IJ}=0$ and hence $(M_J)_I=0$. Thus $M_J\in \Mod(\Id_\Dscr/I)$. Since by construction $\ker (M\r M_J)\in \Mod(\Id_\Dscr/J)$ we are done. - Let $M\in \Mod(\Id_\Dscr/I)\cdot \Mod(\Id_\Dscr/J)$. Then there is an exact sequence $$0\r M_2\r M\r M_1\r 0$$ with $M_2\in \Mod(\Id_\Dscr/J)$, $M_1\in \Mod(\Id_\Dscr/I)$. In the commutative diagram $$\begin{CD} M_2 @>>> M\\ @VVV @VVV\\ (M_2)_J @>>> M_J \end{CD}$$ we have $(M_2)_J=0$ and thus the composition $M_2\r M \r M_J$ is zero. In other words, $M\r M_J$ factors through $M_1$. Thus $M_J$ is a quotient of $M_1$ and hence $M_{IJ}=(M_J)_I=0$. So $M\in \Mod(\Id_\Dscr/IJ)$. Our next aim is to characterize the ideals in $\Id_\Dscr$ in terms of certain subcategories of $\Dscr$. The following is a trivial consequence of Prop. \[ref:3.1.1a\](6) and Prop. \[ref:3.4.2a\]. \[ref:3.5.5a\] Assume that $\Dscr$ is complete and has exact direct products and an injective cogenerator. Then a weak ideal in $\Id_\Dscr$ is an ideal if and only if $\Mod(\Id_\Dscr/I)$ is biclosed. Hence our problems lie in the cases where products are not exact. For simplicity we assume below that $\Dscr$ is complete and has an injective cogenerator. Let us first recall the definition of the derived functors of the product functor. Assume that $J$ is some index set and let $\prod_{j\in J}\Dscr$ be the category whose objects consist of all families $(M_j)_{j\in J}$ with $M_j\in \Dscr$. Then we denote by $\prod$ the functor $$\prod :\prod_{j\in J}\Dscr \r \Dscr:(M_j)_{j\in J}\mapsto \prod_{j\in J}M_j$$ which is obviously left exact. We denote the derived functors of $\prod$ by $(R^i\prod)_i$. These functors are computed in the standard way. Let $(M_j)_{j\in J}$ be in $\prod_{j\in J}\Dscr$. One takes injective resolutions $M_j\r E_j^\cdot$ and then one has $$R^i\prod (M_j)_{j\in J}=H^i\bigl(\prod_{j\in J} E_j^\cdot\bigr)$$ A biclosed subcategory $\Escr\subset \Dscr$ is called well-closed if for all families $(E_j)_{j\in J}$ whose members are injective in $\Escr$ (but not necessarily in $\Dscr$) one has $R^1\prod (E_j)_j=0$. \[ref:3.5.7a\] Let $\Dscr$ be as above. That is, $\Dscr$ is complete and has an injective cogenerator. Let $I\subset \Id_\Dscr$ be a weak ideal. Then $I$ is an ideal if and only if $\Mod(\Id_\Dscr/I)$ is well-closed. - Since $I$ is an ideal, $\Id_\Dscr/I\in \Alg(\Dscr)$. Hence $\Mod(\Id_\Dscr/I)$ is certainly biclosed. Now let $(E_j)_{j\in J}$ be a family on injectives in $\Escr$ and let $(F_j)_{j\in J}$ be the corresponding injective hulls in $\Dscr$, Since $F_j$ is an essential extension of $E_j$ and $E_j$ is injective we have $$\HHom_\Dscr(\Id_\Dscr/I,F_j)=E_j$$ Applying $\HHom_\Dscr(-,\prod_j F_j)$ to the exact sequence in $\Bimod(\Dscr-\Dscr)$ $$\label{ref:3.17a} 0\r I\r \id_\Dscr\r \Id_\Dscr/I\r 0$$ yields an exact sequence $$\label{ref:3.18a} 0\r \prod_j E_j \r \prod_j F_j\r\prod_j \HHom_\Dscr(I,F_j)\r 0$$ which is obviously the direct product of the short exact sequence $$\label{ref:3.19a} 0\r E_j \r F_j \r \HHom_\Dscr(I,F_j)\r 0$$ obtained from applying $\HHom(-,F_j)$ to . Now applying the long exact sequence for $(R^i\prod)_i$ to yields that $R^1\prod (E_j)_j=0$. - According to Prop. \[ref:3.1.1a\](5) it suffices to show that $\HHom_\Dscr(I,-)$ commutes with products, when evaluated on injectives. Let $(F_j)_j$ be a family of injectives in $\Dscr$ and put $E_j=\HHom_\Dscr(\Id_{\Dscr}/I,F_j)$. We have again the short exact sequences given by and using the fact that $R^1\prod (E_j)_j=0$ we obtain the short exact sequence from the long exact sequence for $(R^i\prod)_i$. On the other hand if we apply $\HHom_\Dscr(-,\prod (F_j)_j)$ to the exact sequence we obtain $$\label{ref:3.20a} 0\r \prod_j E_j \r \prod_j F_j \r \HHom_\Dscr(I,\prod F_j)\r 0$$ Comparing and yields what we want. Now we prove an analog for Proposition \[ref:3.4.4a\] for well-closed subcategories. \[ref:3.5.8a\] Assume that $\Cscr$ is complete and has an injective cogenerator. Assume that $\Dscr$ is well-closed in $\Cscr$ and that $\Escr$ is well-closed in $\Dscr$. Then $\Escr$ is well-closed in $\Cscr$. Note that by Prop. \[ref:3.4.3a\] $\Dscr$ is complete and it is clear that $\Dscr$ has an injective cogenerator. So it makes sense to speak of a well-closed subcategory of $\Dscr$. We already know that $\Escr$ is biclosed in $\Cscr$, so we only have to show that $R^1\prod (E_\alpha)_\alpha=0$ for a family of injectives in $\Escr$. Let $$0\r E_\alpha \r F_{\alpha 0} \r F_{\alpha 1} \r F_{\alpha 2}$$ be (truncated) injective resolutions of $E_\alpha$ in $\Dscr$. Furthermore let $$0\r F_{\alpha i}\r G_{\alpha i 0}\r G_{\alpha i 1} \r G_{\alpha i 2}$$ be (truncated) Cartan Eilenberg resolutions for the complexes $F_\alpha^\cdot$. Taking products this yields a diagram of complexes $$\begin{CD} @. @. 0\\ @. @. @VVV\\ 0 @>>> \prod_\alpha E_\alpha @>>> \prod_\alpha F_\alpha^\cdot\\ @. @. @VVV\\ @. @. \prod_\alpha G_\alpha^\cdot\\ \end{CD}$$ By hypotheses the columns of this diagram are exact and so is the first row. Then it easily follows that $$0\r \prod_\alpha E_\alpha \r \prod_\alpha G_{\alpha 00} \r \prod_\alpha G_{\alpha 10}\oplus \prod_\alpha G_{\alpha 01} \r \prod_\alpha G_{\alpha 20}\oplus \prod_\alpha G_{\alpha 11} \oplus \prod_\alpha G_{\alpha 0 2}$$ is exact. Since this is a product of truncated injective resolutions of $E_\alpha$ in $\Cscr$ we are done. We don’t know if well-closedness is compatible with the Gabriel product and hence we don’t know if the product of ideals is an ideal. In order to deal with this problem in the sequel we introduce one more technical notion. \[ref:3.5.9a\] Assume that $\Dscr$ is complete and has an injective cogenerator. A biclosed subcategory $\Escr\subset \Dscr$ is very well-closed if for all families $(M_j)_{j\in J}$ of objects in $\Escr$ one has $R^1\prod (M_j)_j=0$. \[ref:3.5.10a\] Assume that $\Cscr$ is complete and has an injective cogenerator. Assume that $\Dscr$ is well-closed in $\Cscr$ and $\Escr$ is very well-closed in $\Dscr$. Then $\Escr$ is very well-closed in $\Cscr$. This is proved similarly as Prop. \[ref:3.5.8a\], \[ref:3.5.11a\] Assume that $\Dscr$ is complete and has an injective cogenerator. Then a biclosed subcategory $\Escr\subset \Dscr$ is very well-closed if and only if 1. $\Escr$ is well-closed. 2. $\Escr$ has exact direct products. One direction is clear. To prove the other direction we use Proposition \[ref:3.5.10a\] with $\Escr=\Dscr$. \[ref:3.5.12a\] Assume that $\Dscr$ is complete and has an injective cogenerator. Let $\Escr_1,\Escr_2\subset \Dscr$ be very well-closed subcategories. Then $\Escr_1\cdot \Escr_2$ is very well-closed. Let $(M_j)_j$ be a family of objects in $\Escr_1\cdot\Escr_2$. We have exact sequences $$0\r M_{j2}\r M_j \r M_{j1} \r 0$$ with $M_{ji}\in \Escr_i$. From the long exact sequence for $(R^i\prod)_i$ and the hypotheses we deduce that $R^1\prod (M_j)_j=0$. If $\Lscr\in\BIMOD(\Dscr-\Dscr)$ is such that $\HHom_\Dscr(\Lscr,-)$ is an equivalence of categories then we call $\Lscr$ an invertible bimodule. Obviously in that case $\Lscr\in \Bimod(\Dscr-\Dscr)$, and there exist $\Lscr^{-1}\in \Bimod(\Dscr-\Dscr)$ such that $\Lscr\otimes_\Dscr\Lscr^{-1}\cong \Lscr^{-1} \otimes_\Dscr \Lscr\cong \Id_\Dscr$. Assume that $\Lscr$ is invertible. We will call a weak ideal in $\Lscr$ a subobject $I$ of $\Lscr$ in $\BIMOD(\Dscr-\Dscr)$. If $I$ actually lies in $\Bimod(\Dscr-\Dscr)$ then we call $I$ an ideal in $\Lscr$. Clearly $I\mapsto I\otimes_\Dscr\Lscr^{-1}$ and $I\mapsto \Lscr^{-1} \otimes_\Dscr I$ induce bijections between (weak) ideals in $\Lscr$ and (weak) ideals in $\Id_\Dscr$. If $\Lscr,\Mscr$ are invertible $\Dscr-\Dscr$-bimodules and $I\subset \Lscr$, $J\subset \Mscr$ are ideals then we define $IJ$ as the image of $I\otimes_\Dscr J$ in $\Lscr\otimes_\Dscr\Mscr$. \[ref:3.5.13a\] Let $I\subset \Lscr$ be a weak ideal in an invertible bimodule $\Lscr$. Then the Rees algebra $\Dscr(I)$ is the graded weak algebra given by $$\Id_\Dscr\oplus I\oplus I^2\oplus\cdots$$ (with obvious multiplication). Clearly $\Dscr(I)\in \ALG(\Dscr)$. However if $\Mod(\Id_\Dscr/I)$ is very well-closed then by Prop. \[ref:3.5.4a\] and \[ref:3.5.12a\] $\Dscr(I)$ lies in $\Alg(\Dscr)$. It would be useful if we could find weaker conditions under which the Rees algebra of an ideal lies in $\Alg(\Dscr)$. We close this section by introducing some terminology which we will use later. If $I\subset \Id_\Dscr$ is a weak ideal defining a (closed) subcategory $\Escr$ of $\Dscr$ then $I/I^2\in\BIMOD(\Escr-\Escr)$. Following [@rosenberg] we call $I/I^2$ the conormal bundle of $\Escr$ in $\Dscr$. Quasi-schemes {#ref:3.6b} ------------- For us a quasi-scheme $X$ will be a Grothendieck category which we denote by $\Qch(X)$. A morphism $\alpha:Y\r X$ of quasi-schemes will be an additive functor $\alpha^\ast:\Qch(X)\r \Qch(Y)$ commuting with colimits. The quasi-schemes form a category which we denote by $\Qsch$. Actually we will consider $\Qsch$ as a 2-category whose 2-cells correspond to natural isomorphisms (see Appendix \[ref:Aa\]). Commutative diagrams in a 2-category are usually only assumed to be commutative up to an explicit natural isomorphism (see Appendix \[ref:Aa\]). Such diagrams are sometimes called pseudo-commutative diagrams, but we will call them simply “commutative diagrams”. Likewise we will speak of “functors” when we actually mean pseudo-functors (again see Appendix \[ref:Aa\]). If $\alpha:Y\r X$ is a morphism of quasi-schemes then it follows from Theorem \[ref:2.1b\] that the adjoint to $\alpha^\ast$ exists. This adjoint is unique up to unique isomorphism and we will denote it by $\alpha_\ast$. The assignment $\alpha\mapsto \alpha^$ is obviously functorial since formally $\alpha=\alpha^$! However as is explained in Appendix \[ref:Aa\] the assignment $\alpha\mapsto \alpha_\ast$ is also functorial if we work in the setting of 2-categories. Denote by “$\mathrm{Sch}$” the category of quasi-compact, quasi-separated schemes. If $X\in \mathrm{Sch}$ then the category of quasi-coherent sheaves on $X$ is a Grothendieck category [@thomasson]. In that case we define $\Qch(X)$ as the category of quasi-coherent sheaves on $X$. Rosenberg in [@rosenberg1] has proved a reconstruction theorem which allows one to recover $X$ from $\Qch(X)$ (generalizing work of Gabriel in the noetherian case). Furthermore he has also shown that the functor $$\mathrm{Sch}/\Spec \ZZ\r \QSch/\Spec \ZZ$$ which sends a scheme to its associated quasi-scheme is fully faithful in the sense of $2$-categories. Now let $X$ be a quasi-schemes. We write $\Alg(X)=\Alg(\Qch(X))$, $\Bimod(X)=\Bimod(\Qch(X))$, etc…. Below by an algebra on $X$ we will mean an object of $\Alg(X)$ unless otherwise specified. Likewise $\Ascr$-$\Bscr$-bimodules will in general be objects of $\Bimod(\Ascr-\Bscr)$. By $o_X$ we denote the identity functor on $\Qch(X)$ considered as an object of $\Alg(X)$. Obviously $\Mod(o_X)=\Qch(X)$, $\Bimod(o_X)=\Bimod(X)$. If $\Ascr\in\Alg(X)$ then we denote by $\Spec\Ascr$ the object $\Qsch/X$ given by the pair $(\Mod(\Ascr),-\otimes_{o_X}\Ascr)$. If $X$ is a quasi-scheme then we define $\Gralg(X)=\Gralg(\Qch(X))$ and also $\GRALG(X)=\GRALG(\Qch(X))$. Below by a graded algebra on $X$ we will mean an object of $\Gralg(X)$ unless otherwise specified. Likewise graded $\Ascr$-$\Bscr$-bimodules will in general be objects of $\Bigr(\Ascr-\Bscr)$. Related to the notion of relative categories is the notion of enriched quasi-schemes. An enriched quasi-scheme will be a pair $(X,\Oscr_X)$ where $X$ is a quasi-scheme and $\Oscr_X$ is an arbitrary object of $\Qch(X)$. A morphism $(Y,\Oscr_Y)\r (X,\Oscr_X)$ between enriched quasi-schemes is a pair $(\alpha,s)$ where $\alpha:Y\r X$ is a morphism between quasi-schemes and $s$ is an isomorphism $s:\alpha^\ast(\Oscr_X)\r \Oscr_Y$. Note that if $(Y,\alpha)\in \Qscr/X$ and $X$ is an enriched quasi-scheme then $Y$ becomes canonically an enriched quasi-scheme if we put $\Oscr_Y=\alpha^\ast\Oscr_X$. The prototype of an enriched quasi-scheme is $\Spec R=(\Mod(R),R_R)$ for a ring $R$. The following lemma will be used tacitly throughout the paper. Let $R$ be a ring. Then the category $\Qsch/\Spec R$ is equivalent (as a two-category) to the category of $R$-linear enriched quasi-schemes. The equivalence is given by sending $(Y,\alpha)$ to $(Y,\alpha^\ast(R))$. If $(X,\Oscr_X)$ is an enriched quasi-scheme and $\Uscr\in \Qch(X)$ then we put $\Gamma(X,\Uscr)=\Hom_{o_X}(\Oscr_X,\Uscr)$. We say that $\Uscr$ is generated by global sections if $\Uscr$ is a quotient of some $\Oscr_X^{\oplus I}$. If $\Mscr$ is a bimodule on $X$ then we put $\Mscr_{o_X}=\Oscr_X\otimes_{o_X} \Mscr$. We think of $\Mscr_{o_X}$ as the “right structure” of $\Mscr$. Care should be taken however since $\Mscr\mapsto \Mscr_{o_X}$ is apriori not an exact functor. This will not be a problem in our applications. If $\Mscr$ is a bimodule on $X$ then we define $\Gamma(X,\Mscr)=\Gamma(X,\Mscr_{o_X})$. Again one should be careful since $\Gamma(X,-)$ is in general not left exact. It is easy to see that if $\Ascr$ is an algebra on $X$ then $A=\Gamma(X,\Ascr)$ will be a ring. If $\Mscr$ is an $\Ascr$-module then $\Gamma(X,\Mscr)$ will be an $A$-module. A quasi-scheme $X$ will be called noetherian if $\Qch(X)$ is locally noetherian. For an enriched quasi-scheme, we also require that $\Oscr_X$ is noetherian. A morphism $\alpha:X\r Y$ will be called noetherian if $\alpha^\ast$ is noetherian. An algebra $\Ascr$ on $X$ is said to be noetherian if the functor $-\otimes_{o_X}\Ascr$ preserves noetherian objects. Clearly if $X$ is noetherian and $\Ascr$ is noetherian then so is $\Spec\Ascr$. A this point we will introduce a convention that will be in force throughout this paper. If $\Cscr=\text{Xyz}\cdots(\cdots)$ is a category then $\text{xyz}\cdots(\cdots)$ stands for the full subcategory of $\Cscr$ whose objects are the noetherian objects in $\Cscr$. Divisors {#ref:3.7b} -------- We will say that a map $\alpha:Y\r X$ is a biclosed embedding if $\alpha_\ast$ embeds $\Qch(Y)$ in $\Qch(X)$ as a biclosed subcategory. Assume that $X,Y$ are quasi-schemes where $Y$ is embedded in $X$ by a biclosed embedding. For simplicity we also assume that $X$ is noetherian, although that is not strictly necessary. We denote $$o_X(-Y)=\ker(o_X\r o_{Y})$$ We say that $Y$ is a divisor in $X$ if $o_X(-Y)$ is invertible. In the rest of this section we will assume that $Y$ is a divisor in $X$ and we denote the inclusion mapping by $i$. If $\Mscr\in\Qch(X)$ then we write $\Mscr_Y$ for $\Mscr\otimes_{o_X}o_Y$ and $\Mscr(nY)$ for $\Mscr\otimes_{o_X} o_X(nY)$, where $o_X(nY)=o_X(Y)^{\otimes n}$. We denote the inclusion $o_X(-Y)\r o_X$ by $t$ and we do the same with the induced maps $\Mscr(-Y)\r \Mscr$ for $\Mscr\in \Qch(X)$. The normal bundle of $Y$ in $X$ is defined by $\Nscr_{Y/X}=o_X(Y)/o_X$. For use below we define a few categories. $$\begin{aligned} \tors_Y(X)&=\{\Mscr\in\coh(X)\mid \text{There exist $n$ such that the map $t^n:\Mscr(-nY)\r \Mscr$ is zero}\}\\ \iso_Y(X)&=\{\Mscr\in \coh(X)\mid \text{The map $t:\Mscr(-1)\r \Mscr$ is an isomorphism}\}\end{aligned}$$ $\Tors_Y(X)$ and $\Iso_Y(X)$ will be the closures of $\tors_Y(X)$ and $\iso_Y(X)$ under direct unions. $\Tors_Y(X)$ and $\Iso_Y(X)$ are localizing subcategories of $\Qch(X)$. Since $\Qch(X)$ is a locally noetherian category, it is easy to see that it is sufficient to show that $\tors_Y(X)$ and $\iso_Y(X)$ are Serre subcategories in $\coh(X)$. For $\tors_Y(X)$ this is clear, so we concentrate on $\iso_Y(X)$. It is clearly sufficient to show that $\iso_Y(X)$ is closed under taking subobjects. Let $\Mscr\in \iso_Y(X)$ and let $\Nscr\subset\Mscr$. We have isomorphisms $t^n:\Mscr\r \Mscr(nY)$ and these yield an ascending chain of submodules in $\Mscr$ given by $t^{-n}(\Nscr(nY))$. Since $\Mscr$ is noetherian this chain must stop. From this we easily obtain that $\Nscr\in\iso_Y(X)$. We will say that the objects in $\Tors_Y(X)$ are supported on $Y$. If $\Mscr(-Y)\r\Mscr$ is injective then we will say that $\Mscr$ is $Y$-torsion free. We will use the following result. \[ref:3.7.2a\] Let $\Mscr\in \coh(X)$. Then the filtration $$\cdots \subset t^n(\Mscr(-nY))\subset t^{n-1}(\Mscr(-(n-1)Y))\subset \cdots\subset t(\Mscr(-Y))\subset \Mscr$$ satisfies the Artin-Rees condition. Left to the reader. \[ref:3.7.3a\] Let $\Mscr\in\coh(X)$. Then $\Mscr$ contains an $Y$-torsion free submodule $\Nscr$, such that $\Mscr/\Nscr$ is supported on $Y$. Let $\Tscr$ be the maximal submodule of $\Mscr$ supported on $Y$. Since $\Tscr$ is noetherian we will have $t^n(\Tscr(-nY))=0$ for some $n$. By Proposition \[ref:3.7.2a\] there will be some $m$ such that $t^m(\Mscr(-mY))\cap \Tscr\subset t^n(\Tscr(-nY))=0$. Thus $\Nscr=t^m(\Mscr(-mY))$ is $Y$ torsion free and $\Mscr/\Nscr$ is supported on $Y$. If $i:(Y,\Oscr_Y)\r (X,\Oscr_X)$ is a map of enriched quasi-schemes then we say that $i$ makes $Y$ into a divisor in $X$ if the underlying map $Y\r X$ makes $Y$ into a divisor in $X$ in the sense of ordinary quasi-schemes and if in addition the induced map $\Oscr_X(-Y)\r \Oscr_X$ is injective. Proj {#ref:3.8b} ---- Below $X$ will be a noetherian quasi-scheme and $\Ascr=\oplus_n\Ascr_n\in \Gralg(X)$ will be noetherian. The definition of $\Proj \Ascr$ is entirely similar to the ring case [@AZ]. As before $\Gr(\Ascr)$ is the category of $\ZZ$-graded $\Ascr$-modules, $\Tors(\Ascr)$ is the full subcategory of $\Gr(\Ascr)$ consisting of graded modules that are unions of right bounded modules and $$\Qgr(\Ascr)=\Gr(\Ascr)/\Tors(\Ascr)$$ We use similar notations as in [@AZ]. Thus $\pi:\Gr(\Ascr)\r \Qgr(\Ascr)$ is the quotient map. $\omega:\Qgr(\Ascr)\r \Gr(\Ascr)$ is the right adjoint to $\pi$ and $\tau:\Gr(\Ascr)\r \Tors(\Ascr)$ is the functor which associates to every object its maximal torsion subobject. We put $\tilde{(-)}=\omega\pi$. Let $\Pqsch/X$ be the 2-category of triples $(Y,\alpha,s)$ where $Y$ is a quasi-scheme, $\alpha:Y\r X$ a morphism and $s$ an autoequivalence of $\Qch(Y)$. Morphism between triples $(Y,\alpha,s)\r (Z,\beta,t)$ are given by a morphism $\gamma:Y\r Z$, a natural isomorphism $\mu:\gamma^\ast\circ\beta^\ast\r\alpha^\ast$ and a natural isomorphism $\psi:s\circ\gamma^\ast\r \gamma^\ast\circ t$. We leave it to the reader to define natural isomorphisms between such triples. We now define $\Proj \Ascr$ as the object of $\Pqsch/X$ given by the triple $(\Qgr(\Ascr),\pi(-\otimes_\ox\Ascr),s)$ where $s$ is the shift functor on $\Qgr(\Ascr)$ obtained from the canonical shift on $\Gr(\Ascr)$. We will also denote by $\Proj\Ascr$ the object in $\Qsch/X$ obtained by forgetting the shift. If we denote by $\alpha$ the structure map $\Proj\Ascr\r X$ then by definition, $\alpha^\ast$ is given by $\pi(-\otimes_\ox\Ascr)$. Since $\omega$ is the right adjoint to $\pi$ we deduce that $\alpha_\ast$ is given by $\omega(-)_0$. In the sequel it will be convenient to use $\underline{\alpha}_\ast$ as a synonym for $\omega$. Thus $$\underline{\alpha}_\ast(\Mscr)_\ox=\oplus_n \alpha_\ast(\Mscr(n))$$ Below we will generalize some of the results of [@AZ] to our situation since we will need them. Usually we can simply copy the proofs in [@AZ]. \[ref:3.8.1a\] Let $\Ascr$, $\Bscr$ be noetherian graded algebras on $X$, $\Nscr$ a graded $\Ascr$-module, $\Mscr_1$, $\Mscr_2$ graded $\Ascr-\Bscr$ bimodules. 1. Assume that $\Nscr$ is right bounded and $(\Mscr_1)_{o_X}$ is a quotient of $\Ascr\otimes_\ox\Mscr'_1$ where $\Mscr'_1$ is a right bounded graded $\ox-\ox$-module. Then $\Nscr\otimes_\Ascr\Mscr_1$ is right bounded. 2. Assume that $\Mscr_1$ is right bounded. Then $\Nscr\otimes_\Ascr\Mscr_1$ is torsion. 3. Assume that $\phi:\Mscr_1\r\Mscr_2$ is a morphism of graded $\Ascr-\Bscr$ bimodule which is an isomorphism in high degree. Then $\ker,\coker(\Nscr\otimes_\Ascr\Mscr_1\r\Nscr\otimes_\Ascr\Mscr_2)$ are torsion. <!-- --> 1. This is trivial since $\Nscr\otimes_\Ascr\Mscr_1$ is a quotient as graded $\ox$-bimodules of $$\Nscr\otimes_\Ascr(\Ascr\otimes_\ox\Mscr'_1)=\Nscr\otimes_\ox\Mscr'_1$$ 2. This is a special case of (3). 3. Now write $\Nscr$ as a quotient of modules of the form $\Pscr_1\r\Pscr_0$ where $\Pscr_{0,1}$ are direct sums of shifts of modules of the form $\Nscr'\otimes_\ox\Ascr$, $\Nscr'\in\Qch(X)$ (located in degree zero). This is possible by Proposition \[ref:3.1.15a\](8). If $\Nscr=\Nscr'\otimes_\ox\Ascr$ then the lemma is true because $$(\Nscr'\otimes_\ox\Ascr)\otimes_\Ascr\Mscr=\Nscr'\otimes_\ox {}_\ox\!\Mscr$$ for all $\Mscr\in\Bimod(\Ascr-\Bscr)$. Hence the map $\Nscr\otimes_\Ascr\Mscr_1\r\Nscr\otimes_\Ascr\Mscr_2$ is an isomorphism in high degree. The general case follows from the fact that $-\otimes_\Ascr\Mscr_{1,2}$ is compatible with colimits. From this lemma one obtains the following corollary (cfr. [@AZ Prop. 2.5]). \[ref:3.8.2a\] Let $\phi:\Ascr\r\Bscr$ be a morphism of noetherian graded algebras on $X$ such that $\phi$ is an isomorphism in high degree. Then the functors $$\label{ref:3.21a} \Gr(\Ascr)\r\Gr(\Bscr):\Nscr\r \Nscr\otimes_\Ascr\Bscr$$ and $$\label{ref:3.22a} \Gr(\Bscr)\r\Gr(\Ascr):\Mscr\r\Mscr_\Ascr$$ factor to give inverse equivalences between $\Qgr(\Ascr)$ and $\Qgr(\Bscr)$. Furthermore $\Proj(\Ascr)$ and $\Proj(\Bscr)$ are equivalent. Assume that $\phi_n$ is an isomorphism for $n\ge n_0$. Then $\Bscr$ is a quotient of $\Ascr\oplus\Ascr\otimes(\oplus_{n<n_0}\Bscr_n(-n))$. Assume that $\Mscr$ is a $\Bscr$-module, torsion as $\Ascr$-module. Then $\Mscr_\Ascr$ is a quotient of $\oplus_{i\in I}\Mscr_i$ where the $\Mscr_i$ are right bounded $\Ascr$-modules. Hence $\Mscr$ is a quotient of $\oplus_{i\in I}\Mscr_i\otimes_\Ascr\Bscr$. By lemma \[ref:3.8.1a\](1) all the $\Mscr_i\otimes_\Ascr\Bscr$ are right bounded. Thus $\Mscr$ is also torsion as $\Bscr$-module. Thus $\Bscr$-torsion is equivalent to $\Ascr$-torsion and so we will simply speak of torsion. obviously preserves isomorphism mod torsion, so we concentrate on . Assume that $\ker,\coker(\Mscr\xrightarrow{\theta}{\Nscr})$ are torsion for some $\Mscr,\Nscr\in\Gr(\Ascr)$. Then we have the following commutative diagram. $$\begin{CD} \Mscr @>>> \Mscr\otimes_\Ascr\Bscr\\ @V\theta VV @V\theta \otimes 1 VV\\ \Nscr @>>> \Nscr\otimes_\Ascr\Bscr \end{CD}$$ According to lemma \[ref:3.8.1a\](3) the horizontal maps are isomorphisms modulo torsion. Since by hypotheses this is also true for $\theta$, we obtain that $\theta\otimes 1$ is an isomorphism modulo torsion as well. The reader now easily verifies that and are mutual inverses modulo torsion. If $\Ascr$, $\Bscr$ are as in the previous proposition then is clearly compatible with the structure maps and the shift functors on $\Proj\Ascr$ and $\Proj\Bscr$ whence $\Proj\Ascr$ and $\Proj\Bscr$ are equivalent. It follows from the previous proposition that one may restrict oneself to $\NN$-graded algebras since $\Proj\Ascr=\Proj\Ascr_{\ge 0}$. In the sequel all graded algebras will be noetherian and $\NN$-graded, unless otherwise specified. Note that just as in the ring case $\Ascr$ noetherian implies $\Ascr_{\ge 0}$ noetherian. This allows us to use the following technically useful result which we state for further reference. \[ref:3.8.3a\] Assume that $\Ascr$ is noetherian and $\NN$-graded. Then $\Tors(A)$ is closed under essential extensions and hence under injective hulls. Thus $\tau$ is stable in the sense of [@stenstrom] Assume that $\Ascr$ is $\NN$-graded. Then if $\Mscr,\Nscr\in \Gr(\Ascr)$ $$\begin{aligned} \Hom_{\Qgr(\Ascr)}(\pi\Nscr,\pi\Mscr)&= \dirlim_{\Nscr/\Nscr'\text{ torsion}} \Hom_{\Gr(\Ascr)}(\Nscr',\Mscr/\tau(\Mscr))\\ &=\dirlim_{\Nscr/\Nscr'\text{ torsion}} \Hom_{\Gr(\Ascr)}(\Nscr',\Mscr)\end{aligned}$$ where the last equality follows from the fact that $\tau$ is stable (see [@stenstrom]). If $\Nscr\in\gr(\Ascr)$ and if $\Nscr/\Nscr'$ is torsion then it is in fact right bounded, so $\Nscr_{\ge n}\subset \Nscr'$ for $n\ge 0$. So in that case $$\begin{aligned} \Hom_{\Qgr(\Ascr)}(\pi\Nscr,\pi\Mscr)&= \dirlim_n \Hom_{\Gr(\Ascr)}(\Nscr_{\ge n},\Mscr) &=\dirlim\Hom_{\Gr(\Ascr)}(\Nscr_{\ge n},\Mscr_{\ge n})\end{aligned}$$ \[ref:3.8.4a\] Let $\Ascr\in \Gralg(X)$ be noetherian and $\NN$-graded. Then for $\Mscr\in\Gr(\Ascr)$ $$\begin{aligned} \tau(\Mscr)&=\dirlim\underline{\HHom}_\Ascr(\Ascr/\Ascr_{\ge n},\Mscr)\label{ref:3.23a}\\ \tilde{\Mscr}&=\dirlim\underline{\HHom}_\Ascr(\Ascr_{\ge n},\Mscr) \label{ref:3.24a}\end{aligned}$$ By localization theory $$\tau(\Mscr)=\ker(\Mscr\r \tilde{\Mscr})$$ Now from the exact sequence of $\Ascr$-bimodules $$0\r\Ascr_{\ge n}\r \Ascr\r \Ascr/\Ascr_{\ge n}\r 0$$ we obtain a left exact sequence $$0\r\underline{\HHom}_\Ascr(\Ascr/\Ascr_{\ge n},\Mscr)\r \underline{\HHom}_\Ascr(\Ascr,\Mscr)\r \underline{\HHom}_\Ascr(\Ascr_{\ge n},\Mscr)$$ Using the fact that $\underline{\HHom}_\Ascr(\Ascr,\Mscr)=\Mscr$ and exactness of direct limits, it follows that it suffices to prove . Recall that by definition $\tilde{\Mscr}=\omega\pi\Mscr$. So to prove it is sufficient for every $\Nscr\in\Gr(\Ascr)$ to construct a natural isomorphism between $$\label{ref:3.25a} \Hom_{\Gr(\Ascr)}(\Nscr,\omega\pi\Mscr)$$ and $$\label{ref:3.26a} \Hom_{\Gr(\Ascr)}(\Nscr,\dirlim\underline{\HHom}_\Ascr(\Ascr_{\ge n},\Mscr))$$ and since $\Gr(\Ascr)$ is locally noetherian, it suffices to do this in fact for $\Nscr$ noetherian. Now $$\label{ref:3.27a} \eqref{ref:3.25a}=\dirlim_n \Hom_{\Gr(\Ascr)}(\Nscr_{\ge n},\Mscr)$$ On the other hand, again because $\Nscr$ is noetherian. $$\label{ref:3.28a} \eqref{ref:3.26a}=\dirlim\Hom_{\Gr(\Ascr)}(\Nscr\otimes_\Ascr\Ascr_{\ge n},\Mscr)$$ We have to make isomorphic to . Now $\Nscr$ is noetherian and hence left bounded, so by replacing $\Nscr$ by some shift (and doing the same with $\Mscr$) we may assume that $\Nscr$ is in fact $\NN$-graded. Using and lemma \[ref:3.1.11a\] it is now sufficient to show that $$\begin{aligned} \Kscr_n=\ker (\Nscr\otimes_\Ascr\Ascr_{\ge n}\r \Nscr_{\ge n})\\ \Cscr_n=\coker(\Nscr\otimes_\Ascr\Ascr_{\ge n}\r \Nscr_{\ge n})\end{aligned}$$ are torsion inverse systems (see Definition \[ref:3.1.9a\]). Now we write $\Nscr$ as quotient of $\Pscr_1\r\Pscr_0$ where the $\Pscr_i$ are finite direct sums of negative shifts of modules of the form $\Nscr'\otimes_\ox\Ascr$. Working in the abelian category of inverse systems modulo torsion it is sufficient to prove that $\Kscr_n$ and $\Cscr_n$ are torsion inverse systems in the case $\Nscr=\Nscr'\otimes_\ox\Ascr(-m)$. In that case $$\begin{aligned} \Kscr_n&=\ker (\Nscr'\otimes_\ox\Ascr_{\ge n} \r \Nscr'\otimes_\ox\Ascr_{\ge n-m})(-m)\\ \Cscr_n&=\coker (\Nscr'\otimes_\ox\Ascr_{\ge n} \r \Nscr'\otimes_\ox\Ascr_{\ge n-m})(-m)\end{aligned}$$ It is now clear that in this case actually $\Kscr_n=0$ and $\Cscr_n$ lives in degrees $[n,n+m-1]$. So $(\Cscr_n)_n$ is a torsion inverse system as well. Condition [ “$\chi$”]{} and cohomological dimension {#ref:3.9b} --------------------------------------------------- Below $X$ will be a noetherian quasi-scheme and $\Ascr$, $\Bscr$ will be $\NN$-graded noetherian objects in $\Gralg(X)$. $\alpha:\Proj \Ascr\r X$, $\beta:\Proj \Bscr\r X$ will be the structure maps. $\pi$, $\omega$, $\tau$ refer to $\Ascr$. For $\Bscr$ we use notations such as $\pi_\Bscr$, $\omega_\Bscr$, $\tau_\Bscr$. The following is clear. \[ref:3.9.1a\] If $\Mscr\in\Gr(\Ascr)$ then $R^i\tau\Mscr\in\Tors(\Ascr)$. In order to develop a non-commutative version of projective geometry, the following definition was proposed in [@AZ]. $\Ascr$ satisfies $\chi$ if for every $\Mscr\in \gr(\Ascr)$ one has that $(R^i\tau\Mscr)_{\ge 0}\in \tors(\Ascr)$. If $\Mscr\in \Gr(\Ascr)$ then we have a triangle $$R\tau\Mscr\r \Mscr\r R\omega(\pi\Mscr)\xrightarrow{[1]}$$ from which one deduces the exact sequence $$\label{ref:3.29a} 0\r \tau\Mscr\r\Mscr\r \tilde{\Mscr}\r R^1\tau \Mscr\r 0$$ and isomorphisms $$\label{ref:3.30a} R^{i+1}\tau\Mscr=R^{i}\omega(\pi\Mscr)),\qquad i\ge 1$$ So we obtain : \[ref:3.9.3a\] $\Ascr$ satisfies $\chi$ if and only if for all $\Mscr\in\gr \Ascr$ one has 1. $R^i\omega (\pi\Mscr)_{\ge 0}\in \tors(\Ascr)$, $i\ge 0$. 2. $(\tilde{\Mscr})_{\ge 0}/\Mscr\in \tors(\Ascr)$ One should view condition $\chi$ as a kind of ampleness condition. This becomes clearer if one makes the following definitions. \[ref:3.9.4a\] A map $\alpha:Y\r X$ of noetherian quasi-schemes is proper if $R^i\alpha_\ast$ sends $\coh(Y)$ to $\coh(X)$ for all $i$. Let $(Y,\alpha,s)$ be an object in $\Pqsch/X$. Assume that $Y$ is noetherian. Then we say that $s$ is relatively ample if the following two conditions hold for $\Mscr\in\coh(Y)$. 1. For $i>0$ one has $$R^i\alpha_\ast(s^n\Mscr)=0$$ for $n\gg 0$. 2. The adjoint map $$\alpha^\ast \alpha_\ast s^n\Mscr\r s^n\Mscr$$ is surjective for $n\gg 0$. This definition of ampleness is much more restrictive than the one used in [@AZ]. We say that $\Ascr$ is generated in degree one if the multiplication map $\Ascr_m\otimes_{o_X}\Ascr_n\r \Ascr_{m+n}$ is surjective for all $m,n\ge 0$. The following is a consequence of lemma \[ref:3.9.3a\]. \[ref:3.9.7a\] Let $(Y,\alpha,s)=\Proj \Ascr$. If $\Ascr$ satisfies $\chi$ then $\alpha$ is proper and if $\Ascr$ is in addition generated in degree one then $s$ is relatively ample. In the following we will often need the following technical condition on an object $\Mscr\in\BIGR(\Ascr-o_X)$ : - $\Mscr$ has a presentation $\Pscr_1\r \Pscr_0$ where the $\Pscr_i$ are finite sums of shifts of $\Ascr-o_X$-bimodules of the form $\Ascr\otimes_{o_X}\Pscr'$ where the $\Pscr'$ are coherent objects (see \[ref:3.1a\]) in $\BIMOD(o_X-o_X)$ (located in degree zero). We should think of (fin) as a kind of “finite presentation” condition. (fin) is useful because of the following lemma \[ref:3.9.8a\] Assume that $\Mscr\in\BIGR(\Ascr-o_X)$ satisfies (fin) and $\Nscr\in\BIGR(\Bscr-\Ascr)$ (or $\Nscr\in\Gr(\Ascr)$). Assume furthermore that $\Mscr$ is $\NN$-graded. Then the kernel and the cokernel of $$\Nscr_{\ge n}\otimes_\Ascr \Mscr\r (\Nscr\otimes_\Ascr\Mscr)_{\ge n}$$ are torsion inverse systems (cfr. Definition \[ref:3.1.9a\]). Here is another obvious application. \[ref:3.9.9a\] Let $f:\Ascr\r \Bscr$ be a morphism in $\Gralg(X)$ and assume that $\Bscr$ satisfies (fin). Let $\Mscr\in \Gr(\Bscr)$. Then $\tau_{\Ascr}(\Mscr_\Ascr)=(\tau_{\Bscr}(\Mscr))_\Ascr$. Using Proposition \[ref:3.8.4a\] and adjointness this amounts to showing that the inverse systems $\Ascr/\Ascr_{\ge n}\otimes_\Ascr \Bscr$ and $\Bscr/\Bscr_{\ge n}$ are equivalent modulo torsion inverse systems. Since $\Ascr/\Ascr_{\ge n}\otimes_\Ascr \Bscr=\coker (\Ascr_{\ge n}\otimes_\Ascr\Bscr\r \Bscr)$ this follows directly from lemma \[ref:3.9.8a\]. In the sequel we will need a result like . The hypotheses under which we can prove this are unfortunately quite technical and almost certainly too restrictive. \[ref:3.9.10a\] Let $f:\Ascr\r \Bscr$ be a morphism in $\Gralg(X)$. Assume that $\Bscr$ satisfies (fin) and furthermore that there are graded flat $\Ascr-o_X$-bimodules $(\Fscr_i)_i$ satisfying (fin), together with a long exact sequence $$\cdots \r(\Fscr_2)_{\ge p}\r (\Fscr_1)_{\ge p}\r (\Fscr_0)_{\ge p} \r ({}_\Ascr\Bscr_{o_X})_{\ge p}\r 0$$ for a certain $p\in\ZZ$. Then for $\Mscr\in\Gr(\Bscr)$ one has $$\label{ref:3.31a} R^i\tau_{\Ascr}(\Mscr_\Ascr)=(R^i\tau_{\Bscr}(\Mscr))_\Ascr$$ Both sides of are $\delta$-functors. Hence is suffices to show that the two sides of are zero for $i>0$ and naturally isomorphic for $i=0$, when evaluated on injectives. Actually it is not necessary to take all injectives, but only a cogenerating set. Thus we take for our injectives all modules of the form $$F=\underline{\HHom}_\Ascr({}_\Bscr\Bscr_\Ascr,E)$$ where $E\in \Inj(\Ascr)$. Clearly by Proposition \[ref:3.8.4a\], lemma \[ref:3.9.9a\] and adjointness $$\text{RHS}\eqref{ref:3.31a}= \begin{cases} \dirlim \underline{\HHom}_\Ascr(\Ascr/\Ascr_{\le n}\otimes_\Ascr {}_\Ascr\Bscr_\Ascr,E)&\text {if $i=0$}\\ 0&\text{otherwise} \end{cases}$$ On the other hand by the definition of $\underline{\HTor}$ $$\begin{aligned} \text{LHS}\eqref{ref:3.31a}&= \dirlim \underline{\HExt}^i_\Ascr(\Ascr/\Ascr_{\ge n} ,F_{\Ascr})\\ &=\dirlim \underline{\HExt}^i_\Ascr(\Ascr/\Ascr_{\ge n} ,\underline{\HHom}_\Ascr ({}_\Ascr\Bscr_\Ascr,E))\\ &=\dirlim\underline{\HHom}_\Ascr(\underline{\HTor}^{\Ascr}_i(\Ascr/\Ascr_{\le n},{}_\Ascr\Bscr_\Ascr),E)\end{aligned}$$ So the two sides of certainly agree for $i=0$. In general we have to show for $i>0$ that $$\label{ref:3.32a} \underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},{}_\Ascr\Bscr_\Ascr)$$ is a torsion inverse system. From fact that the $\Fscr_i$ are flat we can deduce that, up to inverse systems of the form $$\label{ref:3.33a} \underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr)$$ with $\Tscr_{o_X}$ coherent, is given by the middle homology of $$\label{ref:3.34a} \Ascr/\Ascr_{\ge n}\otimes_\Ascr (\Fscr_{i+1})_{\ge p} \r \Ascr/\Ascr_{\ge n}\otimes_\Ascr (\Fscr_{i})_{\ge p} \r \Ascr/\Ascr_{\ge n}\otimes_\Ascr (\Fscr_{i-1})_{\ge p}$$ By shifting if necessary we may assume that $\Fscr_{i+1}$, $\Fscr_i$, $\Fscr_{i-1}$ are $\NN$-graded. Now assume in general that $\Fscr$ is an $\NN$-graded object in $\BIGR(\Ascr-o_X)$, satisfying (fin). Then $$\Ascr/\Ascr_{\ge n}\otimes_\Ascr \Fscr_{\ge p}=\coker (\Ascr_{\ge n}\otimes_\Ascr \Fscr_{\ge p}\r \Fscr_{\ge p})$$ Now up to inverse systems of the form $$\label{ref:3.35a} \underline{\HTor}_i^\Ascr(\Ascr_{\ge n},\Fscr/\Fscr_{\ge p})$$ we have $$\Ascr_{\ge n}\otimes_\Ascr \Fscr_{\ge p}=\Ascr_{\ge n}\otimes_\Ascr \Fscr$$ Then using lemma \[ref:3.9.8a\] we see that $ \Ascr_{\ge n}\otimes_\Ascr \Fscr $ is, modulo torsion inverse systems, equal to $ \Fscr_{\ge n} $. Assembling everything we find that up to torsion inverse systems and inverse systems of the form and $\Ascr/\Ascr_{\ge n}\otimes_\Ascr \Fscr_{\ge p}$ is equal to $\Fscr_{\ge p}/\Fscr_{\ge n}$ (where we let the inverse systems start with $n=p$). So the middle homology of is, up to torsion inverse systems and inverse systems of the form and , equal to the middle homology of $$\Fscr_{i+1}/(\Fscr_{i+1})_{\ge n} \r \Fscr_{i}/(\Fscr_{i})_{\ge n} \r \Fscr_{i-1}/(\Fscr_{i-1})_{\ge n}$$ Since this is in fact an exact complex we are left with showing that and are torsion. Let us for example consider . It is easy to verify directly from the definitions that $$(``{\invlim}{}"\underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr))_{o_X}= ``{\invlim}{}"\underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr_{o_X})=$$ Now as in the proof of lemma \[ref:3.1.17a\] one verifies that $\underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr_{o_X})$ is coherent as weak $\Ascr-o_X$-bimodule. Hence $\underline{\HHom}_{o_X}(\underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr_{o_X}),-)$ commutes with direct sums. Thus it suffices to show that $$\label{ref:3.36a} \dirlim\underline{\HHom}_{o_X}(\underline{\HTor}^\Ascr_i(\Ascr/\Ascr_{\ge n},\Tscr_{o_X}),E)=0$$ for $E$ an injective object in $\Mod(X)$, viewed as a graded object concentrated in a single degree. Now becomes equal to $$\label{ref:3.37a} \dirlim\underline{\HExt}^i_{\Ascr}(\Ascr/\Ascr_{\ge n},\underline{\HHom}_{o_X}(\Tscr_{o_X},E))$$ Clearly $\underline{\HHom}_{o_X}(\Tscr_{o_X},E)$ is right bounded and hence torsion. Now the fact that $\Tors(\Ascr)$ is stable (lemma \[ref:3.8.3a\]) together with Proposition \[ref:3.8.4a\] easily implies that is zero. Now we translate into more geometrical language. First we indicate when a map $f:\Ascr\r \Bscr$ defines a dual map $\bar{f}:\Proj(\Bscr)\r \Proj(\Ascr)$ \[ref:3.9.11a\] Let $f:\Ascr\r \Bscr$ be a morphism in $\Gralg(X)$ such that ${}_\Ascr\Bscr_{o_X}$ satisfies (fin). Then the functors $$\begin{aligned} \label{ref:3.38a} -\otimes_\Ascr\Bscr&:\Gr(\Ascr)\r \Gr(\Bscr)\\ (-)_\Ascr&:\Gr(\Bscr)\r \Gr(\Ascr) \label{ref:3.39a}\end{aligned}$$ factor through “$\Qgr$” and in this way define respectively $\bar{f}^\ast$ and $\bar{f}_\ast$ for a morphism $\bar{f}:\Proj(\Bscr)\r \Proj (\Ascr)$. In this case $\bar{f}_\ast$ is exact. That factors through “$\Qgr$” as well as the exactness of $\bar{f}_\ast$ is clear so we concentrate on . We have to show that if a map $\Mscr\r \Nscr$ in $\Mod(\Ascr)$ has torsion kernel and cokernel then the same is true for $$\Mscr\otimes_\Ascr\Bscr\r\Nscr\otimes_\Ascr\Bscr$$ Now $\Mod(\Ascr)$ is locally noetherian and $-\otimes_\Ascr\Bscr$ is compatible with direct limits, so we may restrict ourselves to the case where $\Mscr$, $\Nscr$ are noetherian and hence $\Mscr_n\r\Nscr_n$ is an isomorphism for $n\gg 0$. Now using the fact that $\Bscr$ satisfies (fin) it suffices to show that $$\label{ref:3.40a} \Mscr\otimes_\Ascr (\Ascr\otimes_{o_X}\Bscr') \r \Nscr\otimes_\Ascr (\Ascr\otimes_{o_X}\Bscr')$$ has torsion kernel and cokernel for all $\Bscr'\in \BIMOD(o_X-o_X)$. However since $-\otimes_\Ascr(\Ascr\otimes_{o_X}\Bscr')=-\otimes_{o_X}\Bscr'$ it is clear that is an isomorphism in high degree. It remains to verify the adjointness of and when defined on “$\operatorname{QGr}$”. To this end we have to construct for $\Mscr\in\Gr(\Ascr)$ and $\Nscr\in \Gr(\Bscr)$ a natural isomorphism $$\Hom_{\Qgr(\Bscr)}(\pi(\Mscr\otimes_\Ascr\Bscr),\pi\Nscr) \cong \Hom_{\Qgr(\Ascr)}(\pi\Mscr,\pi(\Nscr_\Ascr))$$ and as usual it suffices to do this for $\Mscr$ noetherian. Then $$\begin{aligned} \Hom_{\Qgr(\Bscr)}(\pi(\Mscr\otimes_\Ascr\Bscr),\pi\Nscr) &=\dirlim \Hom_{\Gr(\Bscr)}((\Mscr\otimes_\Ascr \Bscr)_{\ge n},\Nscr)\\ &=\dirlim \Hom_{\Gr(\Bscr)}(\Mscr_{\ge n}\otimes_\Ascr \Bscr,\Nscr)\\ &=\dirlim \Hom_{\Gr(\Ascr)}(\Mscr_{\ge n},\Nscr_\Ascr)\\ &=\Hom_{\Qgr(\Ascr)}(\pi\Mscr,\pi(\Nscr_\Ascr))\end{aligned}$$ The second equality follows from the fact that $\Bscr$ satisfies (fin) as $\Ascr-o_X$-bimodule and thus we can apply lemma \[ref:3.9.8a\] Thus we obtain \[ref:3.9.12a\] Let $f:\Ascr\r\Bscr$ be a morphism in $\Gralg(X)$ and assume that $\Bscr$ is as in Proposition \[ref:3.9.10a\]. Let $\bar{f}:\Proj\Bscr\r\Proj\Ascr$ be defined by $\bar{f}^\ast(\pi\Mscr)=\pi(\Mscr\otimes_\Ascr\Bscr)$ for $\Mscr\in\Gr(\Ascr)$. Then for $\Nscr\in\Qgr(\Bscr)$ $$R^i\beta_\ast(\Nscr)=R^i\alpha_\ast(\bar{f}_\ast\Nscr)$$ Assume $\Nscr=\pi\Nscr'$, $\Nscr'\in\Gr(\Bscr)$. Summing over shifts, it is sufficient to show $$R^i\omega_\Bscr(\pi\Nscr')_\Ascr=R^i\omega_\Ascr(\pi(\Nscr'_\Ascr))$$ If $i\ge 1$ then by this is equivalent to $$R^{i+1}\tau_\Bscr(\Nscr')_\Ascr=R^{i+1}\tau_\Ascr(\Nscr'_\Ascr)$$ which we have shown. The case $i=0$ follows from considering and the 5-lemma. Now we use Proposition \[ref:3.9.10a\] to prove the following result \[ref:3.9.13a\] Assume that we have a map $f:\Ascr\r\Bscr$ of $\NN$-graded algebras on $X$. Assume furthermore that there is an exact sequence $$0\r I\r \Ascr\r{}_\Ascr\Bscr_\Ascr\r 0$$ in $\Bigr(\Ascr-\Ascr)$ with $I$ a graded invertible $\Ascr$-bimodule, satisfying (fin) and living in degree $\ge 1$. Then - If $\Bscr$ is noetherian then so is $\Ascr$. Assume now that $\Bscr$ is noetherian. - If $\Bscr$ satisfies $\chi$ then so does $\Ascr$. - Let “$\cd$” stand for cohomological dimension. Then $$\cd \tau_\Bscr+1\ge \cd\tau_\Ascr \ge \cd \tau_\Bscr$$ 1. We have to show that if $\Mscr\in\Qch(X)$ is noetherian then so is $\Mscr\otimes_{o_X}\Ascr$. Now if $\Nscr\in\Gr(\Ascr)$ is left bounded then a variation of the classical argument by Hilbert shows that if $\Nscr\otimes_\Ascr\Bscr$ is noetherian then the same holds for $\Nscr$. Since by hypotheses $$(\Mscr\otimes_{o_X}\Ascr)\otimes_\Ascr\Bscr=\Mscr\otimes_{o_X}\Bscr$$ is noetherian we are done. 2. We start by observing that $\Bscr$ clearly satisfies the hypotheses of Proposition \[ref:3.9.10a\], so we can use that proposition. In the sequel we will denote by $G$ the functor $-\otimes_\Ascr I$ and $t$ the natural transformation $G\r \Id_{\Gr(\Ascr)}$ coming from the inclusion $I\r \Ascr$. Furthermore $G^n(\Nscr)\xrightarrow {t^n} \Nscr$ is by definition obtained from tensoring $I^n\xrightarrow{t^n}\Ascr$ with $\Nscr$. By $\Nscr t^n$ we denote the image of this map. We say that $\Nscr$ is annihilated by $t^n$ if $\Nscr t^n=0$. $\Nscr$ is $t$-torsion if it is the union of subobjects annihilated by some $t^n$. Similarly $\Nscr$ is $t$-torsion free if multiplication by $t$ is injective. It is easy to see that the functors $$\begin{aligned} -\otimes_\Ascr\Bscr&:\Gr(\Ascr)\r \Gr(\Bscr)\\ (-)_\Ascr&:\Gr(\Bscr)\r \Gr(\Ascr)\end{aligned}$$ induce inverse equivalences between $\Gr(\Bscr)$ and the full subcategory of $\Gr(\Ascr)$ consisting of objects annihilated by $t$. So we will identify these two categories. Let $\Mscr\in\gr(\Ascr)$. As usual $\Mscr$ is an extension $$0\r\Mscr_1\r \Mscr\r \Mscr_2\r 0$$ where $\Mscr_1$ is $t$-torsion and $\Mscr_2$ is $t$-torsion free. Now $\Mscr$ is noetherian and hence so is $\Mscr_1$. Thus $\Mscr_1t^n=0$ for some $n$ and in particular we can write $\Mscr_1$ as an extension of objects annihilated by $t$. We conclude that to verify condition $\chi$ we have to show that $R^i\tau_\Ascr(\Nscr)_{\ge 0}$ is right bounded and noetherian for two classes of noetherian graded $\Ascr$-modules : 1. Those that are annihilated by $t$. 2. Those that are $t$-torsion free. Let us first treat (a). If $\Nscr t=0$ then by the previous discussion $\Nscr=\Nscr'_\Ascr$ for some $\Nscr'\in\gr(\Bscr)$. Thus $$\begin{aligned} R^i\tau_\Ascr(\Nscr)&=R^i\tau_\Ascr(\Nscr'_\Ascr)\\ &=R^i\tau_\Bscr(\Nscr')_\Ascr \qquad(\text{Prop. \ref{ref:3.9.10a}})\end{aligned}$$ which shows what we want. Now consider the case that $\Nscr$ is torsion free. We have an exact sequence in $\Gr(\Ascr)$ $$0\r G(\Nscr)\xrightarrow{t} \Nscr\r \Nscr/\Nscr t\r 0$$ and again $\Nscr/\Nscr t=\Nscr''_\Ascr$ for some $\Nscr''\in\gr(\Bscr)$. This gives us the following exact sequence (using Prop. \[ref:3.9.10a\]) $$\label{ref:3.41a} R^{i-1}\tau_\Bscr(\Nscr'')_\Ascr\r G(R^i\tau_\Ascr(\Nscr)) \xrightarrow{t} R^i\tau_\Ascr(\Nscr) \r R^i\tau_\Bscr(\Nscr'')_\Ascr$$ By lemma \[ref:3.9.1a\] this implies $R^i\tau_A(\Nscr)\in\Tors(\Ascr)$. $R^i\tau_\Ascr(\Nscr)$ is $t$-torsion since $I$ lives purely in positive degree. On the other hand we obtain from and the fact that $\Bscr$ satisfies $\chi$ that $R^i\tau_\Ascr(\Nscr)_{\ge n}$ is $t$-torsion free for $n\gg 0$. Combining this we obtain that $R^i\tau_\Ascr(\Nscr)_{\ge n}=0$ for $n\gg 0$ which shows that $R^i\tau_\Ascr(\Nscr)$ is right bounded. From we obtain exact sequences $$0\r \text{noetherian} \r G(R^i\tau_\Ascr(\Nscr)_{\ge m-1})\xrightarrow{t} R^i\tau_\Ascr(\Nscr)_{\ge m}$$ and by descending induction on $m$ we find that $R^i\tau_\Ascr(\Nscr)_{\ge 0}$ is noetherian. 3. From Proposition \[ref:3.9.10a\] it follows immediately that $\cd \tau_\Ascr\ge \cd \tau_\Bscr$. We therefore concentrate on the other inequality. Assume $\Nscr\in\Gr(\Ascr)$, $p=\cd\tau_\Bscr$. We have to show $R^q\tau_\Ascr(\Nscr)=0$, $q>p+1$. Since $\tau_\Ascr$ commutes with direct limits and $\Gr(\Ascr)$ is locally noetherian, $R^q\tau_\Ascr$ also commutes with direct limits [@Groth1]. Hence we may assume that $\Nscr$ is noetherian. Using the same reduction as in (2) we may assume that either $\Nscr t=0$ or $\Nscr$ is $t$-torsion free, The first case is trivial by Proposition \[ref:3.9.10a\] so we look at the last case. Using for $i=q$ we find that $G(R^q\tau_\Ascr(\Nscr))\r R^q\tau_\Ascr(\Nscr)$ is an isomorphism. Since on the other hand $R^q\tau_\Ascr(\Nscr)$ is $t$-torsion we conclude $R^q\tau_\Ascr(\Nscr)=0$. Proposition \[ref:3.9.13a\](3) can be stated in more geometric terms. Assume that $\Ascr$, $\Bscr$ are noetherian $\NN$-graded algebras which fit in an exact sequence as in Prop. \[ref:3.9.13a\]. Then $$\label{ref:3.42a} \cd\beta_\ast\le \cd \alpha_\ast\le \cd \beta_\ast+1$$ It is easy to see that $$\cd\alpha_\ast=\cd\underline{\alpha}_\ast=\cd\omega_\Ascr$$ and similarly for $\beta$. Furthermore from it follows that $$\label{ref:3.43a} \cd\omega_\Ascr=\max(\cd\tau_\Ascr-1,0)$$ Combining this with Prop. \[ref:3.9.13a\](3) yields what we want. \[ref:3.9.15a\] Assume $\Ascr$ is noetherian and $\NN$-graded. Let $\Tscr$ be a right bounded graded $\Ascr-\Ascr$-bimodule. Then for $\Mscr\in\Gr(\Ascr)$ we have $$\label{ref:3.44a} \uHHom_\Ascr(\Tscr,\Mscr)=\uHHom_\Ascr(\Tscr,\tau(\Mscr))$$ If $\Tscr$ is in fact coherent as $o_X-o_X$-bimodule then $\uHHom_\Ascr(\Tscr,\Mscr)$ is torsion. To prove it is sufficient to show that the right and lefthand side of that equation represent the same functor. This is routine, using the fact that if $\Nscr\in \Gr(\Ascr)$ then $\Nscr\otimes_\Ascr \Tscr$ is torsion by lemma \[ref:3.8.1a\].2. Now we prove the second part of the lemma. By the first part we may clearly assume that $\Mscr$ is torsion. The coherentness of $\Tscr$ also implies that $\Tscr$ is left bounded. We claim that $\Tscr$ is coherent as a graded $\Ascr$-bimodule. In the same way as in \[ref:3.1.6a\] it suffices to show that $-\otimes_\Ascr\Tscr$ preserves $\gr(\Ascr)$. Now since $\Ascr$ is noetherian, any object in $\gr(\Ascr)$ has a presentation consisting of objects of the form $\Pscr\otimes_{o_X}\Ascr$ where the $\Pscr$ are noetherian $o_X$-modules. We now use the fact that $-\otimes_\Ascr\Tscr$ is right exact. Hence $\uHHom_\Ascr(\Tscr,-)$ commutes with direct limits. Therefore we may assume that $\Mscr$ is right bounded. But $\uHHom_\Ascr(\Tscr,\Mscr)$ is contained in $\uHHom_{o_X}(\Tscr,\Mscr)$ which is now clearly also right bounded. This proves what we want. We will need the following variant of Proposition \[ref:3.9.13a\] \[ref:3.9.16a\] Assume that we have a surjective map $f:\Ascr\r \Bscr$ such that $I= \ker f$ is a coherent object in $\Bigr(o_X-o_X)$. Then for $\Mscr\in\Gr(\Bscr)$ we have $$\label{ref:3.45a} R^i\tau_A(\Mscr_A)\cong (R^i\tau_\Bscr(\Mscr))_\Ascr$$ Furthermore if $\Bscr$ satisfies $\chi$ then so does $\Ascr$ and $\cd \tau_\Bscr=\cd \tau_\Ascr$. It is clear that is a special case of Proposition \[ref:3.9.10a\]. Assume that $\Bscr$ satisfies $\chi$. To show that $\Ascr$ also satisfies $\chi$ we have to show that for $\Nscr\in\gr(\Ascr)$ one has $R^i\tau_\Ascr(\Nscr)_{\ge 0}\in\gr(\Ascr)$. Let $\Kscr=\ker(\Nscr\r \Nscr\otimes_\Ascr\Bscr)$. Since $\Nscr$ is noetherian, the same is true for $\Kscr$. Furthermore by Corollary \[ref:3.8.2a\] it follows that $\Kscr$ is torsion. Hence $$R\tau^i_\Ascr(\Kscr)= \begin{cases} \Kscr&\text{if $i=0$}\\ 0&\text{if $i>0$} \end{cases}$$ Thus the long exact sequence for $R^i\tau_\Ascr$ together with yields an exact sequence $$0\r \Kscr \r \tau_\Ascr(\Nscr)\r \tau_\Bscr(\Nscr\otimes_\Ascr\Bscr)\r 0$$ and isomorphisms $$R^i\tau_\Ascr(\Nscr)\r R^i\tau_\Bscr(\Nscr\otimes_\Ascr\Bscr)$$ for $i\ge 1$. This yields that $\Ascr$ satisfies $\chi$ and also that $\cd\tau_\Bscr\ge \cd \tau_\Ascr$. Since implies that $\cd\tau_\Bscr\le \cd \tau_\Ascr$, we are through. Higher inverse images {#ref:3.10b} --------------------- Assume that $\alpha:Y\r X$ is a map of quasi-schemes. As was observed above the definition of $R^i\alpha_\ast$ presents no difficulty since $\Mod(Y)$ has enough injectives. However the definition of $L_i\alpha^\ast$ is more delicate since we have not assumed that $\Mod(X)$ has enough flat objects. As an approximation we define $L_i\alpha^\ast$ as the functor $\MOD(X)\r \MOD(Y)$ given by $$\HHom_{o_Y}(L_i\alpha^\ast \Mscr,E)=\HExt^i_{o_X}(\Mscr,\alpha_\ast E)$$ where $E$ runs through the injectives in $\Qch(Y)$. Furthermore we will define $\cd \alpha^\ast$ as the maximum $i$ such that $L_i\alpha^\ast$, restricted to $\Mod(Y)$ is non-zero. Let $X$ be a quasi-scheme and let $\Ascr\in\Alg(X)$. Put $Y=\Spec \Ascr$ and let $\alpha:Y\r X$ be the structure map. Then one has $L_i\alpha^\ast(-)=\HTor_i^{o_X}(-,\Ascr)$. This lemma is proved in a similar way as the lemma below which covers the graded case. In the graded case we need more hypotheses since we have defined things in less generality. \[ref:3.10.2a\] Let $X$ be a noetherian quasi-scheme and let $\Ascr$ be a noetherian $\NN$-graded algebra on $X$. Put $Y=\Proj\Ascr$. Assume that $\Mscr\in\Mod(X)$ is such that $\underline{\HTor}_i^{o_X}(\Mscr,\Ascr)\in \Gr(\Ascr)$. Then $$L_i\alpha^\ast\Mscr=\pi(\underline{\HTor}_i^{o_X}(\Mscr,\Ascr))$$ We have to show that $$\ \label{ref:3.46a} \Ext^i_{o_X}(\Mscr,\alpha_\ast F)= \Hom_{\operatorname{QGr}(\Ascr)} ( \pi(\underline{\HTor}_i^{o_X}(\Mscr,\Ascr)) ,F)$$ where $F$ runs through the injectives in $\Mod(Y)$. Put $F=\pi E$. It follows from lemma \[ref:3.8.3a\] that $E$ is an injective in $\Gr(\Ascr)$ satisfying $\tilde{E}=E$. Then by adjointness the righthand side of becomes $$\Ext^i_{o_X}(\Mscr,\HHom_{\Gr(\Ascr)}(\Ascr,E)) = \Ext^i_{o_X}(\Mscr,E_0)$$ Hence the assertion we have to prove boils down to $\alpha_\ast \pi E=E_0$. This follows from the fact that $\alpha_\ast(-)=\omega(-)_0$. On the positive part of an algebra strongly graded with respect to a localizing subcategory. {#ref:3.11b} -------------------------------------------------------------------------------------------- In this section we fix the following situation. $X$ is a noetherian quasi-scheme, $\Sscr$ is a localizing subcategory in $\Qch(X)$. $\Ascr$ is a noetherian algebra on $X$, strongly graded with respect to $\Sscr$ (see §\[ref:3.3b\]). With $\Ascr_{\ge 0}$ we denote the positive part of $\Ascr$. It is easy to see that $\Ascr_{\ge 0}$ is also noetherian. We note the following. \[ref:3.11.1a\] Let $\Mscr\in\Gr(\Ascr_{\ge 0})$. Then the $\Ascr_{\ge 0}$-module structure on $\tilde{\Mscr}$ extends in a natural way to an $\Ascr$-module structure. Multiplication defines graded $\Ascr_{\ge 0}$-bimodule maps $$\Ascr_{\ge m}\otimes_{\Ascr_{\ge 0}}\Ascr_{\ge n}\r \Ascr_{\ge m+n}$$ Applying $\uHHom_{\Ascr_{\ge 0}}(-,\Mscr)$ yields a map $$\uHHom_{\Ascr_{\ge 0}}(\Ascr_{\ge m+n},\Mscr)\r \uHHom_{\Ascr_{\ge 0}}(\Ascr_{\ge m},\uHHom_{\Ascr_{\ge 0}}(\Ascr_{\ge n},\Mscr))$$ which by adjointness yields a map $$\uHHom_{\Ascr_{\ge 0}}(\Ascr_{\ge m+n},\Mscr)\otimes_{\Ascr_{\ge 0}}\Ascr_{\ge m} \r \uHHom_{\Ascr_{\ge 0}}(\Ascr_{\ge n},\Mscr)$$ Taking direct limits over $n$, using and letting $m$ go to $-\infty$ we find a map $$\tilde{\Mscr}\otimes_{\Ascr_{\ge 0}}\Ascr\r \tilde{\Mscr}$$ A straightforward, but mildly tedious verification shows that this is an $\Ascr$-module structure. According to Corollary \[ref:3.8.2a\] we have the following inverse equivalences $$\label{ref:3.47a} \Gr(\Ascr)/\Tors(\Ascr) \xymatrix{ \ar@<1ex>[rr]^-{(-)_{\Ascr_{\ge 0}}} & & \ar@<1ex>[ll]^-{-\otimes_{\Ascr_{\ge 0}}\Ascr} } \Gr(\Ascr_{\ge 0})/\Tors(\Ascr_{\ge 0})$$ By lemma \[ref:3.3.1a\] we have inverse equivalences $$\label{ref:3.48a} \Mod(\Ascr_0)/\Sscr \xymatrix{ \ar@<1ex>[rr]^-{-\otimes_{\Ascr_0}\Ascr} & & \ar@<1ex>[ll]^-{(-)_{{0}}} } \Gr(\Ascr)/\Sscr(\Ascr)$$ In order to combine these equivalences we observe that $ \Tors(\Ascr)\subset \Sscr(\Ascr) $. Indeed let $\Mscr\in\Tors(\Ascr)$. Since $\Tors(\Ascr)$ and $\Sscr(\Ascr)$ are closed under direct limits we may assume that $\Mscr$ is right bounded. But then $\Mscr(N)_0=0\in\Sscr$ for $N\gg 0$ and hence according to lemma \[ref:3.3.1a\] we have $\Mscr(N)\in\Sscr(\Ascr)$. Thus the same holds for $\Mscr$. Define $\QSscr(\Ascr_{\ge 0})$ as the image in $\operatorname{QGr}(\Ascr_{\ge 0})$ of $\Sscr(\Ascr_{\ge 0})$ under the quotient map. One notes that $\QSscr(\Ascr_{\ge 0})$ has the following alternative description. $\QSscr(\Ascr_{\ge 0})$ is precisely the image of $\Sscr(\Ascr)$ under $\pi_\Ascr(-)_{\Ascr_{\ge 0}}$. If we look at the commutative diagram $$\begin{CD} \Gr(\Ascr_{\ge 0})@> \pi_{\Ascr_{\ge 0}}>> \operatorname{QGr}(\Ascr_{\ge 0})\\ @A (-)_{\Ascr_{\ge 0}} AA @A (-)_{\Ascr_{\ge 0}} AA\\ \Gr(\Ascr) @>\pi_\Ascr>> \operatorname{QGr}(\Ascr) \end{CD}$$ then we see that it is sufficient to show that the image of $\Sscr(\Ascr)$ under $(-)_{\Ascr_{\ge 0}}$ is $\Sscr(\Ascr_{\ge 0})$ modulo $\Tors(\Ascr_{\ge 0})$. It is clear that this image is indeed contained in $\Sscr(\Ascr_{\ge 0})$. Conversely let $\Mscr\in\Sscr(\Ascr_{\ge 0})$. According to Corollary \[ref:3.8.2a\] we have $\Mscr= (\Mscr\otimes_{\Ascr_{\ge 0}}\Ascr)_{\Ascr_{\ge 0}}$ modulo $\Tors(\Ascr_{\ge 0})$. Since $\Mscr\otimes_{\Ascr_{\ge 0}}\Ascr$ is contained in $\Sscr(\Ascr)$ we are through. Combining , and the previous lemma yields equivalences. $$\Mod(\Ascr_0)/\Sscr \xymatrix{ \ar@<1ex>[rr]^-{-\otimes_{\Ascr_0}\Ascr} & & \ar@<1ex>[ll]^-{(-)_{{0}}} } \Gr(\Ascr)/\Sscr(\Ascr) \xymatrix{ \ar@<1ex>[rr]^-{(-)_{\Ascr_{\ge 0}}} & & \ar@<1ex>[ll]^-{-\otimes_{\Ascr_{\ge 0}}\Ascr} } \operatorname{QGr}(\Ascr_{\ge 0})/\QSscr(\Ascr_{\ge 0})$$ Looking only at the outer categories yields equivalences $$\Mod(\Ascr_0)/\Sscr \xymatrix{ \ar@<1ex>[rr]^-{-\otimes_{\Ascr_0}\Ascr_{\ge 0}} & & \ar@<1ex>[ll]^-{(-\otimes_{\Ascr_{\ge 0}}\Ascr)_0} } \operatorname{QGr}(\Ascr_{\ge 0})/\QSscr(\Ascr_{\ge 0})$$ Define $U=\Spec \Ascr_0$, $Y=\Proj \Ascr_{\ge 0}$ and let the map $\alpha :Y\r U$ be given by $\alpha^\ast=\pi(-\otimes_{\Ascr_0}\Ascr_{\ge 0})$. Let $\Mscr\in\Gr(\Ascr_{\ge 0})$. Modulo $\Sscr$, $\alpha_\ast\pi\Mscr$ is given by $(\Mscr\otimes_{\Ascr_{\ge 0}}\Ascr)_0$. Without loss of generality we may assume that $\Mscr=\tilde{\Mscr}$. Hence in particular by lemma \[ref:3.11.1a\] $\Mscr=\Nscr_{\Ascr_{\ge 0}}$ where $\Nscr$ is a graded $\Ascr$-module. Furthermore $\alpha_\ast\pi\Mscr= \Mscr_0=\Nscr_0$. According to Corollary \[ref:3.8.2a\] we now have that the canonical map $$\Mscr\otimes_{\Ascr_{\ge 0}}\Ascr=\Nscr_{\Ascr_{\ge 0}}\otimes_{\Ascr_{\ge 0}} \Ascr\r \Nscr$$ is an isomorphism modulo $\Tors(\Ascr)\subset \Sscr(\Ascr)$. Hence the restricted map $$(\Mscr\otimes_{\Ascr_{\ge 0}}\Ascr)_0\r \Nscr_0$$ is an isomorphism modulo $\Sscr$. This is precisely what we had to prove. Let us now introduce the more suggestive notation $\alpha^{-1}(\Sscr)$ for $\QSscr(\Ascr_{\ge 0})$. Summarizing everything, we have shown. \[ref:3.11.4a\] Let $U,Y,\alpha,\Sscr$ be as above. Then the functors $\alpha^\ast$, $\alpha_\ast$ factor through $\Sscr$ and $\alpha^{-1}(\Sscr)$ to define inverse equivalences between $\Qch(U)/\Sscr $ and $\Qch(Y)/\alpha^{-1}(\Sscr)$. Veronese subalgebras {#ref:3.12b} -------------------- Let $X$ be a quasi-scheme and let $\Ascr\in\Gralg(X)$ be noetherian and $\NN$-graded. The $n$’th Veronese of $\Ascr$ is the graded subalgebra $\Ascr^{(n)}$ of $\Ascr$ defined by $(\Ascr^{(n)})_m=\Ascr_{nm}$. If $\Mscr\in\Gr(\Ascr)$ then $\Mscr^{(n)}$ is defined similarly. Then we have the following. \[ref:3.12.1a\] Assume that $X$ and $\Ascr$ are noetherian and that $\Ascr$ is generated in degree one (cfr §\[ref:3.9b\]). Then the functors $$\begin{aligned} \Gr(\Ascr)\r \Gr(\Ascr^{(n)}):&\Mscr\r \Mscr^{(n)}\\ \Gr(\Ascr^{(n)})\r\Gr(\Ascr):\Nscr\r \Nscr\otimes_{\Ascr^{(n)}}\Ascr\end{aligned}$$ factor over $\Tors(\Ascr)$ and $\Tors(\Ascr^{(n)})$ and in this way define inverse equivalences between $\Proj\Ascr$ and $\Proj\Ascr^{(n)}$. This is formally similar to the ring case. See for example [@V]. Pseudo-compact rings {#ref:4a} ==================== In the sequel we will study the formal local structure of some specific quasi-schemes. It turns out that this is best described in terms of pseudo-compact rings, so for the convenience of the reader we collect some of the properties of such rings here. Most of this is taken from [@VdBVG] and [@Gabriel]. We refer the reader to these papers for more information. A topological right module $M$ over a topological ring $A$ is pseudo-compact if it is Hausdorf, complete and its topology is generated by right submodules of finite colength. $A$ is said to be a pseudo-compact ring if $A$ is pseudo-compact as a right $A$-module. Left pseudo-compact is defined similarly. By $\operatorname{PC}(A)$, we will denote the category of right pseudo-compact modules over a right pseudo-compact ring. By [@Gabriel] $\operatorname{PC}(A)$ is an abelian category satisfying AB5$^\ast$ and AB3. Let $A$ be a pseudo-compact ring. The dual category of $\operatorname{PC}(A)$ is a locally finite category. That is a Grothendieck category generated by objects of finite length. Conversely, if $\Cscr$ is a locally finite category then $\Cscr$ can be realized as $\operatorname{PC}(A)^\circ$ for some pseudo-compact ring $A$ [@Gabriel]. $A$ is constructed as follows. Let $E$ be an injective cogenerator of $\Cscr$, containing every indecomposable injective at least once and put $A=\End_\Cscr(E)^\circ$. If $S$ is a finite length subobject of $E$ then $$\frak{l}(S)=\{f\in A\mid f(S)=0\}$$ defines a right ideal of finite colength in $A$. We take these right ideal as the basis for a topology on $A$. In this way $A$ becomes a pseudo-compact ring. If $M\in \Cscr$ then we topologize $\Hom_\Cscr(M,E)$ in a similar way. The functor $M\mapsto \Hom_\Cscr(M,E)$ defines a duality between $\Cscr$ and $\operatorname{PC}(A)$. As an application we obtain : \[ref:4.1a\] Inverse limits of projectives are projective in $\operatorname{PC}(A)$. This follows from the fact that $\operatorname{PC}(A)^\circ$ is a locally finite category and hence direct limits of injectives are injective. In nice cases there are good relations between the properties of $\operatorname{PC}(A)$ and $\Mod(A)$. For example by [@Gabriel] it follows that the forgetful functor $\operatorname{PC}(A)\r \Mod(A)$ is faithful and commutes with kernels, cokernels and products. By [@VdBVG Lemma 3.4], an object in $\operatorname{PC}(A)$ is simple in $\operatorname{PC}(A)$ if and only it is simple in $\Mod(A)$. A similar result holds for the property of being noetherian [@VdBVG Cor 3.10]. As usual we denote by $\operatorname{pc}(A)$ the category of noetherian pseudo-compact $A$-modules. An object $M$ in $\operatorname{PC}(A)$ is said to be finitely generated in $\operatorname{PC}(A)$ if $M$ is a quotient of $A^n$ in $\operatorname{PC}(A)$ for some $n$. If $M,N\in\operatorname{PC}(A)$, $M$ finitely generated then according to [@VdBVG Prop.3.5] we have $$\Hom_{\operatorname{PC}(A)}(M,N)=\Hom_{\Mod(A)}(M,N)$$ In particular, $M$ is finitely generated in $\operatorname{PC}(A)$ if and only if it is finitely generated in $\Mod(A)$. From this one can deduce : \[ref:4.2a\] [@VdBVG Cor. 3.8] If $M$ is a finitely generated pseudo-compact $A$-module then $L\subset M$ is open if and only if $M/L$ is of finite length and pseudo-compact when equipped with the discrete topology. (Note that a linear Hausdorf topology on a module of finite length is automatically discrete.) If $A$ is noetherian then it follows from [@VdBVG Prop. 3.19] that the forgetful functor $\operatorname{pc}(A)\r \mod(A)$ is an equivalence of categories. More generally we say that $A$ is locally noetherian if for every primitive idempotent $e$ in $A$ we have that $eA$ is noetherian. Assume that $A$ is locally noetherian. Then by [@VdBVG Cor. 3.15] we have that the forgetful functor $\operatorname{pc}(A)\r\mod(A)$ is fully faithful and closed under extensions. Let $\operatorname{PCFin}(A)$ denote the category of finite length objects in $\operatorname{PC}(A)$. It follows that the objects are precisely the finite length objects in $\Mod(A)$ whose Jordan-Holder quotients are pseudo-compact simples. It also follows that if $M$ is a noetherian $A$-module then the topology on $M$ is simply the cofinite topology. Let $A$ be an arbitrary pseudo-compact ring. Then we denote by $\operatorname{Dis}(A)$ the category of topological $A$-modules which are discrete. It is clear that $\operatorname{Dis}(A)$ is the full subcategory of $\Mod(A)$ consisting of modules $M$ such that for all $m\in M$, $\Ann_A(m)$ is open in $A$, or equivalently that $mA$ is pseudo-compact of finite length (this follows for example from lemma \[ref:4.2a\]). From this we deduce that $\operatorname{Dis}(A)$ is a locally finite category. Clearly $\operatorname{PC}(A)\cap \operatorname{Dis}(A)= \operatorname{PCFin}(A)$ (where the intersection is taken inside the category of topological right $A$-modules). It is interesting to observe that since $\operatorname{Dis}(A)$ is locally finite there must necessarily exist another pseudo-compact ring $A^\ast$ such that $\operatorname{PC}(A^\ast)=\operatorname{Dis}(A)^\circ$. In nice cases we have $A^\ast=A^\circ$ (see for example Proposition \[ref:4.10a\] below). Let $\operatorname{Top}(A)$ be the additive category of topological right $A$-modules. For $M,N\in \operatorname{Top}(A)$ we have functors $$\begin{gathered} \label{ref:4.1b} \Hom_{\operatorname{Top}(A)}(M,-): \operatorname{Dis}(A)\r \Ab\\ \Hom_{\operatorname{Top}(A)}(-,N):\operatorname{PC}(A)^\circ\r \Ab \label{ref:4.2b}\end{gathered}$$ \[ref:4.3a\] Let $(M_i)_{i\in J}$, $(N_j)_{j\in J}$ be respectively an inverse system in $\operatorname{PC}(A)$ and a directed system in $\operatorname{Dis}(A)$. Then $$\label{ref:4.3b} \Hom_{\operatorname{Top}(A)}(\invlim_i M_i,\dirlim_j N_j)=\dirlim_i\dirlim_j\Hom_{\operatorname{Top}(A)}(M_i,N_j)$$ If $M\in\operatorname{PC}(A)$, $N\in\operatorname{Dis}(A)$ then every continuous morphism $f:M\r N$ has an open kernel $M'$. This means that $N'=\im f\cong M/M'$ has finite length. Thus we have the following equalities $$\begin{aligned} \label{ref:4.4a} \Hom_{\operatorname{Top}(A)}(M,N)&=\dirlim_{N'}\Hom_{\operatorname{PC}(A)}(M,N')\\ \label{ref:4.5a} &=\dirlim_{M'}\Hom_{\operatorname{Dis}(A)}(M/M',N)\\ \label{ref:4.6a} &=\dirlim_{M',N'}\Hom_A(M/M',N')\end{aligned}$$ where $M'$ runs trough the open submodules in $M$ and $N'$ runs through the finite length submodules of $N$. Now let us for example show that $\Hom_{\operatorname{Top}(A)}(-,N)$ sends inverse limits to direct limits. We have $$\begin{aligned} \Hom_{\operatorname{Top}(A)}(\invlim_i M_i,N)&= \dirlim_{N'} \Hom_{\operatorname{PC}(A)}(\invlim_i M_i,N')\qquad\text{(eq. \eqref{ref:4.4a})}\\ &=\dirlim_{N'} \Hom_{\operatorname{PC}(A)^\circ}(N',\dirlim_i M_i) \\ &=\dirlim_{i,N'} \Hom_{\operatorname{PC}(A)^\circ}(N',M_i)\\ &=\dirlim_{i,N'}\Hom_{\operatorname{PC}(A)}(M_i,N')\\ &=\dirlim_i \Hom_{\operatorname{Top}(A)} (M_i,N)\qquad\text{(eq. \eqref{ref:4.4a})}\end{aligned}$$ The third equality follows from the fact that $N'$ is a noetherian object in the locally noetherian category $\operatorname{PC}(A)^\circ$. The proof that $\Hom_{\operatorname{Top}(A)}(M,-)$ commutes with direct limits is similar. Since $\operatorname{Dis}(A)$ has enough injectives and $\operatorname{PC}(A)$ has enough projectives, we can define the derived functors of and . Let us temporarily denote them by $\Ext_{I}^i(M,-)$ and $\Ext_{I\! I}^i(-,N)$. $\Ext_I^i$ and $\Ext_{I\!I}^i$ coincide when both are defined. That is if $M\in\operatorname{PC}(A)$ and $N\in \operatorname{Dis}(A)$ then $$\Ext_I^i(M,N)=\Ext_{I\!I}^i(M,N)$$ To prove this we have to show that if $P$ is projective in $\operatorname{PC}(A)$ then $\Hom_{\operatorname{Top}(A)}(P,-)$ is exact on $\operatorname{Dis}(A)$ and if $E$ is an injective in $\operatorname{Dis}(A)$ then $\Hom_{\operatorname{Top}(A)}(-,E)$ is exact when evaluated on $\operatorname{PC}(A)$. Since these are obviously dual statements we only prove the first one. By Proposition \[ref:4.3a\] $$\Hom_{\operatorname{Top}(A)}(P,N)=\dirlim_{N'} \Hom_{\operatorname{PC}(A)}(P,N')$$ where $N'$ runs through the finite length submodules of $N$. Since $\Hom_{\operatorname{PC}(A)}(P,-)$ is exact on $\operatorname{PC}(A)$ and $\dirlim$ is exact on $\Ab$ we have to show that every exact sequence in $\operatorname{Dis}(A)$ $$0\r N_1\r N\r N_2\r 0$$ can be obtained as a direct limit of exact sequences of the form $$0\r N'_1\r N'\r N_2'\r 0$$ where $N'_1,N_2',N'$ are finite length subobjects of $N_1,N_2,N$. That this is true follows easily from the fact that $\operatorname{Dis}(A)$ is locally finite. Henceforth we neglect the distinction between $\Ext_{I}^i$ and $\Ext_{I\!I}^i$, and we simply write $\Ext^i_{\operatorname{Top}(A)}$. We then obtain the following generalization of Proposition \[ref:4.3a\] \[ref:4.5b\] Let $(M_i)_{i\in I}$, $(N_j)_{j\in J}$ be respectively an inverse system in $\operatorname{PC}(A)$ and a directed system in $\operatorname{Dis}(A)$. Then $$\label{ref:4.7a} \Ext^i_{\operatorname{Top}(A)}(\invlim_i M_i,\dirlim_j N_j)=\dirlim_i\dirlim_j\Ext^i_{\operatorname{Top}(A)}(M_i,N_j)$$ Let $N\in \operatorname{Dis}(A)$ and let $E^\cdot$ be an injective resolution of $N$ in $\operatorname{Dis}(A)$. Then $$\begin{aligned} \Ext^i_{\operatorname{Top}(A)}(\invlim_i M_i,N)&=H^i(\Hom_{\operatorname{Top}(A)}(\invlim_i M_i,E^\cdot))\\ &=\dirlim_i H^i(\Hom_{\operatorname{Top}(A)}(M_i,E^\cdot))\qquad \text{(Proposition \ref{ref:4.3a})} \\ &=\dirlim_i \Ext^i(M_i,N)\end{aligned}$$ The fact that $\Ext^i_{\operatorname{Top}(A)}(M,-)$ is compatible with direct limits is proved similarly. Of course the ordinary “$\Ext$” in the abelian categories $\operatorname{PC}(A)$ and $\operatorname{Dis}(A)$ is also defined. Since $\operatorname{PC}(A)$ has enough projectives we clearly have $$\Ext^i_{\operatorname{PC}(A)}(M,N)=\Ext^i_{\operatorname{Top}(A)}(M,N)$$ if $M,N\in \operatorname{PC}(A)$. Similarly if $M,N\in \operatorname{Dis}(A)$ then $$\Ext^i_{\operatorname{Dis}(A)}(M,N)=\Ext^i_{\operatorname{Top}(A)}(M,N)$$ since $\operatorname{Dis}(A)$ has enough injectives. \[ref:4.6b\] Let $M\in\operatorname{PC}(A)$, $N\in \operatorname{Dis}(A)$. We have the following formulas $$\label{ref:4.6c} \operatorname{proj\,dim}_{\operatorname{PC}(A)} M=\sup_i \{i\mid\Ext^i_{\operatorname{PC}(A)}(M,-)\neq 0\}$$ $$\label{ref:4.6d} \injdim_{\operatorname{Dis}(A)} N=\sup_i \{i\mid\Ext^i_{\operatorname{Dis}(A)}(-,N)\neq 0\}$$ If $A$ is locally noetherian and if $M$ is noetherian in $\operatorname{PC}(A)$ then $$\label{ref:4.6e} \operatorname{proj\,dim}_{\operatorname{PC}(A)} M=\sup_i\{i\mid \exists S\text{ simple }: \Ext^i_{\operatorname{PC}(A)}(M,S)\neq 0\}$$ and if $A^\ast$ is locally noetherian and $N$ is artinian in $\operatorname{Dis}(A)$ then $$\label{ref:4.6f} \injdim_{\operatorname{Dis}(A)} N=\sup_i\{i\mid \exists T\text{ simple }: \Ext^i_{\operatorname{Dis}(A)}(T,N)\neq 0\}$$ This is entirely classical. Let us for example prove the fourth equality. By degree shifting this amounts to showing that if $\Ext^1(T,N)=0$ for all simple $T$ and if $N$ is artinian then $N$ is injective. Let $E$ be the injective hull of $N$ and $U=E/N$. Let $q:E\r U$ be the quotient map. By hypotheses $E$ and hence $U$ is artinian. Also by hypotheses, the restriction of $q$ to $\operatorname{Soc}(E)\r \operatorname{Soc}(U)$ is surjective. Since socles are by definition semisimple this last map has a splitting which can be lifted to a map $t:U\r E$. By hypotheses $s=qt$ is the identity on $\operatorname{Soc}(U)$, from which it follows that $s$ is injective. Also $s^n(U)$ is a descending chain of submodules in $U$ and thus $s^n(U)=s^{n+1}(U)$. The injectivity of $s$ yields that $U=sU$ and thus $s$ is an automorphism of $U$. Now $ts^{-1}$ is a splitting of $q$ and hence $N$ is a direct summand of $E$, whence injective. As usual we define $$\begin{aligned} \gldim \operatorname{PC}(A)&=\sup_{M\in \operatorname{PC}(A)} \operatorname{proj\,dim}M\\ \gldim \operatorname{Dis}(A)&=\sup_{N\in \operatorname{Dis}(A)}\injdim N\end{aligned}$$ \[ref:4.7b\] The following holds. 1. $\gldim \operatorname{PC}(A)$ is equal to the supremum of the projective dimensions of the simple pseudo-compact $A$-modules (as objects in $\operatorname{PC}(A)$). 2. $\gldim \operatorname{Dis}(A)$ is equal to the supremum of the injective dimensions of the simple pseudo-compact $A$-modules (as objects in $\operatorname{Dis}(A)$). 3. If $A$ and $A^\ast$ are locally noetherian then $\gldim \operatorname{PC}(A)$ and $\gldim \operatorname{Dis}(A)$ are both equal to the supremum of the $i\in\NN$ such that there exist simple pseudo-compact-$A$-modules $S,T$ such that $\Ext^i(S,T)\neq 0$. 1\. and 2. are true by [@VdBVG Lemma 5.1]. 3. follows from 1. and 2. and the foregoing proposition. An object in $\operatorname{PC}(A)$ is said to be cosemisimple if it is a direct product of simple modules. This is equivalent with being semisimple in the dual category $\operatorname{PC}(A)^\circ$. If $M\in \operatorname{PC}(A)$ then we define $M/\rad(M)$ as the quotient of $M$ which is the socle of $M$ in $\operatorname{PC}(A)^\circ$. By construction $M/\rad(M)$ is the largest cosemisimple object in $\operatorname{PC}(A)$ which is a quotient of $M$. The radical of $M$, denoted by $\rad(M)$, is defined as the kernel of $M\r M/\rad(M)$. Since it is closed, it is pseudo-compact. From the fact that taking socles is left exact it also follows that the functor $M\mapsto M/\rad(M)$ is right exact. It is shown in [@Gabriel] that $\rad(A)$ is a twosided ideal and coincides with the ordinary Jacobson radical of $A$. From the fact that $\rad(A)$ annihilates all cosemisimple objects in $\operatorname{PC}(A)$ we obtain $$\label{ref:4.8a} M\rad(A)\subset \rad(M)$$ \[ref:4.8b\] (Nakayama’s lemma) If $M\in \operatorname{PC}(A)$ then $M=\rad(M)$ if and only if $M=0$. It is easy to see that a non-zero object in $\operatorname{PC}(A)$ maps onto at least one simple object. This proves the lemma. \[ref:4.9a\] The following are equivalent. 1. $M$ is finitely generated. 2. $M/M\rad(A)$ is a finitely generated $A/\rad(A)$-module. 3. $M/\rad(M)$ is a finitely generated $A/\rad(A)$-module. If any one of these conditions holds then $\rad(M)=M\rad(A)$. It is clear that 1. implies 2. From it follows that 2. implies 3. Hence we have to show that 3. implies 1. By lifting the generators of $M/\rad(M)$ we can construct a map $\theta:A^k\r M$ which becomes surjective after applying the functor $T\mapsto T/\rad(T)$. Given the right exactness of this functor we obtain that $C=\rad(C)$ for $C=\coker \theta$. By Nakayama it follows that $C=0$. This proves the first part of the lemma. If condition 2. holds then $M/M\rad(A)$ is a quotient of $(A/\rad(A))^k$ for some $k$, and hence $M/M\rad(A)$ is cosemisimple. We conclude that $\rad(M)\subset M\rad(A)$. For further reference we state the following formula $$\label{ref:4.9b} M/\rad(M)=\prod_{S\text{ simple}} S^{\alpha_{M,S}}$$ where $\alpha_{M,S}=\dim_{\End_{\operatorname{PC}(A)}(S)}\Hom_{\operatorname{PC}(A)}(M,S)$. This is easily proved by looking at the dual statement in $\operatorname{PC}(A)^\circ$. The definition of a pseudo-compact ring is essentially onesided, which is somewhat inconvenient. We now introduce a symmetric notion. Let $k$ be a field. A pseudo-compact ring which is a $k$ algebra is said to be cofinite if all simple pseudo-compact $A$-modules are finite dimensional over $k$. \[ref:4.10a\] Assume that $A$ is cofinite. Then 1. $A/\rad(A)=\prod_{i\in I} M_{n_i}(D_i)$, for finite dimensional division algebras $(D_i)_{i\in I}$. 2. The topology on $A$ is generated by twosided ideals of finite codimension. 3. $A$ is left and right pseudo-compact. 4. $S\mapsto \Hom_k(S,k)$ defines a duality between left and right pseudo-compact $A$-modules of finite length. In particular we can take $A^\ast=A^\circ$. 5. (Matlis-duality) $M\mapsto \Hom_{\operatorname{Top}(k)}(M,k)$ defines a duality between pseudo-compact left (right) $A$-modules and discrete right (left) $A$-modules. Thus $\operatorname{PC}(A)^\circ=\operatorname{Dis}(A^\circ)$ and $\operatorname{Dis}(A)^\circ=\operatorname{PC}(A^\circ)$. <!-- --> 1. According to Gabriel $A/\rad(A)$ is a product of endomorphism rings of vectorspaces over division algebras. The simple pseudo-compact modules over such a ring will be finite dimensional if and only if $A/\rad(A)$ has the indicated form. 2. Let $L\subset A$ be an open ideal. Then $S=A/L$ is a pseudo-compact right $A$-module of finite length. Put $T=\End_k(S)$ and consider $T$ as an $A$-bimodule in the obvious way. Then as right $A$-module, we have $T=S^{\dim_k S}$ and in particular $T$ is pseudo-compact. There is a canonical map of $A$-bimodules $A\r T$, given by the right action of $A$ on $S$. Let $M$ be the kernel of this map. Then $M$ is an open twosided ideal contained in $L$. 3. Since the topology on $A$ is generated by twosided ideals of finite codimension (by 2.) we see that $A$ is also pseudo-compact on the left. 4. This follows again easily from the fact that $S$ is annihilated by an open twosided ideal. 5. This is a consequence of 3. and the fact that objects in $\operatorname{Dis}(A)$ are direct limits of finite length pseudo-compact modules and objects in $\operatorname{PC}(A)$ are inverse limits of finite length pseudo-compact modules. \[ref:4.11a\] For a left and right locally noetherian cofinite pseudo-compact $k$-algebra the numbers $\gldim \operatorname{PC}(A)$, $\gldim \operatorname{Dis}(A)$, $\gldim \operatorname{PC}(A^\circ)$ and $\gldim \operatorname{Dis}(A^\circ)$ all coincide. We call this common value the global dimension of $A$ and denote it by $\gldim A$. This follows from Corollary \[ref:4.7b\].3 and Proposition \[ref:4.10a\].5. \[ref:4.12a\] Let $A,B$ be cofinite $k$-algebras. Let $M$ be a topological $A$-$B$-bimodule. Then $M$ is bi-pseudo-compact if the topology on $M$ is Hausdorf and complete and is generated by subbimodules $M'\subset M$ of finite codimension. It is easy to see that bi-pseudo-compact bimodules form an abelian category satisfying AB3 and AB5${}^\ast$. We denote this category by $\operatorname{PC}(A-B)$. If $A,B,C$ are three cofinite algebras and $M\in \operatorname{PC}(A-B)$, $N\in \operatorname{PC}(B-C)$ then we define $$M\ctimes_B N =\invlim_{M',N'} M/M'\otimes_B N/N'$$ where $M',N'$ run through all open subbimodules in $M$, $N$. By construction $M\ctimes_B N\in \operatorname{PC}(A-C)$. There are analogous definitions when $M$ or $N$ are onesided pseudo-compact modules. One easily obtains $-\ctimes_B-$ is right exact and commutes with inverse limits in both its factors. The following is also standard. \[ref:4.14a\] If $M$ is finitely presented as right $B$-module then $$M\ctimes_B N=M\otimes_B N$$ If $A,B$ are cofinite $k$-algebras then $A\ctimes_k B$ carries a canonical ring structure which makes it into a cofinite $k$-algebra. One has $\operatorname{PC}(A-B)=\operatorname{PC}(A^\circ\ctimes_k B)$. Furthermore the pseudo-compact simple $A^\circ\ctimes B$-modules are of the form $S\otimes_k T$ with $S$ simple in $\operatorname{PC}(A)$ and $T$ simple in $\operatorname{PC}(B)$ (this follows from the corresponding statement for finite dimensional algebras). From this one deduces $A\ctimes_k B/\rad (A\ctimes_k B)=A/\rad(A)\ctimes_k B/\rad(B)$. This observation allows one to prove \[ref:4.15a\] The forgetful functor $\operatorname{PC}(A-B)\r \operatorname{PC}(B)$ preserves projectives. It suffices to prove this for the functor $\operatorname{PC}(A^\circ\ctimes B)\r \operatorname{PC}(B)$. By the above discussion all projectives in $\operatorname{PC}(A^\circ\ctimes B)$ are products of $ Ae\ctimes fB$ where $e,f$ are primitive idempotents in $A$, $B$. These are direct summands of $A\ctimes B$, whence it suffices to show that $(A\ctimes B)_B$ is projective. This follows from the fact that $$\begin{aligned} A\ctimes B& =\invlim_{I,J} (A/I)\otimes (B/J)\\ &=\invlim_I(\invlim_J (A/I)\otimes (B/J))\\ &=\invlim_I A/I\otimes B\end{aligned}$$ where $I,J$ run through the open twosided ideals in $A$ and $B$. Hence $(A\ctimes B)_B$ is an inverse limit of projective $B$-modules, and is therefore itself projective (lemma \[ref:4.1a\]) We now give a structure theorem on cofinite $k$-algebras, which was more or less proved in [@VdBVG] (in slightly greater generality). Below let $(e_i)_{i\in I}$ be a summable set of primitive idempotents in a pseudo-compact ring $A$, having sum 1 (as in [@Gabriel]). \[ref:4.16a\] Let $A$ be cofinite, let $M$ be a pseudo-compact $A$-module and let $N$ be a bipseudo-compact $A$-bimodule. Put $A_{ij}=e_iAe_j$, $M_i=Me_i$, $N_{ij}=e_iNe_j$, equipped with the induced topology. Then $$\begin{aligned} A&=\prod_{i,j} A_{ij}\label{ref:4.10b}\\ M&=\prod_i M_i\label{ref:4.11b}\\ N&=\prod_{i,j} e_iNe_j \end{aligned}$$ These equalities are in fact homeomorphisms if we equip the righthand sides with the product topology. Furthermore the $A_i$ are cofinite with $A_i/\rad(A_i)=D_i$ (where $D_i$ is as in Proposition \[ref:4.10a\]), the $M_i$ are pseudo-compact $A_i$-modules and the $A_{ij}$ and $N_{ij}$ are bipseudo-compact $A_i-A_j$-modules. Similar to [@VdBVG Prop. 4.3]. In the case of a cofinite $k$-algebra, there is a simple test for a pseudo-compact module to be finitely generated. \[ref:4.17a\] Assume that $A$ is cofinite. Then $M\in \operatorname{PC}(A)$ is finitely generated if and only if for every simple pseudo-compact $S$ we have that $\dim_{\End_{\operatorname{PC}(A)}(S)}\Hom_{\operatorname{PC}(A)}(M,S)$ is finite and is bounded independently of $S$. This follows from Lemma \[ref:4.9a\] and . The following is a generalization of lemma \[ref:4.17a\]. \[ref:4.18a\] Assume that $A$ is cofinite. Let $M\in\operatorname{PC}(A)$. Then $M$ has a free resolution of length $n$ in $\operatorname{PC}(A)$ (or, equivalently, in $\Mod(A)$) $$F_n\r F_{n-1}\r \cdots \r F_1\r F_0\r M\r 0$$ where $F_i$ is of finite rank over $A$, if and only if for every $i\in\{0,\ldots,n\} $ and for every simple $S$ the dimension of $\Ext^i_{\operatorname{PC}(A)}(M,S) $ is finite and bounded independently of $S$. This is easily proved from the case $n=0$, by degree shifting. We will need the following result. \[ref:4.19a\] Assume that $A$ is cofinite, $M\in \operatorname{PC}(A-A)$ and $S\in \operatorname{Dis}(A)$. Then $\Ext^i_{\operatorname{Top}(A)}(M,S)\in \operatorname{Dis}(A)$. Here the “$\Ext$” is taken with respect to the right $A$-structure on $M$. Let $P^\cdot$ be a projective resolution of $M$ in $\operatorname{PC}(A-A)$. According to lemma \[ref:4.15a\] the terms of $P^\cdot$ are projective in $\operatorname{PC}(A)$. Hence we find $$\Ext^i_{\operatorname{PC}(A)}(M,S)=H^i(\Hom_{\operatorname{PC}(A)}(P^\cdot,S))$$ Hence it suffices to show that if $P$ is a projective object in $\operatorname{PC}(A-A)$ then $\Hom_{\operatorname{PC}(A)}(P,S)\in \operatorname{Dis}(A)$. Such a projective object is a product of direct summands of $A^\circ\ctimes A$, and since $\Hom_{\operatorname{PC}(A)}(-,S)$ transforms products in the first argument into direct sums (according to Proposition \[ref:4.3a\]) it suffices to show that $\Hom_{\operatorname{PC}(A)}(A\ctimes A,S)$ lies in $\operatorname{Dis}(A)$. Now we have $$\begin{aligned} \Hom_{\operatorname{PC}(A)}(A\ctimes A,S)&=\Hom_{\operatorname{PC}(A)}(\invlim_I A/I\otimes A,S)\\ &=\dirlim_I \Hom_{\operatorname{PC}(A)}(A/I\otimes A,S)\qquad (\text{Proposition \ref{ref:4.3a}})\\ &=\dirlim_I (A/I)^\ast\otimes S\end{aligned}$$ Here $I$ runs through the open twosided ideals in $A$. The right $A$-structure on $(A/I)^\ast\otimes S$ we use is the one on $(A/I)^\ast$. Thus $(A/I)^\ast\otimes S=(A/I)^t$ for some $t$, and we are done. We can define the derived functors of $-\ctimes_{B}-$ in both arguments. For lack of a better notation we denote them by $\Tor_i^{\operatorname{PC}(B)}(-,-)$. Thus if $A,B,C$ are cofinite $k$-algebras and $M\in \operatorname{PC}(A-B)$, $N\in\operatorname{PC}(B-C)$ then we compute $\Tor^{\operatorname{PC}(B)}_i(M,N)$ in the usual way. For example we can start with a projective resolutions of $M$ in $\operatorname{PC}(A-B)$. According to lemma \[ref:4.15a\] this yields a projective resolution of $M$ in $\operatorname{PC}(B)$. We can also start with a projective resolution of $N$ and get the same result. We will need the following. \[ref:4.20a\] Let $E$ be an injective object in $\operatorname{Dis}(C)$. Then $$\Hom_{\operatorname{Top}(C)}(\Tor^{\operatorname{PC}(B)}_i(M,N),E)=\Ext^i_{\operatorname{Top}(B)} (M,\Hom_{\operatorname{Top}(C)}(N,E))$$ This follows easily if we replace $M$ by a projective resolution. Cohen-Macaulay curves embedded in quasi-schemes {#ref:5a} =============================================== Preliminaries {#ref:5.1a} ------------- In the sequel $k$ will be an algebraically closed base field. We will (usually tacitly) assume that all quasi-schemes are in $\QSch/\Spec k$. Note that if $(X,\gamma)\in \QSch/\Spec k$ then $X$ contains the canonical object $\Oscr_X=\gamma^\ast k$ (the “structure sheaf”). However this extra structure on $X$ will not be used until §\[ref:6.6b\]. Below $i:Y\r X$ will be a biclosed embedding of a commutative Cohen-Macaulay curve $Y/k$ as a divisor (in the enriched sense) in a noetherian quasi-scheme $X/k$ (§\[ref:3.7b\]). In the commutative case this hypothesis would imply that $X$ is a surface in a neighborhood of $Y$. Throughout this paper we will impose the following smoothness condition on $X$. Every object in $\Qch(Y)$ has finite injective dimension in $\Qch(X)$. It is easy to see that this is equivalent with the seemingly weaker condition. For every $p$ one has that $\Oscr_p$ has finite injective dimension in $\Qch(X)$. The latter condition is sometimes automatic as can be seen from the following lemma. \[ref:5.1.1a\] Assume that $Y$ is smooth in $p$. Then $\Oscr_p$ has finite injective dimension in $\Qch(X)$. It is easy to see that we have to show that there is some $n$ such that $\Ext^n_{\Qch(X)}(\Fscr,\Oscr_p)=0$ for all $\Fscr\in\coh(X)$. Using the long exact sequence for $\Ext$ we only have to show this in the following two cases. 1. The canonical map $\Fscr(-Y)\r \Fscr$ is injective. 2. $\Fscr\in\coh(Y)$. The lemma now follows from Propositions \[ref:5.1.2a\] below. \[ref:5.1.2a\] Let $\Fscr\in \Qch(X)$, $\Sscr\in \Qch(Y)$. Then 1. If the canonical map $\Fscr(-Y)\r \Fscr$ is an injection then $$\Ext^i_{\Qch(X)}(\Fscr,\Sscr) =\Ext^i_{\Qch(Y)}(\Fscr/\Fscr(-Y),\Sscr)$$ 2. If $\Fscr\in \Qch(Y)$ then there is a long exact sequence $$\begin{gathered} \r \Ext^{i-2}_{\Qch(Y)}(\Fscr,\Sscr(Y)) \r \Ext^i_{\Qch(Y)}(\Fscr,\Sscr)\r \Ext^i_{\Qch(X)}(\Fscr,\Sscr) \\ \r \Ext^{i-1}_{\Qch(Y)}(\Fscr,\Sscr(Y)) \r \Ext^{i+1}_{\Qch(Y)}(\Fscr,\Sscr)\r\end{gathered}$$ This is proved in the same way as if $X$ were a commutative scheme. As an example let us prove 2. Choose an injective resolution $E^\cdot$ for $\Sscr$ in $\Mod(X)$. Write $F^\cdot=\HHom_{o_X}(o_Y,E^\cdot)$. The complex $F^\cdot$ consists of injectives in $\Mod(Y)$. Applying the exact functors $\HHom_{o_X}(-,E_i)$ to $$0\r o_X(-Y)\r o_X\r o_Y\r 0$$ yields an exact sequence of complexes $$\label{ref:5.1b} 0\r F^\cdot \r E^\cdot \r E^\cdot(Y)\r 0$$ Applying the long exact sequence for homology to we see that the homology of $F^\cdot$ is $\Sscr$ and $\Sscr(Y)$ in degrees 0 and 1 respectively. This means that we have a triangle in $\Mod(Y)$ $$\label{ref:5.2a} \Atriangle<1`-1`1; >[\Sscr(Y)[-1]` \Sscr`F^\cdot;``]$$ On the other hand we have $$\begin{aligned} \RHom_{o_X}(\Fscr,\Sscr)&=\Hom_{o_X}(\Fscr,E^\cdot)\\ &=\Hom_{o_Y}(\Fscr,F^\cdot)\\ &=\RHom_{o_Y}(\Fscr,F^\cdot)\end{aligned}$$ Thus applying $\RHom_{o_Y}(\Fscr,-)$ to the triangle we obtain a new triangle $$\Atriangle<1`-1`1; >[R\Hom_{o_Y}(\Fscr,\Sscr(Y)[-1])`\RHom_{o_Y}(\Fscr,\Sscr) `\RHom_{o_X}(\Fscr,\Sscr);``]$$ The long exact sequence for homology of this triangle is precisely 2. Since $\Nscr_{Y/X}$ is an invertible bimodule on $Y$ it follows from [@AZ Prop. 6.8] that we have $\Nscr_{Y/X}=\Nscr_\tau$ for some line bundle $\Nscr$ on $Y$ and an automorphism $\tau$ of $Y$. Recall that by definition $$\label{ref:5.4a} (-\otimes_{o_Y} \Nscr_\tau)=\tau_\ast(-\otimes_{\Oscr_Y}\Nscr)$$ By $\Cscr_f$ we denote the finite length objects in $\Qch(X)$ whose Jordan-Holder quotients lie in $\Qch(Y)$. By $\Cscr$ we denote the corresponding locally finite subcategory of $\Qch(X)$. If $p\in Y$ then we denote by $\Cscr_{f,p}$ the full subcategory of $\Cscr_f$ consisting of objects whose Jordan-Holder quotients are among $(O_{\tau^np})_n$. Again $\Cscr_p$ is the corresponding locally closed subcategory of $\Cscr$. Clearly $\Cscr_f=\oplus_{p\in Y/\langle\tau\rangle}\Cscr_{f,p}$, $\Cscr=\oplus_{p\in Y/\langle\tau\rangle}\Cscr_p$. From the fact that $o_X(Y)/o_X=\Nscr_\tau$ we deduce that $$\label{ref:5.5a} O_q(Y)\cong O_{\tau q}$$ In particular $\Cscr_p$ is stable under $-\otimes_{o_X}o_X(Y)$. From this one deduces ([@VdBVG Prop. 8.4]) \[ref:5.1.3a\] $\Cscr_p$ is closed under injective hulls in $\Qch(X)$. We now translate (and slightly generalize) the main result of [@VdBVG] to our situation. \[ref:5.1.4a\] We have the following. 1. There is a category equivalence $\hat{(-)}_{p}$ between $\Cscr_{p}$ and the category $\operatorname{Dis}(C_p)$ (§\[ref:4a\]) for a certain pseudo-compact ring $C_p$. This ring $C_p$ has the following form : 1. If $\|O_\tau(p)\|=\infty$ then $C_p$ is given by the $\ZZ\times\ZZ$ lower triangular matrices with entries in $\hat{\Oscr}_{Y,p}$. In this case $p$ is regular on $Y$ and thus we have $\hat{\Oscr}_{Y,p}\cong k[[x]]$. 2. If $\|O_\tau(p)\|=n$ then $C_p$ is given by a ring of $n\times n$ matrices of the form $$\begin{pmatrix} R& RU &\cdots &RU\\ \vdots &\ddots& \ddots&\vdots\\ \vdots &&\ddots &RU\\ R&\cdots&\cdots&R \end{pmatrix}$$ where $R$ is a complete local ring of the form $$R=k\langle\langle x,y\rangle\rangle/(\psi)$$ with $$\label{ref:5.6a} \psi=yx-qxy+\mathrm{higher\ terms}$$ for some $q\in k^\ast$, or $$\label{ref:5.7a} \psi=yx-xy-x^2+\mathrm{higher\ terms}$$ $U$ is a regular normalizing element in $\rad(R)$ such that $R/(U)=\hat{\Oscr}_{Y,p}$. If $p$ is not fixed under $\tau$ then $p$ is regular on $Y$ and also $U\not\in\rad^2(R)$. In all cases $R$ carries the usual topology and $C_p$ carries the corresponding product topology. 2. Let $I=\ZZ$ if $\|O_\tau(p)\|=\infty$ and $I=\ZZ/n\ZZ$ if $\|O_\tau(p)\|=n$. In this way the elements of $C_p$ correspond to $I\times I$-matrices. For $i\in I$ let $e_i$ be the corresponding diagonal idempotent. Put $S_i=e_iC_p/\rad(e_iC_p)$. Then $ (\Oscr_{\tau^i p})^\wedge_p=S_i$. 3. Define the following normal element $N$ of $C_p$. 1. If $\|O_\tau(p)\|=\infty$ then $N$ is given by the matrix whose entries are everywhere zero except on the lower subdiagonal where they are one. 2. If $\|O_\tau(p)\|<\infty$ then $$N= \begin{pmatrix} 0&\cdots & 0 & U\\ 1&\ddots && 0\\ \vdots &\ddots & \ddots & \vdots\\ 0&\cdots &1 & 0 \end{pmatrix}$$ Let $\phi=N\cdot N^{-1}$ Then we have the following commutative diagram $$\label{ref:5.8a} \begin{CD} \Cscr_{p} @>-\otimes o_X(-Y)>> \Cscr_{p}\\ @VV\hat{(-)}_{p}V @VV\hat{(-)}_{p}V\\ \operatorname{Dis}(C_{p}) @>(-)_\phi >> \operatorname{Dis}(C_{p}) \end{CD}$$ 4. If $\Fscr$ is an object in $\Cscr_{p}$ then one has the following commutative diagram. $$\begin{CD} (\Fscr(-Y))^\wedge_{p} @> >> (\Fscr)^\wedge_{p}\\ @V\cong VV @\| \\ ((\Fscr)^\wedge_{p})_\phi @>\cdot N>>(\Fscr)^\wedge_{p} \end{CD}$$ where the top arrow is obtained from the inclusion $o_X(-Y)\hookrightarrow o_X$ and the left arrow from . 5. Let $\Fscr$ be a finite length object in $\Qch(Y)$. Then $\hat{\Fscr }_{p}$ is a $C_p/(N)=\prod_{q\in O_\tau(p)}\hat{\Oscr}_{Y,p}$-module and $$\label{ref:5.9a} \hat{\Fscr }_{p}=\prod_{q\in O_\tau(p)} \hat{\Fscr}_{Y,q}$$ where we have written $\hat{(-)}_{Y,q}$ for the ordinary completion at $q$ on $Y$. In [@VdBVG] this theorem was proved with $\Cscr_p$ replaced by $\Cscr_{f,p}$. However it is easy to see that one obtains the current theorem by taking direct limits. Below we write $m$ for the maximal ideal of $R$ and $m_i$ will be the maximal ideal of $C_p$ corresponding to $S_i$. We also use $S_i$ for the bimodule $C_p/m_i$. $R$ is clearly noetherian. $C_p$ is noetherian if the orbit of $p$ is finite, and locally noetherian otherwise. Furthermore it is also clear that $C_p$ is cofinite. The following was proved in [@VdBVG Thm 1.1.4]. Every finite dimensional $C_p$-representation $F$ is in $\operatorname{PCFin}(C_p)$. Hence $F=\prod_i Fe_i$. The ring $C_p$ has another good property, which wasn’t stated in [@VdBVG]. \[ref:5.1.6a\] The ring $C_p$ is coherent. This is clear if the orbit of $p$ is finite, so we assume it to be infinite. We first show that if $B$ is a finitely generated $N$-torsion free right pseudo-compact $C_p$-module then $B$ is finitely presented. We have an exact sequence of pseudo-compact modules $$0\r K\r C_p^m \r B\r 0$$ Tensoring with $C_{Y,p}\overset{\text{def}}{=}C_p/C_pN$ yields an exact sequence $$0\r K/NK \r C_{Y,p}^m \r B/NB\r 0$$ and using Lemma \[ref:4.9a\] we see that it is sufficient to show that $K/NK$ is finitely generated. Now $C_{Y,p}=\prod_i \hat{\Oscr}_{Y,\tau^ip}$, and from the theory of discrete valuation rings we see that the number of generators of a submodule of $C_{Y,p}^m e_i=\hat{\Oscr}_{Y,\tau^ip}^m$ is bounded by $m$. This easily implies what we want. Now we prove that $C_p$ is coherent. We have to show that the kernel of an arbitrary map $\alpha: C_p^m\r C_p$ is finitely generated. Clearly $B=\im \alpha$ is pseudo-compact and finitely generated. We can now apply the result in the previous paragraph. We will also need : \[ref:5.1.7a\] The global dimension of $C_p$ is equal to two. According to Corollaries \[ref:4.11a\] and \[ref:4.7b\] it suffices to show that the projective dimension of each $S_i$ is equal to two. Put $P_i=e_iC_p$. One easily checks that the minimal resolution of $S_i$ is given by $$0\r P_{i-1}\r P_{i-1}\oplus P_i\r P_i\r S_i\r 0$$ This implies what we want. The following result will be used. \[ref:5.1.8a\] Assume that $p$ is a fixed point for $\tau$. Then the multiplicity of $p$ on $Y$ is equal to the largest integer $n$ such that $RU\subset \rad^n R$. Put $S=R/(U)$ and let $m$ be the maximal ideal of $R$. By Theorem \[ref:5.1.4a\].5 $S=\hat{\Oscr}_{Y,p}$. Assume that $U\in m^\mu-m^{\mu+1}$. Equip $R$ and $S$ with the $m$-adic filtration. Since $\gr R$ is a domain (direct verification) it is easy to see that there is an exact sequence $$\label{ref:5.10a} 0\r \gr R(-\mu)\xrightarrow{\cdot U} \gr R\r \gr S\r 0$$ Then from we find that $\mu$ is equal to $\dim_k \rad^u S/\rad^{u+1}S$ for large $u$. Hence $\mu$ is the multiplicity of $p$ in $Y$. Some computations ----------------- Our view point is that $C_p$ encodes the local structure around a point $p\in Y$. In order to blow up $p$ we will consequently need some computations in $C_p$. Our aim here is to prove Corollary \[ref:5.2.4a\] below. This corollary is easy if $p$ is a fixed point for $\tau$ and fairly easy if $p$ has infinite $\tau$-orbit. So the main purpose will be to treat the case where $n=\|O_\tau(p)\|$ satisfies $2\le n<\infty$. However we will develop a formalism which also includes the case $n=\infty$. Perhaps this has some independent interest. Our main result will be Proposition \[ref:5.2.2a\] which is however more elaborate than what we need for the application to Corollary \[ref:5.2.4a\]. Notations will be as in Theorem \[ref:5.1.4a\]. $m$ will be the maximal ideal of $R$ and $m_i$ will be the twosided maximal ideal of $C_p$ corresponding to $S_i$. Note that usually $i$ is a taken modulo $n$ here. An array $a=(a_q)_{q\in\ZZ}$ with entries in $\NN\cup \{+\infty\}$ will be called good if it is non-decreasing, bounded below, and if $a_q$ is infinite for $t\gg 0$ and finite for $t\ll 0$. If $a,b$ are good then $a\ge b$ iff $a_q\ge b_q$ for all $q$. Our aim is to use good arrays as a bookkeeping device in order to study certain right $C_p$-modules in the case that $n\ge 2$. Let $a$ be a good array and let $I$ be as in Theorem \[ref:5.1.4a\]. We distinguish two cases. - The case [ $n< \infty$]{}. Fix an arbitrary $x\in m-m^2$ such that $x-U\not\in m^2$ for $R$. For $s\in \ZZ$ define $$H_{a,s}=\sum_{t\in\ZZ} x^{a_{s-nt}}U^t R$$ where this sum is taken inside the fraction field of $R$ and where by convention $x^\infty=0$. The $H_{a,s}$ are clearly fractional right $R$ ideals. Put $$P_{a}=(H_{a,0},\ldots, H_{a,n-1})$$ - The case [ $n=\infty$]{}. We consider this as a limiting case of the previous case. We define $P_a$ as the row matrices $(P_{a,j})_j$ with $$P_{a,j}=m^{a_j}$$ \[ref:5.2.2a\] In this proposition $a,b,c$ will be good arrays. As above let $n=\|I\|\ge 2$. If $i\in I$ then $$\tilde{\imath}= \begin{cases} \text{the unique element of $\{0,\ldots,n-1\}$ congruent to $i$} &\text{if $n<\infty$}\\ i&\text{if $n=\infty$} \end{cases}$$ 1. The $P_a$ are right $C_p$-modules (with the obvious $C_p$-action). 2. Let $K$ be the fraction field of $R$ and consider the $P_a$ as submodules of $K^I$. Then $$P_{a}\subset P_{b}\iff a\ge b$$ 3. One also has $$\begin{aligned} P_{a}+P_{b}&=P_{\inf(a,b)}\\ \label{ref:5.14a} P_{a}\cap P_{b}&=P_{\sup(a,b)}\end{aligned}$$ 4. If $b\ge a$ and $a_i=\infty\iff b_i=\infty$ then $P_{a}/P_{b}$ has finite length and the composition factors are given by $$\oplus_{i\in\ZZ} S_{\tilde\imath}^{b_i-a_i}$$ (with multiplicity). 5. One has $$P_{a}m_l=P_{c}$$ where $$c_q= \begin{cases} a_q&\text{if $q\not \cong l$ mod $n$}\\ \min(a_{q+1},a_q+1)&\text{if $q \cong l$ mod $n$} \end{cases}$$ 6. One has $$\rad P_a=P_c$$ where $$c_q= \begin{cases} a_q+1&\text{if $a_q\neq a_{q+1}$}\\ a_q&\text{otherwhise} \end{cases}$$ 7. One has $$P_{a}/\rad P_a= \bigoplus_{q\in\ZZ,a_q\neq a_{q+1}} S_{\tilde{q}}$$ 8. For $i\in I$ write $P_i=e_i C_p$. Then $$P_i=P_{c}$$ where $$c=(\ldots,0,\ldots,0,\infty,\ldots)$$ with the first $\infty$ occuring in position $\tilde\imath+1$. All this is fairly easy if $n=\infty$, so we concentrate on the case $n<\infty$. It is easy to see that $x^i U^j$ is a topological $k$-basis for $R$ and from the fact that $U$ is normalizing one obtains the following alternative form of $H_{a,s}$. $$\label{ref:5.11a} H_{a,s}=\prod_{t\in\ZZ}\prod_{i\ge a_{s-nt}} k x^i U^t$$ From the fact that $a_q$ is ascending we obtain $H_{a,s}\subset H_{a,s-1}$. We also have $H_{a,s-n}=H_{a,s}U^{-1}$. This easily implies 1. If $a,b$ are good then implies that $$\label{ref:5.12a} \forall s:H_{a,s}\subset H_{b,s}\iff a\ge b$$ and also $$\begin{aligned} \label{ref:5.13a} H_{a,s}+H_{b,s}&=H_{\inf(a,b),s}\\ \label{ref:5.14a} H_{a,s}\cap H_{b,s}&=H_{\sup(a,b),s}\end{aligned}$$ It is clear that and imply the corresponding properties for $P_a, P_b$. This proves 2. and 3. If $a,a'$ are such that $$a'_q=\begin{cases} a_q &\text{if $q\neq i$}\\ a_q+1&\text{if $q=i$, $a_i\ne \infty$} \end{cases}$$ then it follows from that $$H_{a,s}/H_{a',s}= \begin{cases} R/m&\text{if $s\cong i$ mod $n$} \\ 0&\text{otherwise} \end{cases}$$ We deduce that $$P_{a}/P_{a'}=S_{\bar\imath }$$ This yields 4. We now compute $P_{a}m_l$ for an arbitrary $l\in I$. We find $$\label{ref:5.15a} (P_{a}m_l)_j= \begin{cases} H_{a,j}&\text{if $j\neq l$}\\ H_{a,j}m+H_{a,j+1}&\text{if $j=l$} \end{cases}$$ To compute the righthand side of we use the observation that $U^{-b}xU^b\not\cong U\text{ mod } m^2$ and hence $m=UR+U^{-b}xU^b R$. Thus $$x^aU^bm=x^aU^b(UR+U^{-b}xU^bR)=x^{a+1}U^bR+x^aU^{b+1}R$$ One now easily obtains 5. Items 6. and 7. are consequences of the fact that $\rad P_a=\bigcap_l P_a m_l$. Finally 8. is a simple verification which we leave to the reader. \[ref:5.2.3a\] Let $n=\infty$, $a=(\ldots,0,0,1,1,1,3,\infty,\infty,\ldots)$ where the first $\infty$ occurs in position $2$. Then $$\label{ref:5.16a} P_a/\rad P_a= S_{1}\oplus S_{0}\oplus S_{-3}$$ If $n=3$ and $a$ is the same then $$\label{ref:5.17a} P_a/\rad P_a= S_{\bar{0}}^2\oplus S_{\bar{1}}$$ Note that going from to amounts to introducing periodicity modulo $3$ among the $S_i$. \[ref:5.2.4a\] $\dim_k (C_p/m_0m_{-1}m_{-2}\cdots m_{-p+1})=\frac{p(p+1)}{2}$ This is clear if $n=1$. In the case $n\ge 2$ we use the fact that we have $C_p=\prod_{i\in I} P_i$. The corollary now follows easily from 5. and 8. of the foregoing proposition. Completion of objects in [ $\coh(X)$]{} {#ref:5.3a} --------------------------------------- We have already defined $\hat{(-)}_{p}$ on $\Cscr_p$. In a different direction, it is also possible to extend $\hat{(-)}_{p}$ to $\coh(X)$, but then it loses some of its good properties. For $\Fscr\in\coh(X)$ define $$\hat{\Fscr}_{p}= \invlim_{\Fscr/\Fscr'\in\Cscr_p} (\Fscr/\Fscr')^\wedge_{p}$$ \[ref:5.3.1a\] $(-)^\wedge_{p}$ is an exact functor from $\coh(X)$ to $\operatorname{PC}(C_p)$. The analogs of 3.,4.,5. of Theorem \[ref:5.1.4a\] hold for $\Fscr\in\coh(X)$. If $\Fscr\in\coh(X)$ then $\hat{\Fscr}_{p}$ is finitely presented. Furthermore, $\hat{\Fscr}_{p}$ lies in $\operatorname{pc}(C_p)$ if and only if the intersection of the support of $\Fscr/\Fscr(-Y)\in \Qch(Y)$ and the orbit of $p$ is finite. Let $$0\r \Fscr\xrightarrow{\phi} \Gscr\xrightarrow{\theta} \Hscr\r 0$$ be an exact sequence in $\coh(X)$. We have to show that $$0\r \invlim_{\Fscr/\Fscr'\in\Cscr_p} (\Fscr/\Fscr')^\wedge_{p} \r \invlim_{\Gscr/\Gscr'\in\Cscr_p} (\Gscr/\Gscr')^\wedge_{p} \r \invlim_{\Hscr/\Hscr'\in\Cscr_p} (\Hscr/\Hscr')^\wedge_{p} \r 0$$ is exact. Given the exactness of $(-)^\wedge_{p}$ on finite length objects and the exactness of $\invlim$ on pseudo-compact modules this means we have to show that 1. $\theta(\Gscr')_{\Gscr/\Gscr'\in\Cscr_p}$ is cofinal in $(\Hscr')_{\Hscr/\Hscr'\in\Cscr_p}$ 2. $\phi^{-1}(\Gscr')_{\Gscr/\Gscr'\in\Cscr_p}$ is cofinal in $(\Fscr')_{\Fscr/\Fscr'\in\Cscr_p}$. The first statement is clear. The second statement is the Artin-Rees condition, which is equivalent with $\Cscr_p$ being closed under injective hulls in $\Qch(X)$. This is precisely Proposition \[ref:5.1.3a\]. The fact that the analog of Theorem \[ref:5.1.4a\] holds is easy to see. Now we prove that $\hat{\Fscr}_p$ is finitely presented. According to lemma \[ref:4.18a\] it is sufficient to show that for every $i$ and every $q\in O_\tau(p)$, $\Ext^i_{\operatorname{PC}(C_p)}(\hat{\Fscr}_{p},(\Oscr_q)^\wedge_{p})$ has finite dimension, bounded independently of $q$. By Proposition \[ref:5.3.4a\] below we have $$\Ext^i_{\operatorname{PC}(C_p)}(\hat{\Fscr}_{p},(\Oscr_q)^\wedge_{p}))= \Ext^i_{\Qch(X)}(\Fscr,\Oscr_q)$$ and the dimension of the righthand side of this equation is indeed bounded independently of $q$ by lemma \[ref:5.3.2a\]. Now we concentrate on the second part of the theorem. We have an exact sequence in $\Qch(X)$ $$\Fscr(-Y)\r \Fscr\r \Fscr/\Fscr(-Y)\r 0$$ which yields by the analog of Theorem \[ref:5.1.4a\].5 an exact sequence in $\operatorname{PC}(C_p)$ $$(\hat{\Fscr}_{p})_\phi\xrightarrow{\cdot N} \hat{\Fscr}_{p} \r \prod_{q\in O_\tau(p)}(\Fscr/\Fscr(-Y))^\wedge_{Y,q}\r 0$$ By [@VdBVG Prop 3.22] $\hat{\Fscr}_{p}$ will be noetherian if and only if $\prod_{q\in O_\tau(p)}(\Fscr/\Fscr(-Y))^\wedge_{Y,q}$ is noetherian, that is, if and only if, for almost all $q\in O_\tau(p)$ we have $(\Fscr/\Fscr(-Y))^\wedge_{Y,q}=0$. This is equivalent with the intersection of the support of $\Fscr/\Fscr(-Y)$ with the $\tau$-orbit of $p$ being finite. We now supply the details that were used in the proof of the previous theorem. \[ref:5.3.2a\] Assume that $\Fscr\in \coh(X)$, $q\in Y$. Then $\dim \Ext^i_{\Qch(X)}(\Fscr,\Oscr_q)$ is finite, and bounded independently of $q$. As usual, one reduces to one of the following cases. 1. The canonical map $\Fscr(-Y)\r \Fscr$ is injective. 2. $\Fscr$ is in $\coh(Y)$. Using Proposition \[ref:5.1.2a\], we can then reduce to showing that if $\Gscr\in\coh(Y)$ then $\dim \Ext^i_{\Qch(Y)}(\Gscr,\Oscr_q)$ is finite and bounded independently of $q$. This now follows from the fact that $\Gscr$ has a resolution consisting of vector bundles on $Y$. The following result is useful. \[ref:5.3.3a\] $\hat{\Oscr}_{X,p}\cong C_p$ as right $C_p$-modules. The analogous result for $Y$ is trivially true. We can lift this to $X$ using Nakayama’s lemma (lemma \[ref:4.8b\]). We now have to be able to compare $\Ext$-groups in $\operatorname{PC}(C_p)$ and $\Qch(X)$. In fact we have the following result. \[ref:5.3.4a\] Assume that $\Fscr\in\coh(X)$ and $\Sscr\in \Cscr_{p}$. Then $$\Ext_{\operatorname{Top}(C_p)}^i(\hat{\Fscr}_{p},\hat{\Sscr}_{p})= \Ext_{\Qch(X)}^i(\Fscr,\Sscr)$$ We can reduce to the case $i=0$ by replacing $\Sscr$ with an injective resolution in $\Cscr_p$. We have $$\begin{aligned} \Hom_{\operatorname{Top}(A)}(\hat{\Fscr}_p,\hat{\Sscr}_p)&= \Hom_{\operatorname{Top}(A)}(\invlim_{\Fscr'}(\Fscr/\Fscr')_p^\wedge,\hat{\Sscr}_p)\\ &=\dirlim_{\Fscr'}\Hom_{\operatorname{Dis}(A)}((\Fscr/\Fscr')_p^\wedge,\hat{\Sscr}_p) \qquad \text{(Proposition \ref{ref:4.3a})}\\ &=\dirlim_{\Fscr'} \Hom_{\Qch(X)}(\Fscr/\Fscr',\Sscr)\\ &=\Hom_{\Qch(X)}(\Fscr,\Sscr)\end{aligned}$$ Here $\Fscr'$ runs through the subobjects in $\Fscr$ such that $\Fscr/\Fscr'\in\Cscr_{f,p}$. \[ref:5.3.5a\] Assume $\Fscr\in\coh(X)$. Then $\hat{(-)}_p$ defines a one-one correspondence between open subobjects of $\hat{\Fscr}_p$ and the subobjects $\Fscr'$ of $\Fscr$ such that $\Fscr/\Fscr'\in\Cscr_p$. Assume for example that $H\subset \hat{\Fscr}_p$ is open. Put $S=\hat{\Fscr}_p/ H$. Then $S\in \operatorname{PCFin}(C_p)$. By Theorem \[ref:5.1.4a\] we have $S=\hat{\Sscr}_p$ for some $\Sscr\in \Cscr_{f,p}$. Let $q:\hat{\Fscr}_p\r S$ be the quotient map. According to Proposition \[ref:5.3.4a\], $q$ corresponds to a map $p:\Fscr\r \Sscr$. Define $\Hscr=\ker p$. Then $\hat{\Hscr}_p=H$. Completion of bimodules {#ref:5.4b} ----------------------- If is also possible to define $\hat{(-)}_{p}$ on certain $o_X-o_X$-bimodules. Unless otherwise specified, when we say “bimodule” we mean an object of $\Bimod(o_X-o_X)$. We say that an $o_X-o_X$ bimodule $\Mscr$ is $\Cscr_{p}$-preserving if both $-\otimes_{o_X}\Mscr$ and $\HHom_{o_X}(\Mscr,-)$ preserve $\Cscr_{p}$. In this section we will write $(-)^\sim_{p}:\operatorname{Dis}(C_p)\r \Cscr_{p}$ for the inverse of the functor $(-)^\wedge_{p}$. Assume that $\Mscr$ is a coherent $\Cscr_{p}$ preserving bimodule. Then $-\otimes_{o_X} \Mscr$ preserves $\Cscr_{f,p}$ by \[ref:3.1.6a\]. We define $$\label{ref:5.18a} \hat{\Mscr}_{p}=\invlim_I ((C_p/I)^\sim_{p}\otimes_{o_X} \Mscr)^\wedge_{p}$$ where $I$ runs over the open twosided ideals in $C_p$. \[ref:5.4.2a\] Assume that $\Mscr$ is a coherent $\Cscr_{p}$ preserving $o_X-o_X$-bimodule. Then $\hat{\Mscr}_{p}$ is a bipseudo-compact $C_p$-bimodule. If follows from functoriality that $((C_p/I)^\sim_{p}\otimes_{o_X} \Mscr)^\wedge_{p}$ is annihilated by $I$ on the left. On the right it is in $\operatorname{PCFin}(C_p)$ and hence it is also annihilated by some open twosided ideal $J\subset C_p$. Thus $((C_p/I)^\sim_{p}\otimes_{o_X} \Mscr)^\wedge_{p}$ is bipseudo-compact, and hence so is the inverse limit. We have the following analog of Proposition \[ref:5.3.4a\]. \[ref:5.4.3a\] Assume that $\Mscr$ is a coherent $\Cscr_{p}$-preserving bimodule and $\Sscr\in\Cscr_{p}$. Then $\Ext^i_{\operatorname{Top}(C_p)}(\hat{\Mscr}_{p},\hat{\Sscr}_{p})\in \operatorname{Dis}(C_p)$ and $\HExt^i_{\Qch(X)}(\Mscr,\Sscr)\in \Cscr_p$. Furthermore as objects of $\operatorname{Dis}(C_p)$ $$\Ext^i_{\operatorname{Top}(C_p)}(\hat{\Mscr}_{p},\hat{\Sscr}_{p}) = \HExt^i_{\Qch(X)}(\Mscr,\Sscr)^\wedge_{p}$$ That $\HExt^i_{\Qch(X)}(\Mscr,\Sscr)\in \Cscr_p$ follows from the fact that $\HHom(\Mscr,-)$ is $\Cscr_{p}$ preserving and Proposition \[ref:5.1.3a\]. The fact that $\Ext^i_{\operatorname{Top}(C_p)}(\hat{\Mscr}_{p},\hat{\Sscr}_{p})\in \operatorname{Dis}(C_p)$ is precisely Proposition \[ref:4.19a\]. Below $\Uscr$ will run through $(C_p/I)^\sim_p$, where $I$ is an open twosided ideal in $C_p$. We have $$\label{ref:5.19a} \begin{split} \Ext^i_{\operatorname{Top}(C_p)}(\hat{\Mscr}_{p},\hat{\Sscr}_{p})&= \Ext^i_{\operatorname{Top}(C_p)}(\invlim_\Uscr (\Uscr\otimes_{o_X} \Mscr)^\wedge_{p},\hat{\Sscr}_{p})\\ &= \dirlim_\Uscr \Ext^i_{\operatorname{Dis}(C_p)}((\Uscr\otimes_{o_X} \Mscr)^\wedge_{p}, \hat{\Sscr}_{p})\qquad \text{(Proposition \ref{ref:4.3a})} \\ &=\dirlim_\Uscr \Ext^i_{\Cscr_p}(\Uscr\otimes_{o_X}\Mscr,\Sscr) \qquad\text{(Theorem \ref{ref:5.1.4a})}\\ &=\dirlim_\Uscr \Ext^i_{\Qch(X)}(\Uscr\otimes_{o_X}\Mscr,\Sscr) \quad\text{(Proposition \ref{ref:5.1.3a})}\end{split}$$ Let $E^\cdot$ be an injective resolution of $\Sscr$ in $\Cscr_{p}$. Then $$\begin{aligned} \Ext^i_{\Qch(X)}(\Uscr\otimes_{o_X}\Mscr,\Sscr)& =H^i(\Hom_{\Qch(X)}(\Uscr\otimes_{o_X} \Mscr,E^\cdot)\\ &=H^i(\Hom_{\Qch(X)}(\Uscr,\HHom_{\Qch(X)}(\Mscr,E^\cdot))\\ &= H^i(\Hom_{\operatorname{Dis}(C_p)}(\hat{\Uscr}_p,\HHom_{\Qch(X)}(\Mscr,E^\cdot)^\wedge_p ) \\ &=H^i(\Hom_{\operatorname{Dis}(C_p)}(C_p/I,\HHom_{\Qch(X)}(\Mscr,E^\cdot)^\wedge_p )\end{aligned}$$ Combining this with yields $$\begin{aligned} \Ext^i_{\operatorname{Top}(C_p)}(\hat{\Mscr}_{p},\hat{\Sscr}_{p}) &= \dirlim_I H^i(\Hom_{\operatorname{Dis}(C_p)}(C_p/I, \HHom_{\Qch(X)}(\Mscr,E^\cdot)^\wedge_p )\\ &=H^i(\Hom_{\operatorname{Top}(C_p)}(\invlim_I(C_p/I), \HHom_{\Qch(X)}(\Mscr,E^\cdot)^\wedge_p ))\\ &=H^i(\Hom_{\operatorname{Top}(C_p)}(C_p,\HHom_{\Qch(X)}(\Mscr,E^\cdot)^\wedge_p ))\\ &=\HExt^i_{\Qch(X)}(\Mscr,\Sscr)^\wedge_p \qed\end{aligned}$$ \[ref:5.4.4a\] The functor $(-)_{p}^\wedge$ preserves short exact sequences of coherent $\Cscr_{p}$-preserving bimodules. Let $$0\r \Kscr\r \Mscr\r \Nscr\r 0$$ be an exact sequence of coherent $\Cscr_{p}$ preserving bimodules. We have to show that $$\label{ref:5.20a} 0\r \hat{\Kscr}_p\r \hat{\Mscr}_p \r \hat{\Nscr}_p\r 0$$ is exact. Since the bimodules in are in $\operatorname{PC}(C_p)$ (Proposition \[ref:5.4.2a\]) it suffices to show that $$\label{ref:5.21a} 0\r \Hom_{\operatorname{Top}(C_p)}(\hat{\Nscr}_p,\hat{E}_p)\r \Hom_{\operatorname{Top}(C_p)}(\hat{\Mscr}_p,\hat{E}_p)\r \Hom_{\operatorname{Top}(C_p)}(\hat{\Kscr}_p,\hat{E}_p)\r 0$$ is exact where $E$ is the sum of the injective hulls of the $\Oscr_{\tau^i p}$ (one uses the fact that $\Hom_{\operatorname{Top}(C_p)}(-,\hat{E}_p)$ is exact and faithful on $\operatorname{PC}(C_p)$). Now by Proposition \[ref:5.4.3a\] it follows that is obtained from completing the following exact sequence in $\Cscr_p$. $$0\r \HHom_{\Qch(X)}(\Nscr,E)\r \HHom_{\Qch(X)}(\Mscr,E)\r \HHom_{\Qch(X)}(\Kscr,E)\r 0$$ and hence we are done! \[ref:5.4.5a\] Completion commutes with tensor product in the following sense. Let $\Fscr$ be a coherent object in $\Qch(X)$ and let $\Mscr$, $\Nscr$ be coherent $\Cscr_{p}$ preserving bimodules. Then there are natural isomorphisms 1. $(\Fscr\otimes_{o_X} \Nscr)^\wedge_{p}=\hat{\Fscr}_{p}\ctimes_{C_p} \hat{\Nscr}_{p}$ (note that by Theorem \[ref:5.3.1a\] and lemma \[ref:4.14a\] we may replace “$\ctimes$” by “$\otimes$”). 2. $(\Mscr\otimes_{o_X} \Nscr)^\wedge_{p}=\hat{\Mscr}_{p}\ctimes_{C_p} \hat{\Nscr}_{p}$ <!-- --> 1. By definition we have $$(\Fscr\otimes_{o_X} \Nscr)^\wedge_{p}=\invlim_\Tscr( (\Fscr\otimes_{o_X} \Nscr) /\Tscr)^\wedge_p$$ where $\Tscr$ runs through the subobjects of $\Fscr\otimes_{o_X} \Nscr$ such that $(\Fscr\otimes_{o_X}\Nscr)\Tscr\in \Cscr_p$. Now we claim that for every $\Tscr$ there exists a $\Fscr'\subset \Fscr$ such that $\Fscr/\Fscr'\in\Cscr_p$ and such that the image of $\Fscr'\otimes_{o_X} \Nscr$ in $\Fscr\otimes_{o_X} \Nscr$ is contained in $\Tscr$. Let $\Qscr=(\Fscr\otimes_{o_X} \Nscr) /\Tscr$ and let $\Fscr\otimes_{o_X}\Nscr\r \Qscr$ be the quotient map. By adjointness there is a corresponding map $ \Fscr\r \HHom(\Nscr,\Qscr)$. We define $\Fscr'$ as the kernel of this map. This $\Fscr'$ has the properties we want. We obtain $$\label{ref:5.22a} (\Fscr\otimes_{o_X} \Nscr)^\wedge_{p}=\invlim_{\Fscr'} ((\Fscr/\Fscr')\otimes_{o_X} \Nscr)^\wedge_p$$ For $K\in \operatorname{PC}(C_p)$ we define the functor $$F(K)=\invlim_{K'} ((K/K')^\sim_p \otimes_{o_X} \Nscr)^\wedge_p$$ where $K'$ runs through the open subobjects of $K$. One shows that this functor is right exact and commutes with products. Furthermore, according to we have $F(C_p)=\hat{\Nscr}_p$. By and Corollary \[ref:5.3.5a\] $$\begin{aligned} F(\hat{\Fscr}_p)&=\invlim ((\Fscr/\Fscr')\otimes_{o_X} \Nscr)^\wedge_p\\ &=(\Fscr\otimes_{o_X}\Nscr)^\wedge_p\end{aligned}$$ Now by Theorem \[ref:5.3.1a\] we know that $\hat{\Fscr}_p$ is coherent. Take a presentation $$C^n_p\r C_p^m\r \hat{\Fscr}_p\r 0$$ By right exactness of tensor product we find that the cokernel of $(C^n_p\r C_p^m)\otimes_{C_p} \hat{\Nscr}_p$ is equal to $\hat{\Fscr}_p\otimes_{o_X} \hat{\Nscr}_p$. On the other hand by right exactness of $F$ we find that this cokernel is equal to $F(\hat{\Fscr}_p)=(\Fscr\otimes\Nscr)^\wedge_p$. 2. This follows from 1. We have $$\begin{aligned} (\Mscr\otimes\Nscr)^\wedge_p&=\invlim_I ((C_p/I)^\sim_p\otimes_{o_X} \Mscr\otimes_{o_X}\Nscr)^\wedge_p\\ &=\invlim_I ((C_p/I)_p^\sim\otimes_{o_X} \Mscr)^\wedge_p\ctimes_{C_p} \hat{\Nscr}_p\\ &=\invlim_I ((C_p/I)\ctimes_{C_p} \hat{\Mscr}_p\ctimes_{C_p} \hat{\Nscr}_p\\ &= \hat{\Mscr}_p\ctimes_{C_p} \hat{\Nscr}_p\end{aligned}$$ where we have used the fact that $\ctimes$ commutes with inverse limits. The category [$\tilde{\Cscr}_{f,p}$]{} {#ref:5.5b} -------------------------------------- If $\Gscr\in\coh(Y)$ then we can view $\Gscr$ as a coherent $\Oscr_Y$-bimodule, by defining $\HHom_{o_Y}(\Gscr,-)$ as $\HHom_{\Oscr_Y}(\Gscr,-)$ where the second “$\HHom$” is the ordinary $\HHom$ for sheaves. $\HHom_{\Oscr_Y}(\Gscr,-)$ has a left adjoint given by $-\otimes_{\Oscr_Y} \Gscr$ so $\Gscr\in\Bimod(o_Y-o_Y)$. Hence if $q\in Y$ then we can view $o_q$ as a coherent object in $\Bimod(o_Y-o_Y)$. We also write $o_q$ for the coherent object in $\Bimod(o_X-o_X)$ given by ${}_{o_X} o_q {}_{o_X}$ (cfr. for notations). \[ref:5.5.1a\] 1. Assume that $\Fscr\in\Qch(X)$. Then $$\label{ref:5.23a} \HHom_{\Qch(X)}(o_q,\Fscr)=\Oscr_q\otimes_k \Hom_{\Qch(X)}(\Oscr_q,\Fscr)$$ 2. If $\Fscr\in \coh(X)$ then $$\label{ref:5.24a} \Fscr\otimes_{o_X} o_q=\Oscr_q\otimes_k \Hom_{\Qch(X)}(\Fscr,\Oscr_q)^\ast$$ 3. We have a canonical identification $$\label{ref:5.25a} o_q\otimes_{o_X} o_X(Y)=o_X(Y)\otimes_{o_X} o_{\tau q} \qed$$ <!-- --> 1. Let $\Fscr\in \Qch(X)$. According to we have $$\HHom_{\Qch(X)}(o_q,\Fscr)=\HHom_{\Qch(Y)}(o_q,i^!(\Fscr))_{o_X}$$ and by adjunction we also have $$\Hom_{\Qch(X)}(\Oscr_q,\Fscr)=\Hom_{\Qch(Y)}(\Oscr_q,i^!(\Fscr))$$ Now we use the fact that the analog of holds on $Y$. 2. This is also proved by reduction to $\coh(Y)$, using $$\Fscr\otimes_{o_X} o_q=i_\ast(i^\ast(\Fscr)\otimes_{o_Y} o_q)\qquad \text{(see \eqref{ref:3.14a})}$$ 3. It suffices to show that the left and righthand side of take the same values on arbitrary $\Fscr\in \Qch(X)$. We have $$\begin{split} \label{ref:5.26a} \HHom_{\Qch(X)}(o_q\otimes_{o_X} o_X(Y),\Fscr)&= \HHom_{\Qch(X)}(o_q,\Fscr(-Y))\\ &=\Oscr_q\otimes_k \Hom(\Oscr_q,\Fscr(-Y))\\ &=\Oscr_q\otimes_k \Hom(\Oscr_q(Y),\Fscr)) \end{split}$$ and $$\label{ref:5.27a} \begin{split} \HHom_{\Qch(X)}(o_X(Y)\otimes_{o_X} o_{\tau q},\Fscr) &=\HHom_{\Qch(X)}(o_X(Y),\HHom_{\Qch(X)}(o_{\tau q},\Fscr))\\ &=(\Oscr_{\tau q}\otimes_k \Hom_{\Qch(X)}(\Oscr_{\tau q},\Fscr))(-Y)\\ &=\Oscr_{\tau q}(-Y)\otimes_k \Hom_{\Qch(X)}(\Oscr_{\tau q},\Fscr) \end{split}$$ Now by our conventions $\Oscr_q(Y)\cong \Oscr_{\tau q}$ (see ). From this it follows that the righthand sides in and are isomorphic. The reason that this is canonical is that we use the (non-canonical) identification $\Oscr_q(Y)\cong\Oscr_{\tau q}$ twice, in such a way that the ambiguities cancel. Let $\cohBIMOD(o_x-o_X)$ be the full subcategory of $\BIMOD(o_X-o_X)$ consisting of coherent objects. According to Corollary \[ref:3.1.8a\] this is an abelian subcategory of $\BIMOD(o_X-o_X)$, closed under extensions. $o_q$ is a simple object in $\cohBIMOD(o_X-o_X)$. An object in $\cohBIMOD(o_X-o_X)$ is determined by the values it takes on indecomposable injectives. We have by lemma \[ref:5.5.1a\].1 $$\HHom_{\Qch(X)}(o_q,E)=\begin{cases} \Oscr_q&\text{if $E$ is the injective hull of $\Oscr_q$}\\ 0&\text{otherwise} \end{cases}$$ Hence a proper left exact subfunctor of $\HHom_{\Qch(X)}(o_q,-)$ commuting with direct sums will be the zero-functor. From this proposition we easily deduce The objects $o_q\otimes_{o_X} o_X(nY)$ with $n\in\ZZ$ are simple objects in $\cohBIMOD(o_X-o_X)$. Now we define $\tilde{\Cscr}_{f,p}$ as the full subcategory of $\cohBIMOD(o_X-o_X)$ consisting of finite length objects whose Jordan-Holder quotients are of the form $o_{\tau ^m p}\otimes o_X(nY)$, $m,n\in \ZZ$. \[ref:5.5.4a\] Assuming that $\Cscr$, $\Dscr$ are abelian categories possessing an injective cogenerator. Assume that we have an exact sequence in $\BIMOD(\Cscr-\Dscr)$ $$0\r \Kscr\r \Mscr\r \Nscr\r 0$$ with $\Nscr\in\Bimod(\Cscr-\Dscr)$, such that $R^1\prod_{i\in I} \HHom_\Dscr(\Nscr,E_i)=0$ (see the discussion after Proposition \[ref:3.5.5a\]) for all families of injectives $(E_i)_{i\in I}$ in $\Dscr$. Then $\Kscr\in \Bimod(\Cscr-\Dscr)$ if and only if $\Mscr\in\Bimod(\Cscr-\Dscr)$. We have to show that $\HHom_\Dscr(\Kscr,-)$ commutes products when evaluated on injectives, if and only if the same holds for $\HHom_\Dscr(\Mscr,-)$. Using the hypotheses this is a direct consequence of the five-lemma. \[ref:5.5.5a\] Assume that $\Sscr\in \tilde{\Cscr}_{f,p}$. Then $R^1\prod_{i\in I} \HHom_{\Qch(X)}(\Sscr,E_i)=0$ for all families of injectives $(E_i)_{i\in I}$ in $\Qch(X)$. Using the long exact sequence for $R\prod_{i\in I}$ one sees that it is sufficient that $R^1\prod_{i\in I} \HHom_{\Qch(X)}(o_q,E_i) =0$ for $q\in Y$ and for all families of injectives in $\Qch(X)$. If $\Mscr\in\Qch(X)$ then $\HHom_{\Qch(X)}(o_q,\Mscr)\in \Qch(q)$. By Definition \[ref:3.5.9a\] it now suffices to show that $q$ is very well closed in $X$. Since $q$ is obviously defined by an ideal in $o_Y$, we have that $q$ is well closed in $Y$ by Proposition \[ref:3.5.7a\]. $\Qch(q)$ is of course equivalent with the category of $k$-vectorspaces. In particular $\Qch(q)$ has exact direct products and hence $q$ is very well closed in $Y$ by Corollary \[ref:3.5.11a\]. Furthermore $Y$ is also defined by an ideal inside $X$ and hence by Proposition \[ref:3.5.7a\] it follows that $Y$ is well closed in $X$. Now we invoke Proposition \[ref:3.5.10a\] which yields that $q$ is very well closed in $X$. \[ref:5.5.6a\] If we have an exact sequence in $\BIMOD(o_X-o_X)$ $$0\r \Mscr\r \Nscr\r \Sscr\r 0$$ with $\Sscr\in\tilde{\Cscr}_{f,p}$ then $\Mscr\in \Bimod(o_X-o_X)$ if and only if $\Nscr \in \Bimod(o_X-o_X)$. $\tilde{\Cscr}_{f,p}\subset \Bimod(o_X-o_X)$. Let $\Sscr\in \tilde{\Cscr}_{f,p} $. By induction there is a short exact sequence in $\tilde{\Cscr}_{f,p}$ $$0\r \Sscr_1\r \Sscr\r \Sscr_2\r 0$$ such that $S_i\in \Bimod(o_X-o_X)$. It now suffices to apply Corollary \[ref:5.5.6a\] above. Let us denote by $\operatorname{PCFin}(C_p-C_p)$ the full subcategory of $\operatorname{PC}(C_p-C_p)$ consisting of finite length objects. Thus $\operatorname{PCFin}(C_p-C_p)=\operatorname{PCFin}(C_p^\circ\ctimes C_p)$. Recall that $C_p/\rad(C_p)=k^I$ where $I$ is as in Theorem \[ref:5.1.4a\]. Thus the twosided bipseudo-compact maximal ideals in $C_p$ are naturally indexed by $I$. The corresponding simple modules were denoted by $S_i$ in Theorem \[ref:5.1.4a\]. Below we will view the $S_i$ as $C_p$-bimodules. It is clear that $(C_p)^\circ \ctimes C_p$ has a normalizing sequence given by $N\ctimes 1$ and $1\ctimes N$ and the quotient is given by $(C_p/(N))^\circ\ctimes (C_p/(N))$ which is locally noetherian. It follows from [@VdBVG Prop. 3.23] that $(C_p)^\circ\ctimes C_p$ is locally noetherian. Hence in particular the category $\operatorname{PCFin}(C_p-C_p)$ is given by the category of $C_p$-bimodules which are finite extensions of the $S_i$. The functor $\hat{(-)}_p$ is defined on $\tilde{\Cscr}_{f,p}$. We have \[ref:5.5.8a\] $(o_{\tau^i p})^\wedge_p=S_i$ We use . Let $q=\tau^i p$. $$\begin{aligned} (o_q)^\wedge_p&=\invlim_I ((C_p/I)^\sim_p\otimes_{C_p} o_q)^\wedge_p\\ &=\invlim_I((\Oscr_q)^\wedge_p\otimes_k \Hom_{\Qch(X)}((C_p/I)^\sim_p ,\Oscr_q)^\ast)\qquad (\text{lemma \ref{ref:5.5.1a}}) \\ &=\invlim_I(S_i\otimes_k \Hom_{\operatorname{PC}(C_p)}(C_p/I,S_i)^\ast)\\ &=S_i\end{aligned}$$ where as usual $I$ runs through the open twosided ideals in $C_p$. We have the following result. \[ref:5.5.9a\] The functor $\hat{(-)}_p$ defines an equivalence between $\tilde{C}_{p,f}$ and $\operatorname{PCFin}(C_p-C_p)$. First note that thanks to lemma \[ref:5.5.8a\] and Proposition \[ref:5.4.4a\] the image of $\tilde{C}_{p,f}$ under $(-)^\wedge_p$ is contained in $\operatorname{PCFin}(C_p-C_p)$. To show that this is actually an equivalence we will construct an inverse. Assume $U\in \operatorname{PCFin}(C_p-C_p)$ and $\Fscr\in \coh(X)$. Define $$\label{ref:5.28a} T(\Fscr)=(\hat{\Fscr}_p\otimes_{C_p} U)^\sim_p$$ Since $\hat{\Fscr}_p$ is finitely generated (Theorem \[ref:5.3.1a\]) we find that $T$ defines a right exact additive functor $\coh(X)\r \Cscr_{f,p}\subset \coh(X)$. Hence this functor extends to a functor $\tilde{T}:\Qch(X)\r \Qch(X)$ commuting with colimits. We denote by $\tilde{U}_p$ the right adjoint to this functor. If we view $\tilde{U}_p$ as an object in $\Bimod(o_X-o_X)$ then notationally we have $$\label{ref:5.29a} \Fscr\otimes_{o_X} \tilde{U}_p=T(\Fscr)=(\hat{\Fscr}_p\otimes_{C_p} U)^\sim_p$$ We now show that $\tilde{(-)}_p$ is a left inverse to $\hat{(-)}_p$. Let $\Sscr\in \Cscr_{f,p}$. To show that $(\hat{\Sscr}_p)^\sim_p=\Sscr$ it suffices to construct a natural isomorphism $$\Fscr\otimes_{o_X} (\hat{\Sscr}_p)^\sim_p=\Fscr\otimes_{o_X}\Sscr$$ where $\Fscr\in\coh(X)$. With the help of we compute $$\begin{aligned} \Fscr\otimes_{o_X} (\hat{\Sscr}_p)^\sim_p&= (\hat{\Fscr}_p\otimes_{C_p} \hat{\Sscr}_p)^\sim_p \\ &=((\Fscr\otimes_{o_X} \Sscr)^\wedge_p)^\sim_p\qquad\text{(Proposition \ref{ref:5.4.5a})} \\ &=\Fscr\otimes_{o_X} \Sscr\end{aligned}$$ Thus the composition $$\label{ref:5.30a} \tilde{\Cscr}_{f,p}\xrightarrow{\hat{(-)}_p} \operatorname{PCFin}(C_p-C_p)\xrightarrow{\tilde{(-)}_p} \Bimod(o_X-o_X)$$ is the identity on $\tilde{\Cscr}_{f,p}$. We now claim that the essential image of $\tilde{(-)}_p$ is contained in $\tilde{\Cscr}_{f,p}$. Since $\tilde{\Cscr}_{f,p}$ is closed under extensions in $\Bimod(o_X-o_X)$ and since by $\tilde{(-)}_p$ is at least right exact (it is of course exact) it suffices to show that $\tilde{S}_{i,p}\in \tilde{\Cscr}_{f,p}$. This follows from the fact that $(o_{\tau^i p})^\wedge_p=S_i$ by lemma \[ref:5.5.8a\], whence $\tilde{S}_{i,p}=o_{\tau^i p}$. Now we show that $\tilde{(-)}_p$ is also a right inverse to $\hat{(-)}_p$. Let $S\in \operatorname{PCFin}(C_p-C_p)$. We show that $(\tilde{S}_p)^\wedge_p=S$. By definition $$\begin{aligned} (\tilde{S}_p)^\wedge_p&=\invlim_I ((C_p/I)^\sim_p\otimes_{o_X} \tilde{S}_p)^\wedge_p\\ &=\invlim_I ((((C_p/I)^\sim_p)^\wedge_p\otimes_{C_p} S)^\sim_p)^\wedge_p\qquad\text{(eq. \eqref{ref:5.29a})}\\ &=\invlim (C_p/I)\otimes_{C_p} S\\ &=S\qed\end{aligned}$$ To close this section we discuss $\HTor$ between objects of $\tilde{\Cscr}_{f,p}$. \[ref:5.5.10a\] Let $\Fscr\in \coh(X)$, $\Tscr\in \tilde{\Cscr}_{f,p}$ and let $\Sscr$ be a coherent $\Cscr_p$-preserving $o_X$-bimodule such that $\hat{\Sscr}_p$ is finitely presented on the right. Then $$\begin{aligned} \label{ref:5.31a} \HTor_i^{o_X}(\Fscr,\Tscr)&\in \Cscr_{f,p}\\ \label{ref:5.32a} \HTor_i^{o_X}(\Sscr,\Tscr)&\in\tilde{\Cscr}_{f,p}\\ \label{ref:5.33a} \Tor^{\operatorname{PC}(C_p)}_i(\hat{\Fscr}_p,\hat{\Tscr}_p)&\in \operatorname{PCFin}(C_p)\\ \label{ref:5.34a} \Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p,\hat{\Tscr}_p)&\in \operatorname{PCFin}(C_p-C_p)\end{aligned}$$ Furthermore we have $$\begin{aligned} \label{ref:5.35a} \HTor_i^{o_X}(\Fscr,\Tscr)^\wedge_p& =\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Fscr}_p,\hat{\Tscr}_p)\\ \label{ref:5.36a} \HTor_i^{o_X}(\Sscr,\Tscr)^\wedge_p& =\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p,\hat{\Tscr}_p)\end{aligned}$$ By Proposition \[ref:5.1.6a\] and Theorem \[ref:5.3.1a\] $\hat{\Fscr}_p$ has a resolution by free modules of finite rank. This implies . From the hypotheses we also find that $\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p,\hat{\Tscr}_p)$ lies in $\operatorname{PCFin}(C_p)$. Since it also lies in $\operatorname{PC}(C_p-C_p)$, it must necessarily lie in $\operatorname{PCFin}(C_p-C_p)$. This proves . To prove the other statements of the theorem it suffices to prove $$\begin{aligned} \label{ref:5.37a} \HTor_i^{o_X}(\Fscr,\Tscr) &=\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Fscr}_p,\hat{\Tscr}_p)^\sim_p\\ \label{ref:5.38a} \HTor_i^{o_X}(\Sscr,\Tscr) &=\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p,\hat{\Tscr}_p)^\sim_p\end{aligned}$$ Recall that to prove these equations, we have to show that they represent the same functor on injectives. We’ll only prove . is similar. Let $E$ be an injective object in $\Qch(X)$. Let $E_1$ be the $\Cscr_p$ part of $E$. We have $$\begin{aligned} \HHom_{\Qch(X)}(\Tor^{\operatorname{PC}(C_p)}_i& (\hat{\Sscr}_p,\hat{\Tscr}_p)^\sim_p,E)\\ &=\HHom_{\Qch(X)}(\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p, \hat{\Tscr}_p)^\sim_p,E_1) \\ &=\Hom_{\operatorname{Top}(C_p)} (\Tor^{\operatorname{PC}(C_p)}_i(\hat{\Sscr}_p,\hat{\Tscr}_p),\hat{E}_{1,p})^\sim_p \qquad\text{(Prop. \ref{ref:5.4.3a})} \\ &=\Ext^i_{\operatorname{Top}(C_p)}(\hat{\Sscr}^p,\Hom_{\operatorname{Dis}(C_p)}(\hat{\Tscr}_p ,\hat{E}_{1,p}))^\sim_p\qquad \text{(lemma \ref{ref:4.20a})}\\ &=\HExt^i_{\Qch(X)}(\Sscr,\HHom_{\Qch(X)}(\Tscr,E_1))\qquad \text{(Prop. \ref{ref:5.4.3a})}\\ &=\HExt^i_{\Qch(X)}(\Sscr,\HHom_{\Qch(X)}(\Tscr,E))\\ &=\HHom_{\Qch(X)}(\HTor_i(\Sscr,\Tscr),E)\qquad \text{(eq. \eqref{ref:3.1b})}\end{aligned}$$ Done! Completion of algebras ---------------------- In the sequel we will need the following concept. \[ref:5.6.1a\] A topological $\ZZ$-graded ring $A$ is said to be graded cofinite if $A_0$ is cofinite and if the homogeneous parts of $A$ are bipseudo-compact $A_0$-modules. If $A$ is as above then we denote by $\operatorname{GrTop}(A)$ the category of graded right topological $A$-modules. With $\operatorname{GrDis}(A)$, resp. $\operatorname{GrPC}(A)$ we denote the full subcategories consisting of objects $\oplus_m M_m$ such that $M_m$ is in $\operatorname{Dis}(A_0)$, resp. $\operatorname{PC}(A_0)$. We say that $A$ is (right) locally noetherian if for every indecomposable idempotent in $A_0$ we have that $e_iA$ is noetherian in $\operatorname{GrPC}(A)$. Left locally noetherian is defined similarly. If $A$ is locally noetherian then it easily follows that $\operatorname{GrDis}(A)$ is locally noetherian. By analogy with the definition of $\Proj$ of an ordinary graded ring we define $\Tors(A)$ as the full subcategory of $\operatorname{GrDis}(A)$ consisting of objects which are direct limits of right bounded ones. We then put $\operatorname{QGrDis}(A)=\operatorname{GrDis}(A)/\Tors(A)$. Now let $X,Y,p$ etc…be as in the previous sections. Let $\Ascr=\oplus_n \Ascr_n$ be in $\Gralg(X)$ and assume that the $\Ascr_n$ are coherent and $\Cscr_p$ preserving. Then it makes sense to define $\hat{\Ascr}_p$ as $\oplus_n \hat{\Ascr}_p$. The compatibility of $\hat{(-)}_p$ with tensor product implies that $\hat{\Ascr}_p$ is a cofinite algebra. As in §\[ref:3.3b\] we define $\Cscr_{p}(\Ascr)$ as the graded $\Ascr$-modules $\Mscr$ such $\Mscr_n\in\Cscr_p$ for all $n$. Similarly $Q\Cscr_p(\Ascr)$ is the image of $\Cscr_p(\Ascr)$ in $\Gr(\Ascr)$. It is easy to prove the following. \[ref:5.6.2a\] The functor $\hat{(-)}_p$ defines an equivalence between $\Cscr_p(\Ascr)$ and $\operatorname{GrDis}(\hat{\Ascr}_p)$ and between $Q\Cscr_p(\Ascr)$ and $\operatorname{QGrDis}(\hat{\Ascr}_p)$. This follows easily from compatibility of completion with tensor product (Proposition \[ref:5.4.5a\]). Multiplicities in the case that $\tau$ has infinite order {#ref:5.7b} --------------------------------------------------------- This section is somewhat more special than the previous ones. Our aim is to use the completion functor to attach to certain $\Fscr\in\mod(X)$ an invariant $T_p(\Fscr)$ which, in some sense, is an analogue for the multiplicities of $\Fscr$ in $p$ and in the points infinitely near to $p$. A straightforward treatment seems only to be possible in the case that the $\tau$-orbit of $p$ is infinite, so we assume this throughout this section. Note that this implies in particular that $p$ is smooth on $Y$. To simplify the notations we will write $\Fscr_Y$ for $i^\ast\Fscr=\Fscr/\Fscr(-Y)$. We will say that $\Fscr\in\mod(X)$ is $Y$-transversal if $\Fscr_Y$ has finite length. It then follows automatically that $\ker (\Fscr(-Y)\r \Fscr)$ also has finite length (exercise). The category of $Y$-transversal objects is denoted by $\trans_Y(X)$. If $\Fscr\in \trans_Y(X)$ then, in a slight generalization of [@Ajitabh; @VdB23], we define $\Div(\Fscr)$ as the difference of the divisors on $Y$ associated respectively to the finite length $\Oscr_Y$-modules given by $\Fscr_Y$ and $\ker (\Fscr(-Y)\r \Fscr)$. It is easy to see that $\Div$ is additive on short exact sequences. Furthermore if one has a surjective map $\Fscr\r \Oscr_q$, with kernel $\Gscr$ then the following formula holds [@VdB23] $$\label{ref:5.39a} \Div(\Gscr)=\Div(\Fscr)-(q)+(\tau^{-1} q)$$ The disadvantage of the invariant $\Div(\Fscr)$ is that it depends on $\Fscr$ itself and not only on the image of $\Fscr$ in $\mod(X)/\Cscr_{f,p}$. We now define a better behaving invariant. Assume that $\Fscr\in \trans_Y(X)$. We look at the exact sequence $$\label{ref:5.40a} 0\r \Gscr\r \Fscr(-Y)\r \Fscr\r \Hscr\r 0$$ By definition $\Gscr$, $\Hscr$ are finite length modules on $Y$. The completion of looks like $$\label{ref:5.41a} 0\r \hat{\Gscr}_p\r \hat{\Fscr}_{p,\phi}\xrightarrow{\times N} \hat{\Fscr}_p \r \hat{\Hscr}_p\r 0$$ Now $\hat{\Gscr}_p$, $\hat{\Hscr}_p$ are finite length modules over $C_p/(N)$. It follows from Nakayama’s lemma(lemma \[ref:4.8b\]) that $\hat{\Fscr}_p$ is a quotient of a finite direct sum of $P_i$’s. Using we can decompose as a product of exact sequences of the form $$\label{ref:5.42a} 0\r \hat{\Gscr}_pe_i\r \hat{\Fscr}_p e_{i+1}\xrightarrow{\times N} \hat{\Fscr}_p e_i \r \hat{\Hscr}_p e_i\r 0$$ Since $\hat{\Fscr}_p$ is a quotient of a finite direct sum of $P_i$’s it follows that $\hat{\Fscr}_p e_i=0 $ will be zero if $i\gg 0$. Therefore it follows from that $\hat{\Fscr}_p e_i$ is a finite dimensional $R$-module. Furthermore since $\hat{\Gscr}_pe_i$, $\hat{\Hscr}_pe_i$ are only non-zero for a finite number of $i$, it follows that $ \hat{\Fscr}_p e_{i+1}\xrightarrow{\times N} \hat{\Fscr}_p e_i $ is an isomorphism for large negative $i$. We define $T_p(\Fscr)$ as $\invlim_{i} \hat{\Fscr}_pe_i$ (with transition maps given by $N$). Thus $T_p$ is a functor from $\trans_Y(X)$ to finite dimensional $R$-modules. $T_p$ has the following properties. 1. $T_p$ is exact. 2. $T_p(\Fscr)$ does only depend on the image of $\Fscr$ in $\mod(X)/\Cscr_{f,p}$. <!-- --> 1. This follows from the exactness of the completion functor. 2. This follows from the fact that for a finite dimensional $C_p$-module $G$ one has $Ge_i=0$ for almost all $i$. Let us now indicate a direct way of computing $T_p(\Fscr)$. We say that $\Fscr$ is if $\Fscr$ is $\Cscr_{f,p}$-torsion free and $\tau^ip$ is not in the support of $\Fscr_Y=0$, unless $i=0$. We have the following result. \[ref:5.7.2a\] 1. Every $\Fscr\in \trans_Y(X)$ is equivalent modulo $\Cscr_{f,p}$ to a unique $p$-normalized object in $\mod(X)$. 2. If $\Fscr$ is $p$-normalized then $T_p(\Fscr)=\hat{\Fscr}_{Y,p}$. <!-- --> 1. Let us first consider uniqueness. If we have an inclusion of $p$-normalized objects $\Fscr\subset\Fscr'$ such that $\Fscr'/\Fscr\in\Cscr_{f,p}$ then by $\Fscr=\Fscr'$. Now assume that $\Fscr$ and $\Fscr'$ are $p$-normalized objects which are equivalent modulo $\Cscr_{f,p}$. This means that there exist a $\Cscr_{f,p}$-torsion free $\Gscr$, together with inclusions $$\Fscr\r \Gscr\l \Fscr'$$ whose cokernel lies in $\Cscr_{f,p}$. We now replace $\Fscr$, $\Fscr'$ by their images in $\Gscr$ and we put $\Hscr=\Fscr+\Fscr'$. Since $\Hscr$ is a quotient of $\Fscr\oplus \Fscr'$ it follows that $\Hscr$ is also $p$-normalized. But then by the above it follows that $\Fscr=\Hscr=\Fscr'$. Let us now consider existence. Assume that $\Fscr\in \trans_Y(X)$. Without loss of generality we may assume that $\Fscr$ is $\Cscr_{f,p}$-torsion free. We will then modify $\Div(\Fscr)$ step by step until it satisfies the condition for $p$-normalization. If $\tau^i p\in\Div(\Fscr)$ with $i>0$ then we let $\Fscr'$ be the kernel of the associated surjective map $\Fscr\r \Oscr_{\tau^i p}$. By it follows that $\Fscr'$ is closer to being $p$-normalized than the original $\Fscr$. Assume now $\tau^i p\in\Div(\Fscr)$ with $i<0$. In this case we use the associated injective map $\Oscr_{\tau^{i+1} p}\r\Fscr_Y(Y)$ and we let $\Fscr'$ be the pullback of the maps $$\begin{CD} \Fscr(Y) @>>> \Fscr_Y(Y)\\ @. @AAA\\ @. \Oscr_{\tau^{i+1} p} \end{CD}$$ One now easily checks $$\Div(\Fscr')=\Div(\Fscr)-(\tau^i p)+(\tau^{i+1} p)$$ so that we have again made progress. Repeating these constructions we eventually find a normalized object, which is equivalent to $\Fscr$. 2. Assume that $\Fscr$ is normalized. Then $\hat{\Fscr}_{Y,p}e_i$ is zero, except if $i=0$. Whence we find from the exact sequence $T_p(\Fscr)=\hat{\Fscr}_{Y,p}e_0=\hat{\Fscr}_{Y,p}$. Let $N_p$ be the functor which associates to $\Fscr\in \trans_Y(X)$ its normalization. Then we have proved the following $N_p$ defines an equivalence of categories between $\trans_Y(X)/\Cscr_{f,p}$ and the full subcategory of $\trans_Y(X)$ consisting of $p$-normalized objects. If $\Fscr$ is $p$-normalized then we can reconstruct $\hat{\Fscr}_p$ from $T_p(\Fscr)$. \[ref:5.7.4a\] Assume that $\Fscr\in \trans_Y(X)$ is $p$-normal. Then $\hat{\Fscr}_p$ is isomorphic to the row vector $$(\cdots T_p(\Fscr) \cdots T_p(\Fscr)\ 0 \cdots 0 \cdots)$$ where the right most $T_p(\Fscr)$ occurs in position $0$. Apriori $\hat{\Fscr}_p$ is given by the row vector $(\hat{\Fscr}_p e_i)_i$ . By the previous discussion $\hat{\Fscr}_p e_i=0$ for $i\gg 0$ and multiplication by $N$ is an isomorphism on $\hat{\Fscr}_pe_i$ for $i\neq 0$. It now follows from the definition that $$\hat{\Fscr}_pe_i= \begin{cases} T_p(\Fscr)&\text{if $i\le 0$}\\ 0&\text{if $i>0$} \end{cases} \qed$$ Blowing up a point on a commutative divisor {#ref:6a} =========================================== Some ideals {#ref:6.1a} ----------- Let $X,Y,p$ be as above and let $q$ be an arbitrary point on $Y$. The bimodule $o_q$ defined above is by construction a quotient of $o_Y$ and hence also of $o_X$. We define $$\begin{aligned} m_{Y,q}&=\ker(o_Y\r o_q)\\ m_q&=\ker (o_X\r o_q)\end{aligned}$$ By Corollary \[ref:5.5.6a\], $m_{Y,q}$, $m_q\in \Bimod(o_X-o_X)$. We also define $$\label{ref:6.1b} \begin{split} I_Y&= m_{Y,p}\otimes_{o_X} o_X(Y)\\ &= m_{Y,p}\otimes_{o_Y}\Nscr_{Y/X}\\ {I}&= m_p\otimes_{o_X}o_X(Y) \end{split}$$ Clearly $I_Y\subset \Nscr_{Y/X}$, ${I}\subset \Oscr_X(Y)$. Below we give some properties of ${I^n}\overset{\text{def}}{=} \im(I^{\otimes n}\r o_X(nY))$. It will be clear that suitable analogs for $I_Y^n$ hold. \[ref:6.1.1a\] 1. $o_X(nY)/{I}^n\in \tilde{\Cscr}_{f,p}$. 2. ${I}^n$ is a $\Cscr_p$-preserving coherent bimodule contained in $\Bimod(o_X-o_X)$. <!-- --> 1. We use induction on $n$, the case $n=1$ being clear. Put $\Sscr_n=o_X(nY)/{I^n}$. Tensoring the exact sequence $$\label{ref:6.2a} 0\r {I^n}\r o_X(nY)\r \Sscr_n\r 0$$ with ${I}$ yields an exact sequence $${I}\otimes_{o_X} {I^n}\r {I} \otimes_{o_X} o_X(nY)\r {I}\otimes_{o_X} \Sscr_n\r 0$$ and thus $$\label{ref:6.3a} I^{n+1}=\ker ({I} \otimes_{o_X} o_X(nY)\r {I}\otimes_{o_X} \Sscr_n)$$ Viewing ${I}\otimes o_X(nY)$ and $I^{n+1}$ as subbimodules of $o_X((n+1)Y)$ we obtain from an exact sequence $$0\r {I}\otimes_{o_X} \Sscr_n\r \Sscr_{n+1}\r \Sscr_1(nY)\r 0$$ Hence it suffices to prove that ${I}\otimes_{o_X} \Sscr_n\in \tilde{\Cscr}_{p,f}$. Tensoring $$0\r {I} \r o_X(Y)\r o_{\tau p}\r 0$$ with $\Sscr_n$ yields $$0\r \HTor_1^{o_X}(o_{\tau p},\Sscr_n)\r {I}\otimes_{o_X} \Sscr_n \r o_X(Y)\otimes_{o_X} \Sscr_n\r o_{\tau p}\otimes_{o_X} \Sscr_n\r 0$$ By it follows that indeed ${I}\otimes_{o_X} \Sscr_n\in \tilde{\Cscr}_{f,p}$. 2. From the exact sequence , Proposition \[ref:3.1.8a\] and Corollary \[ref:5.5.6a\] it follows that indeed $I^{n}$ is coherent and is contained in $\Bimod(o_X-o_X)$. Let $\Tscr\in \Cscr_p$. Applying $\Tscr\otimes_{o_X}-$ and $\HHom_{o_X}(-,\Tscr)$ to the sequence yields long exact sequences $$\begin{gathered} 0\r\HTor_1^{o_X}(\Tscr,\Sscr_n)\r \Tscr\otimes_{o_X} {I^n} \r \Tscr(nY)\r \Tscr\otimes_{o_X} \Sscr_n\r 0 \\ 0\r\HHom_{o_X}(\Sscr_n,\Tscr)\r \Tscr(-nY)\r \HHom_{o_X}({I^n},\Tscr)\r \HExt^1_{o_X}(\Sscr_n,\Tscr)\r 0\end{gathered}$$ Thanks to and Proposition \[ref:5.4.3a\] we can conclude that ${I^n}$ is $\Cscr_p$ preserving. The following proposition gives a little additional information on ${I^n}$ which we will need below. \[ref:6.1.2a\] ${I^n}$ has the following additional properties. Assume that $\Mscr\in \coh(X)$. Then $$\HTor_i^{o_X}(\Mscr,{I^n})\qquad \begin{cases} \in \coh(X) &\text{if $i=0$}\\ \in \Cscr_{f,p}&\text{if $i=1$}\\ =0&\text{if $i\ge 2$} \end{cases}$$ The case $i=0$ is already covered by Proposition \[ref:6.1.1a\]. Hence we consider the case $i>0$. Tensoring the exact sequence on the left with $\Mscr$ yields $$\HTor_i^{o_X}(\Mscr,{I^n})=\HTor_{i+1}^{o_X}(\Mscr,\Sscr_n)$$ Hence by Theorem \[ref:5.5.10a\] $\HTor_i^{o_X}(\Mscr,{I^n})\in \Cscr_{f,p}$ for $i\ge 1$. It remains to be shown that $\HTor_j^{o_X}(\Mscr,\Sscr_n)=0$ for $j\ge 3$. Using Lemma \[ref:5.5.8a\] and Theorem \[ref:5.5.10a\], it is sufficient to know that the left projective dimension of $S_i$ is $\le 2$. This is the version for left modules of Proposition \[ref:5.1.7a\]. The following proposition gives a more explicit description of ${I^n}$. \[ref:6.1.3a\] One has the following alternative expression for ${I^n}$. $$\begin{gathered} {I^n}=m_{p}m_{\tau^{-1}p}\cdots m_{\tau^{n-1}p}\otimes_{o_X} o_X(nY)\end{gathered}$$ We have $$\begin{aligned} \label{ref:6.4a} ({I}^n)^\wedge_p&=\hat{I}^n_p\\ \label{ref:6.5a} &=(m_0m_{-1}m_{-2}\cdots m_{-n+1})_{\phi^{-n+1}}\end{aligned}$$ as subobjects of the invertible $C_p$-bimodule $(C_p)_{\phi^{-n+1}}$. Here $\phi$ is as in Theorem \[ref:5.1.4a\] and for $i\in\ZZ$, $m_i$ is the twosided maximal ideal in $C_p$ corresponding to $S_i$. The first statement follows easily from which implies that $$o_X(Y)\otimes_{o_X} m_q=m_{\tau^{-1}q}\otimes_{o_X} o_X(Y)$$ (as subobjects of $o_X(Y)$). Now we prove the second statement. Completing the exact sequence $$0\r m_p \r o_X\r o_{p}\r 0$$ and using lemma \[ref:5.5.8a\] gives $\hat{m}_p=m_0$. Furthermore applying definition together with Theorem \[ref:5.1.4a\].4 yields $o_X(Y)^\wedge_p=(C_p)_{\phi^{-1}}$. Using the compatibility of tensor product with $(-)^\wedge_p$ we then deduce from $$\begin{aligned} \label{ref:6.6a} \hat{I}_p&=(m_0)_{\phi^{-1}}\end{aligned}$$ is easily proved by induction. By we have an exact sequence $$0\r I^{n+1}\r {I}\otimes_{o_X} o_X(nY)\r {I}\otimes_{o_X} \Sscr_n\r 0$$ which yields an exact sequence (using exactness of $\hat{(-)}$ and compatibility with tensor product, see Propositions \[ref:5.4.4a\] and \[ref:5.4.5a\]) $$\label{ref:6.7a} 0\r (I^{n+1})^\wedge_p \r (\hat{I}_p)_{\phi^{-n}} \r \hat{I}_p\otimes_{o_X} \hat{\Sscr}_{n,p}\r 0$$ On the other hand completing the exact sequence $$0\r I^{n}\r o_X(nY)\r \Sscr_n\r 0$$ and tensoring with $\hat{I}_p$ yields by induction an exact sequence $$\label{ref:6.8a} 0\r \hat{I}_p\cdot \hat{I}^{n}_p \r (\hat{I}_p)_{\phi^{-n}} \r \hat{I}_p\otimes_{o_X} \hat{\Sscr}_{n,p}\r 0$$ Comparing and yields what we want. now follows from , and the easily verified fact that $\phi^{-1}(m_i)=m_{i-1}$. Before we continue we make a few remarks on the case that $p$ is a fixed point for $\tau$. In that case it follows from that $$\label{ref:6.9a} m_p\otimes_{o_X} o_X(Y)=o_X(Y)\otimes_{o_X} m_p$$ as subobjects of $o_X(Y)$. Let $\mu$ be the multiplicity of $p$ in $Y$. If $p$ is a fixed point then it is possible that $\mu\ge 1$ by Theorem \[ref:5.1.4a\]. We have the following description of $\mu$ which is completely analogous to the classical case. \[ref:6.1.4a\] Assume that $p$ is a fixed point for $\tau$. Then $\mu$ is the largest integer $n$ such that $o_X(-Y)\subset m_p^n$ as subbimodules of $o_X$. We have to find the largest integer such that $$o_X/m^n_p\r o_X/(o_X(-Y)+m^n_p)$$ is an isomorphism. Using the properties of $\hat{(-)}$ and in particular Theorem \[ref:5.5.9a\] we see that this is equivalent with $$R/m^n\r R/(RU+m^n)$$ being an isomorphism, where $m$ is the maximal ideal of $R$. This in turn is the same as saying that $U\in m^n$. Now lemma \[ref:5.1.8a\] yields what we want. \[ref:6.1.5a\] Let $\mu$ be the multiplicity of $Y$ in $p$ and let $n\in\ZZ$. Then the natural exact sequence $$0\r o_X((n-1)Y)\r o_X(nY) \r o_Y(nY)\r 0$$ restricts to an exact sequence $$\label{ref:6.10a} 0\r I^{n-\mu}((\mu-1)Y) \r {I^n}\r I^n_{Y}\r 0$$ Here $I^n$ for $n<0$ is to be interpreted as $o_X(nY)$. Note that by there is no ambiguity in the notation $I^{n-\mu}((\mu-1)Y)$. First we verify that the lower sequence is well defined. We have $$\label{ref:6.11a} \Oscr(- Y)\subset m_pm_{\tau^{-1} p}\cdots m_{\tau^{\mu-1}p}$$ If $\mu=1$ then this is by definition. If $\mu>1$ then $p$ is a fixed point and we can use lemma \[ref:6.1.4a\]. Tensoring with $o_Y(\mu Y)$ yields an inclusion $$\Oscr((\mu-1)Y)\subset I^\mu$$ Assume $n\ge \mu$. Multiplying with $I^{n-\mu}$ then yields an inclusion $$\label{ref:6.12a} I^{n-\mu}((\mu-1)Y)\subset {I^n}$$ In a similar way one verifies directly from that also holds if $n<\mu$. Hence is indeed well defined. To prove the proposition we look at the following commutative diagram. $$\begin{CD} @.0 @.0 @.0 @.\\ @. @VVV @VVV @VVV @.\\ 0 @>>> I^{n-\mu}((\mu-1)Y) @>\alpha>> {I^n} @>\beta>> I^n_{Y} @>>> 0\\ @. @VVV @VVV @VVV @.\\ 0 @>>> o_X((n-1)Y) @>>> o_X(nY) @>>> o_Y(nY) @>>> 0\\ @. @VVV @VVV @VVV @.\\ @. \Sscr_{n-\mu}((\mu-1)Y) @>\gamma>> \Sscr_n @>\delta >> \Sscr_{Y,n} @>>> 0\\ @. @VVV @VVV @VVV @.\\ @.0 @.0 @.0 @. \end{CD}$$ where as usual $\Sscr_n=o_X(nY)/{I^n}$, $\Sscr_{Y,n}=o_Y(nY)/I^n_{Y,p}$. If $n<0$ then $\Sscr_n$, $\Sscr_{n,Y}$ are defined as zero. In this diagram the middle row is exact, $\alpha$ is injective and $\beta$ is surjective. Elementary diagram chasing shows that the lower row is exact. We now have to show that $\ker \beta=\im\alpha$. A diagram chase shows that this is equivalent to $\gamma$ being injective. This is clear if $n<\mu$ so we may assume that $n\ge \mu$. Injectivity of $\gamma$ is then in turn equivalent to $$\label{ref:6.13a} \operatorname{length}\Sscr_n=\operatorname{length}\Sscr_{n-\mu}+\operatorname{length}\Sscr_{Y,n}$$ By Theorem \[ref:5.5.9a\] this can be checked on $C_p$. By Corollary \[ref:5.2.4a\] we have $$\operatorname{length}{\Sscr}_{n}=\frac{n(n+1)}{2}$$ If $\|O_\tau(p)\|>1$ then by Theorem \[ref:5.1.4a\] $p$ is smooth on $Y$ and hence $$\operatorname{length}\Sscr_{Y,n}=n$$ One sees that in this case is satisfied. Assume now that $\tau p=p$. By Theorem \[ref:5.5.9a\] we may check the injectivity of $\gamma$ after completing. We have $C_p=R$ as in Theorem \[ref:5.1.4a\]. Let $m$ be the maximal ideal of $R$. We then have $\hat{\Sscr}_n=R/m^n$ and the map $\hat{\gamma}_p$ is given by right multiplication by $U$. From lemma \[ref:5.1.8a\] it follows that $U\in m^\mu-m^{\mu+1}$. From this and the fact that $\gr R$ is a domain (for the $m$-adic filtration) we deduce that right multiplication by $U$ defines an injective map $R/m^{n-\mu}\r R/m^n$. Some Rees algebras {#ref:6.2b} ------------------ We will use $I_Y$, ${I}$ to define the following Rees algebras. $$\begin{aligned} \Dscr_Y&=o_Y\oplus I_Y\oplus I^2_{Y}\oplus\cdots\\ \Dscr&=o_X\oplus {I}\oplus I^2 \oplus \cdots\end{aligned}$$ As a corollary to Proposition \[ref:6.1.1a\] we immediately deduce $\Dscr_Y$, $\Dscr\in \Alg(X)$. Now we deduce some other good properties of $\Dscr$. We denote shifting in $\Gr(\Dscr)$ and $\Bigr(\Dscr)$ by $(-)$. As usual the case where $p$ is fixed point for $\tau$ is somewhat peculiar. In that case it follows easily from that $$\label{ref:6.14a} \Dscr(nY)\overset{\text{def}}{=}\Dscr\otimes_{o_X}o_X(nY) =o_X(nY)\otimes_{o_X}\Dscr$$ is a twosided invertible graded bimodule over $\Dscr$. This is clearly false if $\tau p\neq p$. It is sometimes convenient to define a modified Rees algebra $\tilde{\Dscr}$ by $$\label{ref:6.15a} \tilde{\Dscr}_n = \begin{cases} o_X(nY) &\text{if $n<0$}\\ I^n & \text{if $n\ge 0$} \end{cases}$$ $\tilde{\Dscr}_Y$ is defined similarly but with $o_Y,I_Y$ replacing $o_X,I$. \[ref:6.2.2a\] Let $\Dscr$, $\Dscr_Y$ be as above. Let $\mu$ be the multiplicity of $p$ on $Y$. Then 1. There is an exact sequence of graded $\tilde{\Dscr}$-bimodules. $$\label{ref:6.16a} 0\r \tilde{\Dscr}(-\mu)((\mu-1)Y)\r \tilde{\Dscr}\r \tilde{\Dscr}_Y\r 0$$ 2. $\Dscr$ is noetherian. 3. $\Dscr$ satisfies $\chi$. 4. $\cd\tau_{\Dscr}\le 2$. The fact that exists follows directly from Proposition \[ref:6.1.5a\]. Our aim is now to deduce 2., 3. and 4. from the corresponding statements for $\Dscr_Y$. So we first have to handle this case. If $\mu=1$ then $I_Y$ is an invertible bimodule. Then we can apply Proposition \[ref:3.9.13a\] to the exact sequence $$0\r I_Y\otimes_{o_X} \Dscr_Y(-1) \r\Dscr_Y \r \Oscr_{Y}\r 0$$ and we are done. Hence by Theorem \[ref:5.1.4a\] we only have to treat the case that $p$ is a fixed point. But this means that $m_{Y,p}$ is $\tau$-invariant. Let $\Escr_Y$ be the ordinary rees algebra associated to $m_{Y,p}$. $\Escr_Y$ may be viewed as a sheaf of algebras in the ordinary sense and the reader may verify that it satisfies properties 2., 3. and $\cd \tau_{\Escr_Y}\le 1$. Furthermore there is a category equivalence given by $$\label{ref:6.17a} \Gr(\Dscr_Y)\r \Gr(\Escr_Y):\oplus_n\Mscr_n\mapsto \oplus_n\Mscr_n(-nY)$$ From this it is routine to pull the good properties of $\Escr_Y$ back to $\Dscr_Y$. Define now $\Dscr'_{Y}$ by $$\label{ref:6.18a} (\Dscr'_{Y})_n= \begin{cases} \Dscr_Y&\text{if $n\ge \mu$}\\ \Dscr&\text{otherwise} \end{cases}$$ Then gives an exact sequence $$\label{ref:6.19a} 0\r \Dscr(-\mu)((\mu-1)Y)\r \Dscr\r \Dscr'_{Y}\r 0$$ It is easy to see that the fact that $\Dscr_Y$ is noetherian, together with the fact that ${I^n}$ is coherent for $n=0,\ldots,\mu-1$, implies that $\Dscr'_{Y}$ is noetherian. Furthermore $\Dscr_Y$ and $\Dscr'_{Y}$ have the same tails. So from lemma \[ref:3.9.16a\] we obtain that $\Dscr'_{Y}$ satisfies $\chi$ and furthermore $\cd \tau_{\Dscr'_{Y}}=\cd \tau_{\Dscr_Y}=1$. Now applying Proposition \[ref:3.9.13a\] to implies 2., 3. and 4. The inequality in \[ref:6.2.2a\].4 is of course an equality. However we won’t need this. Definition of blowing up {#ref:6.3b} ------------------------ We define ${\tilde{Y}}=\Proj \Dscr_Y$, $\tilde{X}=\Proj \Dscr$ and we call $\tilde{X}$ the blowing up of $X$ in $p$. We denote by $\pi$ resp. $\pi_Y$ the quotient maps $\Gr(\Dscr)\r \Qgr({\Dscr})$ resp. $\Gr(\Dscr_Y)\r \Qgr({\Dscr}_{Y})$. It follows from that we have a commutative diagram of quasi-schemes $$\label{ref:6.20a} \begin{CD} {\tilde{Y}} @>\beta>> Y\\ @V j VV @V i VV \\ \tilde{X} @>\alpha >> X \end{CD}$$ where $i$ is as before, $j$ comes from the quotient map $\Dscr\r \Dscr_Y$ (through Proposition ) and $\alpha^\ast\Mscr=\pi(\Mscr\otimes_{o_X} \Dscr)$, with the analogous definition for $\beta^\ast$. The derived functors $R^i\alpha_\ast$ are defined as usual. $L_i\alpha_\ast$ is defined as in Section §\[ref:3.10b\]. From lemma \[ref:3.10.2a\] together with Proposition \[ref:6.1.2a\] it follows that $L_i\alpha^\ast$ actually defines a functor $\Mod(X)\r \Mod({\tilde{X}})$. The following theorem summarizes some of the main properties of blowing up. \[ref:6.3.1a\] 1. The pair $({\tilde{Y}},\beta)$ is isomorphic in $\Qsch/Y$ to the ordinary commutative blowing up at $p$ of $Y$. In particular $\beta$ is an isomorphism if $Y$ is smooth in $p$. 2. ${\tilde{X}}$ is a noetherian quasi-scheme. 3. The ideal in $o_{{\tilde{X}}}$ defined by ${\tilde{Y}}$ is invertible. 4. $i_\ast\circ R^i\beta_\ast=R^i\alpha_\ast\circ j_\ast$. 5. We have $ \cd \alpha^\ast\le 1$, $\cd \alpha_\ast\le 1$ 6. $\alpha$ is proper 7. $o_{\tilde{X}}({\tilde{Y}})$ is relatively ample for $\alpha$. 8. $j$ makes $\tilde{Y}$ into a divisor in $\tilde{X}$ in the sense of enriched quasi-schemes (cfr §\[ref:3.7b\]). These properties are either straightforward translations of the definitions or they follow from properties of $\Dscr$ which we have already proved. 1. As usual we separate two cases. If $\tau p=p$ then we let $\Escr_Y$ be as in the proof of Theorem \[ref:6.2.2a\]. The category equivalence between $\Gr(\Dscr_Y)$ and $\Gr(\Escr_Y)$ yields in that case what we want. If $p$ is smooth on $Y$ (in particular if $\tau p\neq p$) then $m_{Y,p}$ is invertible. Thus the same holds for $I_Y$ and we have $$\Dscr_Y=o_{Y}\oplus I_Y\oplus I_Y^{\otimes 2}\cdots$$ It now follows from Propositions \[ref:3.11.4a\] (with $\Sscr=0$) that $\operatorname{QGr}(\Dscr_Y)\cong \Qch(Y)$. For use below we state the following formula for $\Nscr\in\Gr(\Dscr_Y)$. $$\label{ref:6.21a} \beta_\ast(\pi_Y\Nscr)=\dirlim \Nscr_n\otimes_{o_Y} I^{\otimes{-n}}_{Y}$$ We leave the easy proof to the reader. 2. This follows from Theorem \[ref:6.2.2a\].2. 3. Using Theorem \[ref:6.2.2a\].1. one easily checks that $$\label{ref:6.22a} \pi \Mscr\otimes_{o_{\tilde{X}}} o_{\tilde{X}}(-{\tilde{Y}})= \pi(\Mscr \otimes_{\Dscr} \Dscr(-\mu)((\mu-1)Y))$$ What we want to prove now follows from the fact that $\Dscr(-\mu)((\mu-1)Y)$ is an invertible graded bimodule over $\Dscr$ (this is trivial if $\mu=1$ and if $\mu\neq 1$ it follows from ). 4. Define $\gamma=i\circ \beta$. Since $i_\ast$ is exact we have $R^i\gamma_\ast=i_\ast\circ R^i\beta_\ast$. Hence we have to show that $$\label{ref:6.23a} R^i\gamma_\ast=R^i\alpha_\ast\circ j_\ast$$ We have ${\tilde{Y}}=\Proj \Dscr'_{Y}$ where $\Dscr'_{Y}$ is as in . shows that $\Dscr'_{Y,p}$ satisfies the conditions for Proposition \[ref:3.9.12a\]. Hence we can employ that proposition to obtain what we want. 5. This follows from Theorem \[ref:6.2.2a\].4, equation and Proposition \[ref:6.1.2a\]. 6. This follows from Theorem \[ref:6.2.2a\].3 together with part of Proposition \[ref:3.9.7a\]. 7. Since $\Dscr$ satisfies $\chi$ it follows from Proposition \[ref:3.9.7a\] that the canonical shift functor $\Mscr\mapsto \Mscr(1)$ is relatively ample. Denote the corresponding bimodule by $o_{\tilde{X}}(1)$. If $p$ is smooth on $Y$ then it follows from that $o_{\tilde{X}}({\tilde{Y}})=o_{\tilde{X}}(1)$. So in that case we are done. If $p$ is singular on $Y$ then by $-\otimes_{o_X} o_X(Y)$ defines an autoequivalence of $\Qch({\tilde{X}})$. Then yields $$\label{ref:6.24a} o_{\tilde{X}}({\tilde{Y}})=o_{\tilde{X}}(\mu)\otimes_{o_X} o_X((1-\mu )Y)$$ This yields what we want since it is clear that $-\otimes_{o_X} o_X(Y)$ commutes with $R^i\alpha_\ast$ and $L_i\alpha^\ast$. Note in passing that also makes sense in the case that $\tau p\neq p$ since then $\mu=1$. 8. Given 3. we only have to show that the map $\Oscr_{\tilde{X}}(-\tilde{Y})\r \Oscr_{\tilde{X}}$ is injective. Now $\Oscr_{\tilde{X}}(-\tilde{Y})$ is by definition $\alpha^\ast\Oscr_X\otimes_{o_{\tilde{X}}} o_{\tilde{X}}(-\tilde{Y})$ which according to is equal to $\pi(\Oscr_X\otimes_{o_X} \Dscr(-\mu)((\mu-1)Y))$. Hence by it suffice to show that $\HTor_{o_X}^i(\Oscr_X,I^n_Y((\mu-1)Y))=0$ for $i>0$, $n\ge 0$. As in the commutative case one checks that $\HTor_{o_X}^i(\Oscr_X,I^n_Y((\mu-1)Y))= \HTor_{o_Y}^i(\Oscr_Y,I^n_Y((\mu-1)Y))$. One now easily shows that $\HTor_{o_Y}^i(\Oscr_Y,I^n_Y((\mu-1)Y))=0$, for example using the fact that $I^n_Y\subset o_Y(nY)$ with the quotient being in $\Cscr_{f,p}$, together with the definition of $\HTor(-,-)$ (cfr. ). The normal bundle {#ref:6.4b} ----------------- The main result of this section (equation ) will be used in §\[ref:6.5b\]. If $t:U\r V$ is a map of schemes, $\Nscr$ is a line bundle on $Y$ and $\Fscr\subset \Nscr$ is a quasicoherent subsheaf then $t^{-1}(\Fscr)$ is defined as the image of $t^\ast(\Fscr)$ in $t^\ast(\Nscr)$. Now let us revert to the notations in use in the previous sections. Recall that according to we have $\Nscr_{Y/X} =\Nscr_\tau$ where $\Nscr$ is an invertible sheaf on $Y$ and $\tau$ is an automorphism of $Y$. Define $\tau':{\tilde{Y}}\r {\tilde{Y}}$ as follows. If $p$ is a fixed point for $\tau$ then $\tau$ extends in a natural way to the blowup of $Y$ in $p$. We denote this extended map by $\tau'$. If $p$ is not a fixed point then the map $\beta:{\tilde{Y}}\r Y$ is an isomorphism and we put $\tau'=\beta^{-1}\tau\beta$. So in all cases we have $$\label{ref:6.25a} \beta\tau'=\tau\beta$$ Our aim in this section is to prove the following formula $$\label{ref:6.26a} \Nscr_{{\tilde{Y}}/{\tilde{X}}}=\beta^{-1}(m_{Y,p}\Nscr)_{\tau'}$$ In view of the commutative case this formula seems quite logical. However there are some pitfalls. The main problem is that a priori ${\tilde{Y}}$ is only a quasi-scheme. Thus, although we can use the ordinary definition of $\beta^{-1}(m_{Y,p}\Nscr)$, the fact that we can consider the result as a bimodule, depends on the “accidental” event that ${\tilde{Y}}$ is commutative. So to make sense of we have to bring in explicitly the identification of ${\tilde{Y}}$ with a commutative scheme (which was given in the proof of Theorem \[ref:6.3.1a\].1). Below we give the necessary computations. The reader is advised to skim through the rest of this section. It is most convenient to separate two cases, depending on whether $p$ is a fixed point or not. [$p$ is not a fixed point for $\tau$.]{} In this case $\mu=1$ and $m_{Y,p}$ is invertible. Thus $\beta^{-1}(m_{Y,p}\Nscr)=\beta^{\ast}(m_{Y,p}\Nscr)$. We have to show that for $\Mscr\in \Qch({\tilde{Y}})$ we have $$\label{ref:6.27a} \Mscr\otimes_{o_{\tilde{Y}}}o_{\tilde{Y}}({\tilde{Y}})=\Mscr\otimes_{o_{\tilde{Y}}}(\beta^{\ast} (m_{Y,p}\Nscr))_{\tau'})$$ In this case the identification of ${\tilde{Y}}$ with a commutative scheme is given by $\beta$. Thus if $\Tscr$ is a quasicoherent $\Oscr_Y$-module then $$\Mscr\otimes_{o_{\tilde{Y}}}\beta^{\ast}(\Tscr)=\beta^\ast(\beta_\ast\Mscr\otimes_{o_Y} \Tscr)$$ We use this in the computation below. $$\begin{aligned} \Mscr\otimes_{o_{\tilde{Y}}}(\beta^{\ast}(m_{Y,p}\Nscr))_{\tau'} &=\tau'_\ast(\Mscr\otimes_{o_{\tilde{Y}}}\beta^{\ast}(m_{Y,p}\Nscr))\qquad \text{(See \eqref{ref:5.4a})}\\ &=(\beta^{-1})_\ast\tau_\ast\beta_\ast\beta^\ast(\beta_\ast\Mscr\otimes_{o_Y} m_{Y,p}\Nscr)\\ &=(\beta^{-1})_\ast\tau_\ast(\beta_\ast\Mscr\otimes_{o_Y}m_{Y,p}\Nscr)\end{aligned}$$ Thus, using , reduces to $$\beta_\ast(\Mscr(1))=\tau_\ast(\beta_\ast\Mscr\otimes_{o_Y}m_{Y,p}\Nscr)$$ Now the righthand side of this equation is equal to $$\beta_\ast\Mscr\otimes_{o_Y}m_{Y,p}\Nscr_\tau= \beta_\ast\Mscr\otimes_{o_Y}I_Y$$ So finally we have to show $$\beta_\ast(\Mscr(1))=\beta_\ast\Mscr\otimes_{o_Y}I_Y$$ But this follows easily from . [$p$ is a fixed point for $\tau$.]{} Now we identify ${\tilde{Y}}$ with the ordinary commutative blowup of $Y$ at $p$. Denote the latter by ${\tilde{Y}}'$. We have ${\tilde{Y}}'=\Proj\Escr_Y$ where $\Escr_Y$ is as in the proof of Theorem \[ref:6.2.2a\]. Let $\beta':{\tilde{Y}}'\r Y$ be the structure map. There is now a commutative diagram $$\begin{CD} {\tilde{Y}} @>\gamma>> {\tilde{Y}}'\\ @V\beta VV @V\beta' VV\\ Y @= Y \end{CD}$$ where $\gamma:{\tilde{Y}}\r {\tilde{Y}}'$ denotes the identification. Note however that $\gamma$ does not commute with the canonical shift functors. Translating we find that $$\label{ref:6.28a} \gamma_\ast(\Mscr(1))=(\gamma_\ast\Mscr)(1)\otimes_{o_{Y}} o_Y(Y)$$ If we now look back at the definition of $\tau'$ then we see that it was in fact defined on ${\tilde{Y}}'$ and then pulled back to ${\tilde{Y}}$ by $\gamma$. Thus $$\tau'=\gamma^{-1}\tau''\gamma$$ where $\tau''$ is the extension of $\tau$ to ${\tilde{Y}}'$. We compute $$\begin{aligned} \Mscr\otimes_{o_{\tilde{Y}}}\beta^{-1}(m^\mu_{Y,p}\Nscr)_{\tau'} &=(\gamma^{-1})_\ast\tau''_\ast(\gamma_\ast\Mscr\otimes_{o_{{\tilde{Y}}'}}\beta^{'-1} m^\mu_{Y,p}\Nscr)\end{aligned}$$ This equality is a formal computation, analogous to the one where $p$ is not a fixed point. Using we now have to show $$\label{ref:6.29a} \gamma_\ast(\Mscr(\mu)\otimes_{o_Y} o_Y((1-\mu)Y))= \tau''_\ast(\gamma_\ast\Mscr\otimes_{o_{{\tilde{Y}}'}}\beta^{'-1} (m^\mu_{Y,p}\Nscr))$$ From it follows that the lefthand side of this equation is equal to $$\label{ref:6.30a} \gamma_\ast\Mscr(\mu)\otimes_{o_Y}o_Y(Y)=(\gamma_\ast\Mscr\otimes_{o_Y} o_Y(Y)) (\mu)$$ Now we translate everything to graded modules. Let $\gamma_\ast\Mscr$ be represented by $\Pscr$. Then the righthand side of is represented by $$\tau_\ast(\Pscr\otimes_{\Escr_Y}(m^\mu_{Y,p} \Escr_Y\otimes_{o_Y}\Nscr))$$ is represented by $$\tau_\ast(\Pscr\otimes_{o_Y}\Nscr)(\mu)=\tau_\ast(\Pscr \otimes_{\Escr_Y}(\Escr_Y(\mu)\otimes_{o_Y}\Nscr))$$ Thus it is sufficient to show that up to right bounded bimodules $$m^\mu_{Y,p} \Escr_Y=\Escr_Y(\mu)$$ This is now clear. Birationality {#ref:6.5b} ------------- Here the notations are as in the previous section. We will show that $X$ and ${\tilde{X}}$ are isomorphic “outside the $\tau$-orbit of $p$”. In particular we may view $X$ and ${\tilde{X}}$ as being birational. Define $\alpha^{-1}(\Cscr_p)$ as in §\[ref:3.11b\]. Thus the objects in $\alpha^{-1}(\Cscr_p)$ are the objects in $\Qch({\tilde{X}})$ that are represented by graded $\Dscr$-modules such that $\Mscr_n\in\Cscr_p$ for all $p$. $\alpha^{-1}(\Cscr_p)$ has the following slightly more intrinsic description. \[ref:6.5.1a\] $\alpha^{-1}(\Cscr_p)$ is the full subcategory of objects $\Mscr$ in $\Qch({\tilde{X}})$ for which one has $\alpha_\ast\Mscr(n{\tilde{Y}})\in\Cscr_p$ for all $n$. From together with it easily follows that $\alpha_\ast\Mscr(n{\tilde{Y}})\in\Cscr_p$ if $\Mscr\in\alpha^{-1}(\Cscr_p)$. Let us now prove the converse inclusion. Assume that $\alpha_\ast\Mscr(n{\tilde{Y}})\in\Cscr_p$ for all $n$. We have to show that $\Mscr\in\alpha^{-1}(\Cscr_p)$. Let $\Mscr=\pi\Nscr$ for $\Nscr\in\Gr(\Dscr)$. It suffices to consider the case that $\Nscr$ is noetherian. We will prove that $\Nscr_n\in\Cscr_p$ for $n\gg 0$. This implies $\Mscr\in\alpha^{-1}(\Cscr_p)$. According to $$\label{ref:6.31a} \Mscr(n{\tilde{Y}})=\pi\Nscr(n\mu)\otimes_{o_X}o_X(n(1-\mu)Y)$$ So if $\mu=1$ then this formula says that $\tilde{\Nscr}_n\in\Cscr_p$ for all $n$. Since $\Dscr$ satisfies $\chi$ (Theorem \[ref:6.2.2a\]) it follows from lemma \[ref:3.9.3a\] that $\tilde{\Nscr}_n=\Nscr_n$ for $n\gg 0$. Hence we are done. This reasoning still works if $\mu>1$ since $-\otimes_{o_X}o_X(Y)$ commutes with $\alpha_\ast$. Hence yields that $\tilde{\Nscr}_{n\mu}\in\Cscr_p$ for all $n$. Then according to lemma \[ref:3.12.1a\] we have modulo $\Tors(\Dscr)$ $$\Nscr=\tilde{\Nscr}=\tilde{\Nscr}^{(\mu)}\otimes_{\Dscr^{(\mu)}}\Dscr$$ Let $\Nscr'$ be the module on the righthand side of this equation. It is clear that $\Nscr'_n\in\Cscr_p$ for all $n$. Since $\Nscr$ and $\Nscr'$ are isomorphic in $\operatorname{QGr}(\Dscr)$ and $\Nscr$ is noetherian, it easily follows that $\Nscr_n\in\Cscr_p$ for $n\gg 0$. \[ref:6.5.2a\] We have 1. $\alpha^\ast$ and $\alpha_\ast$ define inverse equivalences between $\Qch(X)/\Cscr_p$ and $\Qch({\tilde{X}})/\alpha^{-1}(\Cscr_p)$. 2. The functor $\alpha_\ast$ sends $\alpha^{-1}(\Cscr_p)$ to $\Cscr_p$ and the functor $R^1\alpha_\ast$ sends $\Qch({\tilde{X}})$ to $\Cscr_p$. 3. The functor $\alpha^\ast$ sends $\Cscr_p$ to $\alpha^{-1}(\Cscr_p)$ and the functor $L_1\alpha^\ast$ sends $\Qch(X)$ to $\alpha^{-1}(\Cscr_p)$. 4. The functors $R^i\alpha_\ast$ and $L_i\alpha^\ast$ preserve coherent objects. <!-- --> 1. It is clear from Proposition \[ref:6.1.1a\] and Theorem \[ref:5.5.10a\].1 that $\tilde{\Dscr}$ is strongly graded with respect to $\Cscr_p$. What we have to prove now follows from Proposition \[ref:3.11.4a\]. 2. The statement about $\alpha_\ast$ follows from . The statement about $R^1\alpha_\ast$ follows from applying $\alpha_\ast$ to an injective resolution and using the fact that $\alpha_\ast$ is exact modulo $\Cscr_p$ (by Proposition \[ref:3.11.4a\]). 3. This follows from the definition of $L_i\alpha^\ast$ together with Propositions \[ref:6.1.1a\] and \[ref:6.1.2a\]. 4. This follows from Theorem \[ref:6.3.1a\].6 and lemma \[ref:3.1.17a\]. \[ref:6.5.3a\] Assume that $q\in Y$ is not contained in the $\tau$-orbit of $p$. Let $q'$ be the unique point of ${\tilde{Y}}$ such that $\beta(q')=q$. Then $\alpha^\ast$ and $\alpha_\ast$ define inverse equivalences between $\Cscr_q$ and $\Cscr_{q'}$. Using the foregoing proposition it is sufficient to prove the following 1. $\Cscr_q\cap \Cscr_p=0$. 2. $\Cscr_{q'}\cap \alpha^{-1}(\Cscr_p)=0$. 3. $\alpha_\ast(\Cscr_{q'})\subset \Cscr_q$. 4. $\alpha^\ast(\Cscr_q)\subset \Cscr_{q'}$. 1\. is clear. To prove 3. it is sufficient to show that $\alpha_\ast \Oscr_{\tau^{\prime n}q'}\in \Cscr_q$ where $\tau'$ is as in §\[ref:6.4b\]. This follows from the fact that this is obviously true for $\beta$ by . Now we prove 2. Suppose $\Mscr\in \Cscr_{q'}\cap \alpha^{-1}(\Cscr_p)$. Then $\alpha_\ast\Mscr(n{\tilde{Y}})\in \Cscr_q\cap \Cscr_p=0$. From the relative ampleness of $o_{\tilde{X}}(n{\tilde{Y}})$ we deduce that $\Mscr=0$. Finally we prove 4. It is sufficient to show that $$\label{ref:6.32a} \alpha^\ast \Oscr_{\tau^n q}=\beta^\ast \Oscr_{\tau^n q}$$ First note that if $\Nscr\in\Cscr_{f,q}$ then $\hat{\Nscr}_p=0$. Hence if $\Sscr\in\tilde{\Cscr}_{f,p}$ then by we have $\HTor_i^{o_X}(\Nscr,\Sscr)=0$. We deduce that $$\Oscr_{\tau^n q}\otimes_{o_X}{I^n}=\Oscr_{\tau^n q}\otimes_{o_X}o_X(nY)= \Oscr_{\tau^n q}\otimes_{o_Y}o_Y(nY)=\Oscr_{\tau^n q}\otimes_{o_Y}I^n_{Y}$$ We obtain by summing over all $n$ and applying $\pi$. The exceptional curve {#ref:6.6b} --------------------- In this section we will for simplicitly make use of the object $\Oscr_X$ and consequently of the functor $(-)_{o_X}$ which goes from bimodules on $X$ to objects in $\Mod(X)$ (cfr. §\[ref:3.6b\]). From the compatibility of completion with $\Hom$ and tensor product we deduce that the following functors $$\tilde{\Cscr}_{f,p} \xrightarrow{(-)_{o_X}} \Cscr_{f,p} \xrightarrow {\Gamma(X,-)} \mod(k)$$ are faithful and exact. On $o_p$ they act by $ o_p\mapsto \Oscr_p\mapsto k $. In the sequel we will write $\Oscr_L=\alpha^\ast O_{\tau p}$. Thus $\Oscr_L=\pi((\Dscr/m_{\tau p}\Dscr)_{o_X})$. The following lemma gives a description of $\Dscr/m_{\tau p}\Dscr$ as $o_X-\Dscr$-bimodule. There is an exact sequence of $o_X-\tilde{\Dscr}$-bimodules. $$\label{ref:6.33a} 0\r (o_X(-Y)\otimes_{o_X}\tilde{\Dscr})(1)\r \tilde{\Dscr}\r \Dscr/m_{\tau p}\Dscr\r 0$$ From Proposition \[ref:6.1.3a\] (and its proof) it follows easily that $$o_X(-Y)\otimes_{o_X} I^{n+1}=m_{\tau p} I^n$$ This yields by taking direct sums over $n$. It was observed by Smith and Zhang [@SmithZhang] that it is possible to develop a formalism such that $\Oscr_L$ is really the structure sheaf of a “non-commutative curve” $L$. We will say more on this in §\[ref:6.7b\]. In the special case that $p=\tau p$ one has that $m_{\tau p}\Dscr$ is a twosided ideal. Then we can simply define $L=\Proj \Dscr/m_p\Dscr$. In this case we will of course use the notation $o_L$ for the algebra on $\tilde{X}$ corresponding to the identity functor on $\Qch(L)$. We find \[ref:6.6.2a\] Assume that $p=\tau p$. Then 1. $L\cong \PP^1$. 2. $L$ is embedded as a divisor in ${\tilde{X}}$. 3. There is a commutative diagram $$\label{ref:6.34a} \begin{CD} L@>\gamma>> p\\ @Vu VV @V Vv V\\ {\tilde{X}} @>\alpha>> X \end{CD}$$ where $u,v$ are the inclusions and $\gamma$ is isomorphic to the structure map $\PP^1\r \spec k$. <!-- --> 1. Using Proposition \[ref:5.6.2a\] we see that $\Gr(\Dscr/m_p\Dscr)$ is equivalent with $\Gr((\Dscr/m_p\Dscr)^\wedge)$ and a similar result for $\operatorname{QGr}$. Using the compatibility results for $\hat{(-)}$ we find $$\label{ref:6.35a} (\Dscr/m_p\Dscr)^\wedge=R/m\oplus (m/m^2)_{\phi^{-1}}\oplus(m^2/m^3)_{\phi^{-2}}\oplus\cdots$$ where $R$ is as in Theorem \[ref:5.1.4a\] and $m$ is the maximal ideal of $R$. It is easily seen that the ring on the right of is a noetherian two generator quadratic algebra with one relation. It is well known that the $\Proj$ of such a graded ring is $\PP^1$. 2. From it follows that up to right bounded bimodules we have $$m_p\Dscr=\Dscr(1)(-Y)$$ Hence, up to right bounded bimodules, $m_p\Dscr$ is an invertible ideal and in particular $L$ is a divisor in $\tilde{X}$. 3. This is a translation of the commutative diagram of $o_X$-algebras, given by $$\begin{CD} \Dscr/m_p\Dscr @<<< o_X/m_p\\ @AAA @AAA \\ \Dscr @<<< o_X \end{CD}\qed$$ We are now in a position to prove the following result. \[ref:6.6.3a\] Assume $q\in {\tilde{Y}}$ if $\tau p\neq p$ and $q\in {\tilde{Y}}\cup L$ if $\tau p= p$. Then $\Oscr_{q}$ has finite injective dimension in $\Qch({\tilde{X}})$. First note that $q\r \Oscr_q$ is a one-one correspondence between the points on ${\tilde{Y}}$ and the simple objects in $\Qch({\tilde{Y}})$. Similarly for $L$ and $\Qch(L)$. Therefore we interprete $L\cap {\tilde{Y}}$ set theoretically as those $q$ such that $\Oscr_q$ lies both in $\Qch({\tilde{Y}})$ and in $\Qch(L)$. Put $$r=\begin{cases} \beta(q)&\text{if $q\in {\tilde{Y}}$}\\ p&\text{if $q\in L$} \end{cases}$$ Using the diagrams and we see that in all cases $\Oscr_r=\alpha_\ast \Oscr_q$. Thus $r$ is well defined. Consider first the case $\tau p\neq p$. Then $\beta$ is an isomorphism. If $r\not\in O_\tau(p)$ then according to Corollary \[ref:6.5.3a\] $\Cscr_r$ is equivalent with $\Cscr_q$. Since both these categories are stable under injective hulls (Proposition \[ref:5.1.3a\]) we are through in this case. However if $r\in O_\tau(p)$ then according to Theorem \[ref:5.1.4a\] $r$ is smooth on $Y$. Thus $q$ is also smooth on ${\tilde{Y}}$ hence we can apply lemma \[ref:5.1.1a\]. Consider now the case $\tau p=p$. If $p\in L$ then we can apply lemma \[ref:5.1.1a\] again with $L$ as our curve. So assume now that $q\in {\tilde{Y}}-{\tilde{Y}}\cap L$. We then claim that $r\neq p$. Suppose the contrary. Thus there is an isomorphism $\zeta:\Oscr_p\r\alpha_\ast \Oscr_q$. By adjointness we obtain a map $\eta:\alpha^\ast \Oscr_p\r \Oscr_q$. This map must be non-zero since otherwise $\zeta$ had to be zero also. Since $\Oscr_q$ is simple we obtain that $\eta$ is surjective. Now the definition of $L$ implies that $\alpha^\ast \Oscr_p\in \Qch(L)$. Hence since $\Qch(L)$ is closed in $\Qch({\tilde{X}})$ we find $\Oscr_q\in \Qch(L)$. Contradiction. Since $r\not\in O_\tau(p)=\{p\}$ we can now use the same reasoning as in the case $\tau p\neq p$ to conclude that $\Oscr_q$ has finite injective dimension. The structure of $\alpha^{-1}(\Cscr_p)$ {#ref:6.7b} --------------------------------------- In this section we prove Proposition \[ref:6.7.1a\] below. It is a generalization of the main result of [@SmithZhang] in our special case. Throughout $n=\|O_\tau (p)\|$. \[ref:6.7.1a\] Let $L=\Qch(L)$ be the full abelian subcategory of $\Qch(\tilde{X})$ whose objects are direct limits of subquotients of finite sums $\Oscr_L(m_1)\oplus\cdots \oplus \Oscr_L(m_n)$ (note that this use of $L$ is consistent with the use in §\[ref:6.6b\], when $n=1$). Let $S$ be one of the following graded rings. 1. If $n=1$ then $S$ is the twist [@ATV2; @Zhang] of $\gr_F R$ by the automorphism $\phi$, which was introduced in Theorem \[ref:5.1.4a\]. Here $F$ denotes the $m$-adic filtration on $R$. 2. If $2\le n<\infty$ then $S=k[u,v]$ where $\deg u=1$, $\deg v=n$. 3. If $n=\infty$ then $S=k[x]$ with $\deg x=1$. Then $$L\cong \begin{cases} \operatorname{QGr}S &\text{if $n<\infty$}\\ \Gr S &\text{if $n=\infty$} \end{cases}$$ This equivalence is compatible with the natural shift functors and sends $\Oscr_L$ to $S$. The proof of this result depends on $n$. If $n=\infty$ then the result follows from [@SmithZhang], so we treat this case first. The case [ $n=\infty$]{} In this case we have an exact sequence (by ) $$0\r \Oscr_L(-1)\r \Oscr_L\r \Oscr_{L,\tilde{Y}}\r 0$$ and $$\label{ref:6.36a} \Oscr_{L,\tilde{Y}}=j^\ast\alpha^\ast\Oscr_{\tau p}= \Oscr_{\tau \beta^{-1}(p)}$$ (where we use $\alpha j=i\beta$). Proposition \[ref:6.7.1a\] follows from [@SmithZhang] if we can show that $\Oscr_{\tau \beta^{-1}(p)}$ is the only simple quotient of $\Oscr_L$. Hence let $\Oscr_L\r\Sscr$ be such a simple quotient. Tensoring with $\Oscr_{\tilde{Y}}$ we find that $\Sscr_Y$ is a quotient of $\Oscr_{\tau \beta^{-1}(p)}$. Since $\Oscr_{\tau \beta^{-1}(p)}$ is simple, we either have $\Sscr_{\tilde{Y}}=\Oscr_{\tau \beta^{-1}(p)}$ or $\Sscr_{\tilde{Y}}=0$. If we are in the first case then $\Oscr_{\tau \beta^{-1}(p)}$ is a quotient of $\Sscr$ and hence by the simplicity of $\Sscr$ : $\Sscr=\Oscr_{\tau \beta^{-1}(p)}$. Hence assume that we are in the second case. Let $\Tscr$ be some graded quotient of $(\Dscr/m_{\tau p}\Dscr)_{o_X}$ such that $\Sscr=\pi \Tscr$, Then $0=\Sscr_{\tilde{Y}}=\Sscr/\Sscr(-Y)=\Sscr/\Sscr(-1)$ implies that $\Tscr_m\cong \Tscr_{m+1}$ as $o_X$-modules for $m\gg 0$. From the next lemma it follows that for $m\gg 0$, $\Tscr_m$ has a composition series starting with $(\Oscr_{\tau p}(mY),\Oscr_{\tau p}((m-1)Y),\ldots)$, which is clearly incompatible with the isomorphism $\Tscr_m\cong \Tscr_{m+1}$. Hence we have obtained a contradiction. We conclude that $\Sscr=\Oscr_{\tau \beta^{-1} (p)}$ and so we can invoke the results of [@SmithZhang]. The following lemma was used. \[ref:6.7.2a\] Fix $t\in\NN$. Then $(\Dscr/m_{\tau p}\Dscr)_{o_X,t}$ is uniserial of length $t+1$ with composition series. $(\Oscr_{\tau p}(tY),\ldots, \Oscr_{\tau p}(Y),\Oscr_{\tau p})$ (starting from the top). By definition $(\Dscr/m_{\tau p}\Dscr)_t$ is given by $$(m_p\cdots m_{\tau^{1-t}p})/(m_{\tau p} m_p\cdots m_{\tau^{1-t} p})$$ We can compute this by completing (for example using Proposition \[ref:5.2.2a\]). We find $$(\Dscr/m_{\tau p}\Dscr)^\wedge_t\cong(\ldots,0,R/m,\ldots,R/m,0,\ldots,\ldots)$$ where the $R/m$ occur in positions $1,\ldots,1+t$. It is easy to see that this is a uniserial right $C_p$-module with the correct composition series. The case [ $n=1$]{} This follows from the proof of Proposition \[ref:6.6.2a\].1. The case [ $2\le n<\infty$]{} From the viewpoint of computations this is the most interesting case. According to Proposition \[ref:5.6.2a\] we may clearly assume that $X=\Spec C_p$, $Y=\Spec C_p/(N)$. Note that one has $C_p/(N)=R/(U)\oplus\cdots\oplus R/(U)$. We should now analyze the blowup $\tilde{X}$ of $X$ in the point defined by $m_0\subset C_p$ (which according to our current conventions corresponds to $p$). However it turns out that it is slightly more convenient to work with $m_{n-1}$. This does not alter our results in any way since all maximal ideals in $C_p$ are conjugate under the automorphism induced by $N$. By definition $$\begin{aligned} \Dscr&=C_p\oplus m_{n-1} N^{-1} \oplus (m_{n-1} N^{-1})^2\oplus\cdots\\ &=C_p\oplus m_{n-1} N^{-1} \oplus m_{n-1} m_{n-2} N^{-2}\oplus\cdots\end{aligned}$$ The $n$’th Veronese of $\Dscr$ is given by $$\Dscr^{(n)}=C_0\oplus J N^{-n}\oplus J^2 N^{-2n}\oplus \cdots$$ where $J=m_{n-1}m_{n-2}\cdots m_0$. Since $N^n J N^{-n} =J$ we find that conjugation by $N^n$ induces an automorphism of $\Dscr^{(n)}$. Twisting by this automorphism [@ATV2; @Zhang] yields that $\Gr(\Dscr^{(n)})$ is equivalent to $\Gr(\Uscr)$ where $\Uscr$ is the ordinary Rees algebra of the ideal $J\subset C_p$. Now before we continue, we remind the reader that taking Veronese’s and twisting is in general not compatible with the natural shift-functors. However here we have $o_{\tilde{X}}(1)=o_{\tilde{X}}(\tilde{Y})$. That is, the shift functor on $\tilde{X}$ is defined by a divisor. Hence to keep track of this shift functor, we simply have to keep track of $\tilde{Y}$. As we will see this is easy. To make progress we have to compute $J$ explicitly. Using the material in Proposition \[ref:5.2.2a\] or directly we find that $$J= \begin{pmatrix} m &(U) &\cdots &\cdots &(U) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots &(U) \\ m &\cdots &\cdots &\cdots &m \end{pmatrix}$$ Computing the powers of $J$ yields $J^n=m^{n-1} J$. Let $T$ be the Reesring of $R$ associated to $m$. We find. $$\Uscr= \begin{pmatrix} T &RU\oplus TU(-1) &\cdots &\cdots &RU\oplus TU(-1) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots &RU\oplus TU(-1) \\ T &\cdots &\cdots &\cdots &T \end{pmatrix}$$ In particular we have an inclusion $$\label{ref:6.37a} \begin{pmatrix} T &TU(-1) &\cdots &\cdots &TU(-1) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots &TU(-1) \\ T &\cdots &\cdots &\cdots &T \end{pmatrix} \hookrightarrow \Uscr$$ with rightbounded cokernel. Thus $\tilde{X}=\Proj\Dscr=\Proj \Dscr^{(n)}\cong \Proj \Uscr$ is equal to the Proj of the lefthand side of . There is a more elegant way to look at this. Let $X_c=\Spec R$, $Y_c=\Spec R/(U)$, $\tilde{X}_c=\Proj T$. Let $T_Y$ be the Reesring of $R/(U)$ and put $\tilde{Y}_c=\Proj Y_c$. Define $\Ascr$ to be the following algebra on $\tilde{X}_c$ : $$\Ascr= \begin{pmatrix} o_{\tilde{X}_c} &o_{\tilde{X}_c}(-\tilde{Y_c}) &\cdots &\cdots &o_{\tilde{X}_c}(-\tilde{Y_c}) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots & o_{\tilde{X}_c}(-\tilde{Y_c}) \\ o_{\tilde{X}_c} &\cdots &\cdots &\cdots &o_{\tilde{X}_c} \end{pmatrix}$$ Then we obtain $\tilde{X}=\Spec \Ascr$ in $\QSch/X_c$. A similar computation yields that $\tilde{Y}=\Spec \Cscr$ where $\Cscr=\diag(o_{\tilde{Y}_c},\ldots, o_{\tilde{Y}_c})$. Furthermore $o_{\tilde{X}}(-\tilde{Y})$ is given by the twosided ideal $$\label{ref:6.38a} \Jscr=\begin{pmatrix} o_{\tilde{X}_c} (-\tilde{Y_c}) &o_{\tilde{X}_c}(-\tilde{Y_c}) &\cdots &\cdots &o_{\tilde{X}_c}(-\tilde{Y_c}) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots & o_{\tilde{X}_c}(-\tilde{Y_c}) \\ o_{\tilde{X}_c} &\cdots &\cdots &\cdots &o_{\tilde{X}_c}(-\tilde{Y_c}) \end{pmatrix}$$ in $\Ascr$. Now we compute the exceptional curve. By definition this will correspond to $\pi(\Uscr/m_0\Uscr)$. A quick computation reveals that $$\Uscr/m_o\Uscr\cong (\gr R\,\, U\!\gr R(-1)\,\,\cdots\,\, U\!\gr R(-1))$$ where $\gr R$ is the associated graded ring for the $m$-adic filtration on $R$. Let $\Oscr_{L_c}$ be the exceptional curve in $\tilde{X}_c$. Then we find that the exceptional curve in $\tilde{X}$ is given by $$\label{ref:6.39a} \begin{split} \Oscr_L&=(\Oscr_{L_c} \,\,\Oscr_{L_c}(-\tilde{Y}_c)\,\,\cdots\,\, \Oscr_{L_c}(-\tilde{Y}_c))\\ &=(\Oscr_{L_c} \,\,\Oscr_{L_c}(-p'_c)\,\,\cdots\,\, \Oscr_{L_c}(-p'_c)) \end{split}$$ where $p'_c=L_c\cap \tilde{Y}_c$. Tensoring with positive and negative powers of $\Jscr$ we find that $\Qch(L)$ is contained in the category of modules over $$\Bscr=\begin{pmatrix} o_{L_c} &o_{L_c}(-p'_c) &\cdots &\cdots &o_{L_c}(-p'_c) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots & o_{L_c}(-p'_c) \\ o_{L_c} &\cdots &\cdots &\cdots &o_{L_c} \end{pmatrix}$$ The inherited shift functor on $\Mod(\Bscr)$ is given by tensoring with the inverse of $$\Iscr= \begin{pmatrix} o_{L_c}(-p'_c) &o_{L_c}(-p'_c) &\cdots &\cdots &o_{L_c}(-p'_c) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots & o_{L_c}(-p'_c) \\ o_{L_c} &\cdots &\cdots &\cdots &o_{L_c}(-p'_c) \end{pmatrix}$$ To interprete this we have to remember that $L_c\cong \PP^1$. So we can view $\Bscr$ as an ordinary sheaf of algebras on $\PP^1$ given by $$\Bscr=\begin{pmatrix} \Oscr_{\PP^1} &\Oscr_{\PP^1}(-1) &\cdots &\cdots &\Oscr_{\PP^1}(-1) \\ \vdots &\ddots &\ddots & &\vdots \\ \vdots & &\ddots &\ddots &\vdots \\ \vdots & & &\ddots & \Oscr_{\PP^1}(-1) \\ \Oscr_{\PP^1} &\cdots &\cdots &\cdots &\Oscr_{\PP^1} \end{pmatrix}$$ and $\Iscr$ as the corresponding sheaf of twosided ideals. With this new point of view, $\Oscr_L$ is given by $$\label{ref:6.40a} \Oscr_L=(\Oscr_{\PP^1} \,\,\Oscr_{\PP^1}(-1)\,\,\cdots\,\, \Oscr_{\PP^1}(-1))$$ From this explicit interpretation it is now easy to verify that every object in $\Mod(\Bscr)$ is a direct limit of subquotients of direct sums of objects like , tensored with powers of $\Iscr$. Hence $L\cong \Spec \Bscr$. What remains to be shown is that $\Spec \Bscr=\Proj S$. One way to accomplish this is as follows. One considers the triple [@AZ] $(\Mod(\Bscr),\Oscr_L,\Iscr^{-1})$ and one verifies that this triple is ample. The corresponding graded ring $$\Gamma_\ast(\Oscr_L)=\sum \Hom(\Oscr_L,\Oscr_L\otimes_{\Bscr} \Iscr^{-n})$$ is equal to $S$. Hence according to [@AZ] : $\Mod(\Bscr)=\operatorname{QGr}(\Bscr)$. This finishes the proof of Proposition \[ref:6.7.1a\] in all cases. \[ref:6.7.3a\] If $\Mscr \in\alpha^{-1}(\Cscr_p)$ then there is a non-trivial map $\Oscr_L(t)\r \Mscr$ for some $t$. Assume that $\Mscr=\pi\Nscr$ where $\Nscr\in\Gr(\Dscr)$ is torsionfree (for $\Tors(\Dscr)$). Let $u$ be such that $\Nscr_u\neq 0$. Since $\Nscr_u\in\Cscr_p$, $\Nscr_u$ will contain some $\Oscr_{\tau^v p}$. Hence there is a non-trivial map $$(\Oscr_{\tau^v p}\otimes_{o_X} \Dscr)(-u)\r \Nscr$$ Since $\Nscr$ is torsion free this yields a non-trivial map. $$\label{ref:6.41a} \alpha^\ast(\Oscr_{\tau^v p})(-u)\r \Mscr$$ If it happens that $\tau^v p=\tau p$ then $\alpha^\ast(\Oscr_{\tau ^v p})=\Oscr_L$ and we are through. Assume $\tau^v p\neq \tau p$. Then according to Proposition \[ref:8.3.2a\] $$\label{ref:6.42a} \alpha^\ast(\Oscr_{\tau^v p}) =\Oscr_{\tau^v \beta^{-1}(p)}$$ There is a surjective map $$\Oscr_L=\alpha^\ast(\Oscr_{\tau p})\r \beta^\ast(\Oscr_{\tau p})=\Oscr_{\tau \beta^{-1}(p)}$$ Twisting yields a surjective map $$\label{ref:6.43a} \Oscr_L(v-1)\r \Oscr_{\tau \beta^{-1}(p)}(v-1) = \Oscr_{\tau^v \beta^{-1}(p)}$$ Combining yields what we want. \[ref:6.7.4a\] Every object in $\alpha^{-1}(\Cscr_p)\cap \coh(\tilde{X})$ is a finite extension of objects in $\mod(L)$. This follows immediately from lemma \[ref:6.7.3a\]. The strict transform {#ref:6.8b} -------------------- This is a more specialized section. We introduce the notion of a strict transform and its influence on the invariant $T_p(\Fscr)$ introduced in §\[ref:5.7b\]. As in §\[ref:5.7b\] we assume that the $\tau$-orbit of $p$ has infinite order. First note some lemmas. \[ref:6.8.1a\] One has $\alpha^{-1}(\Cscr_p)\cap \mod(\tilde{X})\subset \trans_{\tilde{Y}}(\tilde{X})$. This follows for example from Corollary \[ref:6.7.4a\] together with the fact that up to finite length modules every object in $\mod(L)$ is isomorphic to $\Oscr_L^t$ for some $t$ (see [@SmithZhang] or Proposition \[ref:6.7.1a\]). It then suffices to invoke . The functors $\alpha^\ast$, $\alpha_\ast$, restrict to functors between $\trans_Y(X)$ and $\trans_{\tilde{Y}}(\tilde{X})$. That $\alpha^\ast$ preserves transversality follows immediately from $\alpha j=i\beta$. Let us now look at $\alpha_\ast$. Take an object $\Tscr\in \trans_{\tilde{Y}}(\tilde{X})$. We have to show that $i^\ast\alpha_\ast\Tscr$ has finite length. For this it is sufficient to show that $\beta^\ast i^\ast\alpha_\ast \Tscr=j^\ast\alpha^\ast\alpha_\ast\Tscr$ has finite length. Now $\alpha^\ast\alpha_\ast\Tscr$ is isomorphic to $\Tscr$ modulo $\alpha^{-1}(\Cscr_p)\cap \mod(\tilde{X})$ (Theorem \[ref:6.5.2a\]) so it is in $\trans_{\tilde{Y}}(\tilde{X})$ by lemma \[ref:6.8.1a\]. This proves what we want. If $\Fscr\in \Mod(X)$ then we define $$\begin{aligned} \Fscr I^n&=\im (\Fscr\otimes I^{\otimes n}\r \Fscr(nY))\\ \Fscr\cdot \Dscr&=\oplus_n \Fscr I^n\\ \alpha^{-1}(\Fscr)&=\pi(\Fscr\cdot \Dscr)\end{aligned}$$ We will use related notations such as $\Fscr\cdot \Dscr_Y$, $\beta^{-1}(\Fscr)$ when they apply. The idea is that if $\Fscr\in \trans_Y(X)$ then $\alpha^{-1}(\Fscr)$ should correspond to the strict transform of $\Fscr$ on $\tilde{X}$. Unfortunately, unlike in the commutative case $\alpha^{-1}(\Fscr)$ depends on $\Fscr$ itself and not only on the image of $\Fscr$ in $\mod(X)/\Cscr_{f,p}$. Therefore we modify this definition as follows. Assume that $\Fscr\in \trans_Y(X)$ is $Y$-torsion free. Then the strict transform $\alpha_s^{-1}(\Fscr)$ of $\Fscr$ is defined as $(\alpha^{-1}\circ N_p)(\Fscr)$ (see §\[ref:5.7b\] for the definition of $N_p$). \[ref:6.8.4a\] $\alpha^{-1}(\Fscr)$ and $\alpha^{-1}_s(\Fscr)$ are isomorphic to $\alpha^\ast\Fscr$ modulo $\alpha^{-1}(\Cscr_p)$. This follows from the definition of $\alpha^{-1}(\Fscr)$ together with the fact that $\HTor_1^{o_X}(\Fscr,o_X(nY) /I^n)\in\Cscr_p$ by Theorem \[ref:5.5.10a\]. \[ref:6.8.5a\] Identify $\tilde{Y}$ and $Y$ via the map $\beta$. With this identification we have $R=\hat{\Oscr}_{Y,p}=\hat{\Oscr}_{Y,\beta^{-1}(p)}$. As usual let $m$ be the maximal ideal of $R$. Assume that $\Fscr\in\trans_Y(X)$ is $Y$-torsion free. Then we have the following 1. $\alpha_s^{-1}(\Fscr)$ is $\beta^{-1}(p)$-normalized. 2. One has $$\label{ref:6.44a} T_{\beta^{-1}(p)}(\alpha^{-1}_s(\Fscr))=mT_p(\Fscr)$$ Assume that $\Fscr$ is $p$-normalized. The first claim is that the exact sequence coming from the inclusion $o_X(-Y)\r o_X$ $$0\r \Fscr((n-1)Y)\r \Fscr(nY)\r \Fscr_Y (nY)\r 0$$ restricts to an exact sequence (for $n\ge 1$) $$0\r \Fscr I^{n-1} \r \Fscr I^n \r \Fscr_Y I_{Y}^n\r 0$$ Writing out everything this amounts to checking the exactness and well-definedness of the complex $$0\r \Fscr m_p\cdots m_{\tau^{-n+2} p} (-Y)\r \Fscr m_p\cdots m_{\tau^{-n+1} p} \r \Fscr_Y m_{Y,p}\cdots m_{Y,\tau^{-n+1} p} \r 0$$ That this is indeed a complex is clear. To show it is exact it suffices to show that $$\label{ref:6.45a} \begin{split} \operatorname{length}(\Fscr(-Y)/ \Fscr m_p\cdots m_{\tau^{-n+2} p} (-Y))=& \operatorname{length}(\Fscr/\Fscr m_p\cdots m_{\tau^{-n+1} p} )\\ &\quad+ \operatorname{length}(\Fscr_Y /\Fscr_Y m_{Y,p}\cdots m_{Y,\tau^{-n+1} p}) \end{split}$$ Since the finite length modules involved all lie in $\Cscr_p$ we can check this by completing. $\Fscr$ is $p$-normalized so we have $ \hat{\Fscr}_{Y,p}=(\cdots 0, T_p(\Fscr),0\cdots) $ with the non-zero entry occuring in position $0$. From the structure of $\hat{\Fscr}_p$ which is given by lemma \[ref:5.7.4a\]. we compute that $(\Fscr m_p\cdots m_{\tau^{-n+1} p})\hat{}_p=\hat{\Fscr}_p m_0 m_{-1}\cdots m_{-n+1} $ is equal to $$(\cdots T_p(\Fscr) \ mT_p(\Fscr) \cdots mT_p(\Fscr) \ 0\cdots 0\cdots)$$ where now the first $mT_p(\Fscr)$ occurs in location $-n+1$ and the first $0$ in location $1$. In a similar way we find that $(\Fscr_Y m_{Y,p}\cdots m_{Y,\tau^{-n+1} p})\hat{}_p$ is equal to $$(\cdots 0\ m T_p(\Fscr)\ 0\cdots)$$ Now immediately follows. We conclude that we have a complex of graded $\Dscr$-modules $$\label{ref:6.46a} 0\r \Fscr\cdot \Dscr (-1) \r \Fscr\cdot \Dscr \r \Fscr_Y\cdot \Dscr_Y\r$$ exact in degree $\ge 1$. A simple local computation shows that the canonical map $\Fscr_YI_{Y}\otimes_{o_Y} I^{n-1}_{Y}\r\Fscr_YI^n_{Y}$ is an isomorphism and hence, up to right bounded objects, $\Fscr_Y\cdot\Dscr_Y=(\Fscr_Y I_Y\otimes\Dscr_Y)(-1)$. Applying $\pi$ to we find the exact sequence $$0\r \alpha^{-1}(\Fscr)(-\tilde{Y})\r \alpha^{-1}(\Fscr)\r \beta^\ast(\Fscr_Y I_{Y})(-\tilde{Y})\r 0$$ Now $\Fscr_Y I_Y=\Fscr_Y m_p (Y)$ and using that $\beta$ is an isomorphism together with the fact that twisting by $Y$ and $\tilde{Y}$ on finite length modules simply amounts to applying $\tau$ we obtain $\beta^\ast(\Fscr_Y I_Y)(-\tilde{Y})=\beta^\ast(\Fscr_Y m_p)$. This yields $\alpha^{-1}(\Fscr)_{\tilde{Y}}=\beta^\ast(\Fscr_Y m_p)$ which implies what we have to show. \[ref:6.8.6a\] Assume that $\Fscr\in\trans_Y(X)$ is $Y$-torsion free and that $q$ is not in the $\tau$-orbit of $p$ (but that $q$ also has infinite $\tau$-orbit). Then $T_{\beta^{-1}q}(\alpha_s^{-1}(\Fscr))=T_q(\Fscr)$. This is a local verification as in Proposition \[ref:6.8.5a\], but easier. We would also like to understand $\alpha^{-1}(\Fscr)$ if $\Fscr$ is not normalized. Let us denote the quotient functor $\mod(\tilde{X})\r \mod(\tilde{X})/\Cscr_{f}$ by $\eta$. The relevant result is the following. \[ref:6.8.7a\] Assume $\Fscr\in \trans_Y(X)$ is $Y$-torsion free. Then $\eta\alpha^{-1}_s(\Fscr)$ is the minimal subobject of $\eta\alpha^{-1}(\Fscr)$ such that the quotient lies in the image of $\alpha^{-1}(\Cscr_p)$. First note that if $\Fscr'\subset \Fscr$ such that $\Fscr/\Fscr'\in\Cscr_p$ then $\alpha^{-1}(\Fscr)/\alpha^{-1}(\Fscr')\in \alpha^{-1}(\Cscr_p)$ by lemma \[ref:6.8.4a\] and Proposition \[ref:6.5.2a\]. Let $\Fscr_n$ be the $p$-normalization of $\Fscr$ (see §\[ref:5.7b\]). With a similar local verification as in the proof of Proposition \[ref:6.8.5a\] we find that the inclusion $\Fscr_n(-mY)\subset \Fscr_n$, for $m\ge 0$ induces an isomorphism $\alpha^{-1}(\Fscr_n(-mY))\r \alpha^{-1}(\Fscr_n)$ modulo $\Cscr_{f}$. If we take $m$ large enough then we will have an inclusion $\Fscr_n(-mY)\subset \Fscr$. Together with the result of the previous paragraph this yields an inclusion of $\alpha^{-1}(\Fscr_n)\subset \alpha^{-1}(\Fscr)$, if we work modulo $\Cscr_{f}$. Furthermore since $\Fscr_n$ is isomorphic to $\Fscr$ modulo $\Cscr_p$, this inclusion becomes an isomorphism when viewed modulo $\alpha^{-1}(\Cscr_p)$. Now let us assume that $\Fscr$ is $p$-normal and let $\Gscr$ be a subobject of $\alpha^{-1}(\Fscr)$. Put $\Hscr=\alpha_\ast(\Gscr)$. If we choose $m$ large enough then $\Fscr(-mY)\subset \Hscr$ and hence we have inclusions $\alpha^{-1}\Fscr(-mY)\subset \alpha^{-1}(\Hscr)\subset \Gscr \subset \alpha^{-1}(\Fscr)$. Since as above $\alpha^{-1} \Fscr(-mY)= \alpha^{-1}(\Fscr)$ modulo $\Cscr_{f}$, we conclude that $\Gscr=\alpha^{-1}(\Fscr)$ modulo $\Cscr_{f}$. If $\Fscr\in\trans_Y(X)$ then let us write $l_p(\Fscr)$ for the minimal $n$ such that $m^n T_p(\Fscr)=0$. We have the following result. \[ref:6.8.8a\] Let $\Fscr\in \trans_{\tilde{Y}}(\tilde{X})$ be $\tilde{Y}$-torsion free. Then $l_p(\alpha_\ast(\Fscr))\le l_{\beta^{-1}(p)}(\Fscr)+1$. Let $\Gscr$ be the $p$-normalization of $\alpha_\ast(\Fscr)$. By Proposition \[ref:6.8.7a\] we have modulo $\Cscr_f$ an inclusion $\alpha^{-1}(\Gscr)\subset \alpha^{-1}(\alpha_\ast(\Fscr))\subset \Fscr$. Hence we have $l_p(\alpha_\ast(\Fscr))=l_p(\Gscr)\le l_{\beta^{-1}(p)}(\alpha^{-1} (\Gscr))+1\le l_{\beta^{-1}(p)}(\Fscr)+1 $ (the first inequality follows from ). A similar verification yields \[ref:6.8.9a\] Let $\Fscr\in \trans_{\tilde{Y}}(\tilde{X})$ be $\tilde{Y}$-torsion free and assume that $q$ is not in the $\tau$-orbit of $p$ (but that $q$ also has infinite $\tau$-orbit). Then $T_q(\alpha_\ast(\Fscr))=T_{\beta^{-1}(q)}(\Fscr)$. A result on $K_0$ of some categories {#ref:6.9b} ------------------------------------ In this section we use the results on strict transform to prove a technical result which will be used later. We now assume that the $\tau$-orbit of every point on $Y$ has infinite order. Thus in particular $Y$ is smoooth. For a collection of natural numbers $z=(z_o)_{o\in Y/\langle\tau\rangle}$ let us define $\trans_{Y,z}(X)$ as the full subcategory of $\trans_Y(X)$ consisting of objects $\Fscr$ such that $l_q(\Fscr)\le z_{\bar{q}}$ for $q\in Y$. We define $M_z(X)$ as $\trans_{Y,z}(X)/\Cscr_f$. For $o\in Y/\langle\tau \rangle$ let us define $e_o$ by $(\delta_{o,o'})_{o'\in Y/\langle\tau\rangle}$. Our aim is to prove the following theorem \[ref:6.9.1a\] Let $\tilde{X}$ be the blowup of $X$ in $p$. One has the following relation. $$K_0(M_z(\tilde{X}))\cong \begin{cases} K_0(M_{z+e_{\bar{p}}}(X))\oplus\ZZ &\text{if $z_{\bar{p}}\ge 1 $}\\ K_0(M_{z+e_{\bar{p}}}(X))&\text{if $z_{\bar{p}}=0$} \end{cases}$$ By lemma \[ref:6.8.1a\] $\alpha^{-1}(\Cscr_p)\cap \mod(\tilde{X})$ is contained in $\trans_{\tilde{Y}}(\tilde{X})$. Let us write $\Tscr=(\alpha^{-1}(\Cscr_p)\cap \mod(\tilde{X}))/\Cscr_{f,\beta^{-1}(p)}\subset \trans_{\tilde{Y}}(\tilde{X})/\Cscr_f$ and let us temporarily use the notation ${\bar M}_z(\tilde{X})=M_z(\tilde{X})/ (\Tscr\cap M_z(\tilde{X}))$. We first show that $\alpha^{-1}$, $\alpha_s^\ast$ define inverse equivalence between $M_{z+e_{\bar{p}}}(X)$ and ${\bar M}_z(\tilde{X})$. We know already that $\alpha^\ast$ and $\alpha_\ast$ define inverse equivalences between $\mod(X)/\Cscr_{f,p}$ and $\mod(\tilde{X})/(\alpha^{-1}(\Cscr_p)\cap\mod(\tilde{X}))$. Since $\alpha_s^{-1}$ and $\alpha^\ast$ take the same values modulo $\alpha_s^{-1}(\Cscr_p)$, it follows that $\alpha_s^{-1}$ and $\alpha_\ast$ also define inverse equivalences. Furthermore it follows from Propositions \[ref:6.8.5a\],\[ref:6.8.6a\],\[ref:6.8.8a\],\[ref:6.8.9a\] that $\alpha^{-1}$ and $\alpha_\ast$ restrict to equivalences between $M_{z+e_{\bar{p}}}(X)$ and ${\bar M}_z(\tilde{X})$. We now compute the $K_0(\bar{M}_z(\tilde{X}))$ by the localization sequence. $$\label{ref:6.47a} K_0(\Tscr\cap M_z(\tilde{X}))\r K_0(M_z(\tilde{X}))\r K_0(\bar{M}_z(\tilde{X}))\r 0$$ If $z_{\bar{p}}=0$ then $\Tscr\cap M_z(\tilde{X})=0$ so in fact $M_z(\tilde{X})=\bar{M}_z(\tilde{X})$ and we have the corresponding equality on $K_0$-groups. This proves what we need in the case $z_{\bar{p}}=0$. Let us now consider the case $z_{\bar{p}}\ge 1$. In that case $\mod(L)/\Cscr_{f,\beta^{-1}(p)}\subset M_z(\tilde{X})$ and hence the exact sequence becomes $$K_0(\mod(L)/\Cscr_{f,\beta^{-1}(p)})\xrightarrow{\delta} K_0(M_z(\tilde{X}))\r K_0(\bar{M}_z(\tilde{X}))\r 0$$ where $\delta$ is the natural map. From Proposition \[ref:6.7.1a\] it easily follows that $K_0(\mod(\tilde{X})/\Cscr_{f,\beta^{-1}(p)})=\ZZ$, so it remains to show that $\delta$ is injective. For $\Fscr\in \trans_{\tilde{Y}}(\tilde{X})$ let $t(\Fscr)$ be the degree of $\Div(\Fscr)$ (§\[ref:5.7b\]). Clearly $t$ is additive on short exact sequences and furthermore it factors through $\trans_{\tilde{Y}}(\tilde{X})/ \Cscr_{f,\beta^{-1}(p)}$. The generator of $K_0(\mod(L)/\Cscr_{f,\beta^{-1}(p)})$ is $[\bar{\Oscr}_L]$ and since obviously $t(\bar{\Oscr}_L)=1$ we conclude that the image of $[\bar{\Oscr}_L]$ cannot be zero under $\alpha$. Hence $\alpha$ must be injective. We conclude that if $z_{\bar{p}}\ge 1$ then $K_0(M_z(\tilde{X}))=K_0(\bar{M}_z(\tilde{X}))\oplus \ZZ$. This finishes the proof. Derived categories {#ref:7a} ================== Generalities {#ref:7.1a} ------------ If $Z$ is a quasi-scheme then in the sequel we use the notation $D^\ast(Z)$ with $\ast=\phi,+,-,b$ for the standard derived categories of $\Qch(Z)$. If $Z$ is noetherian then we use $D_f^\ast(Z)$ for the full subcategories of $D^\ast(Z)$ whose objects have coherent cohomology. If $\alpha:Y\r X$ is a map between quasi-schemes then since $\Qch(Y)$ has enough injectives it is trivial to define $R\alpha_\ast$. This is unfortunately not the case for $L\alpha^\ast$. We will say that $\Qch(X)$ has enough acyclic objects for $\alpha^\ast$ if every object in $\Qch(X)$ is a quotient of an object $\Uscr$ in $\Qch(X)$ such that $L_i\alpha^\ast \Uscr=0$ for $i>0$ (cfr §\[ref:3.10b\]). The following lemma is easy (see [@RD]). \[ref:7.1.1a\] Assume that $\Qch(X)$ has enough acyclic objects for $\alpha^\ast$. Then $L\alpha^\ast$ exists on $D^-(X)$ and can be computed by acyclic resolutions. If $\alpha^\ast$ has finite cohomological dimension then the same is true with $D(X)$ replacing $D^-(X)$. Assume that $i:Y\r X$ makes $Y$ into a divisor in $X$ (§\[ref:3.7b\]) and denote by $i^!$ the functor $\HHom(o_Y,-)$. From the resolution of $o_Y$ by two invertible bimodules $$\label{ref:7.1b} 0\r o_X(-Y)\r o_X \r o_Y\r 0$$ it is clear that $i^!$ has cohomological dimension one. We will need the following lemma. Assume that $\Qch(X)$ has enough $Y$-torsion free objects. Then for $\Fscr\in \coh(X)$ we have $$\label{ref:7.2a} Ri^!\Fscr=(Li^\ast\Fscr)(Y)[-1]$$ For $\Fscr$ a torsion free object in $\Qch(X)$ this follows immediately by applying $\Hom(-,\Fscr)$ to . The general case follows from [@RD Prop. 7.4]. Admissible compositions of maps between quasi-schemes ----------------------------------------------------- In this section we point out a few problems with compositions of maps between quasi-schemes which have no equivalent in the commutative case. These problems will be solved in all concrete cases, but they nevertheless represent a nuissance. Assume that we have a composition of maps between quasi-schemes. $$\label{ref:7.3a} Z\xrightarrow{\beta} Y\xrightarrow{\alpha} X$$ From the very definition of a morphism it is clear that $\alpha_\ast\beta_\ast=(\alpha\beta)_\ast$. However in contrast with the commutative case there is no reason why the natural map $$\label{ref:7.4a} R(\alpha\beta)_\ast\rightarrow R\alpha_\ast\,R\beta_\ast$$ should be an isomorphism (for trivial reasons, see the discussion after lemma \[ref:7.2.7a\] below). This leads to the following definition. \[ref:7.2.1a\] A composition $(\alpha,\beta)$ as in is admissible if the identity holds (as functors from $D^+(Z)$ to $D^+(X)$). The following lemma is trivial. \[ref:7.2.2a\] The composition $(\alpha,\beta)$ is admissible if and only if for every injective $E\in\Qch(Z)$ one has that $\alpha_\ast E$ is acyclic for $\beta_\ast$. For simplicity we will use some variations on Definition \[ref:7.2.1a\]. If we work in $\Qsch/X$ (i.e. somewhat imprecisely : “if $X$ is a fixed base quasi-schema”), then we will say that the map $\alpha$ is admissible if holds. Similarly if $\alpha:(Y,\Oscr_Y)\r (X,\Oscr_X)$ is a map of enriched quasi-schemes then we will say that $\alpha$ is admissible if $$R\Gamma(X,-)\circ R\alpha_\ast=R\Gamma(Y,-)$$ These conventions are to a certain extent compatible as seen by the following lemma. Assume that we have maps $$Z\xrightarrow{\beta} Y\xrightarrow{\alpha} \Spec R$$ Put $\Oscr_Y=\alpha^\ast R$, $\Oscr_Z=\beta^\ast \Oscr_Y$. Then the compositon $(\alpha,\beta)$ is admissible if and only if the induced map of enriched quasi-schemes $\alpha:(Z,\Oscr_Z)\r (Y,\Oscr_Y)$ is admissible. Below we will give a few adhoc criteria which will allow us to show that certain maps/compositions are admissible. They are all tautologies. \[ref:7.2.4a\] Assume that we have a composition as in , but this time assume that we are working in $\Qsch/W$ for some base quasi-scheme $W$. If the maps $\alpha$, $\beta$ and the composition $(\alpha,\beta)$ are admissible then so is the map $\alpha\beta$. \[ref:7.2.5a\] Assume that $\alpha:(Y,\Oscr_Y)\r (X,\Oscr_X)$ is a map of enriched quasi-schemes. Then $\alpha$ is admissible if and only if $L_i\alpha^\ast\Oscr_X=0$ for $i>0$. \[ref:7.2.6a\] Let $(X,\Oscr_X)$ be an enriched quasi-scheme. Assume that $\Ascr\in \Alg(X)$. Then $\Spec \Ascr \r X$ is admissible if and only if $\HTor^i_{o_X}(\Oscr_X,\Ascr)=0$ for $i>0$. \[ref:7.2.7a\] Assume that $\alpha:(Y,\Oscr_Y)\r (X,\Oscr_X)$ is a map of enriched quasi-schemes which embeds $Y$ as a divisor in $X$ (§\[ref:3.7b\]). Then $\alpha$ is admisible. This is a special case of the previous corollary, since $Y=\Spec o_Y$, where we consider $o_Y$ as an $o_X$ algebra. This corollary indicates how to make a non-admissible map. Take for example commutative schemes $Y,X$, $Y$ being a Cartier divisor in $X$ and change the structure sheaf of $X$ into one which has $Y$-torsion. Below quasi-schemes are often defined as $\Proj$’s of algebras. The following lemma is the obvious analogue of Corollary \[ref:7.2.6a\]. \[ref:7.2.8a\] Let $(X,\Oscr_X)$ be an enriched quasi-scheme. Assume that $\Ascr$ is a noetherian graded algebra on $X$. Then $\Proj \Ascr \r X$ is admissible if $\HTor^{o_X}_i(\Oscr_X,\Ascr)=0$ is right bounded for $i>0$. \[ref:7.2.9a\] Let $f:\Ascr\r\Bscr$ be as in Proposition \[ref:3.9.10a\]. Let $\bar{f}:\Proj\Bscr\r\Proj\Ascr$ be as in Proposition \[ref:3.9.11a\]. Then $\bar{f}$ defines an admissible morphism of quasi-schemes in $\Qsch/X$. This follows from Proposition \[ref:3.9.12a\]. The following is also standard. Assume that we have quasi-schemes and maps as in . Assume that the composition $(\alpha,\beta)$ is admissible and that the maps $\alpha$, $\beta$, $\alpha\beta$ satisfy the conditions of lemma \[ref:7.1.1a\]. Then we have $L(\alpha\beta)^\ast= L\beta^\ast\,L\alpha^\ast$ (as functors from $D^-(X)$ to $D^-(Z)$). It is easy to see that we have at least a natural transformation $$L\beta^\ast\, L\alpha^\ast\r L(\alpha\beta)^\ast$$ In fact this can be deduced from the very definition of derived functors (see [@RD §5]). Now it is clear that if $E$ runs trough the injectives in $\Qch(Z)$ then $\Hom_{D(Z)}(-,E)$ is a conservative system of functors on $D(Z)$. Let $\Mscr\in D^-(X)$. Then we have by adjunction and admissibility $$\begin{aligned} \Hom_{D(Z)}(L\beta^\ast\, L\alpha^\ast\Mscr,E)&=\Hom_{D(X)} (\Mscr,R\alpha_\ast\, R\beta_\ast E)\\ &=\Hom_{D(X)}(\Mscr,R(\alpha\beta)_\ast E)\\ &=\Hom_{D(Z)}(L(\alpha\beta)^\ast\Mscr,E)\qed\end{aligned}$$ It is easy to see that if we have an admissible composition as in then similar results as the ones presented above remain valid on unbounded derived categories provided suitable functors have finite cohomological dimension. In the sequel we will tacitly use such results, leaving the obvious proofs to the reader. We can now show that some commutative diagrams of quasi-schemes we encountered previously consist of admissible maps and admissible compositions. \[ref:7.2.12a\] All maps and compositions of maps in diagram are admissible. $i$ and $j$ are admissible because they are divisors (see Theorem \[ref:6.3.1a\].8). $\beta$ is admissible since it is a map between commutative schemes. To prove that $\alpha$ is admissible it suffices by lemma \[ref:7.2.8a\] to show that $\HTor_i^{o_X}(\Oscr_X,\Dscr)=0$ for $i>0$. More precisely we have to show that $\HTor_i^{o_X}(\Oscr_X,I^n_p)=0$ for $i>0$. Since $I^n_p\subset o_X(nY)$ and the quotient is in $\tilde{C}_{f,p}$ this follows from the compatibility of $\HTor$ with completion (see Theorem \[ref:5.5.10a\]). \[ref:7.2.13a\] All maps and compositions in diagram are admissible. The admissibility of $(\alpha,u)$ follows from lemma \[ref:7.2.9a\]. The other compositions and maps we leave to the reader. The derived category of a non-commutative blowup {#ref:8a} ================================================ The formalism of semi-orthogonal decompositions {#ref:8.1a} ----------------------------------------------- The material in this section is taken from [@Bondal2]. Let $\Ascr$ be a triangulated category and let $\Bscr$, $\Cscr$ be two strict full triangulated subcategories of $\Ascr$. $(\Bscr,\Cscr)$ is said to be a semi-orthogonal pair if $\Hom_\Ascr(B,C)=0$ for $B\in\Bscr$ and $C\in \Cscr$. Define $$\Bscr^\perp=\{A\in\Ascr\mid \forall B\in\Bscr :\Hom_\Ascr(B,A)=0\}$$ ${}^\perp\Cscr$ is defined similarly. The following result is a slight variation of the statement of [@Bondal2 Lemma 3.1]. \[ref:8.1.1a\] The following statements are equivalent for a semi-orthogonal pair $(\Bscr,\Cscr)$. 1. $\Bscr$ and $\Cscr$ generate $\Ascr$. 2. For every $A\in\Ascr$ there exists a distinguished triangle $B\r A\r C$ with $B\in\Bscr$ and $C\in\Cscr$. 3. $\Cscr=\Bscr^\perp$ and the inclusion functor $i_\ast:\Bscr\r\Ascr$ has a right adjoint $i^!:\Ascr\r \Bscr$. 4. $\Bscr={}^\perp\Cscr$ and the inclusion functor $j_\ast:\Cscr\r \Ascr$ has a left adjoint $j^\ast:\Ascr\r \Cscr$ If one of these conditions holds then the triangles in 2. are unique up to unique isomorphism. They are necessarily of the form $$i^! A\r A \r j^\ast A$$ where the maps are obtained by adjointness from the identity maps $i^!A\r i^! A$ and $j^\ast A \r j^\ast A$. In particular triangles as in 2. are functorial. The notations $(i_,i^!, j_, j^)$ are purely symbolic and shouldn’t be interpreted as direct and inverse images. In fact in the main application below (Theorem \[ref:8.4.1a\]) $i_$ will be given by an inverse image! If a pair $(\Bscr,\Cscr)$ satisfies one of the conditions of the previous lemma then we say that it is a semi-orthogonal decomposition of $\Ascr$. For further reference we note the following diagram of arrows $$\label{ref:8.1b} \Cscr \begin{array}{c} j_\ast\\ \rightarrow\\ \leftarrow \\ j^\ast\end{array} \Ascr \begin{array}{c}\displaystyle i^!\\ \rightarrow\\ \leftarrow \\i_\ast \end{array} \Bscr$$ In the following lemma we give some relations between these arrows. One has : $$\begin{gathered} i^! i_\ast=\Id_\Bscr \\ j^\ast j_\ast=\Id_\Cscr\\ j^\ast i_\ast=0\\ i^! j_\ast=0\\\end{gathered}$$ In the sequel we will slightly extend the meaning of the notion of semi-orthogonality. Assume that we have functors $$\Cscr\xrightarrow{j_\ast} \Ascr \xleftarrow{i_\ast} \Bscr$$ which are fully faithful. Assume furthermore that the essential images of $\Bscr$ and $\Cscr$ in $\Ascr$ are semi-orthogonal in $\Ascr$. Then, if no confusion can arise, we wil also call $(\Bscr,\Cscr)$ a semi-orthogonal pair in $\Ascr$. Similarly for a semi-orthogonal decomposition. Semi-orthogonal decompositions can be constructed starting from a pair of adjoint functors. For an arbitrary functor $F:\Ascr\r \Bscr$ between additive categories let us define $\ker F$ as the full subcategory of $\Ascr$ whose objects satisfy $F(A)=0$. \[ref:8.1.4a\] Assume that we have triangulated categories $\Ascr,\Bscr$ and a pair of adjoint functors $i_\ast:\Bscr\r \Ascr$, $i^!:\Ascr\r \Bscr$ such that $i^! i_\ast=\Id_\Bscr$. Then $i_\ast$ is an embedding of $\Bscr$ in $\Ascr$. The corresponding semi-orthogonal decomposition is given by $(\Bscr,\ker i^!)$. Generalities {#ref:8.2a} ------------ In this section the notations and hypotheses of Section \[ref:6a\] will be in force. Our aim is to give a non-commutative version of a well-known theorem by Orlov [@Orlov] which relates $D(\tilde{X})$ to $D(X)$. To this end we need the adjoint functors $R\alpha_\ast$ and $L\alpha^\ast$. In particular we need that $X$ has enough acyclic objects for $\alpha^\ast$ (by lemma \[ref:7.1.1a\]). Therefore at this point we introduce an extra hypothesis which will always hold in the applications. Let us call an object $\Lscr$ in $\coh(X)$ a line bundle around $Y$ if the map $\Lscr(-Y)\r \Lscr$ is injective and if $\Lscr/\Lscr(-Y)$ is a line bundle on $Y$. Note that $\Oscr_X$ itself is a line bundle on $Y$. We denote the additive category of objects which are direct sums of line bundles on $Y$ by $\Vscr$. Every object in $\Qch(X)$ is a quotient of an object in $\Vscr$. The following lemma is left to the reader. \[ref:8.2.1a\] Assume $\Escr\in \Vscr$. 1. $\HTor_i^{o_X}(\Escr,o_Y)=0$ for $i\neq 0$. 2. $\HTor_i^{o_X}(\Escr,o_q)=0$ for $i\neq 0$. 3. $\HTor_i^{o_X}(\Escr,\Dscr)=0$ for $i\neq 0$. 4. $\HTor_i^{o_X}(\Escr,\Dscr_Y)=0$ for $i\neq 0$. 5. $L_i\alpha^\ast \Escr=0$ for $i\neq 0$. From this lemma together with Prop. \[ref:6.1.2a\] it is easy to see that $-\Lotimes_{o_X} \Dscr$, $-\Lotimes_{o_X} \Dscr_Y$, $R\alpha^\ast$ etc…can be defined in the usual way [@RD] by considering resolutions in $\Vscr$. For further reference we recall the following. \[ref:8.2.2a\] $R\alpha_\ast$ and $L\alpha^\ast$ have finite cohomological dimension and commute with direct sums. The fact that $R\alpha_\ast$ and $L\alpha^\ast$ have finite cohomological dimension is proved in Theorem \[ref:6.3.1a\]. Since $L\alpha^\ast$ is the left adjoint to $R\alpha^\ast$ it is clear that it is compatible with direct sums. Hence let us concentrate on $R\alpha_\ast$. From the discussion in Section \[ref:3.8b\] it follows that $R\alpha_\ast$ is equal to $R\omega(-)_0$ and furthermore for $\Mscr\in\Gr(\Dscr)$ there is a triangle $$\label{ref:8.2b} \Atriangle<1`-1`1; >[\Mscr`R\tau\Mscr`R\omega(\pi\Mscr);``]$$ Thus it is sufficient to show that $R\tau$ commutes with direct sums. According to Proposition \[ref:3.8.4a\] $$\tau \Mscr=\injlim \HHom_{\Dscr}(\Dscr/\Dscr_{\ge n},\Mscr)$$ According to Proposition \[ref:6.1.1a\], $I^n_p$ is a coherent $o_X-o_X$ bimodule §\[ref:3a\] for all $n$. Hence this holds also for $\Dscr/\Dscr_{\ge n}$. But then it is easy to see that $\Dscr/\Dscr_{\ge n}$ is also coherent as $\Dscr-\Dscr$-bimodule. Thus $\HHom_{\Dscr}(\Dscr/\Dscr_{\ge n},-)$ commutes with direct sums and hence so does $\tau$. Now $\tau$ has finite cohomological dimension (by this is the same as $\alpha_\ast$ having finite cohomological dimension). It is now easy to see that the standard way of defining $R\tau$ on $D(X)$ [@RD] is compatible with direct sums. Computation of some derived functors {#ref:8.3a} ------------------------------------ We need the following proposition. \[ref:8.3.1a\] Let $n\ge -1$. 1. $R\alpha_\ast (\Oscr_L(n))= (\Dscr/m_{\tau p}\Dscr)_{n,o_X}$. 2. Assume that $\Escr\in \Vscr$. Then $R\alpha_\ast(( \alpha^\ast \Escr)(n))=\Escr\otimes_{o_X}\tilde{\Dscr}_n$ (cfr. ). Since $\Dscr$ satisfies $\chi$ (Theorem \[ref:6.2.2a\]) it is clear that the proposition is true for $n\gg 0$ (lemma \[ref:3.9.3a\]). Our strategy will now be to use descending induction on $n$. To this end it is convenient to treat the cases $\tau p=p$ and $\tau p\neq p$ separately. In the first case $L$ is divisor in $\tilde{X}$ isomorphic to $\PP^1$ so we can use this. In the second case $p$ is smooth on $Y$ (Thm \[ref:5.1.4a\]) and so $\beta:\tilde{Y}\r Y$ is an isomorphism. It follows from this that a suitable analog of the proposition is trivially true for $Y$. We can then lift this to $\tilde{X}$. In the proof below we will write $\Escr_{\tau p}=\Escr\otimes_{o_X} o_{\tau p}$ and similarly $\Escr_Y=\Escr\otimes_{o_X}o_Y$. $\tau p=p$. By admissibility we have $R\alpha_\ast \Oscr_L(n)=R\gamma_\ast \Oscr_L(n)$ (where $\gamma$ is as in diagram ). Thus 1. follows easily from the corresponding result for $\PP^1$. For 2. we use the exact sequence obtained from $$\label{ref:8.3b} 0\r \alpha^\ast(\Escr(-Y))(1)\r \alpha^\ast \Escr\r \alpha^\ast( \Escr_{\tau p})\r 0$$ Since $\alpha^\ast( \Escr_{\tau p})$ is a direct sum of copies of $\Oscr_L$ we deduce from 1. that $$R^i\alpha_\ast (\alpha^\ast( \Escr_{\tau p}))= \begin{cases} \Escr\otimes_{o_X} (\Dscr/m_p\Dscr)_n &\text{if $i=0$}\\ 0&\text{otherwise} \end{cases}$$ Assume that 2. is true for a certain $n\ge 0$. Since the autoequivalence $-\otimes_{o_X} o_X(Y)$ is compatible with $\alpha_\ast$ and $\alpha^\ast$ ($p$ is a fixed point) 2. will also be true for this $n$ if we replace $\Escr$ by $\Escr(mY)$ for arbitrary $m$. Tensoring with $o_{\tilde X}(n-1)$ and applying the long exact sequence for $R\alpha_\ast$ yields $R^i\alpha_\ast ((\alpha^\ast\Escr)(n-1))=0$ for $i>0$ and for $i=0$ we obtain an exact sequence $$\label{ref:8.4a} 0\r \Escr(-Y)\otimes_{o_X} \Dscr_n \r \alpha_\ast(\alpha^\ast (\Escr )(n-1)) \r \Escr\otimes_{o_X} (\Dscr/m_p\Dscr)_n\r 0$$ On the other hand if we tensor the exact sequence on the left with $\Escr$ and take the part of degree $n$ then we obtain an exact sequence $$0\r \Escr(-Y)\otimes_{o_X} \Dscr_n\r \Escr \otimes_{o_X} \Dscr_{n-1} \r \Escr\otimes_{o_X} (\Dscr/m_p\Dscr)_n\r 0$$ Comparing with using the five lemma completes the proof of the proposition in the case $\tau p=p$. $\tau p\neq p$. Now $p$ is smooth on $Y$ (Thm. \[ref:8.4.1a\]) and the map $\beta:\tilde{Y}\r Y$ is an isomorphism. From we obtain an exact sequence $$\label{ref:8.5a} 0\r \alpha^\ast(\Escr)(-1)\r \alpha^\ast(\Escr) \r \beta^\ast(\Escr_Y)\r 0$$ By admissibility $R^i\alpha_\ast ((\beta^\ast\Escr_Y)(n))= R^i\beta_\ast( (\beta^\ast\Escr_Y)(n))$. For $n\ge 0$ we have $$R^i\beta_\ast((\beta^\ast\Escr_Y)(n))= \begin{cases} \Escr_Y\otimes_{o_Y} \Dscr_{Y,n}& \text{if $i=0$}\\ 0&\text{if $i>0$} \end{cases}$$ The case $i=0$ follows from . The case $i>0$ follows from the fact that $\beta$ is an isomorphism. Assume that 2. is true for a certain $n$. Tensor with $o_{\tilde X}(n)$ and apply $R\alpha_\ast$. We obtain $R^i\alpha_\ast ((\alpha^\ast\Escr)(n-1))=0$ for $i\ge 2$ and an exact sequence $$\label{ref:8.6a} 0\r \alpha_\ast((\alpha^\ast\Escr)(n-1)) \r \Escr\otimes_{o_X}\Dscr_n \r \Escr\otimes_{o_X}\Dscr_{Y,n} \r R^1\alpha_\ast((\alpha^\ast\Escr)(n-1)) \r 0$$ Tensoring on the left with $\Escr$ and taking the part of degree $n$ yields an exact sequence $$0\r \Escr\otimes_{o_X}\Dscr_{n-1}\r \Escr\otimes_{o_X} \Dscr_n\r \Escr\otimes_{o_X} \Dscr_{Y,n}\r$$ Comparing with using the five lemma yields 2. in this case. To prove 1. we twist the exact sequence by $o_{\tilde{X}}(n)$ and put $\Escr=\Oscr_X$. Applying $R\alpha_\ast$ then yields what we want. \[ref:8.3.2a\] Assume that $q\in Y$. If $q\neq \tau p$ then $L\alpha^\ast \Oscr_q=\Oscr_{q'}$ with $q'$ such that $\beta(q')=q$ (such a $q'$ is unique!). For $q=\tau p$ we have $$L^i \alpha^\ast \Oscr_{\tau p} = \begin{cases} \Oscr_L &\text{if $i=0$}\\ \Oscr_L(-1)&\text{if $i=1$}\\ 0&\text{otherwise} \end{cases}$$ If $q\not \in O_\tau (p)$ then the result is an easy excercise, for example using the fact that $I^n\subset o_X(nY)$ with quotient in $\tilde{\Cscr}_{f,p}$ (see Prop. \[ref:6.1.1a\]). Hence assume $q\in O_\tau(p)$. We have by lemma \[ref:3.10.2a\] $$L^i\alpha^\ast \Oscr_q=\pi \HTor^i_{o_X}(\Oscr_q,\Dscr)$$ Now $\HTor^i_{o_X}(\Oscr_q,\Dscr)\in\Cscr_p(\Dscr)$ by Propositions \[ref:6.1.1a\] and \[ref:6.1.2a\]. Since “$\HTor$” is compatible with completion (Theorem \[ref:5.5.10a\]) it suffices to compute $\Tor^i_{C_p}(\hat{\Oscr}_q,\hat{\Dscr})$. This is a mildly tedious calculation which we leave to the reader. One way to proceed is to use Proposition \[ref:5.2.2a\] to construct a projective resolution of $\hat{\Oscr}_q$ of length 2 over $C_p$. \[ref:8.3.3a\] 1. $\Hom_{o_{\tilde{X}}}(\Oscr_L,-)$ has finite cohomological dimension. 2. The functor $\RHom_{o_{\tilde{X}}}(\Oscr_L,-)$ is defined on $D(\tilde{X})$. For $\Mscr\in D(X)$ there is a triangle $$\label{ref:8.7a} \Atriangle<1`-1`1; >[R\alpha_\ast(\Mscr(- 1))(Y)` \Oscr_{\tau p}\otimes_k\RHom_{D(\tilde{X})}(\Oscr_L,\Mscr)`R\alpha_\ast\Mscr;``]$$ 3. $\RHom(\Oscr_L,-)$ commutes with direct sums on $D(X)$. 4. $\RHom(\Oscr_L,\Oscr_L)=k$. <!-- --> 1. Assume that $E$ is an injective object in $\Qch(\tilde{X})$. We have $$\Oscr_{\tau p}\otimes_k \Hom_{o_{\tilde{X}}}(\Oscr_L,E)= \HHom_{\Gr(\Dscr)} (\Dscr/m_{\tau p}\Dscr,\omega E)$$ Now we use the exact sequence . From exactness of $\pi$ it follows that $\omega E$ is injective. Furthermore assume that $\Tscr$ is a right bounded $o_X-\Dscr$-module. Then adjointness yields that $\Hom_{o_X}(-,\underline{\HHom}_{\Dscr}(\Tscr,\omega E))$ vanishes. Thus we deduce that $\underline{\HHom}_{\Dscr}(-,\omega E))$ is zero on right bounded $o_X-\Dscr$-bimodules. Hence we obtain from an exact sequence in $\Qch(X)$. $$0\r \HHom_{\Gr(\Dscr)}(\Dscr/m_{\tau p}\Dscr,\omega E)\r (\omega E)_0\r (\omega E)_{-1}(Y)\r 0$$ Which can be rewritten as $$\label{ref:8.8a} 0\r \Oscr_{\tau p}\otimes_k \Hom_{o_{\tilde{X}}}(\Oscr_L,E) \r \alpha_\ast E\r \alpha_\ast (E(-1))(Y)\r 0$$ The fact that $\alpha_\ast$ has finite cohomogical dimension by \[ref:6.3.1a\] yields what we want. 2. The fact that $\RHHom_{o_{\tilde{X}}}(\Oscr_L,-)$ is defined on the whole of $D(X)$ follows from 1. using [@RD]. Now since $\alpha_\ast$ also has finite cohomological dimension, every object in $\Qch(\tilde{X})$ has a resolution by objects which are acyclic for both $\alpha_\ast$ and $\Hom_{o_{\tilde{X}}}(\Oscr_L,-)$ (using the methods of [@RD]). If $F$ is acyclic for $\Hom_{o_{\tilde{X}}}(\Oscr_L,-)$ then it is clear there will be a short exact sequence as in with $E$ replaced by $F$. By considering resolutions with such acyclic objects one finds the triangle . 3. This follows from 1. together with the construction of $\Hom_{o_{\tilde{X}}}(\Oscr_L,-)$ by acyclic resolutions. 4. This follows from substituting $\Mscr=\Oscr_L$ in the triangle and using Proposition \[ref:8.3.1a\]. The following result is well-known. \[ref:8.3.4a\] Let $\Ascr$ be an abelian category. Let $X\in D^(\Ascr)$ with $=+,-$ be an object such that $\Ext^i(H^j(X),H^{j+1-i}(X))=0$ for $i\ge 2$ and for all $j$. Then $X$ is isomorphic to the direct sum of its homology groups. The same is true for $X\in D(\Ascr)$ if in addition $\Ascr$ satisfies AB4, has enough injectives and $\Hom(H^j(X),-)$ has finite cohomological dimension for all $j$. Assume first that $X\in D^+(\Ascr)$. Write $H(X)=\oplus H^i(X)[-i]$. Note that $H(X)$ is the category theoretic direct sum of the $H^i(X)[-i]$ in $D^+(\Ascr)$. We want to construct a quasi-isomorphism $H(X)\r X$. To this end it is sufficient to construct maps $H^i(X)[-i]\r X$ which induce isomorphisms on the $i$’t cohomology. Since $\tau_{\le i} X\r X$ induces an isomorphism on $H^i$, it is clearly sufficient to show that the canonical map $\tau_{\le i}X \r H^i(X)[-i]$ splits. From the triangle $$\tau_{\le i-1} X\r \tau_{\le i} X \r H^i(X)[-i]\r$$ we find that we have to show that $$\label{ref:8.9a} \Hom(H^i(X)[-i],\tau_{\le i-1} X[1])=0$$ Now $\tau_{\le i-1} X$ is a bounded complex and hence follows easily by induction from the hypotheses. The case $X\in D^-(\Ascr)$ is similar. Now assume $X\in D(\Ascr)$. In this case there are two possible problems with the above reasoning. 1. $\oplus_i H^i(X)[-i]$ is perhaps no longer the direct sum of the $H^i(X)[-i]$ in $D(\Ascr)$. 2. $\tau_{\le i-1} X$ is now an unbounded complex so we can no longer verify by induction. The first difficulty is resolved if we assume that $\Ascr$ satisfies AB4 [@Neeman]. For the second difficulty we have to show that $\Hom(H^i(X),-)$ is zero on $\Dscr(\Ascr)_{\le -N}$ for $N\gg 0$. Now according to [@RD Thm 5.1, Cor. 5.3], if $\Ascr$ has enough injectives and $\Hom(H^i(X),-)$ has finite cohomological dimension then we can compute $\Hom(H^i(X),-)=H^0(\RHom(H^i(X),-))$ by acyclic resolutions. It follows easily that if $\cd \Hom(H^i(X),-)=t$ then an object in $\Dscr(\Ascr)_{\le -N}$ can be represented by an acyclic complex which is non-zero only in degree $\le -N+t$. Hence it suffices to take $N>t$. The reader may verify that the statement about $D(\Ascr)$ in the previous lemma is also true under the hypotheses AB4 and AB4${}^\ast$. This is more elegant, but less useful in practice. \[ref:8.3.6a\] Let $\Nscr$ be the additive subcategory of $\Qch(\tilde{X})$ whose objects are direct sums of copies of $\Oscr_L(-1)$. Then $\Nscr$ is a thick subcategory (closed under extensions). Furthermore the map $$D(k)\xrightarrow{-\otimes \Oscr_L(-1)} D_\Nscr(\Qch(X))$$ is an equivalence. This follows from lemma \[ref:8.3.4a\] together with lemma \[ref:8.3.3a\]. The main theorem {#ref:8.4b} ---------------- We prove the following Theorem \[ref:8.4.1a\]. This is a non-commutative version of a theorem by Orlov [@Orlov]. Our proof is slightly different. \[ref:8.4.1a\] There is a semi-orthogonal decomposition of $D(\tilde{X})$ given by $(D(X),D(k))$. The diagram corresponding to is as follows $$\label{ref:8.10a} D(k) \bfig \putmorphism(0,100)(1,0)[``-\otimes_k \Oscr_L(-1)]{1100}{1}{a} \putmorphism(0,-50)(1,0)[``F]{1100}{-1}{b} \efig D(\tilde{X}) \bfig \putmorphism(0,100)(1,0)[``R\alpha_\ast]{300}{1}{a} \putmorphism(0,-50)(1,0)[``L\alpha^\ast]{300}{-1}{b} \efig D(X)$$ where $F$ is the left adjoint to $-\otimes_k\Oscr_L(-1)$ (whose existence follows by lemma \[ref:8.1.1a\]). In we may replace $D$ by $D^\ast$ where $\ast=\emptyset,+,-,b$. Furthermore we may also replace $D^\ast$ everywhere by $D_f^\ast$. On $D_f^-(\tilde{X})$, $F$ is given by $\RHom(-,\Oscr_L(-1))^\ast$. \[ref:1b\] $R\alpha_\ast L\alpha^\ast=\Id$. Since the objects in $\Vscr$ are acyclic for $\alpha^\ast$, this follows from Proposition \[ref:8.3.1a\].2 on $D^-(X)$. The general case is then routine. If $\Fscr\in D(X)$ then we have to show that the adjunction mapping $\Fscr\r R\alpha_\ast L\alpha^\ast\Fscr$ is a quasi-isomorphism, that is $H^i(\Fscr)\r H^i(R\alpha_\ast L\alpha^\ast\Fscr)$ should be an isomorphism for all $i$. Since $\alpha^\ast$, $\alpha_\ast$ have finite cohomological dimension, we can test this (for a fixed $i$) by replacing $\Fscr$ by some $\tau_{\le N}\Fscr$ for $N\gg 0$. But then $\Fscr\in D^-(X)$ and this case was already covered. \[ref:2a\] The composition of $=-\otimes_k \Oscr_L(-1)$ and $R\alpha_\ast$ is zero. This follows from the fact that $R\alpha_\ast$ commutes with direct sums (lemma \[ref:8.2.2a\]) together with $R\alpha_\ast\Oscr_L(-1)=0$, which was proved in Proposition \[ref:8.3.1a\]. \[ref:3b\] At this point $D(k)$ and $D(X)$ form a semi-orthogonal pair in $D(\tilde{X})$. To show that they form a semi-orthogonal decomposition we invoke lemma \[ref:8.1.4a\]. Thus we have to prove that $\ker R\alpha_\ast= D(k)$. We prove first $\ker R\alpha_\ast=D_\Nscr(\tilde{X})$ ($\Nscr$ as in Cor. \[ref:8.3.6a\]). Then by Corollary \[ref:8.3.6a\] we find $D_\Nscr(\tilde{X})=D(k)$. We claim first that $$\ker R\alpha_\ast\subset D_{\alpha^{-1}(\Cscr_p)}(\tilde{X})$$ To prove this assume that $R\alpha_\ast\Mscr=0$. From the spectral sequence $$R^p\alpha_\ast H^q(\Mscr)\r R^{p+q}\alpha_\ast \Mscr$$ we obtain short exact sequences $$0\r R^1\alpha_\ast H^{i-1}(\Mscr)\r R^i\alpha_\ast\Mscr \r R^0\alpha_\ast H^i(\Mscr)\r 0$$ We conclude that for all $i$ one has $\alpha_\ast H^i(\Mscr)=0$, whence by Proposition \[ref:6.5.2a\] $\Mscr\in D_{\alpha^{-1}(\Cscr_p)}(\tilde{X})$. \[ref:4b\] If $\phi:\Mscr\r\Nscr$ is a map in $D(\tilde{X})$ then we denote by $\operatorname{cone}(\phi)$ the cone of the triangle with base $\phi$. $\operatorname{cone}(\phi)$ is unique up to non-unique isomorphism. We claim $$\label{ref:8.11a} \ker R\alpha_\ast=\{ \operatorname{cone}( L\alpha^\ast R\alpha_\ast\Mscr\r \Mscr )\mid \Mscr\in D_{\alpha^{-1}(\Cscr_p)}(\tilde{X}) \}$$ Since $R\alpha_\ast L\alpha^\ast=\Id$ it is clear that $RHS\eqref{ref:8.11a}\subset LHS\eqref{ref:8.11a}$. The opposite inclusion is obvious since if $R\alpha_\ast\Mscr=0$ then according to Step \[ref:3b\] : $\Mscr\in D_{\alpha^{-1}(\Cscr_p)}(\tilde{X})$. Hence $\Mscr=\operatorname{cone}( L\alpha^\ast R\alpha_\ast\Mscr\r\Mscr )\in \operatorname{RHS}\eqref{ref:8.11a}$. \[ref:5b\] The formation of $\operatorname{cone}(L\alpha^\ast R\alpha_\ast \Mscr\r\Mscr)$ is functorial and compatible with shifts and triangles. According to lemma \[ref:8.1.4a\], we have a semi-orthogonal decomposition of $D(\tilde{X})$ given by $(\ker R\alpha_\ast,D(X))$. Since in the triangle $$\Atriangle<1`-1`1; >[\Cscr`L\alpha^\ast R\alpha_\ast\Mscr`\Mscr;``]$$ $L\alpha^\ast R\alpha_\ast\Mscr$ lies in the essential image of $D(X)$ and $\Cscr\in \ker R\alpha_\ast$ the good behaviour of $\operatorname{cone}(L\alpha^\ast R\alpha_\ast\Mscr\r\Mscr)$ is a consequence of lemma \[ref:8.1.1a\]. According to Step \[ref:4b\] we have to show that if $\Mscr\in D_{\alpha^{-1}(\Cscr_p)}(\tilde{X})$ then $\operatorname{cone}(L\alpha^\ast R\alpha_\ast\Mscr\r\Mscr)\in D_\Nscr(\tilde{X})$. That is $$\label{ref:8.12a} H^i(\operatorname{cone}(L\alpha^\ast R\alpha_\ast\Mscr\r\Mscr))\in\Nscr$$ Now since $L\alpha^\ast$, $R\alpha_\ast$ have finite cohomological dimension the truth of will not be influenced by the cohomology of $\Mscr$ in high degree. Hence to verify we may replace $\Mscr$ by a suitable $\tau_{\le a}\tau_{\ge b}\Mscr$, that is, by a bounded complex. According to Step \[ref:5b\] and induction we may then assume that $\Mscr\in \alpha^{-1}(\Cscr_p)$. Now is a statement about homology of complexes, and it is easy to see that this homology is compatible with direct limits, when $\Mscr\in\Qch(\tilde{X})$. Hence we may assume that $\Mscr\in \alpha^{-1}(\Cscr_p)\cap \coh(\tilde{X})$. Now from lemma \[ref:8.4.2a\] below it will follow that we may in fact assume $\Mscr=\Oscr_L$, $\Oscr_L(-1)$ or $\Oscr_q$ with $q\in \beta^{-1}(O_\tau(p))\setminus \beta^{-1}(\tau p)$. The last two cases are trivial since by Propositions \[ref:8.3.1a\] and \[ref:8.3.2a\] we find that $L\alpha^\ast R\alpha_\ast \Mscr=\Mscr$. Hence assume $\Mscr=\Oscr_L$. Thus we have to compute $$\label{ref:8.13a} \operatorname{cone}(L\alpha^\ast R\alpha_\ast \Oscr_L\r \Oscr_L)=\operatorname{cone}(L\alpha^\ast \Oscr_{\tau p}\r \Oscr_L)$$ Since by Proposition \[ref:8.3.2a\] and $\alpha^\ast\Oscr_{\tau p}=\Oscr_L$, and $\alpha^\ast$ has cohomological dimension one we find that, up to shift, is equal to $L_1\alpha^\ast \Oscr_{\tau p}$, which according to Proposition \[ref:8.3.2a\] is equal to $\Oscr_L(-1)$. Odd and ends. First of all since $R\alpha_\ast$ and $L\alpha^\ast$ have finite cohomological dimension and preserve coherent objects (by Theorem \[ref:6.3.1a\]), it is clear that the above reasoning may be repeated for $D^\ast$ and for $D_f^\ast$ with $\ast=\phi,+,-,b$. To show that $F$ on $D^-_f(\tilde{X})$ is equal to $\RHom(-,\Oscr_L(-1))^\ast$ we have to show that this functor is well-defined and is a left adjoint to $-\otimes \Oscr_L(-1)$. A quick verification shows that we have to check that $\coh(\tilde{X})$ contains enough objects $\Fscr$ such that the homology of $\RHom_{o_X}(\Fscr,\Oscr_L(-1))$ is finite dimensional and concentrated in degree zero. For $\Fscr$ we take objects of the form $\alpha^\ast(\Escr)(-n-1)$ with $n\in\NN$ and $\Escr\in \Vscr\cap \coh(X)$. It is easy to see that there are enough of those. We compute $$\begin{aligned} \RHom_{o_{\tilde{X}}}(\alpha^\ast(\Escr)(-n-1),\Oscr_L(-1)) &=\RHom_{o_X}(\Escr,R\alpha_\ast(\Oscr_L(n)))\\ &=\RHom_{o_X}(\Escr,(\Dscr/m_{\tau p}\Dscr)_{n,o_X})\end{aligned}$$ It now follows for example from Proposition \[ref:5.1.2a\] that the homology of the last line of the previous equation has the properties we want. \[ref:8.4.2a\] $\alpha^{-1}(\Cscr_p)\cap \coh(\tilde{X})$ is the smallest subcategory of $\Qch(\tilde{X})$ containing $\Oscr_L(-1)$, $\Oscr_L$ and $\Oscr_q$ with $q\in \beta^{-1}(O_\tau(p))\setminus \beta^{-1}(\tau p)$, and which is closed under 1. extensions; 2. kernels of surjective maps; 3. cokernels of injective maps. By Corollary \[ref:6.7.4a\] every object in $\alpha^{-1}(\Cscr_p)\cap \coh(\tilde{X})$ is an extension of objects in $\Qch(L)\cap \coh(\tilde{X})=\coh(L)$. Hence we have to verify the statement for $\coh(L)$. But then we may as well verify the corresponding statement for $S$. This is a routine exercise, which we leave to the reader. \[ref:8.4.3a\] If $\Qch(X)$ has finite injective dimension then so does $\Qch(\tilde{X})$. Assume that $\Qch(X)$ has finite injective dimension. We have to show that there exists an $N$ such that $\Hom_{D(\tilde{X})}(\Mscr,\Nscr)=0$ for all $\Mscr\in D^b_{\ge N}(\tilde{X})$ and $\Nscr\in D^b_{\le 0}(\tilde{X})$. For an arbitrary morphism $\phi:\Mscr\r \Nscr$ we have a commutative diagram of triangles $$\begin{CD} L\alpha^\ast R\alpha_\ast \Mscr @>>> \Mscr @>>> F\Mscr \otimes_k \Oscr_L(-1)@>>>\\ @V L\alpha^\ast R\alpha_\ast(\phi) VV @V\phi VV @V F\phi\otimes_k \Oscr_L(-1) VV\\ L\alpha^\ast R\alpha_\ast \Nscr @>>> \Nscr @>>> F\Nscr\otimes_k \Oscr_L(-1) @>>> \end{CD}$$ Since $L\alpha^\ast$ is fully faithful and $\Qch(X)$ has finite injective dimension we find that $L\alpha^\ast R\alpha_\ast(\phi)=0$ if $N$ is large enough. Similarly $D^b(k)$ is a semi-simple category and hence $F\phi=0$ if $N$ is large enough. This then implies that $\phi$ is in fact obtained from a map $\theta:F\Mscr\otimes_k \Oscr_L(-1)\r L\alpha^\ast R\alpha_\ast \Nscr $. However by lemma \[ref:8.3.3a\] we find that $\Hom_{D(\tilde{X})}(\Oscr_L,-)$ has finite cohomological dimension. Hence by making $N$ even larger if necessary we also find that $\theta=0$. This finishes the proof. Some results on graded algebras and their sections {#ref:9a} ================================================== Generalities {#ref:9.1a} ------------ In this section $\gamma:X\r \Spec k$ will be a noetherian quasi-scheme over $k$ which is proper (see Def. \[ref:3.9.4a\]). As usual we put $\Oscr_X=\gamma^\ast k$. We will also assume that $\Oscr_X$ is noetherian. By adjointness we have $\gamma_\ast \Mscr=\Hom_{\Qch(X)}(\Oscr_X,\Mscr)=\Gamma(X,\Mscr)$. Similarly for the derived functors. Hence the properness of $X$ simply means that if $\Mscr\in \coh(X)$ then $H^i(X,\Mscr)$ is a finite dimensional $k$-vector space for all $i$. The fact that $\Oscr_X$ is noetherian implies that $H^i(X,-)$ commutes with direct limits. Assume that $\Escr$ is a noetherian $\NN$-graded algebra on $X$ and let $E=\Gamma(X,\Escr)$ (cfr §\[ref:3.6b\]). Then $(\Escr)_{o_X}$ is an $E-\Escr$-bimodule, in the obvious sense. It is clear that we have adjoint functors $$\label{ref:9.1b} \begin{split} -\otimes_E \Escr_{o_X}:&\Gr(E)\r \Gr(\Escr) \\ \Gamma(X,-):&\Gr(\Escr)\r \Gr (E) \end{split}$$ Our aim will be to relate $\operatorname{QGr}(\Escr)$ to $\operatorname{QGr}(E)$. To this end we introduce the following definition. $\Escr$ is ample if for every object $\Mscr$ in $\coh(X)$ we have for $n\gg 0$ that $\Mscr\otimes_{o_X} \Escr_n$ is generated by global sections and $H^i(X,\Mscr\otimes_{o_X} \Escr_n)=0$ for $i>0$. The following proposition was essentially proved in [@AVdB; @VdB11] (see [@VdB11 Thm 5.2]) under more restrictive hypotheses. \[ref:9.1.2a\] Let $\Escr$ be as above and assume that $\Gamma(X,-)$ has finite cohomological dimension. If $\Escr$ is ample then $E$ is noetherian. The functors in send $\gr(E)$ to $\gr(\Escr)$ and vice-versa. Furthermore these functors define inverse equivalences between $\operatorname{QGr}(E)$ and $\operatorname{QGr}(\Escr)$. Unfortunately ampleness of $\Escr$ is too strong for the applications we have in mind. However we will show that under some extra conditions the functors may still be well behaved, even if $\Escr$ is non-ample. We introduce the following hypotheses. $\Gamma(X,-)$ has finite cohomolical dimension and furthermore there is an injective graded $\Escr$-bimodule map $t:\Escr(-1)\r \Escr$ such that $\Escr/\im t$ is ample. Furthermore the induced map $\Escr_{o_X}(-1)\r \Escr_{o_X}$ is injective. Taking global sections yields a regular central element of $E$ in degree one which we also denote by $t$. We write $\bar{\Escr}=\Escr/\im t$. If $\Nscr\in\Gr (\Escr)$ then $t$ defines a map $\Nscr(-1)\r \Nscr$ in $\Gr(\Escr)$. We put $\bar{\Nscr}=\Nscr/\im t$. We say that $\Nscr$ is annihilated by $t$ if $t:\Nscr(-1)\r \Nscr$ is the zero map. We say that $\Nscr$ is $t$-torsion if $\Nscr$ is a union of objects which are each annihilated by some $t^n$ ($n$ variable). We say that $\Nscr$ is $t$-torsion free if $t$ is injective. Similar conventions apply to $\operatorname{QGr}(\Escr)$, $\Gr(E)$ and $\Qgr(E)$. \[ref:9.1.3a\] Let $\Nscr\in \gr(\Escr)$. Then 1. Multiplication by $t$ on $H^i(X,\Nscr)$ is an automorphism in high degree when $i>0$. 2. The complex $$\Gamma(X,\Nscr(-1))\xrightarrow{t} \Gamma(X,\Nscr)\r \Gamma(X,\bar{\Nscr})\r 0$$ is exact in high degree. 3. $H^i(X,\Nscr)$ is finitely generated for all $i$. Furthermore we also have 4. $\Tor^E_i(-,\Escr)$ sends $\Tors(E)$ to $\Tors(\Escr)$ for all $i$. 5. $E$ is noetherian. - Using the appropriate long exact sequences it follows that it is sufficient to prove this in the case that $\Nscr$ is annihilated by $t$ and in the case that $\Nscr$ is $t$-torsion free. In the first case we have that $\Nscr=\bar{\Nscr}$ and furthermore, for $i>0$, $H^i(X,\bar{\Nscr})$ is right bounded (by ampleness of $\bar{\Escr}$). Hence 1.,2. are trivially true in this case. Let us consider the second case. We apply the long exact sequence for $\Gamma(X,-)$ to $$0\r \Nscr(-1)\r \Nscr \r \bar{\Nscr}\r 0$$ Since $H^i(X,\bar{\Nscr})$ is right bounded, we immediately obtain the statement about $H^i(X,\Nscr)$ for $i>1$. So let us assume $i\le 1$. For $n\gg 0$ we have an exact sequence $$0\r \Gamma(X,\Nscr_{n-1})\xrightarrow{t} \Gamma(X,\Nscr_n)\r \Gamma(\bar{\Nscr}_n)\r H^1(X,\Nscr_{n-1})\xrightarrow{t} H^1(X,\Nscr_n)\r 0$$ So $\dim H^1(X,\Nscr_n)$ is descending and hence must eventually become constant. Thus $t:H^1(X,\Nscr_{n-1})\r H^1(X,\Nscr_n)$ is an isomorphism for $n\gg 0$. This proves 1.,2. for $i=0,1$. - In view of 1., the only non-trivial case is $i=0$. By Proposition \[ref:9.1.2a\] $\Gamma(X,\bar{\Nscr})$ is finitely generated. Since according to 2., $\overline{\Gamma(X,\Nscr)}$ has finite colength in $\Gamma(X,\bar{\Nscr)}$ we find that $\overline{\Gamma(X,\Nscr)}$ is finitely generated. Hence by the graded version of Nakayama’s lemma, $\Gamma(X,\Nscr)$ is also finitely generated. - Since $\Tor$ is compatible with direct limits we may prove the corresponding statement for “$\tors$”. Furthermore, using the appropriate long exact sequences, it follows that it is sufficient to prove that $\Tor_i^E(V,\Escr)$ is right bounded for a finite dimensional $E_0$-module $V$ (considered as $E$-module). Since $V$ is annihilated by $t$, we have as usual $$\label{ref:9.2a} \Tor_i^E(V,\Escr)=\Tor_i^{\bar{E}}(V,\bar{\Escr})$$ Put $\bar{E}'=\Gamma(X,\bar{\Escr})$. According to 2., applied to $\Escr_{o_X}$ we find that the map $\bar{E}'\r \bar{E}$ is injective and has finite dimensional cokernel. Hence by [@AZ Prop. 2.5] and Proposition \[ref:9.1.2a\] one has $\operatorname{QGr}(\bar{E}')=\operatorname{QGr}(\bar{E})=\operatorname{QGr}(\bar{\Escr})$. The functor realizing this equivalence is given by $-\otimes_{\bar{E}'} \bar{\Escr}$. Hence this is in particular an exact functor. Now to compute the righthand side of we take a free resolution of $V$. Since the grading on $V$ is right bounded this free resolution is an exact sequence in $\operatorname{QGr}(\bar{E}')$. Hence it remains exact after applying $-\otimes_{\bar{E}'} \bar{\Escr}$. In this way we obtain that $\Tor_i^E(V,\Escr)\in \Tors(\Escr)$ for all $i$. - By Proposition \[ref:9.1.2a\] it follows that $\bar{E}'$ is noetherian. Since the map $\bar{E}\r \bar{E}'$ is injective and has finite dimensional cokernel, we deduce that $\bar{E}$ is also noetherian. Whence, by a Hilbert-type argument, $E$ is noetherian. For use below we write $W=\Proj \Escr$, $T=\Proj \bar{\Escr}$, $V=\Proj E$, $S=\Proj \bar{E}$. It follows that $S,T$ are S divisors in $W,V$. As usual, we will denote the quotient functors $\Gr(E)\r \operatorname{QGr}(E)$, $\Gr(\Escr)\r \operatorname{QGr}(\Escr)$ by $\pi$. The functors $$\begin{aligned} \delta^\ast:&\Qch(V)\r \Qch(W): \pi M\mapsto \pi(M\otimes_E \Escr)\\ \delta_\ast:&\Qch(W)\r \Qch(V):\pi N\mapsto \pi\Gamma(X,N) \end{aligned}$$ are well defined and form an adjoint pair. In particular $(\delta^\ast,\delta_\ast)$ defines an morphism of quasi-schemes $$\delta:W\r V$$ $\delta$ fits in a commutative diagram of quasi-schemes $$\label{ref:9.3a} \begin{CD} T @>j>> W\\ @V\delta VV @V\delta VV\\ S @>i>> V \end{CD}$$ Here $i,j$ are the inclusion mappings. $\delta:T\r S$ is an isomorphism. If $\Mscr\in \Qch(T)$ then we have $$\label{ref:9.4a} \delta_\ast(\Mscr\otimes_{o_T} \Nscr_{T/W})=\delta_\ast(\Mscr)\otimes_{o_S} \Nscr_{S/V}$$ Thus the normal bundles on $T$ and $S$ correspond to each other under the isomorphism $\delta$. It is obvious that $\delta_\ast$ is well-defined. The fact that $\delta^\ast$ is well-defined follows easily from Proposition \[ref:9.1.3a\].4. We now show that $\delta_\ast$, $\delta^\ast$ are adjoint functors. This means that we have to construct a natural isomorphism between $$\Hom_{\operatorname{QGr}(E)}(\Mscr,\delta_\ast\Nscr)$$ and $$\Hom_{\operatorname{QGr}(\Escr)}(\delta^\ast\Mscr,\Nscr)$$ where $\Mscr=\pi M$, $\Nscr=\pi N$. By taking a presentation of $M$ we reduce to $M=E$. We have $$\begin{aligned} \Hom_{\operatorname{QGr}(E)}(E,\delta_\ast \Nscr)&=\dirlim_n \Hom_{\Gr(E)} (E_{\ge n}, \Gamma(X,N))\\ &=\dirlim_n \Hom_{\Gr(\Escr)}(E_{\ge n}\otimes_E \Escr_{o_X},N)\end{aligned}$$ and we have to show that this is equal to $$\dirlim_{n}\Hom_{\Gr(\Escr)}((\Escr_{o_X})_{\ge n},N)$$ Put $$\begin{aligned} K_n&=\ker (E_{\ge n}\otimes_E \Escr_{o_X}\r (\Escr_{o_X})_{\ge n})\\ C_n&=\coker (E_{\ge n}\otimes_E \Escr_{o_X}\r (\Escr_{o_X})_{\ge n})\end{aligned}$$ It is now sufficient to show that $K_n,C_n$ are torsion inverse systems. That is for all $n$ there exists an $m$ such that the maps $K_{n+m}\r K_n$, $C_{n+m}\r C_n$ are zero. Tensoring the exact sequence $$0\r E_{\ge n}\r E\r E/E_{\ge n} \r 0$$ with $\Escr_{o_X}$ and restricting to degrees $\ge n$ yields that $$\begin{aligned} K_n&=\Tor_1^E(E/E_{\ge n}, \Escr_{o_X})\\ C_n&=(E/E_{\ge n}\otimes_E \Escr_{o_X})_{\ge n}\end{aligned}$$ In particular it follows from Proposition \[ref:9.1.3a\].4 that $K_n$, $C_n$ are right bounded. Furthermore it is clear that $C_n$ is zero in degrees $< n$, and by considering a minimal free resolution of $E/E_{\ge n}$ we see that the same holds for $K_n$. It now follows clearly that $K_n$, $C_n$ are torsion inverse systems. That is a commutative diagram is obvious and the fact that $\delta:T\r S$ is an isomorphism is simply the ampleness of $\bar{\Escr}$. The identity is just a special case of the fact that $\delta_\ast$ commutes with shift. From Proposition \[ref:9.1.3a\] one easily deduces that the shifts of $\pi E$ are acyclic for $\delta^\ast$ (in the sense of [@RD]). There are clearly enough of those, so $L\delta^\ast$ exists. It is easy to verify the following formulas $$\label{ref:9.5a} \begin{split} R^i\delta_\ast \pi N&= \pi H^i(X,N)\\ L_i\delta^\ast \pi M&= \pi \Tor^E_i(M,\Escr) \end{split}$$ $R\delta_\ast$ sends $D^+_f(W)$ to $D^+_f(V)$ and $L\delta^\ast$ sends $D^-_f(V)$ to $D^-_f(W)$. The statement about $L\delta^\ast$ follows from the fact that $E$ and $\Escr$ are noetherian. The statement about $R\delta_\ast$ follows from Proposition \[ref:9.1.3a\].3. Thus in particular $\delta$ is proper in the sense of Definition \[ref:3.9.4a\]. The following is also easy. \[ref:9.1.6a\] If $\Nscr\in \Tors_T(W)$ then $R^i\delta_\ast\Nscr=0$ for $i>0$. Similary if $\Mscr\in \Tors_S(V)$ then $L^i\delta^\ast \Mscr=0$ for $i>0$. Finally $\delta_\ast$, $\delta^\ast$ define inverse equivalences between $\Tors_T(W)$ and $\Tors_S(V)$ (see §\[ref:3.7b\] for notation). Since all functors involved commute with direct limits we may assume that $\Mscr$, $\Nscr$ are coherent. The first two statements are easily seen to be true for objects annihilated by $t$. The general case follows from this by considering the appropriate long exact sequences. In particular we obtain that $\delta_\ast$ and $\delta^\ast$ are exact. We now have to show that the adjunction morphisms are isomorphisms on coherent objects. Again this is true for objects annihilated by $t$ by the hypotheses on $\bar{\Escr}$. The general case now follows by filtering objects in such a way that the associated graded quotients are annihilated by $t$. The following proposition will be true in all our applications for trivial reasons. However it may be interesting to note that one can prove it in this generality. $\Qch(W)$ has enough $T$-torsion free objects. In particular $Lj^\ast$ is defined. It suffices to show that every object in $\mod(W)$ is a quotient of a $T$-torsion free one. Let $\Mscr=\pi M$ where $M\in\gr \Escr$. With an argument as in the proof of lemma \[ref:3.7.3a\] one finds a $M_0\subset M$ which is $t$-torsion free such that $M/M_0$ is $t$-torsion. Assume $t^n(M/M_0)=0$ and consider the following diagram with exact rows. $$\label{ref:9.6a} \begin{CD} 0 @>>> M_0 @>>> M @>>> M/M_0 @>>> 0\\ @. @VVV @VVV @\| \\ 0 @>>> M_0/t^n M_0 @>>> M/t^n M @>>> M/M_0 @>>>0 \end{CD}$$ According to Propositin \[ref:9.1.3a\], $$\label{ref:9.7a} \Gamma(X,M)\r \Gamma(X,M/t^n M)$$ is surjective in high degree. Hence there exists a surjective map $E(-a)^b\r \Gamma(X,M)$ for certain $a,b$ such that the composition with has finite cokernel. Tensoring with $\Escr_{o_X}$ and composing with $\Gamma(X,M)\otimes_E \Escr_{o_X}\r M$ we obtain a map $\Escr_{o_X}(-a)^b\r M$ such that the composition with $M\r M/t^n M$ has right bounded cokernel (using Propositions \[ref:9.1.3a\].4 and \[ref:9.1.6a\]). Applying $\pi$ to the induced map $M_0\oplus \Escr_{o_X}(-a)^b\r M$ together with yields what we want. \[ref:9.1.8a\] All maps and compositions in are admissible. Furthermore one has the following identities. $$\begin{aligned} i^\ast\delta_\ast&=\delta_\ast j^\ast\\ \label{ref:9.9a} i^!\delta_\ast&=\delta_\ast j^!\\ \label{ref:9.10a} Ri^!\,R\delta_\ast&= R\delta_\ast Rj^!\\ \label{ref:9.11a} Li^\ast\, R\delta_\ast &= R\delta_\ast Lj^\ast\end{aligned}$$ Checking admissibility is routine using the explicit formulas . We leave this to the reader. To verify the compatibility $i^\ast\delta_\ast=\delta_\ast j^\ast$ we may work in $\coh(X)$ (since everything is compatible with direct limits). But then it is just a reformulation of Proposition \[ref:9.1.3a\].2. The other compatibility $i^!\delta_\ast=\delta_\ast j^!$ involves only left exact functors, so it may be verified on injectives. Let $F\in \Qch(W)$ be such an injective. We have an exact sequence $$0\r j^! F\r F \r F(T) \r 0$$ Now $j^! F\in \Tors_T(W)$ hence $R^1\delta_\ast j^! F=0$. Therefore we obtain an exact sequence $$\label{ref:9.12a} 0\r \delta_\ast j^! F\r \delta_\ast F \r \delta_\ast (F(T)) \r 0$$ Now $\delta_\ast$ is compatible with shift. That is $\delta_\ast (F(T))=(\delta_\ast F)(S)$. Since we also have an exact sequence $$\label{ref:9.13a} 0\r i^!\delta_\ast F\r \delta_\ast F\r (\delta_\ast F)(S)$$ we are through. follows from provided we show for every injective $E\in \Qch(X)$ that $j^!E$ is acyclic for $\delta_\ast$ and that $\delta_\ast E$ is acyclic for $i^!$. The statement about $j^!E$ is clear since $j^!E$ is supported on $T$. The statement of $\delta_\ast E$ follows from the right exactness of . However is the same as and the latter is right exact. This shows what we want. follows from using . In the next few paragraphs we investigate to what extent the map $\delta:W\r V$ is an isomorphism. \[ref:9.1.9a\] 1. $\Iso_S(V)$ and $\Iso_T(W)$ are mapped to each other under $(\delta^\ast,\delta_\ast)$. 2. $R^i\delta_\ast$ sends $\Qch(W)$ to $\Iso_S(V)$ for $i>0$. 3. $L_i\delta^\ast$ sends $\Qch(V)$ to $\Iso_T(W)$ for $i>0$. 4. $(\delta^\ast,\delta_\ast)$ define inverse equivalences between $\Qch(W)/\Iso_T(W)$ and $\Qch(V)/\Iso_S(V)$. <!-- --> 1. This is clear by functoriality. 2. This is precisely 1. of Proposition \[ref:9.1.3a\]. 3. We may assume that $\Mscr$ is coherent. Furthermore by considering the appropriate long exact sequence it suffices to consider the cases where $\Mscr$ is $t$-torsion and where $\Mscr$ is $t$-torsion free. If $\Mscr$ is $t$-torsion then $L_i\delta^\ast\Mscr=0$ for $i>0$ by Proposition \[ref:9.1.6a\] so the result is true. Hence assume that $\Mscr$ is $t$-torsion free. Then the long exact sequence for $\delta^\ast$ applied to $$0\r \Mscr(-1)\r \Mscr\r \bar{\Mscr}\r 0$$ yields that $t$ is an isomorphism between $L_i\delta^\ast\Mscr(-1)$ and $ L_i\delta^\ast\Mscr$ for $i>0$, proving 3. in this case also. 4. We have to show that the adjunction mappings are isomorphisms. We already know that $\delta_\ast,\delta^\ast$ are exact functors between $\Qch(W)/\Iso_T(W)$ and $\Qch(V)/\Iso_S(V)$ commuting with direct limits. Hence it suffices to show that the adjunction mappings are isomorphisms for objects annihilated by $t$ and for objects which are $t$-torsion free. The case of objects annihilated by $t$ is already covered by Proposition \[ref:9.1.6a\], so we concentrate on $t$-torsion free objects. First assume that $\Nscr\in \coh(W)$ is a $t$-torsion free object. We have an exact sequence $$\label{ref:9.14a} 0\r \Nscr(-1)\r \Nscr \r \bar{\Nscr}\r 0$$ Working modulo $\iso_W(T)$ this yields an exact sequence $$\label{ref:9.15a} 0\r \delta^\ast\delta_\ast\Nscr(-1)\r \delta^\ast\delta_\ast\Nscr \r \delta^\ast\delta_\ast\bar{\Nscr}\r 0$$ Now by Proposition \[ref:9.1.6a\] we already know that $ \delta^\ast\delta_\ast\bar{\Nscr}=\bar{\Nscr}$. Combining and using the snake lemma now yields what we want. The proof that the other adjunction mapping is an isomorphism is entirely similar. The case of a blowing up {#ref:9.2b} ------------------------ Now we let $X,Y,p$ be as in §\[ref:5a\]. Furthermore we also assume that $X/k$ is proper. Up to now we have developed everything without assuming projectivity. However in this section we throw in the towel and we introduce the following hypothesis. $p$ is a smooth point on $Y$. $o_X(Y)$ is ample on $X$ and the invertible $o_Y$-bimodule $I_{Y}$ is ample on $Y$. The fact that $X$ has an ample invertible bimodule means in particular that Hypothesis (\\) is satisfied. Furthermore it is also easy to check that under the hypotheses one has $\cd \Gamma(X,-)\le 2$. This is part of Hypothesis (\\\). The smoothness of $Y$ in $p$ implies that $\tilde{Y}\r Y$ is an isomorphism (see Theorem \[ref:6.3.1a\]). Also because of smoothness, we have by an injective map $t:\Dscr(-1)\r \Dscr$ such that $\bar{\Dscr}=\Dscr/\im t$ is equal to $\Dscr_Y$ in degree $\ge 1$. By lemma \[ref:8.2.1a\] it is also clear that $t$ remains injective after applying the functor $(-)_{o_X}$. Since $I_{Y}$ is ample on $Y$ we find that $\Dscr_Y$ and hence $\bar{\Dscr}$ is ample. Thus Hypothesis (\\\) holds and hence the material in §\[ref:9.1a\] applies to the current situation. Let $D$, $D_Y$ be the global sections of $\Dscr$ and $\Dscr_Y$ respectively and put $V=\Proj D$. We now have a commutative diagram of $k$-quasi-schemes $$\label{ref:9.16a} \Atrianglepair<1`1`1`-1`1;>[Y`X`\tilde{X}`V;i`i`i`\alpha`\delta]$$ where the arrows marked $i$ are divisors. All maps and compositions of maps in this diagram are admissible by lemma \[ref:7.2.12a\], Proposition \[ref:9.1.8a\] and lemma \[ref:7.2.4a\]. Further properties of $\delta$ maybe deduced from §\[ref:9.1a\]. We define $\Oscr_M=\delta_\ast \Oscr_L$ and we will refer to $\Oscr_M$ as the exceptional curve in $V$. Since we are in a specific situation here, we can of course prove stronger results than those in §\[ref:9.1a\]. First of all note that we have the following. \[ref:9.2.1a\] $D$ satisfies $\chi$ [@AZ Def. 3.7]. Denote the global sections of $\Dscr_Y$ by $D_Y$. Since $I_{Y}$ is ample, $D_Y$ satisfies $\chi$ [@AZ]. Furthermore, by Proposition \[ref:9.1.3a\].2 the injective map $\bar{D}\r D_Y$ has finite cokernel. Hence by [@AZ Lemma 8.2], $\bar{D}$ will also satisfy $\chi$. Then we can apply [@AZ Thm 8.8] to obtain that $D$ also satisfies $\chi$. We also have the following important result. \[ref:9.2.2a\] The cohomological dimension of $\delta_\ast$ is less than or equal to one. Furthermore if $\Nscr\in\gr (\Dscr)$ then $R^1\delta_\ast\Nscr$ is a finite extension of quotients of $\Oscr_M$. Let $\Nscr\in \gr(\Dscr)$. By taking a resolution of $\Nscr$ we see that it is sufficient to treat the case $\Nscr=\Mscr\otimes_{o_X} \Dscr$ where $\Mscr\in\coh(X)$. Now define $\Escr=\oplus_{n\ge 0} o_X(nY)$. It is easy to see that $\Dscr$ embeds in $\Escr$ as graded $\Dscr$-bimodule. Put $\Nscr'=\Mscr\otimes_{o_X} \Escr$. Let $K,I,C$ be respectively the kernel, image and cokernel of the induced map $\Nscr\r \Nscr'$. Thus we have exact sequences $$\begin{gathered} 0\r K \r \Nscr\r I\r 0\\ 0\r I\r \Nscr'\r C\r 0\end{gathered}$$ It immediately follows from Theorem \[ref:5.5.10a\] that $K,C$ are in $\Cscr_p(\Dscr)$. Hence from it follows that $R^i\delta_\ast \pi K= R^i\delta_\ast \pi C=0$ for $i>0$. Since $o_X(Y)$ is ample it also follows from that $R^i\delta_\ast\pi\Nscr'=0$ for $i>0$. Plugging this in the above exact sequences we find that $R^i\delta_\ast\pi\Nscr=0$ for $i>1$ and also that $R^1\delta_\ast \pi\Nscr$ is a quotient of $\delta_\ast \pi C$. Now $C$ itself is not necessarily coherent, but it is a union of $C_i$ which are coherent. Since $\delta_\ast$ is compatible with direct limits we deduce that $R^1\delta_\ast \pi\Nscr$ is a quotient of some $\delta_\ast \pi C_i$. From lemma \[ref:6.7.3a\] it follows that $\pi C_i$ has a finite filtration whose associated graded quotients are quotients of $\Oscr_L(t)$. It then follows that $\delta_\ast C_i$ has a finite filtration whose associated graded quotients are quotients of $\Oscr_M(t)$. Since these quotients are in $\iso_Y(V)$ they are invariant under shifting. Hence they are also quotients of $\Oscr_M$. The result now follows for $R^1\delta_\ast \pi\Nscr$. The hypotheses for the following proposition will hold in our applications. \[ref:9.2.3a\] Assume that in addition to Hypotheses (\\\\) we have $R\Gamma(X,\Oscr_X)=k$, $H^1(Y,I^n_{Y,p})=0$ for all $n$ and $\Gamma(Y, \Oscr_Y)=k$. Then 1. $R\Gamma(X,\Dscr_{o_X})=D$. 2. There is an exact sequence $$0\r D(-1)\xrightarrow{t} D \rightarrow D_Y\r 0$$ 3. $R\delta_\ast L\delta^\ast=\Id$ on $D^-(V)$. 4. $R\Gamma(\tilde{X},\Oscr_{\tilde{X}})=k$ 5. $R\Gamma(V,\Oscr_V)=k$. - We have $\Dscr_0=\Oscr_X$ and furthermore by there are exact sequences for $n\ge 1$ $$0\r \Dscr_{n-1}\r \Dscr_n \r I^n_{Y,p}\r 0$$ Looking at the corresponding long exact sequence for $\Gamma(X,-)$ yields what we want by induction. - This follows from 1. since $L\delta^\ast$ can be computed by free resolutions. Note that $R\delta_\ast$ is defined on $D^-(\tilde{X})$ since according to Proposition \[ref:9.2.2a\], $\delta_\ast$ has finite cohomological dimension. - We have $$\begin{aligned} R\Gamma(\tilde{X},\Oscr_{\tilde{X}}) &=R\Gamma(X, R\alpha_\ast \Oscr_{\tilde{X}})\qquad \text{(admissibility)}\\ &=R\Gamma(X,\Oscr_X)\qquad (\text{Proposition \ref{ref:8.3.1a}.2}).\\ &=k \qquad\qquad (\text{by hypothesis})\end{aligned}$$ - We have $$\begin{aligned} R\Gamma(V,\Oscr_V)&=R\Gamma(V,R\delta_\ast \Oscr_{\tilde{X}})\qquad \text{(by 1.)}\\ &=R\Gamma(\tilde{X},\Oscr_{\tilde{X}})\qquad\text{(admissibility)}\\ &=k \qquad\qquad (\text{by 4.})\qed\end{aligned}$$ Quantum plane geometry {#ref:10a} ====================== In this section we fix a “triple” $(Y,\sigma,\Lscr)$ with $Y$ a smooth elliptic curve of degree three in $\PP^2$, $\sigma\in \operatorname{Aut}(Y)$ a translation and $\Lscr=i^\ast \Oscr_{\PP^2}(1)$ where $i:Y\r \PP^2$ denotes the inclusion. We let $A=A(Y,\sigma,\Lscr)$ be the regular algebra associated to this triple [@ATV1] and we put $X=\Proj A$. Since $A$ has the Hilbert series of a three dimensional polynomial algebra it is customary to view $X$ as a non-commutative $\PP^2$. $A$ contains a central element $g$ in degree three (determined up to a scalar) such that $Y=\Proj A/gA$ [@ATV1; @ATV2; @AVdB]. Clearly $Y\subset X$ is a divisor and so the theory of the previous sections applies to it. Since $X=\Proj A$, $\Qch(X)$ carries a canonical shift functor which we denote by $\Mscr\mapsto \Mscr(1)$. The corresponding invertible bimodule is denoted by $o_X(1)$. Thus $o_X(Y)=o_X(1)^{\otimes 3}\overset{\text {not.}}{=} o_X(3)$. The $o_Y$ bimodule $o_Y(1)$ is defined similarly. By construction we have $o_Y(1)=\Lscr_{\sigma}$. In addition we use the functors $\omega$, $\pi$, $\tilde{-}$ which were defined in §\[ref:3.8b\]. Multiplicities of some objects ------------------------------ If $\Fscr\in \mod(X)$ then $\Gamma(X,\Fscr)$ is finite dimensional. We denote by $\Ann_{\Oscr_X} \Fscr$ the kernel of the map $\Oscr_X\r \Gamma(X,\Fscr)^\ast \otimes_k\Fscr$ which is obtained from the map $\Gamma(X,\Fscr)\otimes_k \Oscr_X\r \Fscr$. By using the exactness properties of the completion functor (§\[ref:5a\]) it follows easily that if $\Fscr\in \Cscr_f$ (cfr §\[ref:5a\]) then we have $$\label{ref:10.1a} \operatorname{length}(\Oscr_X/\Ann_{\Oscr_X}\Fscr)=\oplus_{p\in Y/\langle\tau\rangle} \dim_k(C_p/\Ann_{C_p}(\hat{\Fscr}_p))$$ Let $\Fscr\in \trans_Y(X)$ and let $p\in Y$. Then we define numbers $t_{p,n}$ for $n\ge 1$ by $T_p(\Fscr)=\oplus_{n\ge 1} \Oscr_{Y,p}/m^{t_{p,n}}_{Y,p}$ (note that to simplify the notation we have dropped the completion sign). We consider $(t_{p,n})_n$ as a partition of the length of $T_p(\Fscr)$ and we let $(r_{p,n}(\Mscr))_n$ stand for the conjugate partition. In the sequel we will consider the $(p,n)$ as the points which are infinitely near to $p$. The numbers $r_{p,n}(\Mscr)$ should be viewed as the multiplicities of those infinitely near points. We identify $(p,1)$ with $p$. If $\Mscr\in \mod(X)$ then $\Mscr=\pi M$ for some finitely generated $A$-module $M$. As usual [@ATV2] we have $\dim M_n=(e/d!)n^d+f(n)$ for $n\gg 0$ where $d$ and $e$ are respectively the Gelfand-Kirillov dimension and the multiplicity of $M$ and where $f(n)$ is a polynomial of degree $<d$. We write $d(\Mscr)=d-1$ and $e(\Mscr)=e$. We conjecture the following (cfr [@H Cor 3.7]). Let $\Mscr\in \trans_Y(X)$ and assume that the image of $\Mscr$ in $\mod(X)/\Cscr_f$ is simple. Write $e=e(\Mscr)$ and $r_{p,n}=r_{p,n}(\Mscr)$ for $p\in Y$ and $n\ge 1$. Then the following inequality holds: $$\label{ref:10.2a} \sum_{p,n} \frac{r_{p,n}(r_{p,n}-1)}{2}\le \frac{(e-1)(e-2)}{2}$$ Note that the modern proof of the commutative analogue of this conjecture uses resolution of singularities for a curve through blowing up the ambient projective plane, together with the fact that an irreducible curve has non-negative genus. Unfortunately we have not been able to generalize this proof to the non-commutative case. Indeed it is rather the other way round. We would like to use a result as to deduce properties of our non-commutative blowing up. Imitating one of the classical proofs of (see [@Fulton]) as far as possible leads to Proposition \[ref:10.1.2a\] below. \[ref:10.1.2a\] Let $\Mscr\in \trans_Y(X)$ and assume that the image of $\Mscr$ in $\mod(X)/\Cscr_f$ is simple. Write $e=e(\Mscr)$ and $r_{p,n}=r_{p,n}(\Mscr)$ for $p\in Y$ and $n\ge 1$. Assume that $a_{p,n}$ are natural numbers such that $$\label{ref:10.3a} \frac{f(f+3)}{2}\ge \sum_{\bar{p},n}\frac{a_{p,n}(a_{p,n}+1)}{2}$$ for some integer $f<e$. Then $ef\ge \sum_{\bar{p},n} r_{p,n} a_{p,n}$. Here the notation $\bar{p}$ indicates that we take only one representant from each $\tau$-orbit. The proof consists of several steps. \[ref:1c\] Assume that $M\in \gr(A)$ contains no finite dimensional submodules and let $\Mscr=\pi M$. Then $$\label{ref:10.4a} \dim \{x\in A_n\mid M_0 x=0\}\ge \dim A_n-\dim \Gamma(X,(\Oscr_X/\Ann\Mscr)(n))$$ To prove this we note that the lefthand side of is the degree $n$ part of the kernel of the canonical map $A\r M^\ast_0\otimes_k M$. If we denote this kernel by $L$ then we have to prove $$\dim_k (A/L)_n\le \dim \omega(\Oscr_X/\Ann\Mscr)_n$$ Now note that there is a canonical map $M_0\r \Gamma(X,\Mscr)$. If we then look at the composition $$\Oscr_X\r \Gamma(X,\Mscr)^\ast \otimes_k\Mscr\r M^\ast_0\otimes_k \Mscr$$ we find $\Ann\Mscr\subset \pi L$. Hence it suffices to prove that $$\dim_k(A/L)_n\le \dim_k\omega(\Oscr_X/\pi L)_n=\dim_k (A/L)\tilde{}_n$$ Now $A/L$ is a graded submodule of $M^\ast_0\otimes_k M$ and hence $A/L$ contains no finite dimensional submodules. Therefore $A/L\hookrightarrow (A/L)\tilde{}$. This proves what we want. \[ref:2b\] Without loss of generality we may (and we will) assume that $a_{p,n}=0$ if $r_{p,n}=0$. Furthermore by definition we have $r_{p,1}\ge r_{p,2}\ge \cdots$ so by permuting the $a_{p,n}$ we may also assume that $a_{p,1}\ge a_{p,2}\ge a_{p,3}\ge$, because in this way remains valid and $\sum_{p,n} r_{p,n}a_{p,n}$ does not decrease. \[ref:3c\] Assume $\underline{a}_p=(a_{p,1},a_{p,2},\ldots)$ is a non-increasing set of numbers, zero for large $n$. We associate to such $\underline{a}_p$ an object in $\Cscr_{f,p}$. Let $(b_{p,n})_{n\ge 1}$ be the partition which is conjugate to $a_{p,n}$. We put $K_p(\underline{a}_p)=\prod_n R/m^{b_{p,-n+1}}$. This definition is to be understood as defining a set of row vectors with the obvious right $C_p$-action. Then we define $\Kscr_p(\underline{a}_p)$ as the object in $\Cscr_{f,p}$ such that $\hat{\Kscr}_p(\underline{a}_p)=K_p(\underline{a}_p)$. Below we need the following formula. $$\label{ref:10.5a} \dim (C_p/\Ann_{C_p}K_p(\underline{a}_p))=\sum_n \frac{a_{p,n}(a_{p,n}+1)}{2}$$ We leave the obvious verification to the reader. \[ref:5c\] Now define $\Kscr=\oplus_p \Kscr_p(\underline{a}_p)$ and $K=\omega \Kscr$. For $f<e$ we are going to bound the dimension of $$\{x\in A_f\mid K_{m} x=0\}$$ for large $m$. This can be done by the method exhibited in Step \[ref:1c\]. We find that this dimension is bigger than $$\dim_k A_f-\dim_k \omega (\Oscr_X/\Ann \Kscr)_{m+f}$$ Now $$\dim_k A_f=\frac{(f+1)(f+2)}{2}$$ and $$\begin{aligned} \dim _k \omega (\Oscr_X/\Ann \Kscr)_{m+f}&=\operatorname{length}(O_X/\Ann \Kscr)\\ &=\sum_p\operatorname{length}(\Oscr_X/\Ann \Kscr_p(\underline{a}_p))\\ &=\sum_p \dim_k (C_p /\Ann_{C_p}(K_p(\underline{a}_p))\\ &=\sum_{p,n} \frac{a_{p,n}(a_{p,n}+1)}{2}\end{aligned}$$ where we have used . So ultimately we obtain for large $m$ $$\dim_k\{x\in A_f\mid K_{m} x=0\} \ge \frac{(f+1)(f+2)}{2}-\sum_{p,n} \frac{a_{p,n}(a_{p,n}+1)}{2}$$ \[ref:6b\] Let $\Mscr$ be as in the statement of the proposition. Since $(r_{p,n})_{p,n}$ is unaffected by $\Cscr_{f}$ we may and we will assume that $\Mscr$ is $p$-normalized where $p$ runs through a fixed set of representatives for the $\tau$-orbits in $Y$. Then by lemma \[ref:5.7.4a\] we can recover the structure of $\hat{\Mscr}_p$ from $T_p(\Mscr)$. Assume that $T_p(\Mscr)=\oplus_n R/m^{t_{p,n}}$. Then it follows that $\oplus_n K_p(a_{p,1},\ldots,a_{p,t_{p,n}})$ is a quotient of $\hat{\Mscr}_p$. Hence $\oplus_{p,n} \Kscr_p(a_{p,1},\ldots,a_{p,t_{p,n}})$ is a quotient of $\Mscr$. This yields that $L=\oplus_{p,n} \omega \Kscr_p(a_{p,1},\ldots,a_{p,t_{p,n}})$ is a quotient of $M=\omega\Mscr$ for large $m$. \[ref:7b\] Since we have chosen $\Mscr$ to be normalized we have in particular that $\Mscr$ contains no subobject supported on $Y$. Since the only modules of GK-dimension one of $A$ are coming from objects supported on $Y$ [@ATV2] it follows that $M=\omega \Mscr$ contains no submodules of GK-dimension one. Since $\Mscr$ was also supposed to be simple modulo $\Cscr_f$ it follows that $M$ is critical. I.e. $M$ contains no non-trivial submodules of multiplicity strictly smaller than $e$. If $0\neq x\in A_f$ then $e(A/xA)=f$. So it follows that $\Hom_A(A/xA,M)=0$ which is the same as saying that multiplication by $x$ is injective. Since $\operatorname{GKdim}M=0$ we have for large $m$: $\dim M_m=em+s$ for some constant $s$. It follows that for large $m$ one has $$\label{ref:10.6a} \dim (M/Mx)_m=\dim M_m-\dim M_{m-f}=(em+s)-(e(m-f)+s)=ef$$ Now by hypotheses $$\frac{f(f+3)}{2}+1=\frac{(f+1)(f+2)}{2}> \sum_{p,n} \frac{a_{p,n}(a_{p,n}+1)}{2}$$ Let $m$ be large. By virtue of Step \[ref:5c\] there exists a non-zero $x\in A_f$ such that $K_m x=0$. Now by construction every indecomposable summand of $L$ is a quotient of $K$. Thus also $L_mx=0$. Now by the Step \[ref:6b\] there is a map $M\r L$, surjective in high degree. Tensoring with $A/Ax$ yields a map $M/Mx\r L/Lx$ with the same property. We conclude $$ef=\dim_k(M/Mx)_m\ge \dim_k (L/xL)_m=\dim_k L_m=\sum_{p,n} \dim K_p(a_{p,1},\ldots, a_{p,t_{p,n}})$$ Now an easy verification shows that $$\sum_{n} \dim K_p(a_{p,1},\ldots, a_{p,t_{p,n}})=\sum_n a_{p,n} r_{p,n}$$ which finishes the proof. The following proposition would be a trivial consequence of . With a lot more work we can also deduce it from Proposition \[ref:10.1.2a\]. Assume that $\Mscr\in \trans_Y(X)$ represents a non-zero simple object in $\mod(X)/\Cscr_f$. Let $r=(r_{p,n}(\Mscr))_{p,n}$, $e=e(\Mscr)$. If the number of non-zero entries of $r$ is $\le 6$ then there are (up to permutation) only two possibilities for $e,r$. $$\begin{aligned} e=1,\qquad &r=(1,1,1)\\ e=2,\qquad &r=(1,1,1,1,1,1)\end{aligned}$$ We will apply Proposition \[ref:10.1.2a\]. Since this is obviously a numerical criterion we will write $r=(r_i)_{i\in I}$ where $i$ runs through the pairs $(p,n)$ such that $r_{p,n}(\Mscr)\neq 0$. We will consider the entries of $r$ up to permutation. Put $m=\|I\|$. We will apply Proposition \[ref:10.1.2a\] with suitably chosen $(a_i)_{i\in I}$. First observe that Proposition \[ref:10.1.2a\] remains true if we allow the $(a_i)_{i\in I}$ to be negative. Indeed if $$\frac{f(f+3)}{2}\ge \sum_{i\in I} \frac{b_i(b_i+1)}{2}$$ with $b_i\in \ZZ$ then we define $$a_i= \begin{cases} b_i&\text{if $b_i\ge 0$}\\ -b_i-1&\text{if $b_i<0$} \end{cases}$$ Then $\sum_i b_i(b_i+1)/2=\sum_i a_i(a_i+1)/2$ and $a_i\ge b_i$. By Proposition \[ref:10.1.2a\] we have $ef\ge \sum_i r_i a_i$ and thus $ef\ge \sum_i r_i b_i$. Now we can reformulate Proposition \[ref:10.1.2a\] as follows: the intersection of the closed ball $$B=\{(a_i)_{i\in I}\mid \sum \frac{a_i(a_i+1)}{2} \le \frac{f(f+3)}{2}\}$$ and the open half space $$H=\{(a_i)_{i\in I}\mid \sum_i r_i a_i > ef\}$$ contains no integral point. Now $B\cap H$ certainly contains an open ball with diameter equal to the radius of $B$ minus the distance of $H$ to the center of $B$. The center of $B$ is $(-\frac{1}{2},\ldots,-\frac{1}{2})$ and the radius of $B$ is given by $$\sqrt{f(f+3)+\frac{m}{4}}$$ The distance of $H$ to the center of $B$ is given by $$\label{ref:10.7a} \frac{ef+\frac{\sum_i r_i}{2}}{\\|r\\|}$$ where $\\|r\\|=\sqrt{\sum_i r^2_i}$. If we use the fact that $3e=\sum_i r_i$ then becomes equal to $$\frac{e(f+\frac{3}{2})}{2}$$ Thus $B\cap H$ contains an open ball of diameter $$\sqrt{f(f+3)+\frac{m}{4}}- \frac{e(f+\frac{3}{2})}{\\|r\\|}$$ Now it is easy to see that an open ball of diameter $>\sqrt{m}$ must contain a point with integral coordinates. Thus Proposition \[ref:10.1.2a\] yields $$\label{ref:10.8a} \sqrt{f(f+3)+\frac{m}{4}}- \frac{e(f+\frac{3}{2})}{\\|r\\|}\le\sqrt{m}$$ From $\sum_i r_i=3e$ we obtain the estimate $\\|r\\|\ge 3e/\sqrt{m}$. Combining this with yields $$\sqrt{f(f+3)+\frac{m}{4}}\le\left( \frac{f}{3}+\frac{3}{2}\right)\sqrt{m}$$ We now take $f=e-1$. Then $$(e-1)(e+2)+\frac{m}{4}\le \left( \frac{e}{3}+\frac{7}{6}\right)^2m$$ which yields $$\frac{(e-1)(e+2)}{ \left( \frac{e}{3}+\frac{7}{6} \right)^2 - \frac{1}{4}} \le m$$ If we combine this with the hypotheses $m\le 6$ then we obtain $e\le 13$. Explicit enumeration of all possibilities for $e,r$ (using a computer) yields that $B\cap H$ always contains an integral point for $e\ge 3$. So the remaining possibilities are $e=1,2$. If $e=2$ then if follows easily from Proposition \[ref:10.1.2a\] that $r=(1,1,1,1,1,1)$. However if $e=1$ then Proposition \[ref:10.1.2a\] gives no information whatsoever. But is $e(\Mscr)=1$ then $\Mscr=\pi M$ where $M$ is a so-called “line-module” [@ATV2]. We will analyze this case directly. According to Proposition \[ref:5.7.2a\] we may change $\Mscr$ in such a way that $\Mscr$ is $p$ normal, where $p$ runs through a set of representatives of the $\tau$-orbits on $Y$. Then we can read of $r$ from the decomposition of $\Mscr/\Mscr(-Y)$ into uniserial $\Oscr_Y$-modules. As indicated above $\Mscr=A/xA$ where $x\in A_1$. But then $\Mscr/\Mscr(-Y)$ is given by the zeroes of $x\in \Gamma(Y,\Lscr)$. Since $\Lscr$ defines an embedding of $Y$ in $\PP^2$ this can be interpreted as the scheme theoretic intersection of $Y$ and $V(x)$. This is a union of uniserial schemes, such that every point in $Y$ occurs at most once in the reduced locus. Hence we obtain that $r=(1,1,1)$. Classification of lines and conics ---------------------------------- The cases $e(\Mscr)=1$ and $e(\Mscr)=2$ represent lines and conics. The first type of object has been studied in [@ATV2] and the second type of object has been studied in [@Ajitabh]. If $\Mscr\in \trans_Y(X)$ then let us denote by $\overline{\Div}(\Mscr)$ the image of $\Div(\Mscr)$ in $\NN(Y/\langle\tau\rangle)$ under the map which sends $p\in Y$ to its $\tau$-orbit. Let us call $M\in \Gr(A)$ standard if $M_0\neq 0$ and $M_i=0$ if $i<0$. Let us call $M\in \gr(A)$ Cohen-Macaulay of dimension $d$ if $M$ has GK-dimension $d$ and if $\Ext^s_A(M,A)=0$ unless $s=3-d$. The following is easy to see and will be used below (cfr [@Aj1]). Assume that $M\in \gr(A)$ has GK-dimension two and that $M$ contains no submodules of GK-dimension one. Then $\tilde{M}$ is Cohen-Macaulay. According to [@Ajitabh] and [@ATV2] there are the following three classes of critical Cohen-Macaulay modules of GK-dimension two and multiplicity $1$ or $2$. 1. Lines: modules of the form $A/xA$, $x\in A_1-\{0\}$. 2. Conics of the first kind: critical graded $A$-modules of the form $A/xA$, $x\in A_2-\{0\}$. 3. Conics of the second kind: critical graded $A$-modules with minimal resolution $$0\r A(-1)^2\r A^2\r M\r 0$$ Shifts of such modules will be called shifted lines and conics (of the first and the second kind). They exhaust all critical Cohen-Macaulay modules of GK-dimension two and multiplicity $\le 2$ [@Ajitabh]. Below $L$ will be an effective divisor on $Y$ such that $\Lscr=\Oscr_Y(L)$. The following results from [@Ajitabh] describe the divisor classes of modules of multiplicity $\le 2$. \[ref:10.2.2a\] 1. If $M=N(s)$ where $N$ is a line module then $\Div\pi M$ is of the form $\sigma^s L$ for some $s$. 2. If $M=Q(s)$ where $Q$ is a conic of the first kind then $\Div \pi M\sim \sigma^{s} L+ \sigma^{s-1}L$. 3. If $M=Q(s)$ where $Q$ is a conic of the second kind then $\Div \pi M\sim\sigma^{s} L+ \sigma^s L$. If $D\in \NN(Y/\langle\tau\rangle)$ and $H$ is a divisor on $E$ then we say that $D$ is compatible with $H$ if there are $(p_i)_{i=1,\ldots,l}$ with $l=\deg H$ such that $D=\sum_{i=1}^{l} \overline{p}_i$ and $H\sim \sum_{i=1}^l p_i$. From the previous proposition we deduce the following: \[ref:10.2.3a\] Assume that $M$ is a Cohen-Macaulay module of GK-dimension 2. If $e(M)=1$ then $\overline{\Div}(\pi M)$ is compatible with $L$. If $e(M)=2$ then $\overline{\Div}(\pi M)$ is compatible with $2L$. This follows from the fact that the image of $\sigma^s L$ in $\NN(Y/\langle\tau\rangle)$ is compatible with $L$. This follows in turn from the fact that $\tau=3\sigma$ and $\deg L=3$. We also deduce the following lemma. \[ref:10.2.4a\] Assume that $M\in\gr(A)$ is a Cohen-Macaulay module of GK-dimension 2. Then the fact whether it is a shifted conic of the first or second kind can be recognized from $\Div \pi M$. On the other hand we can’t recognize the kind of a conic from its image in $\trans_Y(X)$. In fact we have the following easy lemma. \[ref:10.2.5a\] Assume that $\Nscr\in \trans_Y(X)$ is such that $e(\Nscr)=2$. Then $\Nscr$ is equivalent modulo $\Cscr_f$ to an object of the form $\pi(A/xA)(m)$. We may assume that $\Nscr$ is $Y$-torsion free. Let $N=\omega(\Nscr)$. If $N$ is a shifted conic of the first kind then we are done. Assume that this is not the case. Let $p$ be a point in the support of $\Div(\Nscr)$ and let $\Nscr'$ be the kernel of the associated map $\Nscr\r \Oscr_p$. Put $N'=\omega\Nscr'$. Then by the formula together with the discussion preceding lemma \[ref:10.2.4a\] we deduce that $N'$ is of the first kind. \[ref:10.2.6a\] The map $\overline{\Div}$ defines a bijection between 1. simple objects in $\trans_Y(X)/\Cscr_f$ of multiplicity $1$ and elements of $\NN(Y/\langle \tau\rangle)$ compatible with $L$; 2. simple objects in $\trans_Y(X)/\Cscr_f$ of multiplicity $2$ and elements of $\NN(Y/\langle\tau\rangle)$ compatible with $2L$ that are not the sum of two elements of $\NN(Y/\langle\tau\rangle)$ compatible with $L$. For the proof we need the following easy lemma. \[ref:10.2.7a\] Assume that $\Mscr\in \trans_Y(X)$ is $Y$-torsion free and assume that $\Div(\Mscr)=\sum^l_{i=1} p_i$ where the $(q_i)_i$ are in the $\tau$-orbit of the $(p_i)_i$. Then there exists $\Nscr\in \trans_Y(X)$ which is $Y$-torsion free and which is equivalent with $\Mscr$ modulo $\Cscr_f$ such that $\Div(N)=\sum_i q_i$. To prove this we first replace $\Mscr$ by $\Mscr(mY)$, $m$ large and then apply the formula . Note that by lemma \[ref:10.2.3a\] if $\Mscr\in \trans_Y(X)$ then $\overline{\Div}(\Mscr)$ is compatible with $L$ or $2L$ if $e(\Mscr)$ is $1$ or $2$. We first prove that $\overline{\Div}$ is a bijection between simple objects in $\trans_Y(X)/\Cscr_f$ of multiplicity $1$ and elements of $\NN(Y/\langle \tau\rangle)$ compatible with $L$. Surjectivity is obvious so we consider injectivity. Assume that $\Nscr_{1,2}\in \trans_Y(X)$ have multiplicity $2$ and $\overline{\Div}( \Nscr_1)=\overline{\Div}(\Nscr_2)$. According to lemma \[ref:10.2.7a\] and the previous discussion we can assume that $\Nscr_1=\pi (A/xA)(m)$, $\Nscr_2=(A/yA)(n)$ in such a way that $\Div(\Nscr_1)=\Div(\Nscr_2)$. Now consider $x,y$ as global sections of $\Lscr$ on $Y$. Then $\Div \Nscr_1=\sigma^m\operatorname{div}(x)$, $\Div \Nscr_2=\sigma^n \operatorname{div}(y)$. Since $\sigma$ has infinite order this implies $m=n$ and $x=y$ (up to a scalar). Now we prove the second part of the theorem. We consider injectivity first. Assume that $\Nscr_{1,2}\in \trans_Y(X)$ have multiplicity $=2$ and are such that $\overline{\Div}(\Nscr_1)=\overline{\Div}(\Nscr_2)$. Using lemmas \[ref:10.2.5a\],\[ref:10.2.7a\] we may assume that $\Nscr_1=\pi(A/xA)(m)$ and $\Div(\Nscr_1)=\Div(\Nscr_2)$, where $\Nscr_2$ is in addition $Y$-torsion free. But then by lemma \[ref:10.2.4a\] we find that $\Nscr_2$ is obtained from a shifted conic of the first kind, whence $\Nscr_2=\pi((A/yA)(n))$. Now exactly as in the case $e=1$ this implies $x=y$, $m=n$. Now we have to describe the image of $\overline{\Div}$. First let $D\in \NN(Y/\langle\tau\rangle)$ be compatible with $2L$ and not be the sum of two elements compatible with $L$. Then according to lemma \[ref:10.2.3a\] if $\overline{\Div}(A/xA)=D$ then $A/xA$ is critical. Assume on the other hand that $\Mscr\in\trans_Y(X)$ is such that $e(\Mscr)=2$ and $\overline{\Div}(\Mscr)=D_1+D_2$, $D_{1,2}$ compatible with $L$. We claim that $\Mscr$ is not simple modulo $\Cscr_f$. Without loss of generality we may assume that $\Mscr$ is $Y$-torsion free. Using lemma \[ref:10.2.7a\] we may assume that $\Div \Mscr=E_1+E_2$ with $E_1\sim L$, $E_2\sim \sigma^{-1}L$. Put $M=\omega\Mscr$. It follows from Proposition \[ref:10.2.2a\] that $M=A/xA$ with $x\in A_2-\{0\}$. Let $N$ be the line module corresponding to $E_1$. Since the divisor of $x$ contains $E_1$ it follows that $A/xA$ maps surjectively to $N/Ng$. Since $x\in A_2$ this implies that $M$ maps surjectively to $N$. Hence $A/xA$ is not critical. Blowing up $n$ points in the elliptic quantumplane {#ref:11a} ================================================== Derived categories {#ref:11.1a} ------------------ In this section our notations and conventions will be as in Section §\[ref:10a\] except that we do not assume that $\sigma$ has infinite order. We choose $n$ points $p_1,\ldots,p_n\in Y$ where $n\le 8$ (at this point not necessarily distinct). We define quasi-schemes $X_j$, $j=1,\ldots,n+1$, $\tilde{X_j}$, $j=1,\ldots,n$ containing $Y$ as a divisor. We do this as follows : 1. $X_1=X$. 2. $\tilde{X_j}$ is the blowup of $X_j$ in the point $p_j$. 3. $X_{j+1}$ is derived from the pair $(X_j,p_j)$ in the same way as $V$ was derived from $(X,p)$ in §\[ref:9.2b\]. Of course we are only allowed to make this construction if Hypothesis (\\\\) of §\[ref:9.2b\] holds for the triples $(Y,X_j,p_j)$. Let us check this now by induction. Smoothness of the $p_j$ is by hypotheses. Ampleness of $o_{X_j}(Y)$ for $j\ge 2$ follows from the fact that $X_j$ is the $\Proj$ of a graded algebra which satisfies $\chi$ (Proposition \[ref:9.2.1a\]) and $Y$ is defined by a central element of degree one. This argument also works for $X_1$ if we note that $X_1=\Proj A^{(3)}$ where $A^{(3)}$ denotes the $3$-Veronese of $A$. Hence the only thing that remains to be checked is that $I_{Y,p_j}$ is ample (we have now included the point $p_j$ in the prior notation $I_Y$). We check ampleness by induction. Assume that $I_{Y,p_{t}}$ is ample for $t<i$. By definition we have $I_{Y,p_j}=m_{Y,p_j} \Nscr_{Y/X_j}$. By we have that $\Nscr_{Y/X_j}=\Nscr_{Y/\tilde{X}_{j-1}}$ and by we have $\Nscr_{Y/\tilde{X}_{j-1}}=m_{p_{j-1}} \Nscr_{Y/X_{j-1}}$. Since $\Nscr_{Y/X}$ is obtained by the functor $-\otimes Ag^{-1}$ and $Ag^{-1}=A(3)$ one deduces from [@AVdB] that $$\Nscr_{Y/X}=(\Lscr_\sigma)^{\otimes 3} =(\Lscr\otimes_{\Oscr_Y} \sigma^\ast(\Lscr)\otimes_{\Oscr_Y} \sigma^{2\ast}(\Lscr))_{\sigma^3}$$ Thus $\Nscr_{Y/X}=\Nscr_\tau$ with $\tau=\sigma^3$ and $\Nscr=(\Lscr\otimes_{\Oscr_Y} \sigma^\ast(\Lscr)\otimes_{\Oscr_Y} \sigma^{2\ast}(\Lscr))_{\sigma^3}$. In particular $\deg\Nscr=9$. Define $$\label{ref:11.1b} \Nscr_j=(m_{Y,p_1}\cdots m_{Y,p_{j-1}}\Nscr)_\tau$$ By the above discussion we obtain $\Nscr_{Y/X_{j}}=\Nscr_{Y/\tilde{X}_{j-1}}=(\Nscr_j)_\tau$ and $I_{Y,p_j}=(\Nscr_{j+1})_\tau$. Since $\deg\Nscr_{j}=9-j+1$ we obtain that $\Nscr_{j+1}$ has positive degree if $j\le 9$. This is true by the restriction $n\ge 8$. Hence $I_{Y,p_j}$ is ample. With our current notations the following diagram replaces $$\label{ref:11.2a} \Atrianglepair<1`1`1`-1`1;>[Y`X_j`\tilde{X_j}`X_{j+1};i`i`i`\alpha_j`\delta_j]$$ We will denote the exceptional curves in $\tilde{X}_j$ by $\Oscr_{L_j}$ and their direct images in $X_{j+1}$ by $\Oscr_{M_{j+1}}$. \[ref:11.1.1a\] The hypotheses for Proposition \[ref:9.2.3a\] hold for $X_j,Y,p_j$. In particular we have : 1. $R\Gamma(X_j,\Oscr_{X_j})=R\Gamma(\tilde{X}_j,\Oscr_{\tilde{X}_{j}})=k$. 2. $R\delta_{{j},\ast} \,L\delta^\ast_{j}$ is the identity on $D^{-}(X_{{j}+1})$. The fact that the hypotheses for Proposition \[ref:9.2.3a\] hold is verified by induction, starting from the easy fact that $R\Gamma(X,\Oscr_X)=k$. We borrow the following definition from the commutative case. \[ref:11.1.2a\] Assume that $(q_{j})_{{j}=1,\ldots,n}$, $n\le 8$ are points in $\PP^2$. The points $(q_{j})_{j}$ are in general position if and only if - All points are different. - No three points lie on a line. - No six points lie on a conic. - If $n=8$ then not all points lie on a singular cubic divisor in such a way that one of the points lies on the singularity. The following will be our main theorem. \[ref:11.1.3a\] Assume that the points $(\sigma p_{j})_{j}$ are in general position (with respect to the embedding of $Y\subset \PP^2$, fixed in the beginning of this section). Then the following holds for all ${j}$ : 1. $\delta_{{j}\ast},\delta^\ast_{j}$ have cohomological dimension $\le 1$. 2. $L\delta_{j}^\ast$, $R\delta_{{j}\ast}$ are inverse equivalences between $D(\tilde{X}_{j})$ and $D(X_{{j}+1})$. 3. $\Qch(X_{j})$, $\Qch(\tilde{X}_{j})$ have finite injective dimension. Without loss of generality we may assume that ${j}=n$. 1.,2. of the theorem will be proved directly but for 3. we will need induction. That is, we will assume that $\Qch(X_{j})$ has finite injective dimension for ${j}<n+1$. Note that this is clearly true if ${j}=1$ since $X_1$ is the $\Proj$ of a graded algebra of finite global dimenion. By Corollary we find that $\Qch(\tilde{X}_{j})$ also has finite injective dimension for ${j}<n+1$. We already know by lemma \[ref:11.1.1a\] that $R\delta_{n,\ast}\, L\delta^\ast_n$ is the identity on $D^{-}(X_{n+1})$. We will start by showing that $L\delta^\ast_n\, R\delta_{n,\ast}$ is the identity on $D^{-}_f(\tilde{X}_{n})$. By lemma \[ref:8.1.4a\] this means that we have to show that $\ker R\delta_{n,\ast}=0$. Hence assume that $\Mscr\in D^-_f(\tilde{X}_n)$ is such that $R\delta_{n,\ast}\Mscr=0$. We will construct $\Qscr_{j}\in D^-_f(X_{j})$ for ${j}=1,\ldots, n$ in such a way that the following holds : - $R\Gamma(X_{j},\Qscr_{j})=0$ (note that $R\Gamma(X_{j},-)$ is well defined on $D^-(X_{j})$ because $\Qch(X_{j})$ has finite injective dimension by the induction hypotheses). - There are triangles $$\begin{aligned} \label{ref:11.3a} &\Atriangle<1`-1`1;>[\Uscr_{n}`L\alpha^\ast_{n}\Qscr_n`\Mscr;``]&\\ \label{ref:11.4a} &\Atriangle<1`-1`1;>[\Uscr_{{j}}`L\alpha^\ast_{{j}}\Qscr_{j}`L\delta^\ast_{j}\Qscr_{{j}+1};``] \qquad \text{for ${j}=1,\ldots n-1$}\end{aligned}$$ where the $\Uscr_{j}$ are direct sums of shifts in the derived category of copies of $\Oscr_{L_{j}}(-Y)$. To do this we define $$\begin{aligned} \Qscr_n&=R\alpha_{n,\ast}\Mscr\\ \Qscr_{j}&=R\alpha_{{j},\ast}L\delta^\ast_{{j}} \Qscr_{{j}+1}\end{aligned}$$ By adjointness we have maps $$\begin{aligned} L\alpha^\ast_n\Qscr_n&\r \Mscr\\ L\alpha^\ast_{j} \Qscr_{j}&\r L\delta^\ast_{{j}} \Qscr_{{j}+1}\end{aligned}$$ which become the identity after applying $R\alpha_{{j},\ast}$, ${j}=1,\ldots,n$. The existence of the triangles now follows from Theorem \[ref:8.4.1a\]. By admissibility we also have $$R\Gamma(X_n,\Qscr_n)=R\Gamma(\tilde{X}_n,\Mscr)=R\Gamma(X_{n+1}, R\delta_{\ast,n}\Mscr )=0$$ and $$R\Gamma(X_{j},\Qscr_{j})=R\Gamma(\tilde{X}_{j}, L\delta^\ast_{{j}} \Qscr_{{j}+1})= R\Gamma(X_{{j}+1},\Qscr_{{j}+1})$$ (for the last equality we have used lemma \[ref:11.1.1a\].2). Hence by induction $R\Gamma(X_{j},\Qscr_{j})=0$. This finishes the proof of (a) and (b) above. To continue we use restriction to $Y$. Applying $Li^\ast$ to and yields triangles $$\begin{aligned} \label{ref:11.5a} &\Atriangle<1`-1`1;>[Li^\ast\Uscr_{n}`Li^\ast\Qscr_n`Li^\ast\Mscr;``]\\ &\Atriangle<1`-1`1;>[Li^\ast\Uscr_{{j}}`Li^\ast\Qscr_{j}`Li^\ast\Qscr_{{j}+1};``] \qquad \text{for ${j}=1,\ldots n-1$} \label{ref:11.6a}\end{aligned}$$ where the $Li^\ast\Uscr_{j}$ are direct sums of shifts in the derived category of $\Oscr_{\tau p_{j}}(-Y)=\Oscr_{p_{j}}$. Now by hypotheses $R\delta_{\ast,n} \Mscr=0$. Thus by we obtain $ 0= Li^\ast R\delta_{\ast,n} \Mscr=Li^\ast\Mscr $. Hence $Li^\ast \Qscr_n=Li^\ast\Uscr_n$. Taking into account that $\RHom_{D(Y)}(\Oscr_{p_{j}},\Oscr_{p_t})=0$ if ${j}\neq t$ we deduce that the map $Li^\ast\Qscr_n\r Li^\ast \Uscr_{n-1}$, must be the zero map in . Hence $Li^\ast \Qscr_{n-1}=Li^\ast \Uscr_{n-1}[-1]\oplus Li^\ast\Qscr_n=Li^\ast\Uscr_{n-1}[-1]\oplus Li^\ast\Uscr_n$. Continuing yields $$\label{ref:11.7a} Li^\ast \Qscr_1=Li^\ast \Uscr_1[-n+1]\oplus\cdots \oplus Li^\ast\Uscr_n$$ In particular $Li^\ast\Qscr_1$ is a direct sum of shifts in the derived category of copies of $\Oscr_{p_{j}}$. Now $\Qscr_1$ lives on $X_1=X$ and this is a very well understood quasi-scheme. In fact it follows from lemma \[ref:11.1.4a\] below that necessarily $\Qscr_1=0$. But then from we immediately deduce that $\Uscr_{j}=0$. From the triangle we then find that $L\delta_1^\ast\Qscr_2=0$. Applying $R\delta_{\ast,1}$, using lemma \[ref:11.1.1a\] yields that $\Qscr_2=0$. Continuing by induction we eventually find that $\Qscr_n=0$, and hence by we finally obtain $\Mscr=0$. At this point we have partially proved 2. Let us now prove that $\delta^\ast_n$ has cohomological dimension $\le 1$. Let $\Mscr\in \Qch(X_n)$ and put $\Nscr=L\delta_n^\ast\Mscr$. From the spectral sequence $$R^p\delta_{n,\ast} H^q(\Nscr)\r R^{p+q}\delta_{n,\ast} \Nscr$$ we obtain exact sequences $$\label{ref:11.8a} 0\r R^1\delta_{n,\ast} H^{i-1}(\Nscr)\r R^i\delta_{n,\ast}\Nscr \r R^0\delta_{n,\ast} H^i(\Nscr)\r 0$$ Since $R^i\delta_{n,\ast}\Nscr=R^i\delta_{n,\ast}L\delta^\ast_n\Mscr=H^i(\Mscr)$ we deduce that $R^i\delta_{n,\ast}\Nscr=0$ for $i<0$. By this yields that $R^1\delta_{n,\ast} H^{i}(\Nscr)=0$ for $i<-1$ and $R^0\delta_{n,\ast} H^i(\Nscr)$ for $i<0$. In particular $R\delta_{n,\ast} H^i(\Nscr)=0$ for $ i<-1$. By the part of 2. already proved, this implies $H^i(\Nscr)=$ for $i<-1$. Hence $\cd \delta^\ast_n\le 1$. This finishes the proof of 1. Since it now follows that $L\delta^\ast_n$ is defined on the unbounded derived category, we can prove 2. completely. We have to show that the adjunction mappings are isomorphisms on the unbounded derived category. Take for example $\Mscr\in D(\tilde{X}_n)$. We have to show that $\operatorname{cone}(L\delta^\ast_n R\delta_{n,\ast} \Mscr\r \Mscr)$ is zero. In the same way as in Step 6 of the proof of Theorem \[ref:8.4.1a\] we reduce to the case $\Mscr\in \coh(\tilde{X}_n)$. But then $\Mscr\in D^-_f(\tilde{X}_n)$ and this case was already handled. The proof that the other adjunction morphism is an isomorphism is entirely similar. By the induction hypotheses (and the discussion in the first paragraph of this proof) we know that $\Qch(\tilde{X}_n)$ has finite injective dimension. By an argument similar to the proof of Corollary \[ref:8.4.3a\] we then find that $\Qch({X}_{n+1})$ has finite injective dimension (using 1. and 2.). This finishes the proof. \[ref:11.1.4a\] Let $Y,X$ and $(p_{j})$ be as above. Assume that the $\sigma p_{j}$ are in general position. Let $\Tscr\in D_f^-(X)$ be an object such that 1. $R\Gamma(X,\Tscr)=0$ 2. The homology of $Li^\ast\Tscr$ is a direct sum of copies of $\Oscr_{p_{j}}$. Then $\Tscr=0$. Define $\Escr=\Oscr_X(-2)\oplus \Oscr_X(-1)\oplus \Oscr_X$ and $H=\End_{o_X}(\Escr)$. Thus $$H=\begin{pmatrix} k& 0 & 0\\ A_1 & k & 0\\ A_2 & A_1 & k \end{pmatrix}$$ By a standard generalization of [@Beilinson] it follows that the functor $$\label{ref:11.9a} F:D^-_f(X)\r D^-_f(H): \Tscr=\RHom(\Escr,\Tscr)$$ is an equivalence. Let $e_i$, $i=1,2,3$ be the diagonal idempotents in $H$ and let $P_i=e_iH$ be the corresponding projectives. Under $F$ we have the following correspondences $$\Oscr_X(-2)\leftrightarrow P_1\qquad \Oscr_X(-1)\leftrightarrow P_2 \qquad \Oscr_X\leftrightarrow P_3$$ From we deduce that $$R\Gamma(X,\Tscr)=0\iff (F\Tscr)e_3=0$$ Put $H'=(1-e_3)H(1-e_3)=\bigl(\begin{smallmatrix} k&0\\ A_3 & k\end{smallmatrix} \bigr)$. Right modules over $H$ can be written in block form as row vectors $(M_1, M_2, M_3)$. Similarly $H'$-modules can be written as $(M_1,M_2)$. Sending $(M_1,M_2)\mapsto (M_1,M_2,0)$ defines an exact functor $I:\Mod(H')\r \Mod(H)$ which extends to an exact functor $I:D^-_f(H')\r D^-_f(H)$. It is easy to see that this functor defines an equivalence of $D^-_f(H')$ with the full subcategory of $D^-_f(H)$ consisting of objects $T$ such that $Te_3=0$. Hence we find in particular that $\ker R\Gamma(X,-)$ is equivalent to $ D^-_f(H')$. Now $H'$ is hereditary, and hence by lemma \[ref:8.3.4a\] an element of $D^-_f(H')$ is the sum of its homology. Assume that $T\in\mod(H')$. Then $T$ has a minimal resolution of length $2$ $$0\r W\otimes_k P_1\r U\otimes_k P_2 \oplus V\otimes_k P_1 \r T\r 0$$ The complex $W\otimes_k P_1\r U\otimes_k P_2 \oplus V\otimes_k P_1$ splits as a direct sum of $W\otimes_k P_1\r U\otimes_k P_2$ and $V\otimes_k P_1$. Now we view $T$ as a $H$-module via the functor $I$ defined above. We find that $F^{-1}T$ is a sum of $$\label{ref:11.10a} W\otimes_k \Oscr_X(-2)\r U\otimes_k \Oscr_X(-1)$$ and $V\otimes_k \Oscr_X(-2)$. We conclude that $\Tscr$ is a direct sum of shifts in the derived category of complexes of the form and of complexes of the form $\Oscr_X(-2)$. By hypotheses we know in addition that the homology of $Li^\ast\Tscr$ is given by direct sums of copies of $\Oscr_{p_{j}}$. This yields that $Li^\ast\Tscr$ must be a direct sum of complexes $$W\otimes_k \Oscr_Y(-2)\xrightarrow{\phi} U\otimes_k \Oscr_Y(-1)$$ where $\phi$ is injective and $\coker \phi$ is a direct sum of copies of $\Oscr_{p_{j}}$. In particular $\dim W=\dim U$. For convenience we tensor with $o_Y(2)$. Taking into account that $\Oscr_Y(1)=\sigma_\ast(\Lscr)$ we obtain a complex $$W\otimes_k \Oscr_Y\rightarrow U\otimes_k \sigma_\ast(\Lscr)$$ whose cokernel which is a direct sum of copies of $\Oscr_{p_{j}}(2)=\Oscr_{\sigma^{2} p_{j}}$. Now we invoke Lemma \[ref:11.1.6a\] below with $\Mscr=\sigma_\ast(\Lscr)$ and $q_{j}=\sigma^2 p_{j}$. By hypotheses the $q_{j}$ are in general position with respect to the embedding defined by $\Mscr$. We obtain $W=U=0$ and the proof is done. The above proof was based on lemma \[ref:11.1.6a\] below. To prove this lemma we need some notions of the theory of commutative blowing up. Assume that $Z$ is a smooth (commutative!) surface let $Y\subset Z$ be a divisor. Let $p\in Y$ be a smooth point and let $\psi:\tilde{Z}\r Z$ and $\tilde{Y}$ be respectively the blowup of $Z$ in $p$ and the strict transform of $Y$. As usual $\tilde{Z}=\Proj (\oplus_n m_p^n)$, where $m_p$ is the maximal ideal of $\Oscr_Z$ defined by $p$. Let $L\subset \tilde{Z}$ be the exceptional curve. Let $\Pscr$ be a one-dimensonal coherent Cohen-Macaulay module on $Z$. We define the strict transform $\psi^{-1}(\Pscr)$ of $\Pscr$ as $\pi (\oplus_n m_p^n\Pscr)$. The following lemma is easily proved. \[ref:11.1.5a\] Let the notation be as above. Assume that $Y$ is not contained in the support of $\Pscr$ and furthermore that the $p$-primary component of $i^\ast \Pscr$ is of the form $\Oscr_p^{\oplus u}$. Then $\Supp(\psi^{-1}\Pscr)\cap \tilde{Y}\cap L=\emptyset$. We use this lemma to prove the following result. \[ref:11.1.6a\] Let $Y$ be embedded as a cubic divisor in $\PP^2$ through a very ample line bundle $\Mscr$ of degree three. Assume that $(q_{j})_{{j}=1,\ldots,n}\in Y$, $n\le 8$, are smooth points in general position. Then it is impossible to have a map $\Oscr_Y^u\xrightarrow{\phi} \Mscr^u$ whose cokernel is a direct sum of copies of $\Oscr_{q_{j}}$, unless $u=0$. Let $t:Y\r \PP^2$ be the embedding. The map $\phi$ can be uniquely lifted to a map $\Oscr_{\PP^2}^u\xrightarrow{\mu} \Oscr_{\PP^2}^u$ such that $\phi=t^\ast \mu$. Let $\Pscr=\coker \mu$. Then $\Pscr$ is a one dimensional Cohen-Macaulay module on $\PP^2$ such that $t^\ast\Pscr$ is a direct sum of copies of $\Oscr_{q_{j}}$. We claim that this is impossible unless $\Pscr=0$. To do this we perform the blowup of $\PP^2$ in $q_1,\ldots,q_n$. Let $\tilde{\Pscr}$ and $\tilde{Y}$ be the (iterated) strict transforms of $\Pscr$ and $Y$. Then according to lemma \[ref:11.1.5a\] we have $\Supp \tilde{\Pscr}\cap \tilde{Y}=\emptyset$. On the other hand it follows from the Nakai criterion that if the $q_{j}$ are in general position then $\tilde{Y}$ is ample. This is clearly a contradiction. Exceptional simple objects -------------------------- If $Y\subset X$ is a commutative curve contained as a divisor in a quasi-scheme $X$ then we will call a simple object in $\mod(X)$ exceptional if it is not of the form $\Oscr_p$ for $p\in Y$. One of the aims of these notes is to count the exceptional simple objects in the quasi-schemes $X_n$ which were introduced in the previous section. So far we have only been able to do this under some additional hypotheses as can be seen from our main result below (Theorem \[ref:11.2.1a\]). In this section we use the same notations and hypothese as in sections §\[ref:10a\] and §\[ref:11.1a\]. We assume in addition that $\tau$ has infinite order. Throughout we will choose a group law on $Y$ in such a way that if $p,q,r\in Y$ lie on a line then $p+q+r=0$. Furthermore we choose a fixed set $F\subset E$ of representatives for the $\tau$-orbits on $Y$. We partially order $\NN F$ by putting $y\le z$ if $y_p\le z_p$ for all $p\in F$. If $z\in\NN F$ then we write $\|z\|=\sum_{p\in F} z_p$, $N(z)=\sum_{p\in F} z_p p$. For $n\in \NN$ we define $$H_n=\{y\in \NN F\mid \|y\|=n,N(y)\in \ZZ\tau\}$$ and for $z\in \NN F$ we also define $$\begin{gathered} A_z=\{y\in H_3\mid y\le z\}\\ B_z=\{y\in H_6\mid y\le z, \text{$y$ is not the sum of two elements of $H_3$}\}\end{gathered}$$ We now have the following theorem. \[ref:11.2.1a\] Let $p_1,\ldots,p_n\in Y$, $n\le 6$ be such that $(\sigma p_i)_i$ are in general position (Def. \[ref:11.1.2a\]). For $p\in F$ let $z_p$ be the cardinality of the intersection of $\{p_1,\ldots,p_n\}$ with the $\tau$-orbit of $p$. Let $O$ be the number of non-zero $z_p$. Then the number of non-isomorphic exceptional simple objects in $\mod(X_{n+1})$ is equal to $ n+\|A_z\|+\|B_z\|-O $. The proof of this theorem will follow rather easily from our previous results. We start with the following lemma. \[ref:11.2.2a\] Assume that $\Ascr$, $\Bscr$ are two abelian categories and $\Cscr\subset\Ascr$, $\Dscr\subset \Bscr$ are two abelian subcategories closed under subquotients. Assume that there are inverse equivalence $F$, $G$ between $D^b(\Ascr)$ and $D^b(\Bscr)$. Assume furthermore that for all $i$, $H^iF$ sends $\Cscr$ to $\Dscr$ and $H^iG$ sends $\Dscr$ to $\Cscr$. Then the maps $$\label{ref:11.11a} \begin{gathered} \bar{F}:\Cscr\r\Dscr:[C]\mapsto \sum (-)^i [H^i(FC)]\\ \bar{G}:\Dscr\r \Cscr:[D]\mapsto \sum (-)^i [H^i(GD)] \end{gathered}$$ define isomorphisms between the Grothendieck groups of $\Cscr$ and $\Dscr$. Let $\bar{\Cscr}$ and $\bar{\Dscr}$ be the closures of $\Cscr$ and $\Dscr$ under extensions. Then clearly $F$ and $G$ define inverse equivalences between $D^b_{\bar{\Cscr}}(\Ascr)$ and $D^b_{\bar{\Dscr}}(\Bscr)$. Hence we obtain (using standard isomorphisms for Grothendieck groups) $$K_0(\Cscr) \cong K_0(\bar{\Cscr})\cong K_0(D^b_{\bar{\Cscr}}(\Ascr)) \cong K_0(D^b_{\bar{\Dscr}}(\Bscr))\cong K_0(\bar{\Dscr})=K_0(\Dscr)$$ The composition of these isomorphisms (and their inverses) is given by . We deduce the following result. Let $p_1,\ldots,p_n\in Y$, $n\le 8$ be such that $(\sigma p_i)_i$ are in general position. Let $z\in \NN(Y/\langle \tau\rangle)$. Then $K_0(M_z(\tilde{X_i}))$ and $K_0(M_z(X_{i+1}))$ are isomorphic for $i\le n$ (see §\[ref:6.9b\] for notations). By Theorem \[ref:11.1.3a\] we know that $L\delta_{i}^\ast$ and $R\delta_{i,\ast}$ are inverse equivalences between $D^b(\tilde{X}_{i})$ and $D^b(X_{{i+1}})$. Now using the explicit construction of the functors $T_p$ in Proposition \[ref:5.7.2a\] it is easy to verify that for $\Mscr\in \trans_Y(X_{{i}+1})$ we have $T_p(\Mscr)=T_p(\delta_{i}^\ast\Mscr)$. Similarly using Proposition \[ref:7.2.4a\] we have for $\Nscr\in \trans_Y(\tilde{X})$ the identity $T_p(\delta_{{i},\ast}\Nscr)=T_p(\Nscr)$. Finally using Theorem \[ref:9.1.9a\] we find that the higher derived functors of $\delta^\ast_{i}$ and $\delta_{{i},\ast}$ map $\mod(X_{{i}+1})$ to $\iso_Y(\tilde{X}_{{i}})$ and $\mod(\tilde{X}_{i})$ to $\iso_Y(X_{i+1})$. It follows that we can apply lemma \[ref:11.2.2a\] to obtain an isomorphism between $M_z(\tilde{X}_{i})$ and $M_z(X_{{i}+1})$. We now obtain. \[ref:11.2.4a\] Let $p_1,\ldots,p_n\in Y$, $n\le 8$ be such that $(\sigma p_i)_i$ are in general position. Let $z$ and $O$ be as in the statement of Theorem \[ref:11.2.1a\] (where we identify $\NN F$ with $\NN(Y/\langle\tau\rangle)$). Then the number of non-isomorphic exceptional simple objects in $\mod(X_{n+1})$ is equal to $ n+\rk K_0(M_z(X_0))-O $. By iterating Theorem \[ref:6.9.1a\] and lemma \[ref:11.2.4a\] we easily find $$\rk K_0(M_0(X_{n+1}))=n+\rk K_0(M_z(X_1))-O$$ It is now easy to see that $M_0(X_{n+1})$ is equivalent to $\iso_Y(X_{n+1})$. Clearly the exceptional simple objects in $\mod(X_{n+1})$ coincide with the simple objects in $\iso_Y(X_{n+1})$. It now suffices to check that $\iso_Y(X_{n+1})$ is a finite length category. Note that by construction $X_{n+1}=\Proj A_{n+1}$ where $A_{n+1}$ is a noetherian graded ring containing a regular central element $t$ in degree one (see §\[ref:9.1a\] and in particular Prop. \[ref:9.1.3a\]). If $M$ is a graded $A_{n+1}$-module and $\Mscr=\pi M$ then multiplication by $t$ corresponds to the map $\Mscr(-Y)\r \Mscr$. It easily follows that if $M$ is a graded $A_{n+1}$-module of Gelfand-Kirillov dimension $>1$ then $\pi M$ is not in $\iso_Y(X_{n+1})$. Hence objects in $\iso_Y(X_{n+1})$ come from graded $A$-modules of Gelfand-Kirillov dimension one. By considering multiplicity one finds that the latter form a finite length category when viewed in $\Proj A_{n+1}$. By lemma \[ref:11.2.4a\] we have to compute $K_0(M_z(X_0))$ in the special case that $\|z\|\le 6$ (by our assumption that $n\le 6$). Remember that $X_0=\Proj A$ where $A$ is a three dimensional elliptic Artin-Schelter regular algebra. By considering the action of the central element in degree three it is easily seen that every object in $\trans_Y(X_0)$ is of the form $\pi M$ with $\operatorname{GKdim}M\le 2$. Since the graded $A$-modules with $\operatorname{GKdim}\le 1$ correspond to objects in $\Cscr_f$ [@ATV1] it follows by considering multiplicity that $\trans_Y(X_0)/\Cscr_f$ is a finite length category. The same holds for $M_z(X_0)$ so a basis for $K_0(M_z(X_0))$ is given by the isomorphism classes of simple objects. These simple objects have been classified in Theorem \[ref:10.2.6a\]. There are in 1-1 correspondence with the following two sets. $$\begin{aligned} A'_z&=\{D\in \NN(Y/\langle\tau\rangle)\mid D\le z, \text{$D$ is compatible with $L$}\}\\ B'_z&=\{D\in\NN(Y/\langle\tau\rangle)\mid D\le z,\text{$D$ is compatible with $2L$}\\ &\qquad \text{ and $D$ is not a sum $D_1+D_2$ with $D_i$ compatible with $L$}\}\end{aligned}$$ where $L$ represents the divisor of a line in $\PP^2$. Thus $\rk K_0(M_z(X_{n+1}))=\|A'_z\|+\|B'_z\|$. It is now easy to see that $A_z$ is in bijection with $A'_z$ and similarly $B_z$ is in bijection with $B'_z$. This finishes the proof. Non-commutative cubic surfaces {#ref:12a} ============================== In this section we recycle notations and assumptions from section §\[ref:11a\]. We will however assume in addition that $n=6$. Thus will fix six points $(p_{j})_{j}$ on $Y$ and our aim will be to study $X_7$. Since we will make no use of Theorem \[ref:11.1.3a\], we will not assume that the points $(\sigma p_{j})$ are in general position. Furthermore we will also not assume that $\tau$ has infinite order. By construction $X_7=\Proj F$ for a certain graded $k$-algebra $F$. Since the hypotheses for Proposition \[ref:9.2.3a\] hold on $X_6$ (lemma \[ref:11.1.1a\]) we find by that proposition that $F$ contains a regular central element $t$ in degree one such that $\bar{F}=F/tF$ is the twisted homogeneous coordinate ring [@AVdB] associated to the triple $(Y,\Nscr_7,\tau)$. Here $\Nscr_7$ is a line bundle of degree $9-6=3$, defined by . From these data we can compute the Hilbert-series of $F$. We find $H(F,s)=(1-s^3)/(1-s)^4$. This suggests that $X_7$ should be viewed as a non-commutative cubic surface. In this section we substantiate this intuition by showing that there exists a 4-dimensional Artin-Schelter regular algebra $P$ [@AS], containing a normal element $C$ in degree three such that $F=P/(C)$. We can then view $\Proj P$ as a quantum $\PP^4$ which contains $X_7$ as a cubic divisor. Of course the commutative analogue of this is well-known. If one blows up 6 points in general position in $\PP^2$ then one obtains a cubic surface in $\PP^4$ [@H]. However the reader may wonder why, in the non-commutative case, we don’t need that our points are in general position. The explanation is of course that $X_7$ is not a straight blowing up of $X_1$, but is constructed by repeatedly applying the constructions $X_{j}\mapsto \tilde{X}_{{j}}\mapsto X_{{j}+1}$. In the commutative case, if the points are in general position, then $\delta_{j}$ is an isomorphism between $\tilde{X}_{j}$ and $X_{{j}+1}$ and so in that case $X_7$ is indeed a straight blowing up of $X_1$. This will in general not be true in the non-commutative case, except in a derived sense. See Theorem \[ref:11.1.3a\]. The construction of $P$ is easy. It follows from [@ATV1] that $\bar{F}$ has a (minimal) presentation $$\bar{F}=k[x_1,x_2,x_3]/(r_1,r_2,r_3,C_3)$$ where $\deg x_1=1$, $\deg r_i=2$, $\deg C_3=3$. One deduces that $F$ has a presentation $$F=k[x_1,x_2,x_3,t]/(r'_1,r'_2,r_3',C_3',[t,x_1],[t,x_2],[t,x_3])$$ where $r'_i$, $C'_3$ are homogeneous liftings of $r_i$, $C_3$. We now put $$P=k[x_1,x_2,x_3,t]/(r'_1,r'_2,r'_3,[t,x_1],[t,x_2],[t,x_3])$$ and we will show that $P$ is Artin-Schelter regular. To this end we make use of the fact that by [@ATV1] one knows that $\bar{P}=P/tP=k[x_1,x_2,x_3]/(r_1,r_2,r_3)$ is a three-generator three-dimensional Artin-Schelter regular algebra [@AS]. We can then use the criterion [@LSV Cor. 2.7]. To state this criterion we let $R_P$ and $R_{\bar{P}}$ stand for the relations of degree two in $P$ and $\bar{P}$. Suppose we have the following 1. Left and right multiplication by $t$ is injective on $P_1$. 2. The image of $(P_1\otimes R_P)\cap (R_P\otimes P_1)$ under the natural map $P^{\otimes 3}\r \bar{P}^{\otimes 3}$ is $(\bar{P}_1\otimes R_{\bar{P}})\cap (R_{\bar{P}}\otimes \bar{P}_1)$. Then according to [@LSV Cor. 2.7], $P$ will be Artin-Schelter regular with Hilbert series $1/(1-s)^4$. We now verify conditions 1.,2. Condition 1. follows from the observations that $P_{\le 2}=F_{\le 2}$ and that $F$ is a domain since $\bar{F}$ is a domain. Hence we concentrate on condition 2. To simplify the notations we put $V=\sum_i kx_i$, $R_2=\sum_i kr_i\subset V^{\otimes 2}$, $R_3=kC_3\subset V^{\otimes 3}$, $W=V\oplus kt$, $S_2=(\sum_i kr'_i)+ (\sum_ik[t,x_i])\subset W^{\otimes 2}$, $S_3=kC_3'\subset W^{\otimes 3}$. With these notations $$\begin{aligned} \bar{F}&=k[V]/(R_2\oplus R_3)\\ F&=k[W]/(S_2\oplus S_3)\\ P&=k[W]/(S_2)\\ \bar{P}&=k[V]/(R_2)\end{aligned}$$ Note that $R_P=S_2$, $R_{\bar{P}}=R_2$. We first claim that the following complexes are exact in degrees $\le 3$. $$\label{ref:12.1a} 0\r (V\otimes R_2\cap R_2\otimes V)\otimes \bar{F}\r (R_2\oplus R_3)\otimes \bar{F} \r V\otimes \bar{F}\r \bar{F}\r k \r0$$ $$\label{ref:12.2a} 0\r (W\otimes S_2\cap S_2\otimes W)\otimes F\r (S_2\oplus S_3)\otimes {F} \r W\otimes {F}\r {F}\r k \r0$$ Let us first consider . The only place where exactness is non-obvious is at $(R_2\oplus R_3)\otimes \bar{F}$. Hence it is sufficient to show that the alternating sum of the Hilbert series of the terms in is zero in degrees $\le 3$. This easily follows from the fact $$\dim (V\otimes R_2\cap R_2\otimes V)=1$$ which is true because $\bar{P}$ is Koszul. We use the same method to check the exactness of . This time we need $$\dim(W\otimes S_2\cap S_2\otimes W)=4$$ Now we know that the dimensions of $$\begin{aligned} F_2&=W^{\otimes 2}/ (S_2)\\ F_3&=W^{\otimes 3} /(S_3+W\otimes S_2+ S_2\otimes W)\end{aligned}$$ are equal to $10$ and $19$ respectively. This yields that $$\begin{aligned} \dim S_2&=16-10=6\\ \dim (S_3+W\otimes S_2+ S_2\otimes W)&=64-19=45\end{aligned}$$ Since by [@ATV1] we have $R_3\cap (V\otimes R_2+R_2\otimes V)=0$, it also follows that $ S_3\cap(W\otimes S_2+S_2\otimes W)=0 $. Thus $$\dim (W\otimes S_2+S_2\otimes W)=44$$ Hence we obtain that $$\dim(W\otimes S_2\cap S_2\otimes W)=4\times 6+6\times 4-44=4$$ This proves what we want. Now we tensor with $\bar{F}$ and we combine the result with to form the following commutative diagram. $$\begin{CD} @. @. W\otimes \bar{F} @>\gamma>> k\otimes \bar{F} @. @. @.\\ @. @. @V\alpha VV @V\beta VV @. @. @.\\ 0 @>>> (W\otimes S_2\cap S_2\cap W)\otimes \bar{F} @>>> (S_2\oplus S_3)\otimes \bar{F} @>>> W\otimes \bar{F} @>>> \bar{F} @>>> k@>>> 0\\ @. @VVV @VVV @VVV @VVV @.\\ 0 @>>> (V\otimes R_2\cap R_2\cap V)\otimes \bar{F} @>>> (R_2\oplus R_3)\otimes \bar{F} @>>> V\otimes \bar{F} @>>> \bar{F} @>>> k @>>> 0\\ @. @. @VVV @VVV @. @. @.\\ @. @. 0 @. 0 @. @. @.\\ \end{CD}$$ $\alpha$, $\beta$ and $\gamma$ are defined by $$\begin{aligned} \alpha(w\otimes 1)&=(tw-wt)\otimes 1\\ \beta(1\otimes 1)&=t\otimes 1\\ \gamma(w\otimes 1)&=1\otimes \bar{w}\end{aligned}$$ The homology of the middle complex is given by $\Ext^i_F(k,\bar{F})$ in degrees $\le 3$. In particular this complex is exact at $(S_2\oplus S_3)\otimes \bar{F}$. It is also clear that $\gamma$ is surjective in degrees $\ge 1$. A trivial diagram chase now shows that $$W\otimes S_2\cap S_2\otimes W\r V\otimes R_2\cap R_2\otimes V$$ is surjective. This completes the proof of conditions 1. and 2. above. So at this point we know that $P$ is Artin-Schelter regular with Hilbert series $1/(1-s)^4$. We still have to show that $C'_3$ is a regular normalizing element in $P$. By looking at Hilbert series it is clear that $t$ is a regular central element in $P$. Hence since $\bar{P}$ is a domain by [@ATV1], the same holds for $P$. In particular $C'_3$ is regular in $P$. Looking at Hilbert series of $P$ and $F$ reveals that the twosided ideal $(C_3')$ in $P$ must be free of rank one on the left and on the right. Hence $(C_3')=C_3'P=PC_3'$ and thus $C_3'$ is normalizing. This completes the proof. Two-categories {#ref:Aa} ============== A 2-category is a category where the homsets themselves are categories. The objects of such a category are called 0-cells, the arrows are called 1-cells and the arrows between arrow are called 2-cells. Such 2-cells are drawn as follows $$\begin{picture}(40,20) \put(2,10){\makebox(0,0){$A$}} \put(38,10){\makebox(0,0){$B$}} \put(20,2){\makebox(0,0){$g$}} \put(20,18){\makebox(0,0){$f$}} \put(20,10){\makebox(0,0){$\nu$}} \put(10,5){\vector(1,0){20}} \put(10,15){\vector(1,0){20}} \put(5,10){\circle{2}} \put(35,10){\circle{2}} \end{picture}$$ As usual arrows can be composed, and so can 2-cells. It turns out that 2-cells even have two compositions. Vertical ones $$\begin{picture}(30,20) \put(5,0){\vector(1,0){20}} \put(5,10){\vector(1,0){20}} \put(5,20){\vector(1,0){20}} \put(0,10){\circle{2}} \put(30,10){\circle{2}} \put(15,5){\makebox(0,0){$\mu$}} \put(15,15){\makebox(0,0){$\nu$}} \end{picture}$$ denoted by $\mu\cdot \nu$, which come from the composition in $\Hom(A,B)$ and horizontal ones $$\begin{picture}(60,10) \put(0,5){\circle{2}} \put(30,5){\circle{2}} \put(60,5){\circle{2}} \put(15,5){\makebox(0,0){$\mu$}} \put(45,5){\makebox(0,0){$\nu$}} \put(5,0){\vector(1,0){20}} \put(35,0){\vector(1,0){20}} \put(5,10){\vector(1,0){20}} \put(35,10){\vector(1,0){20}} \end{picture}$$ denoted by $\mu\nu$ which come from the fact that the pairing $\Hom(B,C)\times \Hom(A,B)\r \Hom(A,C)$ has to be a bifunctor. Between those two compositions there is a natural compatibility. Assume that one has the following diagram $$\begin{picture}(60,20) \put(5,0){\vector(1,0){20}} \put(5,10){\vector(1,0){20}} \put(5,20){\vector(1,0){20}} \put(0,10){\circle{2}} \put(30,10){\circle{2}} \put(15,5){\makebox(0,0){$\mu_1$}} \put(15,15){\makebox(0,0){$\nu_1$}} \put(35,0){\vector(1,0){20}} \put(35,10){\vector(1,0){20}} \put(35,20){\vector(1,0){20}} \put(60,10){\circle{2}} \put(45,5){\makebox(0,0){$\mu_2$}} \put(45,15){\makebox(0,0){$\nu_2$}} \end{picture}$$ then one has $\mu_1\mu_2\cdot\nu_1\nu_2=(\mu_1\cdot\nu_1)(\mu_2\cdot\nu_2)$. Of course since a set is a category with only the identity arrows we can consider every category trivially as a 2-category. The archetypical example of a 2-category is “$\mathbf{Cat}$,” the category of all categories. In this case the objects are the categories (living in some universe), the arrows are the functors and the 2-cells are the natural transformations. It is therefore not surprising that the standard properties of categories and functors can be mimicked inside a 2-category. For example an arrow $A\xrightarrow{f}B$ is a left adjoint of an arrow $B\xrightarrow{g}A$ if there is a unit $\eta:\Id_{A}\r gf$ and a counit $\epsilon:fg\r \Id_B$ satisfying the standard associativity conditions. As usual $g$ is determined by $f$ up to unique isomorphism. If in this situation the $\eta$ and $\nu$ are isomorphisms then we call $f,g$ inverse equivalences and we say that $A$ and $B$ are equivalent. In a 2-category it is natural not only to consider ordinary commutative diagrams (so-called “strict” commutative diagrams) but also pseudo-commutative diagrams. These diagrams are commutative up to explicit isomorphism. For example the notation $$\label{ref:A.1a} \begin{picture}(40,40) \put(2,20){\makebox(0,0){$h$}} \put(20,2){\makebox(0,0){$j$}} \put(38,20){\makebox(0,0){$f$}} \put(20,38){\makebox(0,0){$g$}} \put(20,20){\makebox(0,0){$\nu$}} \put(10,5){\vector(1,0){20}} \put(5,30){\vector(0,-1){20}} \put(10,35){\vector(1,0){20}} \put(35,30){\vector(0,-1){20}} \put(5,5){\circle{2}} \put(35,5){\circle{2}} \put(5,35){\circle{2}} \put(35,35){\circle{2}} \put(12.5,32.5){\vector(-1,-1){5}} \end{picture}$$ means that there is an isomorphism $fg\xrightarrow{\nu}jh$. Often $\nu$ is clear from the context. In such a case we sometimes tacitly ignore $\nu$ and treat as a real commutative diagram. The naturality of such pseudo-commutative diagrams is reflected in the definition of a pseudo-functor [@KS] between 2-categories, which we give below. Assume that $\Cscr$, $\Dscr$ are 2-categories. A pseudo-functor $T:\Cscr\r \Dscr$ associates to every object of $\Cscr$ an object of $\Dscr$, to every arrow $f:A\r B$ of $\Cscr$ an arrow $T(f):T(A)\r T(B)$ of $\Dscr$ and to every 2-cell $\nu:f\r g$ a two-cell $T(\nu):T(f)\r T(g)$. If $T$ were an ordinary functor then we would require that for compositions of arrows $fg$ one has $T(fg)=T(f)T(g)$. However for a pseudo-functor we only require the existence of isomorphisms $\eta_{f,g}:T(f)T(g)\r T(fg)$ which we consider as being part of the description of $T$. The data describing $T$ has to satisfy a list of compatibilities which may be summarized by saying that every diagram that can commute must commute. One can go on and define natural transformations between pseudo-functors and even natural transformations between natural transformations (“modifications”). In this way pseudo-functors between 2-categories form themselves a 2-category and the category of all 2-categories is a 3-category! We will call a pseudo-functor $S:\Cscr\r \Dscr$ an equivalence if for every object $D$ in $\Dscr$ there exists an object $C$ in $\Cscr$ such that $D$ is equivalent to $S(C)$ (essential surjectivity) and for all objects $A,B$ in $\Cscr$ the canonical map $$\Hom_\Cscr(A,B)\r \Hom_\Dscr(SA,SB)$$ is an equivalence of categories. As usual $\Cscr$ and $\Dscr$ are said to be equivalent if there exists an equivalence $S:\Cscr\r \Dscr$. One verifies that such an $S$ has a quasi-inverse (in an appropriate sense) and hence “equivalence of two-categories is symmetric”. An example of a pseudo-functor between 2-categories is given by adjunction. Assume that $\Cscr$ is a 2-category in which every arrow $f$ possesses a right adjoint $Rf$. Then $R$ defines a pseudo-functor $R:\Cscr\r \Cscr^{\text{opp}}$. If every arrow $f$ has also a left adjoint $Lf$ then $R$, $L$ are inverse equivalences of 2-categories between $\Cscr$ and $\Cscr^{\text{opp}}$. One more bit of notation. If $\Cscr$ is a 2-category and $A$ is an object of $\Cscr$ then the relative category $\Cscr/A$ is the 2-category of pairs $(B,f)$ where $B$ is an object of $\Cscr$ and $f:B\r A$ is an arrow. An arrow $(B,f)\r (C,g)$ in $\Cscr/A$ is given by an arrow $h:B\r C$ together with an isomorphism $\mu:f\r gh$. A 2-cell $(h,\mu)\r (h',\mu')$ is given by an isomorphism $\nu:h\r h'$ such that $(\Id_g\nu)\cdot \mu=\mu'$. \#1\#2[0to 0[\#1]{}]{} Summary of notations ==================== =Section\ $k$ an algebraically closed field §\[ref:1a\]\ $\Inj(\Dscr)$injective objects in $\Dscr$ §\[ref:3.1a\]\ $\Lscr(\Dscr,\Cscr)$left exact functors from $\Dscr$ to $\Cscr$ §\[ref:3.1a\]\ $\BIMOD(\Cscr-\Dscr)$the opposite categorie of $\Lscr(\Dscr,\Cscr)$ §\[ref:3.1a\]\ $\Bimod(\Cscr-\Dscr)$objects in $\BIMOD(\Cscr-\Dscr)$ that have a left adjoint §\[ref:3.1a\]\ $\otimes$composition of bimodules §\[ref:3.1a\]\ $\MOD(\Cscr)$the category $\BIMOD(\Ab-\Cscr)$ §\[ref:3.1a\]\ $\HHom_\Cscr(\Mscr,-)$ the left exact functor represented by the bimodule $\Mscr$ §\[ref:3.1a\]\ $\HExt$the derived functor of $\HHom$ §\[ref:3.1a\]\ $\HTor$a kind of derived functor of “$\otimes$” §\[ref:3.1a\]\ $\ALG(\Dscr)$the algebra objects in $\BIMOD(\Dscr-\Dscr)$ §\[ref:3.1a\]\ $\Alg(\Dscr)$the algebra objects in $\Bimod(\Dscr-\Dscr)$ §\[ref:3.1a\]\ $\Mod(\Ascr)$the module category of the algebra $\Ascr$ §\[ref:3.1a\]\ ”$\invlim$”the virtual inverse limit §\[ref:3.1a\]\ $\BIGR(\Cscr-\Dscr)$graded “weak” $\Cscr$-$\Dscr$ bimodules §\[ref:3.2b\]\ $\Bigr(\Cscr-\Dscr)$graded $\Cscr$-$\Dscr$ bimodules §\[ref:3.2b\]\ $\GRALG(\Dscr)$the algebra objects in $\BIGR(\Dscr-\Dscr)$ §\[ref:3.2b\]\ $\Gralg$the algebra objects in $\Bigr(\Dscr-\Dscr)$ §\[ref:3.2b\]\ $\Gr(\Ascr)$the graded modules over the algebra $\Ascr$ §\[ref:3.2b\]\ $\Sscr(\Ascr)$a certain Serre subcategory of $\Gr(\Ascr)$ §\[ref:3.3b\]\ $i_\ast$, $i^\ast$, $i^!$functors associated to an inclusion $i$ of quasi-schemes §\[ref:3.4b\]\ $\Mod(X)$the category of objects associated to a quasi-scheme $X$ §\[ref:3.6b\]\ $\Oscr_X$a distinguished object for an “enriched” quasi-scheme. §\[ref:3.6b\]\ $\Gamma(X,-)$notation for the functor $\Hom_X(\Oscr_X,-)$ §\[ref:3.6b\]\ $o_X$the bimodule on $X$ represented by the identity functor §\[ref:3.6b\]\ $\Bimod(X)$notation for $\Bimod(\Mod(X)-\Mod(X))$ §\[ref:3.6b\]\ $\Alg(X)$notation for $\Alg(\Mod(X)$ §\[ref:3.6b\]\ $\Sch$the category of quasi-compact, quasi-separated schemes §\[ref:3.6b\]\ $\QSch$the category of quasi-schemes §\[ref:3.6b\]\ $\QSch/X$the category of quasi-schemes over $X$ §\[ref:3.6b\]\ $\Spec \Ascr$a quasi-scheme with module category $\Mod(\Ascr)$ §\[ref:3.6b\]\ $o_X(-Y)$a subbimodule of $o_X$ associated to a divisor $Y\subset X$ §\[ref:3.7b\]\ $o_X(nY)$notation for $o_X(-Y)^{\otimes -n}$ §\[ref:3.7b\]\ $\Mscr(nY)$notation for $\Mscr\otimes o_X(nY)$ §\[ref:3.7b\]\ $\Nscr_{Y/X}$the “normal bundle” of $Y$ in $X$ §\[ref:3.7b\]\ $\Tors_Y(X)$, $\Iso_Y(X)$certain categories associated to $Y\subset X$ §\[ref:3.7b\]\ $\Tors(\Ascr)$torsion modules over a graded algebra $\Ascr$ §\[ref:3.8b\]\ $\operatorname{QGr}(\Ascr)$the category $\Gr(\Ascr)/\Tors(\Ascr)$ §\[ref:3.8b\]\ $\tau$the “torsion functor” §\[ref:3.8b\]\ $\pi$the quotient functor $\Gr(\Ascr)\r\operatorname{QGr}(\Ascr)$ §\[ref:3.8b\]\ $\omega$the right adjoint to $\pi$ §\[ref:3.8b\]\ $(\tilde{-})$the composition $\omega\pi$ §\[ref:3.8b\]\ $\Proj \Ascr$A quasi-scheme whose category is $\operatorname{QGr}(\Ascr)$ §\[ref:3.8b\]\ $\Pqsch/X$“projective” quasi-schemes over $X$ §\[ref:3.8b\]\ $Q\Sscr(\Ascr)$the image of $\Sscr(\Ascr)$ in $\operatorname{QGr}(\Ascr)$ §\[ref:3.11b\]\ $\alpha^{-1}(\Sscr)$an alternative notation for $Q\Sscr(\Ascr)$ §\[ref:3.11b\]\ $\Ascr^{(n)}$the $n$’th Veronese of $\Ascr$ §\[ref:3.12b\]\ $L_i\alpha^\ast$a kind of derived functor to $\alpha^\ast$ §\[ref:3.10b\]\ $\operatorname{PC}(A)$the category of pseudo-compact $A$-modules §\[ref:4a\]\ $\operatorname{Top}(A)$the category of topological $A$-modules §\[ref:4a\]\ $\operatorname{Dis}(A)$the category of discrete $A$-modules §\[ref:4a\]\ $\operatorname{PCFin}(A)$the category of pseudo-compact finite length modules §\[ref:4a\]\ $\operatorname{PC}(A-B)$the category of pseudo-compact $A$-$B$-bimodules §\[ref:4a\]\ $\ctimes$completed tensor product §\[ref:4a\]\ $X$usually a fixed quasi-scheme §\[ref:5.1a\]\ $Y$usually a fixed commutative curve which is a divisor in $X$ §\[ref:5.1a\]\ $p$usually a fixed point on $Y$ §\[ref:5.1a\]\ $\tau$an automorphism of $Y$ associated to $\Nscr_{Y/X}$ §\[ref:5.1a\]\ $\Nscr_\tau$the twisting of the line bundle $\Nscr$ by the automorphism $\tau$ §\[ref:5.1a\]\ $O_\tau(p)$the $\tau$-orbit of $p$ §\[ref:5.1a\]\ $\Oscr_p$the object in $\Mod(X)$ corresponding to $p\in Y$ §\[ref:5.1a\]\ $\Cscr_f$finite length objects supported on $Y$ §\[ref:5.1a\]\ $\Cscr$direct limits of objects in $\Cscr_f$ §\[ref:5.1a\]\ $\Cscr_p$objects in $\Cscr$ supported on the $\tau$-orbit of $p$ §\[ref:5.1a\]\ $\Cscr_{f,p}$the finite length objects in $\Cscr_p$ §\[ref:5.1a\]\ $C_p$the pseudo-compact ring associated to $\Cscr_p$ §\[ref:5.1a\]\ $N$a canonical normal element in $C_p$ §\[ref:5.1a\]\ $\hat{(-)}_p$the completion functor §\[ref:5.1a\],§\[ref:5.3a\],§\[ref:5.4b\]\ $R$the completion of $\Oscr_Y$ at $p$ §\[ref:5.1a\]\ $m$the maximal ideal of $R$ §\[ref:5.1a\]\ $U$the generator of the maximal ideal of $R$ §\[ref:5.1a\]\ $m_i$the $i$’th maximal ideal of $C_p$ §\[ref:5.1a\]\ $S_i$the $i$’th simple $C_p$-module §\[ref:5.1a\]\ $e_i$the $i$’th diagonal primitive idempotent in $C_p$ §\[ref:5.1a\]\ $P_i$the $i$’th pseudo-compact projective of $C_p$ §\[ref:5.1a\]\ $o_q$the bimodule on $X$ corresponding to $q\in Y$ §\[ref:5.1a\]\ $\cohBIMOD(o_X-o_X)$coherent bimodules on $X$ §\[ref:5.5b\]\ $\tilde{\Cscr}_{f,p}$finite extensions of $o_{\tau^i p}$ §\[ref:5.5b\]\ $\trans_Y(X)$objects “transversal” to $Y$ §\[ref:5.5b\]\ $\Div(\Fscr)$the divisor on $Y$ associated to an object in $\mod(X)$ §\[ref:5.5b\]\ $\Fscr_Y$notation for $\Fscr/\Fscr(-Y)$ §\[ref:5.7b\]\ $T_p(\Fscr)$an invariant for $\Fscr\in\trans_Y(X)$ §\[ref:5.7b\]\ $N_p(\Fscr)$the $p$-normalization of $\Fscr$ §\[ref:5.7b\]\ $m_q$the ideal in $o_X$ associated to $o_q$ §\[ref:6.1a\]\ $m_{Y,q}$the ideal in $o_Y$ associated to $o_q$ §\[ref:6.1a\]\ $I$notation for $m_p(Y)$ §\[ref:6.1a\]\ $I_Y$notation for $m_{Y,p}(Y)$ §\[ref:6.1a\]\ $\mu$the multiplicity of $p$ on $Y$ §\[ref:6.1a\]\ $\Dscr$the Rees algebra for $I$ §\[ref:6.2b\]\ $\Dscr_Y$the Rees algebra for $I_Y$ §\[ref:6.2b\]\ $\tilde{X}$the blowup of $X$ in $p$ §\[ref:6.3b\]\ $\tilde{Y}$the strict transform of $Y$ in $\tilde{X}$ §\[ref:6.3b\]\ $\alpha$, $\beta$, $i$, $j$maps in a diagram relating $X$,$\tilde{X}$, $Y$,$\tilde{Y}$ §\[ref:6.3b\]\ $\tau'$the lifting of $\tau$ to $\tilde{Y}$ §\[ref:6.3b\]\ $\alpha^{-1}(\Cscr_p)$analog for the objects supported on the exceptional curve §\[ref:6.5b\]\ $L$the exceptional curve §\[ref:6.6b\]\ $\Mod(L)$the objects on $L$ §\[ref:6.6b\]\ $\Oscr_L$the “structure sheaf” on $L$ §\[ref:6.6b\]\ $\alpha^{-1}$the non-normalized strict transform §\[ref:6.8b\]\ $\alpha_s^{-1}$the normalized strict transform §\[ref:6.8b\]\ $l_p(\Fscr)$the Loewy length of $T_p(\Fscr)$ §\[ref:6.8b\]\ $M_z(X)$A certain subcategory of $\trans_Y(X)/\Cscr_f$ §\[ref:6.9b\]\ $D^\ast(Z)$, $D^\ast_f(Z)$derived categories of a quasi-scheme $Z$ §\[ref:10a\]\ $(Y,\Lscr,\sigma)$a “triple” in the sense of [@ATV1] §\[ref:10a\]\ $A$the Artin-Schelter regular algebra associated to $(Y,\Lscr,\sigma)$ §\[ref:10a\]\ $g$the canonical central element in $A$ §\[ref:10a\]\ $o_X(1)$the invertible bimodule corresponding to the shift on $\Gr(A)$ §\[ref:10a\]\ $r_{p,n}(\Mscr)$the multiplicities of $\Mscr$ at points infinitely near to $p$ §\[ref:10a\]\ $e(\Mscr)$the multiplicity of $\Mscr$ §\[ref:10a\]\ $\overline{\Div}(\Mscr)$a variant on $\Div(\Mscr)$ §\[ref:10a\] [10]{} K. Ajitabh, [Modules over regular algebras and quantum planes]{}, Ph.D. thesis, MIT, 1994. [to3em]{}, [Modules over elliptic algebras and quantum planes]{}, Proc. London Math. Soc. (3) [72]{} (1996), 567–587. K. Ajitabh and M. Van den Bergh, [Presentation of critical modules of gk-dimension 2 over elliptic algebras]{}, to appear in Proc. Amer. Math. Soc., 1998. M. Artin and J. T. Stafford, [Graded rings of [G]{}elfand-[K]{}irillov dimension two]{}, Invent. Math. [122]{} (1995), 231–276. M. Artin and J. J. Zhang, [Noncommutative projective schemes]{}, Adv. in Math. [109]{} (1994), no. 2, 228–287. M. Artin and W. Schelter, [Graded algebras of global dimension 3]{}, Adv. in Math. [66]{} (1987), 171–216. M. Artin, J. Tate, and M. Van den Bergh, [Some algebras associated to automorphisms of elliptic curves]{}, The Grothendieck Festschrift, vol. 1, Birkhäuser, 1990, pp. 33–85. [to3em]{}, [Modules over regular algebras of dimension 3]{}, Invent. Math. [ 106]{} (1991), 335–388. M. Artin and M. Van den Bergh, [Twisted homogeneous coordinate rings]{}, J. Algebra [188]{} (1990), 249–271. M. Artin, [A conjecture about graded algebras of dimension $3$]{}, to appear in the proceedings of the colloquim in honor of Maurice Auslander, 1996. A. Beilinson, [Coherent sheaves on [$\PP^n$]{} and problems of linear algebra]{}, Functional Anal. Appl. [12]{} (1978), 214–216. M. B[ö]{}ckstedt and A. Neeman, [Homotopy limits in triangulated categories]{}, Compositio Math. [86]{} (1993), 209–234. A. I. Bondal and A. Polischuk, [Homological properties of associative algebras: the method of helices]{}, Russian Acad. Sci. Izv. Math [42]{} (1994), 219–260. A. I. Bondal, [Representations of associative algebras and coherent sheaves]{}, Math. USSR-Izv. [34]{} (1990), no. 1, 23–42. W. Fulton, [Algebraic curves]{}, W. A. Benjamin, Inc, New York, Amsterdam, 1969. P. Gabriel, [Des catégories abéliennes]{}, Bull. Soc. Math. France [ 90]{} (1962), 323–448. A. Grothendieck, Bourbaki Exposé 234. A. Grothendieck, [Sur quelques points d’algèbre homologiques]{}, T[ô]{}hoku Math. J. (2) [9]{} (1957), 119–221. R. Hartshorne, [Residues and duality]{}, Lecture notes in mathematics, vol. 20, Springer Verlag, Berlin, 1966. [to3em]{}, [Algebraic geometry]{}, Springer-Verlag, 1977. G. M. Kelly and R. Street, [Review of the elements of 2-categories]{}, Category Seminar, vol. 420, Springer Verlag, 1972, pp. 75–103. L. Le Bruyn, S. P. Smith, and M. Van den Bergh, [Central extensions of three-dimensional [A]{}rtin-[S]{}chelter regular algebras]{}, Math. Z. [222]{} (1996), 171–212. L. Le Bruyn, [Conformal [$\mathfrak{sl}_2$]{} enveloping algebras]{}, Comm. Algebra [23]{} (1995), no. 4, 1325–1362. V. A. Lunts and A. L. Rosenberg, [Localization for quantum groups]{}, MPI preprint, 1996. S. MacLane, [Natural associativity and commutativity]{}, Rice Univ. Stud., vol. 49, Rice University, 1963. C. Nastacescu and F. Van Oystaeyen, [Graded ring theory]{}, North-Holland, 1982. D. O. Orlov, [Projective bundles, monoidal transformations and derived functors of coherent sheaves]{}, Russian Acad. Sci. Izv. Math [41]{} (1993), no. 1, 133–141. A. L. Rosenberg, [Non-commutative algebraic geometry and representations of quantized algebras]{}, Mathematics and its Applications, vol. 330, Kluwer Academic Publishers, Dordrecht, 1995. [to3em]{}, [The spectrum of abelian categories and reconstruction of schemes]{}, MPI preprint, 1996. S. P. Smith and J. J. Zhang, [Curves on non-commutative schemes]{}, preprint, University of Washington, 1997. B. Stenstr[ö]{}m, [Rings of quotients]{}, Die [G]{}rundlehren der mathematischen [W]{}issenschaften in [E]{}inzeldarstellungen, vol. 217, Springer Verlag, Berlin, 1975. R. Thomason and T. Trobaugh, [Higher algebraic [$K$]{}-theory of schemes and of derived categories]{}, The Grothendieck Festschrift, vol. 3, Birkhäuser, 1990, pp. 247–435. M. Van den Bergh, [A translation principle for [S]{}klyanin algebras]{}, J. Algebra [184]{} (1996), 435–490. [to3em]{}, [Abstract blowing down]{}, to appear in Proc. Amer. Math. Soc., 1998. M. Van den Bergh and M. Van Gastel, [Graded modules of [G]{}elfand-[K]{}irillov dimension one over three-dimensional [A]{}rtin-[S]{}chelter regular algebras]{}, to appear in J. Algebra. A. B. Verevkin, [On a non-commutative analogue of the category of coherent sheaves on a projective scheme]{}, Amer. Math. Soc. Transl. [151]{} (1992), 41–53. J. J. Zhang, [Twisted graded algebras and equivalences of graded categories]{}, Proc. London Math. Soc. (3) [72]{} (1996), no. 2, 281–311.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Co/BaTiO$_3$(001) is one of the most interesting multiferroic heterostructures as it combines different ferroic phases, setting this way the fundamentals for innovative technical applications. Various theoretical approaches have been applied to investigate the electronic and magnetic properties of Co/BaTiO$_3$(001). Here we determine the magnetic properties of 3 ML Co/BaTiO$_3$ by calculating spin-polarized electron diffraction as well as angle-resolved photoemission spectra, with both methods being well established as surface sensitive techniques. Furthermore, we discuss the impact of altering the BaTiO$_3$ polarization on the spectra and ascribe the observed changes to characteristic details of the electronic structure.' author: - Stephan Borek - Jürgen Braun - Ján Minár - Dimitry Kutnyakhov - 'Hans-Joachim Elmers' - Gerd Schönhense - Hubert Ebert bibliography: - 'spleed\_co\_bto\_new.bib' title: 'Determination of surface and interface magnetic properties for the multiferroic heterostructure Co/BaTiO$_3$ using SPLEED and ARPES' --- Introduction ============ The combination of two ferroic phases may lead to new materials with interesting behaviour and corresponding applications which are based on the electronic and magnetic properties of their constituents. This special class of materials is named multiferroics and was introduced first in the 1960’s by Smolenskii and Venevtsev [@SC82; @VG94]. In 1994 the interest in this topic has been renewed by H. Schmid [@schmidt1], who introduced first the synomym multiferroics. For this class of materials one has to distinguish between the combination of two or more ferroic properties in one single phase and the so-called multiferroic heterostructures. The latter ones are very interesting for technical applications because their material properties can be tuned by shaping the interface between the ferroic phases according to the technical requirements. An essential step in realizing such systems was the tremendous progress in developing crystal grow techniques which allow the creation of defined interfaces [@dawber1]. Among the various multiferroic heterostructures which have been studied so far one promising candidate is the Co/BaTiO$_3$ (Co/BTO) system [@HBM+15]. In contrast to the well-established multiferroic heterostructure 3 ML Fe/BTO the 3 ML Co/BTO system has the advantage of ferromagnetic stability at room temperature [@HBM+15]. In a recent publication it was shown that the interface of Co/BTO is very similar to that of the intensively studied Fe/BTO [@HBM+15]. It was discovered in detail that the coupling of the involved ferroic phases is caused by nearest neighbor interaction at the interface. For the investigated systems of 1, 2 and 3 ML Co/BTO the strongest coupling was predicted for 2 ML Co. Similar to the multiferroic heterostructure Fe/BTO the system Co/BTO opens the possibility to control the magnetic properties of the Co layer changing the electric polarization of BTO. The ferroelectric polarization could be altered using external voltage or stress. On the other hand it is possible to change the magnetic moments (direction or magnitude) of the Co layers affecting this way the ferroelectric polarization [@fechner1]. For the various properties predicted by theory a corresponding verification by experiment is mandatory. A well established way to do so is the use of spectroscopy. For example, it was shown that using x-ray magnetic circular dichroism (XMCD) and x-ray magnetic linear dichroism (XMLD), it is possible to detect the altering of the BTO polarization by investigating the magnetic properties of Co [@HBM+15; @BMF+12]. By comparing both methods it was shown that XMLD reacts more sensitive on a change of the electrical polarization. However, there exist other suitable methods for studying magnetic and electronic properties at surfaces, as for example, angle-resolved photoemission spectroscopy (ARPES) or spin-polarized low energy electron diffraction (SPLEED). In this work we performed first principles calculation for both methods to investigate the surface magnetic properties of 3 ML Co/BTO affected by an altering of the BTO polarization. The surface sensitive SPLEED technique allows for a somewhat indirect analysis of the coupling mechanisms at the interface. ARPES on the other hand enables a detailed investigation of the surface and interface electronic structure. Especially changes of the dispersion relation which occur during a polarization change of the BTO can be quantitatively monitored. To account for the high surface sensitivity of both methods we used a fully relaxed surface and interface structure for 3 ML Co/BTO. Our calculations have been done by use of fully relativistic multiple scattering techniques in the framework of spin-polarized density functional theory [@SPR-KKR6.3]. Therefore, all effects originating from spin-orbit coupling and exchange interaction are treated on the same level of accuracy [@ebert1; @SPR-KKR6.3]. The paper is organized as follows: In Sec. \[sec\_theory\] we briefly describe the theoretical methods used in this work, i.e. the SPLEED and ARPES formalism. In Sec. \[sec\_discussion\] we discuss our theoretical results and in Sec. \[sec\_summary\] we summarize our findings. Theoretical application \[sec\_theory\] ======================================= As described in our previous work we use the layer-KKR approach to the calculate the corresponding SPLEED pattern [@feder2; @BBM+15]. Within the scope of this method a semi-infinite surface system is treated as a stack of atomic layers using the so-called layer doubling method [@Pendry_SPLEED]. For every specific atomic layer the multiple scattering of the incident electron has to be determined. The potentials needed for this computational step have been taken from previous works [@Bor13; @HBM+15]. Based on these potentials the single-site scattering matrices can be evaluated, which are needed for the calculation of the so called Kambe X-matrix [@Kambe], which determines the multiple scattering in a specific layer. In a last step inter-layer scattering has to be considered. With the determination of both scattering mechanisms one can calculate the so-called bulk reflection matrix, which represents the scattering of an electron from a semi-infinite stack of atomic layers [@feder1; @feder2]. The self-consistent calculation of the electronic structure has been done for a half-space which consists out of four unit cells of BTO, i.e. beyond this bulk properties have been assumed. With respect to a transition from the inner BTO potentials to the surface the electronic structure was allowed to relax. This guarantees a realistic description of the electronic structure at the interface and surface. Similar to our investigations done for the multiferroic heterostructure Fe/BTO the calculations are based on the tetragonal distorted structure of BTO. In this phase (P4mmm) the BTO has a permanent electrical polarization originating from a shift of the Ti and O atoms [@CBS+11]. For BTO we assumed a lattice constant of 3.943 Å [@Bor13]. The relaxed crystal structure at the interface was determined using the VASP code [@KF96; @KJ99; @HBM+15]. We applied the resulting crystal structure and the the corresponding self-consistent potentials as input quantities for our fully relativistic multiple scattering formalism to calculate the density of states (DOS), the SPLEED pattern and ARPES intensities [@SPR-KKR6.3]. The scattering process itself includes different parameters. The most important one is the location of the scattering plane with respect to the sample surface. The scattering plane is defined by the wave vectors of the incident and scattered electrons. Additionally, the direction of the surface magnetization and the polarization of the incident electron have to be defined. In our calculation both have been aligned perpendicular to the scattering plane. As can be shown by symmetry considerations for this setup exchange and spin orbit scattering contribute both [@tamura1]. With respect to the used scattering configuration one has to calculate four reflectivities for all combinations of the electron polarization and the surface magnetization ($I_{\mu}^{\sigma}$) [@feder1; @tamura1]. The indices $\mu$ and $\sigma$ represent the surface magnetization direction and the electron polarization, respectively. Based on these quantities the effective reflectivity $I_{eff}$ can be calculated via: $$I_{eff}=\frac{1}{4}(I_+^++I_+^-+I_-^++I_-^-) \label{eq:effective_reflectivity}$$ The reflectivities ($I_{\mu}^{\sigma}$) enable the calculation of additional observables measured in the experiment, i.e. spin-orbit asymmetry, exchange asymmetry and figure of merit. The exchange asymmetry is defined as: $$A_{ex,+}=\frac{I_{+}^{+}-I_{+}^{-}}{I_{+}^{+}+I_{+}^{-}}. \label{eq:exchange_asymmetry}$$ The plus index at $A_{ex,+}$ indicates here a fixed magnetization direction along \[100\]. For all following calculations the magnetization was oriented along this direction. Additional quantities which are important for the calculation of SPLEED patterns determine the escape of the electron into the vacuum, i.e. the work function and the surface potential barrier. For the work function we assumed the value 4.7 eV, which has shown to be reasonable for diffraction calculations on similar systems [@BBE+15]. This choice is close to the experimental value of 5.0 eV [@haynes]. Concerning the surface barrier we have taken a parameterization based on Rundgren and Malmström [@skriver1; @rundgren1]. This type of surface barrier was applied in the past successfully to a variety of transition metal systems. Using the method described so far we have calculated SPLEED patterns for a wide range of kinetic energies and polar angles for both polarization directions of BTO (P$_{\text{up}}$, P$_{\text{down}}$). The ARPES calculations have been performed in the framework of the so-called one-step model [@Pendry_SPLEED; @Braun]. In contrast to the three-step model which treats the electron emission in different steps in the electron emission, i.e. excitation of the electron, travel of the electron through the solid to the surface, emission of the electron into the vacuum, all excitation steps are explicitly included in the one-step model. Beside the three-step model neglects several effects important for the theoretical description of the photoemission process. As for example the interference of surface and bulk emission, the interaction of the photoelectron with atoms during the transport to the surface and a proper description of the transition process of the electron into the vacuum. For a more detailed description of the ARPES calculations see for example [@braun1; @hueffner1]. We investigated a fully relaxed surface and interface structure with an (001) orientation of BTO. The crystal and electronic structure has been investigated in detail in previous works [@HBM+15]. Here we summarize only some essential structural parameters for the 3 ML Co/BTO interface (see Tab. \[crystal\_structure\]). In Fig. \[crystal\_structure\] the corresponding structure at the interface of the Co/BTO system is shown schematically. ![Crystal structure of 3 ML Co/BTO. The non-equivalent atomic sides have been indicated by name (not shown Co$\rm_{Ti}$ located on top of the Ti atom, color online).[]{data-label="fig:crystal_structure"}](cell2.eps) P$\rm_{up}$ P$\rm_{down}$ ----------------- ------------- --------------- O-Co$\rm_O$ 1.829 1.827 I-II 1.179 1.183 II-III 1.153 1.154 Ti-Co$\rm_{Ti}$ 3.014 3.094 : Atomic layer distances for the interface of 3 ML Co/BTO. The distances are given in units of (Å) [@HBM+15]. Both polarization directions of BTO are shown. BTO polarization pointing in surface direction (P$\rm_{up}$). BTO polarization pointing in substrate direction (P$\rm_{down}$). (I: Co layer on top of BTO, II: second Co layer, III: third Co layer representing the surface).[]{data-label="crystal_structure"} The crystal structure of the Co/BTO system shows many similarities when compared to the Fe/BTO system, which was investigated in previous studies [@fechner2; @BMF+12; @BBM+15]. The first Co layer (Co$_{\text{O}}$) is placed on top of the O atoms of the Ti-O terminated BTO surface. In the second Co layer two non-equivalent Co positions occur. The first position is on top of Ti (Co$_{\text{Ti}}$) and the second position is on top of Ba (Co$_{\text{Ba}}$). According to previous studies the largest changes when switching the BTO polarization result for Ti at the interface inducing a large displacement of the Co$_{\text{Ti}}$ atom [@MFO]. Therefore, the interaction of the Co$_{\text{Ti}}$ atom with the BTO substrate dominates the coupling between the ferroic materials and gives the main contribution to changes in the electronic structure when altering the BTO polarization. The third Co layer (Co$_{\text{I}}$) is placed on top of the interfacial O atoms. Concerning theses structural details 3 ML Co form an unusual tetragonal distorted based centered crystal structure with electronic properties that are closely related to 3 ML Fe/BTO [@fechner2]. Results and Discussion \[sec\_discussion\] ========================================== Due to the exchange scattering the magnetic properties at the Co surface are essential for our SPLEED calculations. The magnetic moments at the surface are affected by the reduced number of nearest neighbors and by hybridization effects with atoms at the interface. It has been shown that the magnetic properties depend on the number of Co layers on top of BTO [@HBM+15]. For the systems investigated so far (1, 2, 3 ML Co/BTO) a ferromagnetic ground state occurs. In actual investigations an in-plane configuration for 2 ML Co/BTO was predicted [@HBM+15]. Beside this it was shown that the Co spin magnetic moments are strongly affected by the geometry of the Co films [@HBM+15]. Especially, the Co spin magnetic moments are influenced by the hybridization of the Co 3$d$ states with states of the substrates atoms. It was shown that the spin magnetic moment for 3 ML Co/BTO are quenched for all three atomic layers which is based on the lower volume of the Co atoms according to the relaxed crystal structure [@HBM+15]. In Fig. \[spleed\] the SPLEED patterns for reflectivity (top row) and exchange asymmetry (bottom row) are shown. Both quantities have been calculated for an upward (in surface direction) oriented polarization of BTO (left column) and a downward oriented polarization of BTO (right column). ![SPLEED pattern for 3 ML Co/BTO. Reflectivity (top panel) and exchange asymmetry (bottom panel, values given in (%)). The plane of incidence was aligned along the \[010\] direction the magnetization of Co along \[100\]. Left: The polarization of the BTO is directed towards the surface (P$\rm_{up}$). Right: The polarization of the BTO pointing in opposite direction (P$\rm_{down}$) (color online).[]{data-label="spleed"}](reflectivity_asym.eps) As shown in Fig. \[spleed\] a dependency of the SPLEED pattern on the surface magnetization (i.e. the BTO polarization) occurs. For kinetic energies less than 6 eV the reflectivity contains large changes when switching the BTO polarization. This is in contrast to higher kinetic energies where the reflectivity is less sensitive to a BTO altering. This result is related to the coupling of incident electrons to bands which disperses perpendicular to the Co(001) surface. Depending on the band dispersion along the normal vector of the (001)-plane and the kinetic energy of the incident electron a coupling is possible or not [@BOC99]. Without coupling a large reflectivity results, with vanishing coupling the intensity of the reflected electrons decreases. The reflectivity pattern decompose in two main parts. Below 6 eV a high reflectivity occurs whereas for higher kinetic energies the reflectivity decreases. In addition, below 6 eV switching the BTO polarization has a higher impact on the effective reflectivity. This is caused by a higher sensitivity of electrons with lower kinetic energy to changes in the magnetic properties at the surface. According to Fig. \[spleed\], electrons with kinetic energies above 8 eV can couple with states inside the crystal resulting in a decrease of the effective reflectivity. In addition electrons with higher kinetic energy are less sensitivity to changes of the surface magnetization. For the Co/BTO system one can therefore expect that both effects contribute to the resulting SPLEED patterns. In Fig. \[spleed\] (bottom panel) the exchange asymmetries are presented. The exchange asymmetry gives a more detailed picture with additional information concerning the magnetic properties. This is due to the loss of information when calculating the effective reflectivity averaging over the individual spin-dependent electron beams. In the exchange asymmetry pattern, especially for higher kinetic energies (12-16 eV), pronounced changes in the polarization are visible. Additionally, the exchange asymmetry alters for low kinetic energies ($\le$ 4 eV). Therefore, the incident electron is affected by a change of the surface magnetization over the total kinetic energy range. In principle the occurrence of an exchange asymmetry indicates an energetically split unoccupied band structure [@OYY+08]. These exchange-split unoccupied bands are distinguishable via SPLEED. In Fig. \[spleed\] the red areas represent a parallel alignment of electron spin and surface magnetization whereas the blue areas stand for an antiparallel alignment. For special regions with positive (red) or negative (blue) exchange asymmetry the incident electrons can couple to majority or minority bands. According to the definition of the exchange asymmetry positive values correspond to a reflection of the incident electron from the majority (spin up) states. For example, with respect to kinetic energy and polar angle a parallel alignment of the electron spin and the surface magnetization (red areas) results in a coupling of the incident electron to the minority bands whereas the coupling to the majority bands is less pronounced. Therefore, the reflectivity of electrons with parallel aligned spin and surface magnetization is higher than for an antiparallel alignment resulting in a positive exchange asymmetry. At band gaps and band crossings the exchange asymmetry is zero. Based on this information it becomes extractable which spin projected bands have been affected significantly by a change of the BTO polarization. In Fig. \[spleed\] it can be seen that for kinetic energies of 14 eV and a polar angle of 60 deg a change of the BTO polarization affects the exchange asymmetry strongly. In this energy range the scattering is mainly due to majority bands (red color). For kinetic energies of 12 eV the same is visible for scattering at minority bands (blue color). Due to the fact that the splitting of the unoccupied states are related to the exchange splitting of the occupied states the SPLEED exchange pattern gives a fingerprint of the magnetic properties at the surface influenced by the BTO polarization. In addition we calculated angle-resolved photoemission spectra for 3 ML Co/BTO to visualize the impact of the BTO polarization for a sample path along the $\overline{\text{H}}$-$\overline{\Gamma}$-$\overline{\text{H}}$ direction along the surface Brillouin zone. The corresponding results for the two polarization directions of the BTO are shown in Fig. \[pes\]. ![One-step calculation of photoemission spectra for 3 ML Co/BTO. The calculations assume normal incidence, a photon energy of 60 eV and a \[100\] magnetization direction for the Co layers. Left: The polarization of the BTO is directed along the surface normal (P$\rm_{up}$). Right: The polarization of the BTO is directed in opposite direction (P$\rm_{down}$).[]{data-label="pes"}](E_k_intensity_up.eps) The different polarizations at the interface strongly influence the band dispersion. The main changes in the dispersion relation by altering the BTO polarization are visible at the Fermi energy. This reflects the changes of the magnetic properties due to a coupling between Co and BTO. Below -5 eV the band dispersion includes bands with quadratic $\textbf{k}$ dependence as expected from free electron-like behaviour. Above -5 eV the bands show the typical dispersion of $d$ bands for transition metals, i.e. a small band width resulting in high photoemission intensities. The band dispersion is closely related to the effective mass of the electrons representing the interaction with ion cores. For a large interaction the effective mass becomes large with respect to the free electron mass resulting in a small bandwidth [@christman]. These band dispersions are visible above -5 eV. Additionally, to the ARPES band dispersions we calculated the DOS for the Co/BTO system. In Fig. \[dos\] we plotted the results for atoms at the interface (BTO P$\rm_{up}$). ![DOS for Co/BTO(001) for the various atomic types near the interface. The polarization of the BTO is directed along the surface normal (P$\rm_{up}$). (color online)[]{data-label="dos"}](dos_interface.eps) Comparing the calculated ARPES spectra (left side, P$\rm_{up}$) with the plotted DOS it is visible that the main contribution to the bands in the energy range from -2 eV to -4 eV originates from O $p$ states, Co $d$ states and a small contribution of Ti $d$ states. A hybridization between these states is visible. In contrast the bands at the Fermi level are mainly Co $d$ states. The appearance of Co $d$ states from -6 eV to 1 eV is due to the exchange splitting, i.e. the spin magnetic moment of Co. With respect to Fig. \[dos\] the main contribution to the unoccupied states arises from $d$ states of Ti ($\approx$ 3 eV) and $d$, $f$ states of Ba ($\ge$ 6 eV). It is also visible that the magnetic properties of the Co/BTO system can be related to the Co atoms where for Ti and O due to hybridization effects a small magnetic moment is induced [@HBM+15]. As has been shown in our previous work on Fe/BTO it is reasonable to assume that the incident electron interacts only with the first atomic layer of Co. Therefore, the induced magnetic moments for Ti and O can be neclegted concerning the interpretation of the characteristic structures in the SPLEED pattern. In previous works the coupling mechanism at the interface was investigated via XMCD and XMLD [@HBM+15; @Bor13; @BMF+12]. Similar to these methods SPLEED probes the unoccupied states [@OYY+08]. In contrast to XAS where the transitions involve electronic states just above the Fermi energy in SPLEED the electrons have to eject into the vacuum passing the surface potential. In Fig. \[pes\] this corresponds to an energy of 4.7 eV above the Fermi level. As visible in the ARPES spectra above 5 eV large changes occur in the band dispersion when switching the BTO polarization. This is the origin of the changes in the exchange asymmetry enabling a high sensitivity for the investigation of the ferroic coupling at the interface. Summary \[sec\_summary\] ======================== We have used a fully relativistic multiple-scattering technique to calculate SPLEED pattern and ARPES intensities for the multiferroic system 3 ML Co/BTO(001). We have investigated the impact of switching the BTO polarization on the exchange scattering and on the resulting ARPES spectra. We have shown that both methods are sensitive to the BTO polarization, giving the opportunity for experimental investigations on the coupling mechanisms for this multiferroic heterostructure. Acknowledgement =============== We thank the BMBF (05K13WMA), the DFG (FOR 1346), CENTEM PLUS (LO1402), and the COST Action MP 1306 for financial support. SB thank the SFB 762: Functionality of Oxide Interfaces (Prof. I. Mertig, Dr. habil. A. Chassé) providing the framework for the calculation of the self-consistent potentials.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The classical theorems about singular perturbation reduction (due to Tikhonov and Fenichel) are concerned with convergence on a compact time interval (in slow time) as a small parameter approaches zero. For unbounded time intervals Hoppensteadt gave a convergence theorem, but his criteria are generally not easy to apply to concrete given systems. We state and prove a convergence result for autonomous systems on unbounded time intervals which relies on criteria that are relatively easy to verify, in particular for the case of a one-dimensional slow manifold. As for applications, we discuss several reaction equations from biochemistry.' author: - \| Christian Lax, Katrin Seliger, Sebastian Walcher\ Lehrstuhl A für Mathematik\ RWTH Aachen\ 52056 Aachen, Germany title: 'A coordinate-independent version of Hoppensteadt’s convergence theorem' --- $$\begin{aligned} &\text{\textbf{MSC2010:} 34E15, 92C45, 34C45}\\ &\text{\textbf{Keywords:} singular perturbations, reduction, reaction system, Lyapunov}\end{aligned}$$ Introduction ============ Singular perturbation phenomena occur frequently in the modelling and analysis of chemical or biological systems, in particular for reaction equations, and are highly relevant for reducing the dimension of a problem. For reaction equations (involving a small parameter $\varepsilon$), such phenomena may often be interpreted in the context of quasi-steady state (QSS) or partial equilibrium approximations (PEA). In many instances, the classical work of Tikhonov [@tikh] and Fenichel [@fenichel] provides a method to obtain a reduced equation.\ The theorems of Tikhonov and Fenichel guarantee convergence on some fixed compact time interval as $\varepsilon\to 0$. But beyond this result, in many applications one expects convergence for all positive times after a short initial phase, i.e. with slow time ranging in $[\tau_0,\infty)$ for some $\tau_0>0$. (In general convergence does not hold on an unbounded interval; see Fenichel [@fenichel], p. 68 for a well-known example involving the van der Pol equation).\ Hoppensteadt [@Hoppensteadt] stated and proved a convergence theorem for singularly perturbed systems which guarantee convergence on unbounded intervals, essentially resolving the matter up to coordinate transformations. However, in many potential applications these transformations (which effect a separation of variables into “slow” and “fast”) cannot be determined explicitly, and the hypotheses of the theorem are difficult to verify. In fact, even if a system is given in slow-fast coordinates, Hoppensteadt’s crucial conditions may not be readily verifiable. In the literature one finds some applications of this theorem where an explicit coordinate transformation is determined and the validity of Hoppensteadt’s conditions is verified directly. Thus, Cavallo and Natale [@cavnat], Teel et al. [@teeletal], and Back and Shim [@bashi] discuss applications to control theory. For the classical Michaelis-Menten enzyme model (with low enzyme concentration), no doubt was ever expressed about the validity of the reduction for all positive times, and the direct estimates given in Segel and Slemrod [@SSl] do imply convergence for the case of irreversible product formation. For the reversible case, it seems that a convergence proof was given only relatively recently, in [@nw11]. This proof uses Hoppensteadt’s criteria, and the crucial part invokes explicit knowledge of a first integral for the fast system. The argument cannot be extended to familiar variants of Michaelis-Menten, e.g. those including inhibition or cooperativity. One purpose of the present work is to provide more easily applicable criteria for reaction systems.\ The paper is organized as follows. We start the main section (Section 2) with an auxiliary result on Lyapunov functions and asymptotic stability. Then we proceed to a version of Hoppensteadt’s theorem for autonomous systems that are written in slow-fast coordinates with special properties. Following this, we do not only specialize (and thus simplify) Hoppensteadt’s conditions for autonomous systems but we also replace one of the conditions with another that is somewhat stronger, but readily verifiable. We next recall Tikhonov-Fenichel reduction for singularly perturbed systems with no a priori separation of slow and fast variables [@gw2]. Finally we give additional conditions which guarantee convergence to solutions of the reduced system on unbounded intervals, with Theorem \[hshsatz\] the main result. Some of the conditions we impose (e.g. eigenvalue conditions) are relatively easy to verify in applications, but for others verification may still be problematic. In particular this concerns the existence of a global parameterization of the asymptotic slow manifold, and the existence of a Lyapunov function for the reduced system on this manifold. But at least for the case of a one-dimensional slow manifold, which is highly relevant for QSS in biochemistry, these problems can be resolved completely, and one obtains readily applicable criteria. In Section 3 we discuss a number of examples.\ Some of the results presented are based on work in the theses [@selma] and [@laxdiss]. The Appendix (Section 5) contains a list of Hoppensteadt’s conditions and the corresponding theorem for easy reference. Hoppensteadt’s theorem for autonomous systems ============================================= The goal of this section is to state and prove a version of Hoppensteadt’s convergence theorem [@Hoppensteadt] for autonomous systems which is readily applicable to the investigation of a reasonably large class of differential equations. In the main result we will not require any a priori separation of fast and slow variables, and we will focus on conditions that are amenable to explicit verification.\ We first prove an auxiliary result on Lyapunov functions, and then an autonomous version of the convergence theorem for special slow-fast coordinates, before turning to a general coordinate-free version. The result is, in particular, easily applicable to systems with a one-dimensional slow manifold.\ Let $U\subset \mathbb R^m$ be open, and $p\in C^1(U;\mathbb R^m)$. We consider the differential equation $$\label{odep} x^\prime=p(x)$$ on $ U$, with the prime (here and in the following) denoting differentiation with respect to the independent variable $\tau$. An auxiliary result ------------------- \[lyapunov\] Let $Y\subset \mathbb R^m$ be a submanifold and $\widehat K\subset U$ compact such that $Y\cap\widehat K$ is positively invariant with respect to . Assume there exists a neighborhood $S\subset U$ of $Y\cap \widehat K$ and a function $\varphi\in C^1(S)$ that satisfies the following conditions: (i) The inequality $\varphi(x)\geq0$ holds for all $x\in Y\cap \widehat K$, and there exists exactly one $z\in Y\cap\widehat K$ such that $\varphi(z)=0$. (ii) Given a norm ${\ifx\relax\relax \ensuremath{\left\Vert\cdot \right\Vert} \else \ensuremath{\left\Vert\cdot \right\Vert_{\relax}}\fi}$ on $\mathbb R^m$ there exist $c_1,c_2>0$, a positive integer $a$, and $\rho>0$ such that for all $x\in Y\cap B_{\rho}(z)$ the inequalities $$c_1{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}^a\leq \varphi(x)\leq c_2 {\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}^a$$ are satisfied. (Here $z$ is from (i), and $B_{\rho}(z)$ denotes the open ball with center $z$ and radius $\rho$.) (iii) There exist $\nu>0$ and $k\geq 1$ such that the Lie derivative of $\varphi$ with respect to $p$ satisfies $$L_{p}(\varphi)(x)\leq -\nu \varphi(x)^k$$ for all $x\in Y\cap \widehat K$. (Recall $L_p(\varphi)(x)=D\varphi(x)\cdot p(x)$ for all $x$.) Then there exists $c>0$ such that for all $x_0\in Y\cap \widehat K$, $x_0\not=z$ the solution $\Phi(\tau,x_0)$ of the initial value problem $x^\prime=\frac{dx}{d\tau}=p(x)$, $x(0)=x_0$ satisfies the inequality $${\ifx\relax\relax \ensuremath{\left\Vert\Phi(\tau,x_0)-z\right\Vert} \else \ensuremath{\left\Vert\Phi(\tau,x_0)-z\right\Vert_{\relax}}\fi}\leq c{\ifx\relax\relax \ensuremath{\left\Vertx_0-z\right\Vert} \else \ensuremath{\left\Vertx_0-z\right\Vert_{\relax}}\fi}\gamma(\tau),$$ with $$\gamma(\tau)=\left\{ \begin{array}{lcl} e^{-\nu \tau/a} &\text{ for }& k=1,\\ ((k-1)\nu \tau \varphi(x_0)^{k-1}+1)^{1/[a(1-k)]}&\text{ for }& k>1 \end{array}\right.$$ strictly decreasing to $0$ as $\tau\to\infty$.\ The function $$\frac{\varphi(x)}{{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}^a}$$ is continuous on $\widehat K\cap \{x\in\mathbb R^m;\, {\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}\geq \rho\}$ and therefore bounded below and above by positive constants. Hence there exist $0<c_1^<c_2^$ such that for all $x\in Y\cap \widehat K$ the inequalities $$\label{cineq} c_1^{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}^a\leq \varphi(x)\leq c_2^{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}^a$$ hold.\ Now use condition (iii) and recall a result on differential inequalities (e.g. Amann [@Amann], Lemma 16.4): Since the initial value problem $$w^\prime = -\nu w^k,\quad w(0)= \varphi(x_0)>0$$ in $\mathbb R$ is solved by $$\widetilde \gamma(\tau)=\left\{ \begin{array}{lcl} \varphi(x_0)e^{-\nu \tau} &\text{ for }& k=1,\\ ((k-1)\nu \tau +\varphi(x_0)^{1-k})^{1/(1-k)}&\text{ for }& k>1, \end{array}\right.$$ the solution $\Phi(\tau,x_0)$ satisfies $$\varphi(\Phi(\tau,x_0))\leq \widetilde \gamma(\tau)\leq c_2^{\ifx\relax\relax \ensuremath{\left\Vertx_0-z\right\Vert} \else \ensuremath{\left\Vertx_0-z\right\Vert_{\relax}}\fi} ^a \gamma(\tau)$$ due to (ii). By virtue of we obtain $${\ifx\relax\relax \ensuremath{\left\Vert\Phi(\tau,x_0)-z\right\Vert} \else \ensuremath{\left\Vert\Phi(\tau,x_0)-z\right\Vert_{\relax}}\fi}\leq \left(\tfrac1{c_1^}\varphi(\Phi(\tau,x_0))\right)^{1/a}\leq \left(\tfrac{c_2^}{c_1^}\right)^{1/a}{\ifx\relax\relax \ensuremath{\left\Vertx_0-z\right\Vert} \else \ensuremath{\left\Vertx_0-z\right\Vert_{\relax}}\fi}\gamma(\tau).$$ The assertion follows. Systems in Tikhonov standard form {#subs22} --------------------------------- Here we will prove an intermediate result for systems written in special coordinates. With the exception of (ASII) below, the conditions are patterned after Hoppensteadt [@Hoppensteadt], conditions (I) through (VII).\ In the following denote by ${\ensuremath{\left\vert\cdot\right\vert}}_1$ the 1-norm, let $s$ and $r$ be positive integers and $m=s+r$. For $R>0$ we define $$\label{tubeone} \begin{array}{rcl} S_R&:=&\{y=(y_1,y_2)\in\mathbb R^{s+r},\ {\ensuremath{\left\verty\right\vert}}_1={\ensuremath{\left\verty_1\right\vert}}_1+{\ensuremath{\left\verty_2\right\vert}}_1\leq R\},\\ S_{1,R}&:=&\{y_1\in\mathbb R^{s},\, {\ensuremath{\left\verty_1\right\vert}}_1\leq R\},\\ S_{2,R}&:=&\{y_2\in\mathbb R^{r},\,{\ensuremath{\left\verty_2\right\vert}}_1\leq R\}.\\ \end{array}$$ We consider a singularly perturbed autonomous system in Tikhonov standard form $$\begin{aligned} &y_1'= f(y_1,y_2,\varepsilon)\label{tnfhs1}\\ &y_2'=\varepsilon^{-1}g(y_1,y_2,\varepsilon)\label{tnfhs2} \end{aligned}$$ as well as its counterpart in fast time $t=\tau/\varepsilon$, viz. $$\begin{aligned} &\dot y_1= \varepsilon f(y_1,y_2,\varepsilon)\\ &\dot y_2=g(y_1,y_2,\varepsilon) \end{aligned}$$ subject to the following conditions: 1. Both $f$ and $g$ are $C^2$ functions in an open subset $\widetilde U$ of $\mathbb R^{s}\times \mathbb R^r\times \mathbb R$, and $\widetilde U$ contains $S_R\times [0,\varepsilon_0)$ for some $R>0$ and $\varepsilon_0>0$. 2. The system $$\begin{aligned} &y_1'=f(y_1,y_2,0) \label{redSystem1}\\ &0=g(y_1,y_2,0)\label{redSystem2} \end{aligned}$$ admits the stationary point $0\in\mathbb R^s\times \mathbb R^r$. 3. For all $y_1\in S_{1,R}$ one has $g(y_1,\,0,\,0)=0$, and there is a constant $\nu>0$ such that all eigenvalues of $D_2g(y_1,0,0)$ have real part $\leq -\nu$.\ (Here and in the following $D_i$ denotes the partial derivative with respect to $y_i$.) 4. The hypotheses of Lemma \[lyapunov\] hold with $p(y)=f(y_1,0,0)$, $Y=\{(y_1,0)\in S_R\}$ and $z=0$. Invoking a compactness argument, it would suffice in (ASII) to require that the eigenvalues of $D_2g(y_1,0,0)$ have real part $<0$ for all $y_1\in S_{1,R}$. \[hopstandard\] Whenever assumptions [(AS0)]{} through [(ASIII)]{} hold, there exists a compact neighborhood $K\times [0,\varepsilon_0^]\subseteq S_R\times [0,\varepsilon_0)$ of $0\in\mathbb R^{s+r+1}$ with the following properties: Given $y_0:=(y_{1,0},\,y_{2,0})\in K$ and $\varepsilon\in (0,\varepsilon_0^]$, the solution $\Phi(\tau,y_0)$ of – exists for $0\leq \tau<\infty$. As $\varepsilon\to 0$, this solution converges uniformly on all closed subsets of $(0,\infty)$ to the solution of – with respect to the initial value $y_1(0)=y_{1,0}$. One has to verify conditions (I) through (VII) in Hoppensteadt [@Hoppensteadt]; for the reader’s convenience these are recalled in Section \[appsec\] below. Clearly (AS0) and (ASI), together with the fact that the system is autonomous, ensure that conditions (I) through (V) hold. (In the autonomous case the uniformity requirements follow readily by continuity and compactness.) Condition (VI) is a consequence of (ASIII) and Lemma \[lyapunov\].\ There remains to show that (ASII) implies the validity of (VII). In principle one could refer to Fenichel [@fenichel], but we give a proof with some details here. (i) From $g(y_1,0,0)=0$ for all $y_1\in S_{1,R}$ and by Hadamard’s Lemma (see e.g. Nestruev [@nestruev2003smooth], Lemma 2.8) there exists a $C^1$ function $\widehat R$ such that $$g(y_1,y_2,0)=\widehat R(y_1,y_2)y_2$$ for all $(y_1,y_2)\in S_R$. Furthermore $$\widehat R(y_1,y_2)y_2=\left(A(y_1)+R(y_1,y_2)\right)y_2$$ where $y_1\mapsto A(y_1)=\widehat R(y_1,0) \in \mathbb R^{r\times r}$ is $C^1$ and $\\|R(y_1,y_2)\\|\to 0$ as $\\|y_2\\|\to 0$, uniformly in $y_1$. According to (ASII), for $y_1\in S_{1,R}$ all eigenvalues of $A(y_1)$ have real part $\leq -\nu<0$. (ii) We denote by $\phi(\cdot,\cdot)$ the standard Euclidean scalar product on $\mathbb R^r$, and thus have $$\phi(y_2,y_2)=\\|y_2\\|_2^2.$$ Denote by $C$ the unit sphere in $\mathbb R^r$ with respect to $\\|\cdot\\|_2$. For every $y_1$ there exists a $\theta(y_1)>0$ such that $$\begin{array}{rccl} 2\phi(y_2,A(y_1)y_2)&\leq & -2\theta(y_1)\\|y_2\\|_2^2& \text{ for all }y_2\in\mathbb R^r,\\ 2\phi(y_2,A(y_1)y_2)&\leq & -2\theta(y_1)& \text{ for all }y_2\in C. \end{array}$$ The proof of the first inequality follows by the arguments in Walter [@Walter] (§30, IV(d) and proof of §29, VIII). These imply that there exists some positive definite symmetric bilinear form $\psi$ such that $$2\psi(y_2,A(y_1)y_2)\leq -\nu/2\cdot\psi(y_2,y_2) \text{ for all }y_2,$$ and the assertion follows by the equivalence of all norms on $\mathbb R^r$. The second inequality is a simple consequence but it shows that one can choose $-2\theta(y_1)$ as the maximum of the left hand side function on $C$. (iii) Given $y_1^\in S_{1,R}$ there exists a neighborhood $U(y_1^)$ such that $$\phi(y_2,A(y_1)y_2)\leq -\theta(y_1^)\\|y_2\\|_2^2 \text{ for all }y_2\in\mathbb R^r, \,y_1\in U(y_1^).$$ This follows by the estimates in (ii), the continuity of the map $$S_{1,R}\times C\to \mathbb R,\quad (y_1,y_2)\mapsto \phi(y_2, A(y_1)y_2),$$ and the homogeneity of $\phi$. In conjunction with the compactness of $S_{1,R}$ this estimate implies the existence of some $\beta>0$ such that $$\phi(y_2,A(y_1)y_2)\leq -2\beta \\|y_2\\|_2^2 \text{ for all } (y_1,y_2)\in S_R.$$ (iv) Moreover there exists $\rho>0$ such that $$\|2\phi(y_2,R(y_1,y_2)y_2)\|\leq \beta\cdot\\|y_2\\|_2^2 \text{ for all } y_1\in S_{1,R},\,\,y_2\in S_{2,\rho}$$ due to uniform convergence with respect to $y_1$ (again, see Walter [@Walter], loc.cit.). Altogether one obtains $$L_g(\phi)(y_2)= 2\phi(y_2, A(y_1)y_2)+2\phi(y_2,R(y_1,y_2)y_2)\leq -\beta\\|y_2\\|_2^2$$ for all $(y_1,y_2)\in S_{1,R}\times S_{2,\rho}$. (v) Denote the solution of $ x_2^\prime=g(y_1,x_2,0)$ with initial value $y_2$ by $\Gamma(y_1,y_2,\tau)$. Then by Amann [@Amann], Lemma 16.4 one finds $$\phi(\Gamma(y_1,y_2,\tau))\leq \\|y_2\\|_2^2\cdot\exp(-\beta \tau),$$ hence $$\\|\Gamma(y_1,y_2,\tau)\\|\leq \\|y_2\\|\cdot\exp(-\beta \tau/2).$$ Hoppensteadt’s condition (VII) follows via the equivalence of all norms on $\mathbb R^r$. Hoppensteadt [@Hoppensteadt] develops his conditions (for non-autonomous systems) from a more general setting in a step-by-step manner, with certain normalizations (that can not necessarily be carried out explicitly) being invoked at various stages; see Section \[appsec\]. Eventually his crucial conditions require the special setting which we consider above (and furthermore restrict to autonomous systems). General systems --------------- The goal of this subsection is to extend Proposition \[hopstandard\] to settings where no a priori separation of “slow” and “fast” variables is given. Thus we consider a system $$\label{evolution} \dot x=h(x,\varepsilon)=h^{(0)}(x)+\varepsilon h^{(1)}(x)+\varepsilon^2 h^(x,\varepsilon)$$ with right-hand side $C^2$ in $(x,\varepsilon)$, and $(x,\varepsilon)$ is in some open subset of $\mathbb R^m\times\mathbb R$ that contains $(x_0,0)$ for some $x_0$. We will also work with the time-scaled version ($\tau=\varepsilon t$ as in subsection \[subs22\]), thus $$\label{evolutionscaled} x^\prime= \frac{d x}{d\tau}=\varepsilon^{-1}h(x,\varepsilon)=\varepsilon^{-1}h^{(0)}(x)+ h^{(1)}(x)+\ldots$$ of this equation. We first recall a coordinate-free version of standard (Tikhonov-Fenichel) singular perturbation reduction from [@gw2], Theorem 1. (The theorem was stated for systems with rational right-hand side, but as noted in [@gw2], Remark 2, sufficient differentiability already guarantees existence.) The following conditions are relevant. - There exists a point $x_0$ in the zero set $\mathcal V(h^{(0)})$ such that $\operatorname{rank}Dh^{(0)}(x)=r<m$ for all $x\in \mathbb R^m$ in some neighborhood of $x_0$. By the implicit function theorem, (TF0) implies the existence of a neighborhood $U$ of $x_0$ such that $V:=U\cap \mathcal V(h^{(0)})$ is a $(m-r)$-dimensional submanifold. - There is a direct sum decomposition $$\mathbb R^m=\ker Dh^{(0)}(x) \oplus \operatorname{im}Dh^{(0)}(x)$$ for all $x\in V$. (In other words, one requires that algebraic and geometric multiplicity of the eigenvalue zero of $Dh^{(0)}(x)$ are equal.) For details and proofs concerning the next two results we refer to [@gw2]. \[decompred\] Let [(TF0)]{} and [(TFI)]{} be given. Then the following hold. (a) (Product decomposition) On some neighborhood $\widetilde U\subseteq U$ of $x_0$ there exist $C^1$ maps $$P\colon \widetilde U\to \mathbb R^{m\times r}\quad \text{and} \quad \mu\colon\widetilde U\to \mathbb R^r$$ with $\operatorname{rank}P(x_0)=\operatorname{rank}D\mu(x_0)=r$, such that $$h^{(0)}(x)=P(x)\mu(x),\quad x\in \widetilde U.$$ Moreover, the zero set $Y$ of $\mu$ satisfies $Y=V\cap \widetilde U= \mathcal V(h^{(0)})\cap\widetilde U$. The entries of $\mu$ may be taken as any $r$ entries of $h^{(0)}$ that are functionally independent at $x_0$. (b) The system $$\label{Grenzsystem} x^\prime=q(x):=Q(x)\cdot h^{(1)}(x)$$ with $$Q(x):=Id-P(x)(D\mu(x)P(x))^{-1}D\mu(x),$$ is defined in $\widetilde U$, and the manifold $Y$ is an invariant set of . Moreover, every entry of $\mu$ is a first integral of . We will call the Tikhonov-Fenichel reduction of . The result holds for every connected component of $Y$, hence we may and will assume that $Y$ is connected. \[remarkred\] (a) Conditions (TF0) and (TFI) ensure the existence of a coordinate transformation that puts into [Tikhonov]{} standard form, and the reduced system corresponds to the familiar reduction with slow and fast variables; see [@nw11]. (b) As was shown in [@gw2], for rational $h^{(0)}$ one may choose $P$ and $\mu$ rational, and the decomposition can be obtained constructively by methods of algorithmic algebra. The next condition guarantees local convergence of solutions to solutions of the reduced system. - All nonzero eigenvalues of $Dh^{(0)}(x)$, $x\in Y$, have negative real part. With these assumptions one can state a coordinate-free local version of (Tikhonov’s and) Fenichel’s reduction theorem; see [@gw2], Theorem 1. \[zitatgoeke\] Assume that [(TF0)]{}, [(TFI)]{} and [(TFII)]{} hold. Then there exists $T>0$ and a neighborhood $U^\subset U$ of $Y$ such that solutions of starting in $U^$ converge uniformly on $[\tau_0,T]$ to solutions of the reduced system on $Y$ as $\varepsilon\to 0$, for any $\tau_0$ with $0<\tau_0<T$. \[zitatremark\] (a) The submanifold $V$ is called the [asymptotic slow manifold]{} (or [critical manifold]{}). (b) Concerning the question of finding the appropriate initial values on $Y$ (which was in principle also settled by Fenichel [@fenichel], Theorem 9.1), we briefly summarize the discussion in [@gw2] Proposition 2: The system $\dot x=h^{(0)}(x)$ admits $m-r$ independent first integrals in a neighborhood of $x_0$, and the intersection of a common level set of the first integrals with $Y$ consists (locally) of a single point. To project an initial value of system to an initial value of on $Y$, choose the corresponding intersection point. Thus, a solution of starting at $x_0\in U^$ converges to the solution of starting at the projected initial value. (c) In the situation of Proposition \[zitatgoeke\] we sometimes call a convergent Tikhonov-Fenichel reduction of ; in contrast to a [formal]{} reduction whenever only (TF0) and (TFI) hold. (d) The proof of Proposition \[zitatgoeke\] (see [@nw11] Proposition 2.3 and [@gw2] Theorem 1) shows that the coordinate transformation which puts into [Tikhonov]{} standard form (see Remark \[remarkred\] (a)) also maps solutions of the reduced system to solutions of the corresponding reduced system – in [Tikhonov]{} standard form. Up to this point we focussed on conditions which ensure convergence of singular perturbation reduction on some compact subinterval of $(0, \infty)$. We now introduce additional conditions to guarantee validity of the reduction on unbounded intervals. The first of these conditions could be weakened, but it is convenient for applications and it is satisfied for many relevant systems, in particular reaction systems. - There exists a compact neighborhood $K\subseteq \widetilde U$ of $x_0$ which is positively invariant for all differential equations with $0<\varepsilon<\varepsilon_0$. By continuous dependence one obtains: Under the assumptions of Proposition \[zitatgoeke\] and given condition (CIS), the set $K\cap Y$ is positively invariant for the reduced system . Next come the crucial conditions. - There exists a contractible open subset $W$ of $\mathbb R^{s}$, $s=m-r$, and a global injective $C^2$ immersion $\Lambda^\colon W\to Y$. (a) We introduce condition (GP) to make the reasoning more transparent, and in order to state the following Lemma in a more general context. But below we will introduce a further condition (LC) which actually implies (GP). Indeed, as an argument in the proof of Theorem \[hshsatz\] will show, by global asymptotic stability there exists a flow on $W$ which contracts to a point (see also Corollary \[graphcor\]). (b) We (may and) will assume that $S_{1,R}\subseteq W$ for some $R>0$. (c) In the statement of (GP) one may replace $Y$ by a relatively open neighborhood of $Y\cap K$ in $Y$. Under the assumptions of Proposition \[zitatgoeke\] and given conditions (CIS) and (GP) there is a compact set $K^\supseteq K\cap Y$ with nonempty interior, some $0<\rho\leq R$ and a $C^2$-diffeomorphism $$\Lambda\colon S_\rho\to K^, \quad \text{with } \Lambda\|_{S_{1,\rho}\times \{0\}}= \Lambda^\|_{S_{1,\rho}\times \{0\}}.$$ (Thus, there are open neighborhoods of $S_\rho$ resp. $K^$ that are mapped to each other by $\Lambda$ and its inverse.) The normal bundle $N$ of $Y$ is trivial, since $W$ is simply connected; see Hirsch [@hirsch], Ch. 4, Corollary 2.5. Now the assertion follows by injectivity of $\Lambda^$ and Hirsch [@hirsch], Ch. 4, Theorem 5.1. The final condition we require is as follows. - The reduced system admits one and only one stationary point $z$ in $Y\cap K$, and the conditions in Lemma \[lyapunov\] are satisfied. With these assumptions the convergence statement of Proposition \[hopstandard\] carries over. \[hshsatz\] For system assume that [(TF0)]{}–[(TFII)]{} as well as [(CIS)]{}, [(GP)]{} and [(LC)]{} are satisfied. Then there exist a compact $\widetilde K\subseteq \Lambda(S_\rho)$ with nonempty interior, $z\in \widetilde K$, and $\varepsilon_0^>0$ with the following properties: Given $y_0\in \widetilde K$ and $0<\varepsilon<\varepsilon_0^$, the solution $\Phi(t,y_0)$ of – exists for $0\leq t<\infty$. As $\varepsilon\to 0$, this solution converges uniformly on all closed subsets of $(0,\infty)$ to the solution of with initial value according to Remark \[zitatremark\] b). Use $\Lambda\colon S_\rho\to K^$ to define $$\widetilde h(x,\varepsilon):=D\Lambda(x)^{-1}h(\Lambda(x,\varepsilon)).$$ By construction, the diffeomorphism $\Lambda$ sends solutions of $\dot x =\widetilde h(x,\varepsilon)$ to solutions of $\dot x=h(x,\varepsilon)$; and it is sufficient to verify conditions (AS0) through (ASIII) for the former system.\ (i) For $\varepsilon=0$ one has the identity $$D\Lambda(x)\widetilde h^{(0)}(x)=h^{(0)}(\Lambda(x)),$$ with $\widetilde h^{(0)}(x):=\widetilde h(x,0)$. Let $x=(x_1,x_2)$, with $x_1\in \mathbb R^s$ and $x_2\in\mathbb R^r$. Since $\Lambda$ extends $\Lambda^$, we have $$\widetilde h^{(0)}((x_1,0))=D\Lambda((x_1,0))^{-1}h^{(0)}(\Lambda((x_1,0)))=0,$$ and we obtain the conjugacy property $$D\widetilde h^{(0)}((x_1,0))=D\Lambda((x_1,0))^{-1}Dh^{(0)}(\Lambda((x_1,0)))D\Lambda((x_1,0))$$ by differentiation, noting that the second term on the right hand side vanishes due to $h^{(0)}(\Lambda((x_1,0)))=0$. (ii) Moreover $\dot x=\widetilde h^{(0)}(x)$ may be assumed to be in the particular form $$\begin{array}{rcl} \dot x_1&=&0\\ \dot x_2&=& g^{(0)}(x_1,x_2) \end{array}$$ with $ g^{(0)}(x_1,0)=0$. In other words, $\dot x =\widetilde h(x,\varepsilon)$ is in Tikhonov standard form and (AS0), (ASI) hold.\ To verify this, note that (TFII) holds and use Fenichel [@fenichel], Lemma 5.3. (A different proof for analytic systems makes use of the fact that the differential equation $\dot x =\widetilde h^{(0)}(x)$ admits $s$ independent first integrals in the neighborhood of any stationary point; see [@nw11], Proposition 2.2.) (iii) By conjugacy of Jacobians, (TFII) and compactness one sees that (ASII) is satisfied. (iv) There remains to verify the existence of a Lyapunov function so that (ASIII) holds. The reduced system corresponding to $\widetilde h$ will be called $x^\prime =\widetilde q(x)$. It has the special form $$\begin{array}{rcl} x_1^\prime&=& f^{(1)}(x_1,0)\\ x_2^\prime &=& 0 \end{array}$$ with the slow manifold being given by $x_2=0$. Due to Remark \[zitatremark\] (d), the map $\Lambda$ sends solutions of $\dot x=\widetilde q(x)$ to solutions of $\dot x=q(x)$. Now the Lyapunov function $\varphi$ for $q$ satisfies $L_q(\varphi)\leq -\nu\cdot\varphi^k$, and with the well-known identity $$L_{\widetilde q}(\varphi\circ\Lambda)=L_q(\varphi)\circ\Lambda$$ one obtains that $\widetilde\varphi:=\varphi\circ\Lambda$ is a Lyapunov function for $\widetilde q$, and that the inequality $$L_{\widetilde q}(\widetilde \varphi)\leq -\nu \widetilde \varphi^k$$ holds. Moreover, obviously $\widetilde\varphi\geq 0$ with $0$ the only zero. Finally, since $\Lambda^{-1}$ is a diffeomorphism and its derivative is bounded on the compact set $S_\rho$, the mean value estimate shows the existence of positive constants $k_1$ and $k_2$ such that $$k_1{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}\leq {\ifx\relax\relax \ensuremath{\left\Vert\Lambda^{-1}(x)-\Lambda^{-1}(z)\right\Vert} \else \ensuremath{\left\Vert\Lambda^{-1}(x)-\Lambda^{-1}(z)\right\Vert_{\relax}}\fi}\leq k_2{\ifx\relax\relax \ensuremath{\left\Vertx-z\right\Vert} \else \ensuremath{\left\Vertx-z\right\Vert_{\relax}}\fi}$$ for all $x\in S_\rho$. This implies condition (ii) from Lemma \[lyapunov\]. \[graphcor\] If $Y$ is the graph of some smooth function $\Gamma\colon W\to \mathbb R^r$, with $W\subseteq R^{m-r}$ contractible and open, and [(CIS)]{} and [(LC)]{} are satisfied in addition to [(TF0)]{}–[(TFII)]{}, then [(GP)]{} and thus the conclusion of Theorem \[hshsatz\] hold. \[lyappar\] In the setting of Theorem \[hshsatz\], it suffices to require the existence of a Lyapunov function $\varphi$ for $q$ on $Y\cap K$ (rather than in some neighborhood of $Y$), since $\widetilde \varphi$ can be extended to $S_\rho$ by setting $\widehat \varphi(x_1,\,x_2):=\widetilde\varphi(x_1)$. One-dimensional slow manifolds ------------------------------ For a general system the verification of condition (LC) on $Y$ a priori requires explicit knowledge of a Lyapunov function. However, for one-dimensional slow manifolds a simple condition will imply (LC). We first note a property of differential equations on real intervals which is essentially common knowledge (since differential equations in $\mathbb R$ are gradient systems); a proof is included for the reader’s convenience. \[1dlyapunov\] Let $U\subseteq \mathbb R$ be an open interval containing $0$, and $p\in C^1(U)$ with $p(0)=0$. Moreover let $K\subset U$ be compact with $0\in K$, and $0$ the only stationary point of $x'=p(x)$ in $K$. If $0$ is linearly asymptotically stable (i.e. $p'(0)<0$) then $$\widetilde\varphi(x)=-\int_{0}^xp(y)\: dy$$ is a Lyapunov function of $x'=p(x)$ which satisfies the hypotheses of Lemma \[lyapunov\] on $K$, with $a=2$ and $k=1$. By Hadamard’s lemma $$p(x)= x\cdot \widehat p(x),\text{ with }\widehat p \text { continuous and }\widehat p(0)=-\theta<0;$$ moreover $\widehat p$ is negative throughout $K$. This implies that $K$ is positively invariant and that $\widetilde \varphi$ is nonnegative, with $0$ its only zero. In some neighborhood $\widetilde U$ of $0$ one has the estimates $$-2\theta\cdot x\leq p(x)\leq -\theta/2\cdot x \text{ for }x>0,\quad -\theta/2\cdot x\leq p(x)\leq -2\theta\cdot x \text{ for }x<0,$$ which imply $$\theta/2\cdot x^2\leq \widetilde\varphi(x)\leq 2\theta\cdot x^2$$ for all $x\in \widetilde U$. Therefore condition (ii) from Lemma \[lyapunov\] holds on $K$, since $\widetilde\varphi$ and $L_p(\widetilde\varphi)$ are continuous and the complement of $\widetilde U$ in $K$ is compact. By construction $L_{p}(\widetilde\varphi)(x)=-p^2(x)$, and the above estimate shows $$L_{p}(\widetilde\varphi)(x)=-x^2 \cdot 4\theta^2$$ in $\widetilde U$, whence conditions (ii) and (iii) in Lemma \[lyapunov\] hold in $\widetilde U$ with $a=2$ and $k=1$, and (by compactness and continuity arguments) on all of $K$. Now we can state our result. \[1dprop\] For system assume that [(TF0)]{}–[(TFII)]{} and [(CIS)]{} are satisfied. Moreover assume that $Y\cap K$ is one-dimensional, connected, contains exactly one stationary point $z$, and that the linearization of the reduced equation $\dot x=q(x)$ at $z$ admits a negative eigenvalue. Then the conclusion of Theorem \[hshsatz\] holds. We first note that the one-dimensional compact and connected manifold is homeomorphic to a compact interval or to a circle; see e.g. Milnor [@milnor]. But the latter is incompatible with the existence of a single stationary point that is asymptotically stable. The curve $Y$ admits a global parameterization by curve length, and therefore (GP) is satisfied. According to Proposition \[decompred\], there are $m-1$ functionally independent defining equations for $Y$ near any of its points, and these are first integrals for the reduced equation (see also Proposition \[zitatgoeke\]). Therefore the linearization at $z$ admits the eigenvalue $0$ with geometric multiplicity $\geq m-1$ at $z$, and the eigenspace of the nonzero (negative) eigenvalue must be equal to the tangent space to $Y$ at $z$. Now Lemma \[1dlyapunov\] and Theorem \[hshsatz\] apply. Examples ======== In this section we will discuss some reaction equations, with an emphasis on one-dimensional slow manifolds. Let $\mathbb R_+^n$ be the set of all vectors $x\in\mathbb R^n$ with nonnegative entries. Moreover, the concentration of a chemical species $Z$ will be denoted with a lowercase letter $z$. Michaelis-Menten reaction and variants -------------------------------------- The well-known (reversible) Michaelis-Menten reaction is defined by the reaction scheme $$E+S \xrightleftharpoons[k_{-1}]{k_{1}} C \xrightleftharpoons[k_{-2}]{k_{2}} E+P;$$ see Michaelis and Menten[@michaelismenten], and also Briggs and Haldane [@briggshaldane]. The concentrations of each chemical species will be denoted by the corresponding lower-case letter. By mass-action kinetics, using the linear first integrals $e+c$ and $s+c+p$ from stoichiometry and assuming that initially no complex $C$ or product $P$ are present, one obtains the following two-dimensional problem: $$\begin{aligned} &\dot s=-k_1e_0s+(k_1s+k_{-1})c\\ &\dot c=k_1e_0s-(k_1s+k_{-1}+k_2)c+k_{-2}(e_0-c)(s_0-s-c). \end{aligned}$$ ### Small enzyme concentration The standard approach takes the assumption of small initial enzyme concentration $e_0=\varepsilon e_0^$. By the results in [@nw11] and [@gswz], there exists a convergent Tikhonov-Fenichel reduction to $$s^\prime=-e_0^\frac{(k_1k_2+k_{-1}k_{-2})s-k_{-1}k_{-2}s_0}{k_1s+k_{-1}+k_2+k_{-2}(s_0-s)},$$ the slow manifold being $V=\{(s,0)\in {\mathbb R}^2_+\}$. Using an explicit transformation to Tikhonov standard form and employing some straightforward but elaborate computations, convergence on unbounded time intervals was already proven in [@nw11]. (As mentioned before, Segel and Slemrod [@SSl] gave a proof by direct estimates for the irreversible case $k_{-2}=0$.) We use this example only to illustrate how Proposition \[1dprop\] greatly simplifies convergence proofs. Indeed, the validity of (TF0)–(TFII) and (CIS) is easy to verify, and clearly, the reduced equation admits exactly one stationary point $$s^=\frac{k_{-1}k_{-2}s_0}{k_1k_2+k_{-1}k_{-2}}$$ which is linearly asymptotically stable. Thus, Proposition \[1dprop\] shows convergence (with the irreversible case included for $k_{-2}=0$.) ### Slow product formation for the irreversible system There are other choices for a “small parameter” that yield convergent Tikhonov-Fenichel reductions of the Michaelis-Menten reaction (see [@gwz] for an exhaustive discussion). All of these admit one-dimensional slow manifolds and it is easy to check the validity of the hypotheses of Proposition \[1dprop\]. For instance, assuming slow, irreversible product formation (i.e., $k_2=\varepsilon k_2^$ small and $k_{-2}=0$), one finds the reduction $$s^\prime = \frac{-(k_1s+k_{-1})k_1k_2^e_0s}{k_1k_{-1}e_0+(k_1s+k_{-1})^2},$$ on $V=\{(s,c)\in{\mathbb R}^2_+,\ k_1e_0s=(k_1s+k_{-1})c\}$. Again, Proposition \[1dprop\] is applicable to show convergence on $[\tau_0,\,\infty)$ for any $\tau_0>0$. ### Competitive inhibition {#compinone} As an extension of the Michaelis-Menten model we discuss an irreversible enzyme reaction with inhibition (see e.g. Keener and Sneyd [@KeenerSneyd]). The reaction scheme is given by $$\begin{aligned} &E+S \xrightleftharpoons[k_{-1}]{k_{1}} C_1 \xrightarrow{k_{2}} E+P\\ &E+I \xrightleftharpoons[k_{-3}]{k_{3}} C_2. \end{aligned}$$ We assume the usual initial values $c_1(0)=c_2(0)=0$, $s(0)=s_0>0$, $e(0)=e_0>0$ and $i(0)=i_0>0$. Using the linear first integrals $$\begin{aligned} &\psi_1(e,s,c_1,c_2,p,i)=e+c_1+c_2,\quad \psi_2(e,s,c_1,c_2,p,i)=s+c_1+p,\\ &\psi_3(e,s,c_1,c_2,p,i)=i+c_2 \end{aligned}$$ from stoichiometry, one obtains a three-dimensional system. Again we assume that the initial enzyme concentration is low, thus $e_0=\varepsilon e_0^$, and obtain the differential equation $$\begin{array}{cclcl} \dot s&=&k_{-1}c_1+k_1s(c_1+c_2)&-&\varepsilon e_0^* k_1s \\ \dot c_1&=&-k_1s(c_1+c_2)-(k_{-1}+k_2)c_1 &+& \varepsilon e_0^* k_1s\\ \dot c_2&=&-k_3(c_1+c_2)(i_0-c_2)-k_{-3}c_2&+&\varepsilon e_0^k_3(i_0-c_2) \end{array}$$ The reduction was computed in [@gswz], subsection 3.2; in particular (TF0)-(TFII) (GP) and (CIS) hold for the slow manifold $Y$ defined by $c_1=c_2=0$, and the reduced equation $$s^\prime=-e_0^\, \frac{k_1k_2k_{-3}s}{k_{-3}(k_1s+k_{-1}) +(k_{-1}+k_2)k_3i_0 +k_2k_{-3}}$$ admits the only stationary point $0$, which is linearly asymptotically stable. By Proposition \[1dprop\] we obtain convergence on every interval $[\tau_0,\,\infty)$, with $\tau_0>0$. (The same holds true for reversible product formation.)\ For this system the method employed in [@nw11] is not feasible, since an explicit transformation to Tikhonov standard form (in particular the requisite first integrals) seems to be unavailable. Maltose transport ----------------- In order to further illustrate the range of applicability of Proposition \[1dprop\], we discuss an example which is less straightforward from a computational perspective. Thus, we continue the discussion in [@gw2], Section 4, of a reaction equation proposed by Stiefenhofer [@sti] for maltose transport. According to the model, in order to pass through the cell membrane, a maltose molecule $X$ first reacts with a binding protein $Z$ to a complex $Y_1$. The latter reacts with the membrane-bound receptor $R$, forming a complex $Y_2$, which subsequently degrades, releasing maltose into the cell. This last process is modelled by a reaction involving the maltose concentration $X_i$ in the interior of the cell. Moreover, Stiefenhofer assumes a direct reaction between the binding protein and the membrane receptors, modelled by a further reaction. Altogether, the transport mechanism is modelled by the network $$\begin{aligned} &Y_2 \xrightarrow{k_{1}} R+Z+X_i,\quad Z+X\xrightleftharpoons[k_{-2}]{k_{2}} Y_1\\ &Y_1+R\xrightleftharpoons[k_{-3}]{k_{3}} Y_2,\quad Z+R\xrightleftharpoons[k_{-4}]{k_{4}} Y_3. \end{aligned}$$ In order to reduce notational and computational complexity, we follow Stiefenhofer by setting all rate constants equal to to 1, except for $k_1=\varepsilon$. Moreover we define $v:=(x,z,r,\xi,y_1,y_2,y_3)$, $\bar v:=(\xi,y_1,y_2,y_3)$ and assume $y_1(0)=y_2(0)=y_3(0)=0$. Now we can use the stoichiometric first integrals $$\begin{aligned} &\psi_1(v)=z+y_1+y_2+y_3,\quad \psi_2(v)=r+y_2+y_3,\quad \psi_3(v)=x+\xi+y_1+y_2 \end{aligned}$$ to write the reaction rates in the form $$\begin{aligned} \hat E_1(v)&=-y_2=:E_1(\bar v),\\ \hat E_2(v)&=y_1-zx\\ &=y_1-(z_0-(y_1+y_2+y_3))(x_0+\xi_0-(\xi+y_1+y_2))=:E_2(\bar v)\\ \hat E_3(v)&=y_2-y_1r\\ &=y_2-y_1(r_0-(y_2+y_3))=:E_3(\bar v)\\ \hat E_4(v)&=y_3-zr\\ &=y_3-(z_0-(y_1+y_2+y_3))(r_0-(y_2+y_3))=:E_4(\bar v). \end{aligned}$$ Thus the reaction is described by the following system $$\begin{aligned} &\dot \xi=-\varepsilon E_1(\bar v)\\ &\dot y_1=-E_2(\bar v)+E_3(\bar v)\\ &\dot y_2=\varepsilon E_1(\bar v)-E_3(\bar v)\\ &\dot y_3=-E_4(\bar v). \end{aligned}$$ As proven in [@gw2], there exists a (formal) reduction to $$\begin{aligned} &\xi^\prime=y_2\\ &y_1^\prime=\frac{y_2(y_1+y_2+y_3-z_0)}{n(\bar v)}\\ &y_2^\prime=-y_2-\frac{y_2(\xi-\xi_0+2(y_1+y_2)+y_3-(x_0+z_0+1))}{n(\bar v)}\\ &y_3^\prime=\frac{y_2((y_2+y_3)(y_1+y_2+y_3-r_0-z_0)+r_0(z_0-y_1))}{n(\bar v)} \end{aligned}$$ where $$n(\bar v)=\xi_0-\xi+(y_1+y_2+y_3-z_0)(y_2+y_3-r_0-1)-(y_1+y_2)+1+x_0.$$ The system is given on $K:=L\cap Y$, where $$L:=\{\bar v\in{\mathbb R}^4_+,\ y_1+y_2+y_3\leq z_0,\ y_2+y_3\leq r_0,\ \xi+y_1+y_2\leq \xi_0+x_0\}$$ is the chemically relevant region (determined by stoichiometry) and the curve $$Y:=\{\bar v\in{\mathbb R}^4_+, E_2(\bar v)=E_3(\bar v)=E_4(\bar v)=0\}$$ is the slow manifold.\ We first complete the discussion in [@gw2] by showing that all nonzero eigenvalues of the Jacobian have negative real parts; in particular we have convergence of the reduction. To this end, note that with $\varepsilon=0$ the Jacobian of the reaction equation can be written as $$\begin{pmatrix} 0&0&0&0\\ &-1-a-b-c& -a-b+1+d&-a+d\\ & c& -1-d&-d\\ & -c&-b-c& -1-b-c\\ \end{pmatrix}$$ with $$\begin{array}{rcl} a&:=&x_0+\xi_0-(\xi+y_1+y_2)\\ b&:=&z_0-(y_1+y_2+y_3)\\ c&:=&r_0-(y_2+y_3)\\ d&:=& y_1 \end{array}$$ all of which are nonnegative in view of the first integrals $\psi_1$, $\psi_2$ and $\psi_3$ and nonnegativity of concentrations. The characteristic polynomial $$x^3+A_1x^2+A_2x+A_3$$ of the lower right $3\times 3$ minor has all roots with negative real parts if (and only if) $A_1>0$, $H_2:=A_1A_2-A_3>0$ and $A_3>0$; see the Hurwitz-Routh criterion (Gantmacher [@gantmacher], Ch. V, §6). A straightforward computation (using the [Maple]{} software package) shows $$\begin{array}{rcl} A_1&=&3+2b+2c+d+a\\ H_2&=&{a}^{2}b+{a}^{2}c+{a}^{2}d+3\,a{b}^{2}+7\,abc+4\,abd+3\,a{c}^{2}+4\,acd\\ &+&a{d}^{2}+2\,{b}^{3}+7\,{b}^{2}c +3\,{b}^{2}d+7\,b{c}^{2}+6\,bcd\\ &+&b{d}^{2}+2\,{c}^{3}+3\,{c}^{2}d +c{d}^{2}+2\,{a}^{2}+10\,ba+9\,ca+6\,da+10\,{b}^{2}\\ &+&21\,cb+10\,db+9\,{c}^{2}+10\,cd+2\,{d}^{2}+8\,a+16\,b+14\,c+8\,d+8\\ A_3&=&{b}^{2}c+b{c}^{2}+bcd+ba+ca+da+{b}^{2}+2\,cb+db+a+2\,b+c+d+1 \end{array}$$ and the nonnegativity of $a,\ldots, d$ implies positivity of $A_1$, $H_2$ and $A_3$. Thus condition (TFII) holds.\ Now we address global convergence. No explicit parameterization of $Y$ seems to be known. Nonetheless in [@gw2] the existence of exactly one stationary point in $K$ was shown, and also its linear asymptotic stability on $Y$ and global asymptotic stability on $K$. By Theorem \[1dprop\] we obtain the desired convergence result. A two-dimensional slow manifold ------------------------------- Some variants of Michaelis-Menten may lead to two-dimensional slow manifolds, depending on the parameters. We consider again competitive inhibition with irreversible product formation (see subsection \[compinone\]), but now with small parameters $k_1=\varepsilon k_1^$, $k_{-1}=\varepsilon k_{-1}^$ and $k_2=\varepsilon k_2^$. The differential equation is $$\begin{aligned} &\dot s=\varepsilon\left[k_{-1}^c_1-k_1^s(e_0-c_1-c_2)\right]\\ &\dot c_1=\varepsilon\left[k_1^s(e_0-c_1-c_2)-(k_{-1}^+k_2^)c_1\right]\\ &\dot c_2=k_3(e_0-c_1-c_2)(i_0-c_2)-k_{-3}c_2 \end{aligned}$$ on the (chemically relevant) positively invariant compact set $$L:=\{(s,c_1,c_2)\in {\mathbb R}^3_+,\ c_1+c_2\leq e_0,\ s+c_1\leq s_0,\ c_2\leq i_0\}.$$ A short computation (see [@gwz], p. 1175 f.) shows that (TFII) is satisfied, and that there exists a convergent Tikhonov-Fenichel reduction to $$\begin{aligned} &s'=k_{-1}^c_1-k^_1s(e_0-c_1-c_2)\label{inhib1}\\ &c_1'=k_1^s(e_0-c_1-c_2)-(k_{-1}^+k_2^)c_1\label{inhib2}\\ &c_2'=\frac{-(i_0-c_2)[k_1^s(e_0-c_1-c_2)-(k_{-1}^+k_2^)c_2]}{\kappa+e_0+i_0-c_1-2c_2} \end{aligned}$$ in $L\cap Y$ with the two-dimensional asymptotic slow manifold $$Y:=\{(s,c_1,c_2)\in {\mathbb R}^3_+,\ (e_0-c_1-c_2)(i_0-c_2)-\kappa c_2=0\}$$ and $\kappa:=\frac{k_{-3}}{k_3}$. (Note that $\kappa+e_0+i_0-c_1-2c_2>0$ on $L\cap Y$.)\ Thus, with $$\label{inhibman} c_2=\vartheta(c_1):=\frac{\kappa+e_0+i_0-c_1-\sqrt{(\kappa+e_0+i_0-c_1)^2-4i_0(e_0-c_1)}}{2},$$ $L\cap Y$ is contained in the graph of a function of $s$ and $c_1$, and it suffices to analyze – in the positively invariant compact set $$\widetilde L:=\{(s,c_1)\in {\mathbb R}^2_+,\ c_1+\vartheta(c_1)\leq e_0,\ s+c_1\leq s_0,\, \vartheta(c_1)\leq i_0\}.$$ We will abbreviate this system as $$\begin{pmatrix} s\\ c_1\end{pmatrix}^\prime =q\left(\begin{pmatrix} s\\ c_1\end{pmatrix}\right).$$ It is easy to see that $(0,0)$ is the only stationary point in $\widetilde L$. We construct a suitable Lyapunov function. Let $\alpha>0$ and note that $$\begin{array}{rcccl} \varphi_1&:=&s+c_1&\mbox{ satisfies }& L_q(\varphi_1)= -k_2^c_1,\\ \varphi_2&:=&\alpha s&\mbox{ satisfies }& L_q(\varphi_2)= \alpha(k_{-1}^+k_1^s)c_1 -\alpha k_1^s(e_0-\vartheta(c_1)).\\ \end{array}$$ On $L\cap Y$ we have $0\leq k_{-1}^+k_1^s\leq k_{-1}^+k_1^s_0$, and furthermore $e_0-c_2>0$. The first of these assertions is obvious. To verify the second, note that $e_0-c_1-c_2\geq 0$ by stoichiometry, whence $e_0-c_2=0$ implies $c_1=0$ and, by the defining equation for $Y$, $$0=(e_0-c_1-c_2)(i_0-c_2)=\kappa c_2=\kappa e_0>0,$$ a contradiction.\ By compactness, there exists $\beta >0$ such that $e_0-c_2\geq \beta$ for all points in $L$. Now choose $\alpha>0$ such that $$\alpha\cdot(k_{-1}^+k_1^s_0)<k_2^,$$ and define $\varphi:=\varphi_1+\varphi_2$. Then, by the previous estimates we have on $\widetilde L$: $$\begin{array}{rcl} L_q(\varphi)(s,c_1)&\leq& -(k_2^-\alpha\cdot(k_{-1}^+k_1^s_0))c_1 -\alpha k_1^\beta s\\ &\leq & -\nu\cdot ((1+\alpha)s+c_1) =-\nu\cdot \varphi \end{array}$$ for some $\nu>0$. Thus (LC) holds, and by Theorem \[hshsatz\] convergence to the reduced system holds for all positive times. Concluding remarks ================== We finish with a few observations, and some comments on possible extensions and generalizations. - The results of the present paper will be hardly surprising to application-oriented readers (e.g. with a background in biochemistry) although the underlying mathematical argument (essentially due to Hoppensteadt [@Hoppensteadt]) is far from trivial, and not easy to transfer to applications. This gap between intuition (where matters may seem obvious) and rigorous proofs (which may require an extensive technical build-up) can be observed quite frequently. The authors’ main goal was to facilitate the applicability of Hoppensteadt’s theorem to relevant settings. - By its nature Theorem \[hshsatz\] is local, but the domain of attraction of $Y$ may properly contain $\widetilde K$. There remains, however, the question how fast a solution approaches the slow manifold.\ In the setting of reaction equations it is known from the work of Horn and Jackson [@horja] and Feinberg [@feinberg] that global Lyapunov functions often exist; for instance this is the case for deficiency zero and complex balanced systems. Given a system with slow and fast reactions, a Lyapunov function for the fast subsystem could imply a condition akin to Hoppensteadt’s condition (VII), with reasoning similar to (and extending) Lemma \[lyapunov\]. But the Lyapunov functions from the cited papers do not generally satisfy the hypotheses of the Lemma; thus a case-by-case analysis would be in order. - In Lemma \[1dlyapunov\], if $K$ is a neighborhood of $0$ then it suffices that $p$ changes sign at $0$, with the lowest (necessarily odd) order nonzero derivative at $0$ being negative. By analogous estimates one obtains conditions (ii) and (iii) of Lemma \[lyapunov\], with exponents $a>2$ and $k>1$. In order to transfer this to systems with one-dimensional slow manifold one would have to generalize Proposition \[1dprop\] by requiring suitable properties of the Poincaré-Dulac normal form at $z$. Appendix: Hoppensteadt’s conditions {#appsec} =================================== For the reader’s convenience we recall here the conditions and the main result from Hoppensteadt’s original paper [@Hoppensteadt]. Recall the notation $S_R$ from . Hoppensteadt considers a non-autonomous system that is given in Tikhonov standard form $$\begin{aligned} &y_1'= f(\tau,y_1,y_2,\varepsilon)\label{tnfhs1na}\\ &y_2'=\varepsilon^{-1}g(\tau,y_1,y_2,\varepsilon)\label{tnfhs2na} \end{aligned}$$ with $f$ and $g$ defined on an open set $$[0,\infty)\times D \times [0,\varepsilon_0)\subseteq [0,\infty)\times\mathbb R^s\times \mathbb R^r\times [0,\varepsilon_0)\to \mathbb R^r$$ which satisfies $S_R\subseteq D$ for some $R>0$, and $f,\,g$ having values in $\mathbb R^s$ and $\mathbb R^r$, respectively. Assume that the following conditions hold: - The system $$\begin{aligned} y_1'&=f(\tau,y_1,y_2,0) \label{redSystem1na}\\ 0&=g(\tau,y_1,y_2,0)\label{redSystem2na} \end{aligned}$$ admits a solution $Y:\,[0,\infty)\to \mathbb R^{s+r}$, $\tau\mapsto Y(\tau)$. With a suitable transformation of – one may assume that $\widetilde Y\equiv0$ is a solution of the transformed system. From here on, it will be assumed that – admits the solution $\widetilde Y\equiv 0$. - The functions $f$, $g$ and their partial derivatives with respect to $\tau$, $y_1$, $y_2$ respectively satisfy $$f,g,D_1f,D_2f,\partial_{\tau}g,D_1g,D_2g \in C([0,\infty)\times S_R\times [0,\varepsilon_0]).$$ - There exists an isolated and bounded $C^2$-solution $Y_2=Y_2(\tau,y_1)$ of the implicit equation $$g(\tau,y_1,Y_2(\tau,y_1),0)=0$$ for all $\tau\in [0,\infty)$ and $y_1\in S_1$. By a transformation $\widetilde y_1=y_1$ and $\widetilde y_2+Y(\tau,y_1)=y_2$, one may obtain that $Y_2(\tau,\widetilde y_1)=0$ for all $(\tau,\widetilde y_1)\in [0,\infty)\times S_1$. This will be assumed for the original system – in the following. - $f(\cdot,\cdot,0,0)$ is uniformly continuous in $[0,\infty)\times S_{1,R}$, and moreover $f(\cdot,\cdot,0,0)$ and $D_1f(\cdot,\cdot,0,0)$ are bounded in $[0,\infty)\times S_{1,R}$. - $g(\cdot,\cdot,\cdot,0)$ is uniformly continuous in $[0,\infty)\times S_R$, and moreover $g(\cdot,\cdot,\cdot,0)$, $\partial_{\tau}g(\cdot,\cdot,\cdot,0)$, $D_1g(\cdot,\cdot,\cdot,0)$ $D_2g(\cdot,\cdot,\cdot,0)$ are bounded in $[0,\infty)\times S_R$. - The solution $\widetilde Y\equiv0$ of $$\label{(D)} y'=f(\tau,y,0,0)$$ is uniformly asymptotically stable in the following sense: There exist a continuous, strictly increasing function $$d\colon [0,\infty)\to[0,\infty)\quad \text{with }d(0)=0$$ and a continuous, strictly decreasing function $\sigma\colon [0,\infty)\to[0,\infty)$ with $\lim_{s\to\infty}\sigma(s)=0$, such that for every solution $\Phi(\tau,z)$ of with initial value $y_1(0)=z\in S_{1,R}$ and all $\tau\geq0$ one has $${\ensuremath{\left\vert\Phi(\tau,z)\right\vert}}_1\leq d({\ensuremath{\left\vertz\right\vert}}_1)\cdot\sigma(\tau).$$ - For all $(\alpha,\beta)\in[0,\infty)\times S_{1,R}$ the solution $\widetilde Y\equiv0$ of $$\dot x=g(\alpha,\beta,x,0)$$ is uniformly asymptotically stable in the following sense: There exist a continuous, strictly increasing function $$e\colon [0,\infty)\to[0,\infty)\quad \text{with }e(0)=0$$ and a continuous, strictly decreasing function $\rho\colon [0,\infty)\to[0,\infty)$ with $\lim_{s\to\infty}\rho(s)=0$, such that for every solution $\Psi(t,x_{0};\alpha,\beta)$ of the equation with initial value $x(0)=x_{0}\in S_{2,R}$ and parameters $(\alpha,\beta)\in[0,\infty)\times S_{1,R}$ and all $t\geq0$ one has $${\ensuremath{\left\vert\Psi(t,x_0;\alpha,\beta)\right\vert}}_1\leq e({\ensuremath{\left\vertx_0\right\vert}}_1)\cdot\rho(t).$$ Given these assumptions, Hoppensteadt’s main result [@Hoppensteadt] can be stated as follows: \[HoppensteadtSatz\] There exists a compact neighborhood $K\subset S_R$ of $0$ and $\varepsilon_0^\in (0,\varepsilon_0)$ such that the solution $\Phi(t,y_0,\varepsilon)$ of – with initial value $y(0)=y_0:=(y_{1,0},y_{2,0})\in K$ at $\tau=0$ exists for all positive times provided that $0<\varepsilon<\varepsilon_0^$. Moreover $\Phi(t,y_0,\varepsilon)$ converges uniformly on all closed subsets of $(0,\infty)$ towards the solution of – with initial value $y_1(0)=y_{1,0}$, as $\varepsilon\to0$. [99]{} H. Amann: [Ordinary Differential Equations. An Introduction to Nonlinear Analysis]{}. Walter de Gruyter, Berlin - New York (1990). J. Back, H. Shim: [Adding robustness to nominal output-feedback controllers for uncertain nonlinear systems: A nonlinear version of disturbance observer.]{} Automatica [44]{}, 2528–2537 (2008). G. E. Briggs and J. B. S. Haldane. A note on the kinetics of enzyme action. , 19:338–339, 1925. A. Cavallo, C. Natale: [Output feedback control based on a high-order sliding manifold approach.]{} IEEE Transactions on Automatic Control [48]{}, 469–472 (2003). M. Feinberg: [The existence and uniqueness of steady states for a class of chemical reaction networks.]{} Arch. Ration. Mech. Anal. [132]{}, 311–370 (1995). N. Fenichel: [Geometric singular perturbation theory for ordinary differential equations]{}. [J. Differential Equations]{} [31]{}, [53–98]{} (1979). F.R. Gantmacher: [Applications of the theory of matrices.]{} Dover Publ., Mineola, NY (2005). A. Goeke, C. Schilli, S. Walcher, E. Zerz: [Computing quasi-steady state reductions]{}. [J. Math. Chem.]{} [50]{}, [1495–1513]{} (2012). A. Goeke, S. Walcher: [A constructive approach to quasi-steady state reductions]{}. [J. Math. Chem.]{} [52]{}, [2596–2626]{} (2014). A. Goeke, S. Walcher, E. Zerz: [Determining “small parameters” for quasi-steady state]{}. [J. Differential Equations]{} [259]{}, [1149–1180]{} (2015). M.W. Hirsch: [Differential Topology.]{} Springer, New York (1976). F. Horn, R. Jackson: [General mass action kinetics.]{} Arch. Ration. Mech. Anal. [47]{}, 8–116 (1972). F.C. Hoppensteadt: [Singular Perturbations on the Infinite Interval]{}. [Trans. Amer. Math. Soc.]{} [123]{}, 521–535 (1966). J. Keener, J. Sneyd: [Mathematical physiology I: Cellular physiology]{}, Second Ed. Springer-Verlag, New York (2009). C. Lax: [Analyse und asymptotische Analyse von Kompartimentsystemen.]{} Doctoral dissertation, RWTH Aachen (2016). L. Michaelis and M. L. Menten. . , 49:333–369, 1913. J.W. Milnor: [Topology from the Differentiable Viewpoint.]{} Princeton University Press, Princeton (1997). J. Nestruev: [Smooth manifolds and observables]{}. Springer, New York (2003). L. Noethen, S. Walcher: [Tikhonov’s theorem and quasi-steady state]{}. [Discrete Contin. Dyn. Syst. Ser. B]{} [16]{}, [945–961]{} (2011). L.A. Segel, M. Slemrod: [The quasi-steady-state assumption: A case study in perturbation.]{} SIAM Review [31]{}, 446 - 477 (1989). K. Seliger: [Singuläre Störungen auf unbeschränkten Intervallen.]{} Master’s thesis, RWTH Aachen (2015). M. Stiefenhofer: [Quasi-steady-state approximation for chemical reaction networks]{}. [J. Math. Biol.]{} [36]{}, [593–609]{} (1998). A.R. Teel, L. Moreau, D. Nesic: [A unified framework for input-to-state-stability in systems with two time scales.]{} IEEE Transactions on Automatic Control [48]{}, 1526–1544 (2003). A.N. Tikhonov: [Systems of differential equations containing a small parameter multiplying the derivative (in Russian)]{}. [Math. Sb.]{} [31]{}, [575–586]{} (1952). W. Walter: [Ordinary differential equations.]{} Springer-Verlag, New York (1998).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Deep learning (deep structured learning, hierarchical learning or deep machine learning) is a branch of machine learning based on a set of algorithms that attempt to model high-level abstractions in data by using multiple processing layers with complex structures or otherwise composed of multiple non-linear transformations. In this paper, we present the results of testing neural networks architectures on H2O platform for various activation functions, stopping metrics, and other parameters of machine learning algorithm. It was demonstrated for the use case of MNIST database of handwritten digits in single-threaded mode that blind selection of these parameters can hugely increase (by 2-3 orders) the runtime without the significant increase of precision. This result can have crucial influence for opitmization of available and new machine learning methods, especially for image recognition problems.' author: - title: Comparative Performance Analysis of Neural Networks Architectures on H2O Platform for Various Activation Functions --- deep learning; neural networks; classification; single-threaded mode; H2O Introduction ============ Machine learning (ML) is a subfield of Artificial Intelligence (AI) discipline. This branch of AI involves the computer applications and/or systems design that based on the simple concept: get data inputs, try some outputs, build a prediction. Nowadays, ML has driven advances in many different fields [@Kochura] like pedestrian detection, object recognition, visual-semantic embedding, language identification, acoustic modeling in speech recognition, video classification, fatigue estimation [@Gordienko], generation of alphabet of symbols for multimodal human-computer interfaces [@Hamotskyi], etc. This success is related to the invention of more sophisticated machine learning models and the development of software platforms that enable the easy use of large amounts of computational resources for training such models [@Sibi]. H2O is one of such open source deep learning platform. We have tested the H2O system by using the publicly available MNIST dataset of handwritten digits. This dataset contains 60,000 training images and 10,000 test images of the digits 0 to 9. The images have grayscale values in the range 0:255. Figure \[fig1\] gives an example images of handwritten digits that were used in testing. We have trained the net by using the host with Intel Core i7-2700K insight. The computing power of this CPU approximately is 29.92 GFLOPs. In this paper, we present testing results of various net architectures by using H2O platform for single-threaded mode. Our experiments show that net architecture based on cross entropy loss function, tanh activation function, logloss and MSE stopping metrics demonstrates better efficiency by recognition handwritten digits than other available architectures for the classification problem. This paper is structured as follows. Section \[Motivation\] describes the motivation for this idea. Section \[Common Information\] introduces us with activation and loss functions and describes parameters of deep neural nets that were used in the experiments. In Section \[Experimental Results\], we present our experimental results where we apply different activation functions and stopping metrics to the classification problem with use case in single-threaded mode. Section \[Conclusion\] contains the conclusions of the work. ![Example Images of Handwriten Digits[]{data-label="fig1"}](example.png){width="3in"} Motivation {#Motivation} ========== A motivation for providing testing of various neural network architectures on H2O platform has related to lacking publicly available similar publications, especially with an emphasize on influence of some parameters of machine learning method like activation function, stopping metrics, etc. This paper is a logical continuation of the previous paper [@Kochura] and it aims to learn the main features of H2O platform for classification problem. Common Information {#Common Information} ================== The Activation Functions ------------------------ Activation functions also known as transfer functions are used to map input nodes to output nodes in certain fashion [@AF] (see the conceptual scheme of an activation function in Figure \[fig1\_1\]). We are considering here most common activation functions that are widely using for deep learning. ![The Role of Activation Function in the Process of Learning Neural Net [@ANN][]{data-label="fig1_1"}](AF.png){width="3in"} The rectifier activation function is defined as $$\label{rectifier} f(x) = \max(0, x)$$ where $x$ is the input to a neuron. When given an input $x \in R^d$, a maxout activation function defins as folows: $$\label{maxout} f_i(x) = \max_{j \in [1,k]}(x_{ij})$$ where $x_{ij} = x^T \cdot W_{ij} + b_{ij}$, $W \in R^{d \times m \times k}$ and $b_{ij} \in R^{m \times k}$ are the learned parameters. The tanh activation function is defined as $$\label{tanh} f(x) = tanh(x) = \frac{2}{1 + \exp^{-2 \cdot x}} - 1$$ Functions with dropout are used for reducinng overfitting by preventing complex co-adaptations on training data. This technique is known as regularization. Figure \[fig2\] demonstrate the difference between standatd neural net and neural net affter applying dropout [@GHinton]. Constant Parameters of the Training Model ----------------------------------------- We have used the network model with such paraeters, namely: - Response variable column is C785 - Hidden layer sise is \[50,50\] - Epochs are 500 - Seed for random numbers is 2 - Adaptive learning rate is false - Initial momentum at the beginning of taining is 0.9 - Final momentum after the ramp is 0.99 - Input layer dropout ratio for improving generalization is 0.2 - Stopping criterion for classification error fraction on training data is disable - Early stopping based on convergence of stopping metric is 3 - Relative tolerance for metric-based stopping criterion is 0.01 - Compute variable impotances for input features is true - Sparse data handling is true - Rorce reproducibility on small data is true Variable Parameters of the Training Model ----------------------------------------- - Activation function: Tanh, TanhWithDropout, Maxout, MaxoutWithDropout, Rectifier, RectifierWithDropout - Metric to use for early stopping: logloss, misclassification, MAE, MSE, RMSE and RMSLE - Loss function: Cross Entropy Loss function is a function that used to measure the degree of fit. The cross entropy loss function for the distributions $p$ and $q$ over a given set is defined as follows: $$\label{H} H(p, q) = H(p) + D_{KL}(p\|\|q)$$ where $H(p) $ is the entropy of $p$, and $D_{KL}(p\|\|q)$ is the Kullback–Leibler divergence of $q$ from $p$ (also known as the relative entropy of $p$ with respect to $q$ [@CE]. Cross entropy is always larger than entropy. Experimental Results {#Experimental Results} ==================== We trained neural networks for classification problems on publicly available MNIST dataset of handwritten digits with use case in single-threaded mode. We found that generalization performance has very strong dependence on activation function and very slight dependence on stopping metric. Tables \[duration\] and \[duration2\] give give the values of runtime for the models used. Figure \[fig3\] shows the runtime values on the logarithm scale obtained for these different architectures as training progresses. [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ Tables \[logloss\] and \[logloss2\] contain the information of training errors for various activation function and stopping metrics. Figure \[fig4\] demonstrates the effectiveness of using tanh activation function for all stopping metrics that considered in this paper. In the case of the learning net based on the tanh activation function, MAE and RMSLE stopping metric hase achieved the logloss value of 0.0104. This architectures demonstrate better training prediction ability than others but take much time for building model. In order to find the best neural net architecture for digits recognition just needs to look at the behavior of models on unknown data should be checked. Tables \[Vlogloss\] and \[Vlogloss2\] give the information that can help us to find out the net architecture that provides the best performance in case of using single-threaded mode. Figure \[fig5\] shows the validation error rates for different architectures that are considered here. We see, the best digits recognition results were achieved in the case of tanh activation function. The type of stopping metric is very slightly effects on the values of the validation error but it does very much on the runtime of building model. [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ [lllll]{} & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ ![Runtime of Learning Nets Different Architectures[]{data-label="fig3"}](1.png){width="3.5in"} ![Training Logloss of Learning Nets Different Architectures[]{data-label="fig4"}](2.png){width="3.5in"} ![Validation Logloss of Learning Nets Different Architectures[]{data-label="fig5"}](3.png){width="3.5in"} Conclusion {#Conclusion} ========== H2O is widely used open source deep learning platform.In this paper, we present the results of testing neural networks architectures on H2O platform for various activation functions, stopping metrics, and other parameters of machine learning algorithm. It was demonstrated for the use case of MNIST database of handwritten digits in single-threaded mode that blind selection of these parameters can hugely increase (by 2-3 orders) the runtime without the significant increase of precision. This result can have crucial influence for opitmization of available and new machine learning methods, especially for image recognition problems. During the process of testing H2O, we found out that generalization performance has very strong dependence on activation function and very slight dependence on stopping metric. The best results of recognition digits were achieved in case of using nets architecture based on tanh activation function, logloss and MSE stopping metrics. [99]{} Yu. Kochura, S. Stirenko, A. Rojbi, O. Alienin, M. Novotarskiy, Yu. Gordienko “Comparative Analysis of Open Source Frameworks for Machine Learning with Use Case in Single-Threaded and Multi-Threaded Modes,” XII-th International Scientific and Technical Conference “Computer Science and Information Technologies”., in press. N. Gordienko, S. Stirenko, Yu. Kochura, O. Alienin, M. Novotarskiy, Yu. Gordienko, A. Rojbi (2017), Deep Learning for Fatigue Estimation on the Basis of Multimodal Human-Machine Interactions, XXIX IUPAP Conference on Computational Physics (CCP2017) (Paris, France). S. Hamotskyi, A. Rojbi, S. Stirenko, and Yu. Gordienko (2017), Automatized Generation of Alphabets of Symbols for Multimodal Human-Computer Interfaces, Proc. of Federated Conference on Computer Science and Information Systems, Prague (FedCSIS-2017) (Prague, Czech Republic). P. Sibi, S. A. Jones, P. Siddarth, (2013). Analysis of different activation functions using back propagation neural networks. Journal of Theoretical and Applied Information Technology, 47(3), ‎1264-1268. Activation Functions. URL: https://medium.com/towards-data-science/activation-functions-in-neural-networks-58115cda9c96 Artificial Neural Networks/Activation Functions. URL: https://en.wikibooks.org/wiki/Artificial\_Neural\_Networks/ Activation\_Functions. N.Srivastava, \[et al.\] “Dropout: A Simple Way to Prevent Neural Networks from Overfitting”, Journal of Machine Learning Research, 2014, pp. 929-1958. Cross entropy. URL: https://en.wikipedia.org/wiki/Cross\_entropy.	{ "pile_set_name": "ArXiv" }
--- author: - 'Liu Xiong, Li Baode, Qiu Xiaoli and Li Bo$^\ast$\' date: title: 'Estimates for Parametric Marcinkiewicz Integrals on Musielak-Orlicz Hardy Spaces ' --- Introduction\[s1\] ================== Suppose that $S^{n-1}$ is the unit sphere in the $n$-dimensional Euclidean space $\rn \ (n\ge2)$. Let $\Omega$ be a [homogeneous function of degree zero]{} on $\rn$ which is locally integrable and satisfies the cancellation condition $$\begin{aligned} \label{e1.1} \int_{S^{n-1}}\Omega(x')\,d\sigma(x')=0,\end{aligned}$$ where $d\sigma$ is the Lebesgue measure and $x':=x/{\|x\|}$ for any $x\neq{\mathbf{0}}$. For a function $f$ on $\rn$, the parametric [Marcinkiewicz integral]{} $\mu^\rho_\Omega$ is defined by setting, for any $x\in\rn$ and $\rho\in(0,\,\infty)$, $$\mu^\rho_\Omega(f)(x):=\lf( \int_{0}^{\fz}\lf\|F^\rho_{\Omega,\,t}(f)(x)\r\|^2\,\frac{dt}{t^{2\rho+1}}\r)^{1/2},$$ where $$F^\rho_{\Omega,\,t}(f)(x):=\int_{\|x-y\|\leq t}\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}f(y)\,{dy}.$$ When $\rho:=1$, we shall denote $\mu^1_\Omega$ simply by $\mu_\Omega$, which is reduced to the classic Marcinkiew- icz integral. In 1938, Marcinkiewicz [@m38] first defined the operator $\mu_\Omega$ for $n=1$ and $\Omega(t):={\rm{sign}}\,t$. The Marcinkiewicz integral of higher dimensions was studied by Stein [@s58] in 1958. He showed that, if $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$ with $\alpha\in(0,\,1]$, then $\mu_\Omega$ is bounded on $L^p(\rn)$ with $p\in(1,\,2]$ and bounded from $L^1(\rn)$ to weak $L^1(\rn)$. On the other hand, in 1960, Hörmander [@h60] proved that, if $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$ with $\alpha\in(0,\,1]$, then $\mu^\rho_\Omega$ is bounded on $L^p(\rn)$ provided that $p\in(1,\,\fz)$ and $\rho\in (0,\,\infty)$. Notice that all the results mentioned above hold true depending on some smoothness condition of $\Omega$. However, in 2009, Jiang et al. [@sj09] obtained the following celebrated result that $\mu^\rho_\Omega$ is bounded on $L^p_\omega(\rn)$ without any smoothness condition of $\Omega$, where $\omega\in A_p$ and $A_p$ denotes the Muckenhoupt weight class. [[Theorem A.]{}]{} [[ Let $\rho\in(0,\,\infty)$, $q\in(1,\,\fz)$, $q':=q/(q-1)$ and $\Omega\in L^q(S^{n-1})$ satisfying . If $\omega^{q'}\in A_p$ with $p\in(1,\,\fz)$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu^\rho_\Omega(f)\r\\|_{L^p_\omega(\rn)}\le C\\|f\\|_{L^p_\omega(\rn)}.$$ ]{}]{} It is now well known that Hardy space $H^p(\rn)$ is a good substitute of the Lebesgue space $L^p(\rn)$ with $p\in(0,\,1]$ in the study for the boundedness of operators and hence, in 2007, Lin et al. [@ll07] proved that the $\mu_\Omega$ is bounded from weighted Hardy space to weighted Lebesgue space under weaker smoothness condition assumed on $\Omega$, which is called $L^q$-Dini type condition of order $\alpha$ with $q\in[1,\,\fz]$ and $\az\in(0,\,1]$ (see Section \[s4\] below for its definition). In 2016, Wang [@w16] discussed the boundedness of $\mu^\rho_\Omega$ from weighted Hardy space to weighted Lebesgue space or to weighted weak Lebesgue space if $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$ with $\az\in(0,\,1]$. More conclusions of Marcinkiewicz integral are referred to [@ak14; @gahk16; @lll17]. On the other hand, recently, Ky [@k14] studied a new Hardy space called Musielak-Orlicz Hardy space $H^\vz(\rn)$, which generalizes both of the weighted Hardy space (cf. [@st89]) and the Orlicz Hardy space (cf. [@j80; @jy10]), and hence has a wide generality. Apart from interesting theoretical considerations, the motivation to study $H^\vz(\rn)$ comes from applications to elasticity, fluid dynamics, image processing, nonlinear PDEs and the calculus of variation (cf. [@d05; @d09]). More Musielak-Orlicz-type spaces are referred to [@lhy12; @hyy13; @ly13; @ccyy16; @lffy16; @lsl16; @fhly17; @ylk17]. In light of Lin [@ll07], Wang [@w16] and Ky [@k14], it is a natural and interesting problem to ask whether parametric Marcinkiewicz integral $\mu^\rho_\Omega$ is bounded from $H^\vz(\rn)$ to $L^\vz(\rn)$ (resp. $WL^\varphi(\rn)$) under weaker smoothness condition (resp. some Lipschitz condition) assumed on $\Omega$. In this paper we shall answer this problem affirmatively. Precisely, this paper is organized as follows. In Section \[s2\], we recall some notions concerning Muckenhoupt weights, growth functions, Musielak-Orlicz space $L^\vz(\rn)$ and weak Musielak-Orlicz space $WL^\vz(\rn)$. Then we establish the completeness of $L^\vz(\rn)$ and $WL^\vz(\rn)$ (see Theorems \[lwbx\] and \[wlwbx\] below). Section \[s3\] is devoted to establishing two boundedness criterions of operators from $H^\vz(\rn)$ to $L^\vz(\rn)$ or from $H^\vz(\rn)$ to $WL^\vz(\rn)$ (see Theorems \[yt\] and \[yt2\] below). In the process of the proofs of Theorem \[yt\] and Theorem \[yt2\], the Aoki-Rolewicz theorem (see Lemma \[ardl\] below) and the weak type superposition principle (see Lemma \[dj\] below) play indispensable roles, respectively. In Section \[s4\], we obtain the boundedness of $\mu^\rho_\Omega$ from $H^\vz(\rn)$ to $L^\vz(\rn)$ (resp. $WL^\varphi(\rn)$) under weaker smoothness condition (resp. some Lipschitz condition) assumed on $\Omega$ (see Theorem \[dingli.1\], Theorem \[dingli.2\], Corollary \[tuilun.1\], Theorem \[dingli.3\] and Theorem \[dingli.4\] below). In the process of the proof of Theorem \[dingli.1\], it is worth pointing out that, since the space variant $x$ and the time variant $t$ appeared in $\vz(x,\,t)$ are inseparable, we can not directly use the method of Lin [@ll07]. This difficulty is overcame via establishing a more subtle pointwise estimate for $\mu^\rho_\Omega(b)$ (see Lemma \[lemma.1\] below for more details), where $b$ is a multiple of an atom. Finally, we make some conventions on notation. Let $\zz_+:=\{1,\, 2,\,\ldots\}$ and $\nn:=\{0\}\cup\zz_+$. For any $\bz:=(\bz_1,\ldots,\bz_n)\in\nn^n$, let $\|\bz\|:=\bz_1+\cdots+\bz_n$. Throughout this paper the letter $C$ will denote a positive constant that may vary from line to line but will remain independent of the main variables. The symbol $P\ls Q$ stands for the inequality $P\le CQ$. If $P\ls Q\ls P$, we then write $P\sim Q$. For any sets $E,\,F \subset \rn$, we use $E^\complement$ to denote the set $\rn\setminus E$, $\|E\|$ its [[$n$-dimensional Lebesgue measure]{}]{}, $\chi_E$ its characteristic function and $E+F$ the [algebraic sum]{} $\{x+y:\ x\in E,\,y\in F\}$. For any $s\in\rr$, $\lfloor s\rfloor$ denotes the unique integer such that $s-1<\lfloor s\rfloor\le s$. If there are no special instructions, any space $\mathcal{X}(\rn)$ is denoted simply by $\mathcal{X}$. For instance, $L^2(\rn)$ is simply denoted by $L^2$. For any index $q\in[1,\,\fz]$, $q'$ denotes the [[conjugate index]{}]{} of $q$, namely, $1/q+1/{q'}=1$. For any set $E$ of $\rn$, $t\in[0,\,\infty)$ and measurable function $\vz$, let $\vz(E,\,t):=\int_E\vz(x,\,t)\,dx$ and $\{\|f\|>t\}:=\{x\in\rn: \ \|f(x)\|>t\}$. As usual we use $B_r$ to denote the ball $\{x\in\rn:\ \|x\|<r\}$ with $r\in(0,\,\fz)$. Completeness of $L^\vz$ and $WL^\vz$ {#s2} ==================================== In this section, we first recall some notions concerning Muckenhoupt weights, growth functions, Musielak-Orlicz space $L^\vz$ and weak Musielak-Orlicz space $WL^\vz$, and then establish the completeness of $L^\vz$ and $WL^\vz$. Recall that a nonnegative function $\vz$ on $\rn\times[0,\,\fz)$ is called a [Musielak-Orlicz function]{} if, for any $x\in\rn$, $\vz(x,\,\cdot)$ is an Orlicz function on $[0,\,\fz)$ and, for any $t\in[0,\,\fz)$, $\vz(\cdot\,,t)$ is measurable on $\rn$. Here a function $\phi: [0,\,\fz) \to [0,\,\fz)$ is called an [Orlicz function]{}, if it is nondecreasing, $\phi(0) = 0$, $\phi(t) > 0$ for any $t\in(0,\,\fz)$, and $\lim_{t\to\fz} \phi(t) = \fz$. Given a Musielak-Orlicz function $\vz$ on $\rn\times[0,\,\fz)$, $\vz$ is said to be of [[uniformly lower]{}]{} (resp. [[upper]{}]{}) [[type]{}]{} $p$ with $p\in(-\fz,\,\fz)$, if there exists a positive constant $C:=C_{\vz}$ such that, for any $x\in\rn$, $t\in[0,\,\fz)$ and $s\in(0,\,1]$ (resp. $s\in[1,\,\fz)$), $$\begin{aligned} \vz(x,\,st)\le C s^p\vz(x,\,t).\end{aligned}$$ The [critical uniformly lower type index]{} and the [critical uniformly upper type index]{} of $\vz$ are, respectively, defined by $$\begin{aligned} \label{e2.1} i(\vz):=\sup\{ p\in(-\fz,\,\fz): \vz \mathrm{ \ is \ of \ uniformly\ lower\ type \ {\it p}} \},\end{aligned}$$ and $$\begin{aligned} \label{e2.1.1} I(\vz):=\inf\{p\in(-\fz,\,\fz):\vz \mathrm{ \ is \ of \ uniformly\ upper\ type \ {\it p}} \}.\end{aligned}$$ Observe that $i(\vz)$ or $I(\vz)$ may not be attainable, namely, $\vz$ may not be of uniformly lower type $i(\vz)$ or of uniformly upper type $I(\vz)$ (see [@lhy12 p.415] for more details). \[d2.2\] 1. Let $q\in[1,\,\fz)$. A locally integrable function $\vz(\cdot\,,t): \rn \to [0,\,\fz)$ is said to satisfy the [uniform Muckenhoupt condition]{} $\aa_q$, denoted by $\vz\in \aa_q$, if there exists a positive constant $C$ such that, for any ball $B\subset\rn$ and $t\in(0,\,\fz)$, when $q=1$, $$\frac{1}{\|B\|}\int_{B} \vz(x,\,t)\,dx\lf\{\esup_{x\in B} [\vz(x,\,t)]^{-1}\r\}\le C$$ and, when $q\in(1,\fz)$, $$\frac{1}{\|B\|}\int_{B}\vz(x,\,t)\,dx \lf\{\frac{1}{\|B\|}\int_{B}[\vz(x,\,t)]^{-\frac{1}{q-1}}\,dx\r\}^{q-1} \le C.$$ 2. Let $q\in(1,\,\fz]$. A locally integrable function $\vz(\cdot\,,t): \rn \to [0,\,\fz)$ is said to satisfy the [[uniformly reverse Hölder condition]{}]{} $\mathbb{RH}_q$, denoted by $\vz\in \mathbb{RH}_q$, if there exists a positive constant $C$ such that, for any ball $B\subset\rn$ and $t\in(0,\,\fz)$, when $q\in(1,\fz)$, $$\lf\{\frac{1}{\|B\|}\int_B\vz(x,\,t)\,dx\r\}^{-1}\lf\{\frac{1}{\|B\|}\int_B[\vz(x,\,t)]^q\,dx\r\}^{1/q}\leq C$$ and, when $q=\fz$, $$\lf\{\frac{1}{\|B\|}\int_B\vz(x,\,t)\,dx\r\}^{-1}\esup_{x\in B} \vz(x,\,t)\,\leq C.$$ Define $\aa_\fz:=\bigcup_{q\in[1,\,\fz)} \aa_q$ and, for any $\vz\in\aa_\fz$, $$\begin{aligned} \label{e2.4} q(\vz):=\inf\{q\in[1,\,\fz):\ \vz\in\aa_q\}.\end{aligned}$$ Observe that, if $q(\vz)\in(1,\,\fz)$, then $\vz\notin\aa_{q(\vz)}$, and there exists $\vz\notin\aa_1$ such that $q(\vz)=1$ (cf. [@jn87]). \[d2.3\][[@k14 Definition 2.1]]{} A function $\vz: \rn\times[0,\,\fz) \to [0,\,\fz)$ is called a [[growth function]{}]{} if the following conditions are satisfied: 1. $\vz$ is a [Musielak-Orlicz function]{}; 2. $\vz\in\aa_\fz$; 3. $\vz$ is of uniformly lower type $p$ for some $p\in(0,\,1]$ and of uniformly upper type $1$. Suppose that $\vz$ is a Musielak-Orlicz function. Recall that the Musielak-Orlicz space $L^{\vz}$ is defined to be the set of all measurable functions $f$ such that, for some $\lambda\in(0,\,\fz)$, $$\int_\rn \vz\lf(x,\, \frac{\|f(x)\|}{\lz}\r)\, dx<\fz$$ equipped with the Luxembourg-Nakano (quasi-)norm $$\\|f\\|_{L^\vz}:=\inf\lf\{ \lz\in(0,\,\fz):\ \int_\rn \vz\lf(x,\, \frac{\|f(x)\|}{\lz}\r)\, dx\le 1\r\}.$$ Similarly, the weak Musielak-Orlicz space $WL^{\vz}$ is defined to be the set of all measurable functions $f$ such that, for some $\lambda\in(0,\,\fz)$, $$\sup_{t\in(0,\,\fz)} \vz\lf(\{\|f\|>t\},\, \frac{t}{\lz}\r)<\fz$$ equipped with the quasi-norm $$\\|f\\|_{WL^\vz}:=\inf\lf\{ \lz\in(0,\,\fz):\ \sup_{t\in(0,\,\fz)}\vz\lf(\{\|f\|>t\},\,\frac{t}{\lz}\r)\le 1\r\}.$$ \[r2.11\] Let $\omega$ be a classic Muckenhoupt weight and $\phi$ an Orlicz function. 1. If $\vz(x,\,t):=\omega(x)t^{p}$ for all $(x,\,t)\in \rn\times[0,\,\infty)$ with $p\in(0,\,\fz)$, then $L^\varphi$ (resp. $WL^\vz$) is reduced to weighted Lebesgue space $L_\omega^p$ (resp. weighted weak Lebesgue space $WL^p_\omega$), and particularly, when $\omega\equiv 1$, the corresponding unweighted spaces are also obtained. 2. If $\vz(x,\,t):=\omega(x)\phi(t)$ for all $(x,\,t)\in \rn\times[0,\,\infty)$, then $L^\varphi$ (resp. $WL^\vz$) is reduced to weighted Orlicz space $L_\omega^\phi$ (resp. weighted weak Orlicz space $WL^\phi_\omega$), and particularly, when $\omega\equiv 1$, the corresponding unweighted spaces are also obtained. Throughout the paper, we always assume that $\varphi$ is a growth function. In order to obtain the completeness of ${L}^\varphi$, we need the following several lemmas, which are some properties of growth functions. \[qcytj\][[[@k14 Lemma 4.2]]{}]{} Let $\varphi$ be a growth function as in Definition \[d2.3\]. Then the following hold true: 1. for any $f\in {L}^\varphi$ satisfying $f\not\equiv 0$, $$\int_{\rn}\vz\lf(x,\,\frac{\|f(x)\|}{\\|f\\|_{L^\vz}}\r)\,dx=1;$$ 2. $\mathop{\lim}\limits_{k\rightarrow\fz}\\|f_k\\|_{{L}^\varphi}=0$ if and only if $\mathop{\lim}\limits_{k\rightarrow\fz}\int_{\rn}\vz\lf(x,\,\|f_k(x)\|\r)\,dx=0.$ The following lemma comes from [[@k14 Lemma 4.1]]{}, and also can be found in [@ylk17]. \[ckj\] Let $\vz$ be a growth function as in Definition \[d2.3\]. Then there exists a positive constant $C$ such that, for any $(x,\,t_j)\in\rn \times [0,\,\fz)$ with $j\in\zz_+$, $$\vz\lf(x,\,\sum_{j=1}^\fz t_j\r)\le C\sum_{j=1}^\fz\vz\lf(x,\,t_j\r).$$ \[lwbx\] The space $L^\varphi$ is complete. In order to prove the completeness of $L^\vz$, it suffices to prove that, for any sequence $\{f_j\}_{j\in\zz_+}\subset L^\varphi$ satisfying $\\|f_j\\|_{L^\varphi}\leq 2^{-j}$, the series $\{\sum_{j=1}^k f_j\}_{k\in\zz_+}$ converges in $L^\vz$. By the uniformly lower type $p$ property of $\vz$ and Lemma \[qcytj\](i), we see that, for any $j\in\zz_+$, $$\begin{aligned} \label{ee.1} \int_{\rn}\vz(x,\,\|f_j(x)\|)\,dx\le\int_{\rn}\vz\lf(x,\,2^{-j}\frac{\|f_j(x)\|}{\\|f_j\\|_{L^\vz}}\r)\,dx\ls2^{-jp}.\end{aligned}$$ Noticing that the series $\{\sum_{j=1}^k f_j\}_{k\in\zz_+}$ is a Cauchy sequence in $L^\varphi$, we have $$\lim_{k,\,m\rightarrow\fz}\lf\\|{\sum_{j=1}^k f_j}-{\sum_{j=1}^m f_j}\r\\|_{L^\varphi}=0,$$ which, together with Lemma \[qcytj\](ii), implies that $$\begin{aligned} \label{eea.1} \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\int_{\rn} \vz\lf(x,\,\lf\|{\sum_{j=1}^k f_j(x)}-{\sum_{j=1}^m f_j(x)}\r\|\r)\,dx=0.\end{aligned}$$ By the uniformly lower type $p$ and the uniformly upper type $1$ properties of $\varphi$, and , we know that, for any $\sigma\in(0,\,\infty)$, $$\begin{aligned} &\mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\lf\|{\sum_{j=1}^k f_j}-{\sum_{j=1}^m f_j}\r\|>\sigma\r\},\,1\r)\\ &\hs=\mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\lf\|{\sum_{j=1}^k f_j}-{\sum_{j=1}^m f_j}\r\|>\sigma\r\},\,\frac{1}{\sigma}\,\sigma\r)\\ &\hs\lesssim \max\lf\{\sigma^{-1},\sigma^{-p}\r\} \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\lf\|{\sum_{j=1}^k f_j}-{\sum_{j=1}^m f_j}\r\|>\sigma\r\},\,\sigma\r)\\ &\hs\lesssim \max\lf\{\sigma^{-1},\sigma^{-p}\r\}\mathop{\lim}\limits_{k,\,m\rightarrow\fz}\int_{\lf\{\lf\|{\sum_{j=1}^k f_j}-{\sum_{j=1}^m f_j}\r\|>\sigma\r\}} \vz\lf(x,\,\lf\|{\sum_{j=1}^k f_j(x)}-{\sum_{j=1}^m f_j(x)}\r\|\r)\,dx\\ &\hs\lesssim \max\lf\{\sigma^{-1},\sigma^{-p}\r\} \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\int_{\rn} \vz\lf(x,\,\lf\|{\sum_{j=1}^k f_j(x)}-{\sum_{j=1}^m f_j(x)}\r\|\r)\,dx\thicksim0.\end{aligned}$$ Hence, there exists some $f$ such that $\sum_{j=1}^k f_j$ converges to $f$ as $k\rightarrow\infty$ in measure. From this and using Riesz’s theorem, we deduce that there exists a subsequence $\sum_{j=1}^{k_i} f_{j}\rightarrow f$ as $i\rightarrow \infty$ almost everywhere. By this, Lemma \[ckj\] and , we obtain that $$\begin{aligned} \int_{\rn}\vz\lf(x,\,\lf\|f(x)-{\sum_{j=1}^{k_i} f_j(x)}\r\|\r)\,dx \lesssim\sum_{j\geq\,{k_i}+1}\int_{\rn}\vz\lf(x,\,\lf\|f_j(x)\r\|\r)\,dx \lesssim\sum_{j\geq\,{k_i}+1}2^{-jp}\rightarrow 0\end{aligned}$$ as $i\rightarrow\fz$. From Lemma \[qcytj\](ii) again, it follows that $\mathop{\lim}\limits_{i\rightarrow\fz}\\|f-{\sum_{j=1}^{k_i} f_j}\\|_{L^\varphi}=0.$ On the other hand, noticing that $\{\sum_{j=1}^k f_j\}_{k\in\zz_+}$ is a Cauchy sequence in $L^\vz$, then it is easy to see that $\lim_{k\rightarrow\fz}\\|\sum_{j=1}^k f_j-f\\|_{L^\vz}=0$ and $f\in L^\vz$. This finishes the proof of Theorem \[lwbx\]. In order to obtain the completeness of ${WL}^\varphi$, we need the following several lemmas, which are some properties of growth functions. \[jxsqj\] Let $\varphi$ be a growth function as in Definition \[d2.3\]. Then the following hold true: 1. [[[@lyj16 Lemma 3.3(ii)]]{}]{} for any $f\in {WL}^\varphi$ satisfying $f\not\equiv 0$, $$\sup_{t\in(0,\,\infty)}\varphi\lf(\{\|f\|>t\},\,\frac{t}{\\|f\\|_{{WL}^\varphi}}\r)=1;$$ 2. $\mathop{\lim}\limits_{k\rightarrow\fz}\\|f_k\\|_{{WL}^\varphi}=0$ if and only if $\mathop{\lim}\limits_{k\rightarrow\fz}\mathop{\sup}\limits_{t\in(0,\,\infty)} \varphi\lf(\{\|f_k\|>t\},\,t\r)=0.$ We only prove $\rm{(ii)}$ of Lemma \[jxsqj\] since $\rm{(i)}$ of Lemma \[jxsqj\] was proved in [[[@lyj16 Lemma 3.3(ii)]]{}]{}. By the uniformly lower type $p$ and the uniformly upper type $1$ properties of $\varphi$, we conclude that, for any $x\in\mathbb{R}^n$, $s\in(0,\,\infty)$ and $t\in(0,\,\infty)$, $$\begin{aligned} \label{max} \varphi\lf(x,\,st\r)\lesssim \max\lf\{s,s^p\r\}\varphi\lf(x,\,t\r)\end{aligned}$$ and $$\begin{aligned} \label{min} \varphi\lf(x,\,st\r)\gtrsim \min\lf\{s,s^p\r\}\varphi\lf(x,\,t\r).\end{aligned}$$ Thus, from and Lemma \[jxsqj\]$\rm{(i)}$, we deduce that $$\begin{aligned} &\mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,t\r) \\ &\hs=\mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,{\\|f_k\\|_{{WL}^\varphi}}\frac{t}{\\|f_k\\|_{{WL}^\varphi}}\r) \\ &\hs\lesssim \max\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\} \mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,\frac{t}{\\|f_k\\|_{{WL}^\varphi}}\r)\\ &\hs\sim \max\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\}.\end{aligned}$$ On the other hand, by and Lemma \[jxsqj\]$\rm{(i)}$, we obtain that $$\begin{aligned} &\mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,t\r) \\ &\hs=\mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,{\\|f_k\\|_{{WL}^\varphi}}\frac{t}{\\|f_k\\|_{{WL}^\varphi}}\r) \\ &\hs\gtrsim \min\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\} \mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,\frac{t}{\\|f_k\\|_{{WL}^\varphi}}\r)\\ &\hs\sim\min\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\}.\end{aligned}$$ From the above two inequalities, it follows that $$\min\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\}\lesssim \mathop{\sup}\limits_{t\in(0,\,\infty)}\varphi\lf(\{\|f_k\|>t\},\,t\r)\lesssim \max\lf\{{\\|f_k\\|_{{WL}^\varphi}},{\\|f_k\\|^p_{{WL}^\varphi}}\r\},$$ which implies that $\rm{(ii)}$ of Lemma \[jxsqj\] holds true. This finishes the proof of Lemma \[jxsqj\]. \[jxxqj\][[[@g09c p.10]]{}]{} Let $\varphi$ be a growth function as in Definition \[d2.3\]. If $\mathop{\liminf}\limits_{k\rightarrow\fz}\|f_k\|=\|f\|$ almost everywhere, then, for any $t\in(0,\,\infty)$, $$\varphi\lf(\{\|f\|>t\},\,t\r)\leq \mathop{\liminf}\limits_{k\rightarrow\fz} \varphi\lf(\{\|f_k\|>t\},\,t\r).$$ \[wlwbx\] The space ${WL}^\varphi$ is complete. To prove that ${WL}^\varphi$ is complete, take $\{f_k\}_{k\in\zz_+}\subset{{WL}^\varphi}$ such that $\mathop{\lim}\limits_{k,\,m\rightarrow\fz}\\|f_k-f_m\\|_{{WL}^\varphi}=0$. By Lemma \[jxsqj\](ii), we know that, for any chosen positive number $\varepsilon$, however small, there exists a positive integer $K$ such that, whenever $k,\,m\in[K,\,\fz)\cap\zz_+$, then $$\begin{aligned} \label{eea.2} \sup_{t\in(0,\,\fz)}\vz(\{\|f_k-f_m\|>t\},\,t)<\varepsilon.\end{aligned}$$ By the uniformly lower type $p$ and the uniformly upper type $1$ properties of $\varphi$, and , we know that, for any $\sigma\in(0,\,\infty)$, $$\begin{aligned} \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\|f_k-f_m\|>\sigma\r\},\,1\r) &= \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\|f_k-f_m\|>\sigma\r\},\,\frac{1}{\sigma}\,\sigma\r)\\ &\lesssim \max\lf\{\sigma^{-1},\sigma^{-p}\r\} \mathop{\lim}\limits_{k,\,m\rightarrow\fz}\vz\lf(\lf\{\|f_k-f_m\|>\sigma\r\},\,\sigma\r)\thicksim0.\end{aligned}$$ Hence, there exists some $f$ such that $f_k\rightarrow f$ as $k\rightarrow\infty$ in measure, which, together with Riesz’s theorem, implies that some subsequence $$\begin{aligned} \label{ee.2} f_{k_s}\rightarrow f \ {\rm{as}} \ s\rightarrow\infty \ {\rm{almost \ everywhere}}.\end{aligned}$$ For the $K$ mentioned above, take $J\in\zz_+$ such that, for any $j\in[J,\,\fz)\cap\zz_+$, the positive integer $k_j\ge K$. By and Lemma \[jxxqj\], we know that, there exists a positive integer $J$ such that, whenever $j\in[J,\,\fz)\cap\zz_+$, then $$\begin{aligned} \sup_{t\in(0,\,\fz)}\vz(\{\|f_{k_j}-f\|>t\},\,t) &=\sup_{t\in(0,\,\fz)}\vz\lf(\lf\{\lim_{s\rightarrow\fz}\|f_{k_j}-f_{k_s}\|>t\r\},\,t\r) \\ &\le\lim_{s\rightarrow\fz}\sup_{t\in(0,\,\fz)}\vz\lf(\lf\{\|f_{k_j}-f_{k_s}\|>t\r\},\,t\r)<\varepsilon,\end{aligned}$$ that is to say, $$\begin{aligned} \lim_{j\rightarrow\fz}\sup_{t\in(0,\,\fz)}\vz(\{\|f_{k_j}-f\|>t\},\,t)=0.\end{aligned}$$ Applying Lemma \[jxsqj\](ii) again, we conclude that $$\lim_{j\rightarrow\fz}\\|f_{k_j}-f\\|_{WL^\vz}=0.$$ On the other hand, noticing that $\{f_k\}_{k\in\zz_+}$ is a Cauchy sequence in $WL^\vz$, then it is easy to see that $\lim_{k\rightarrow\fz}\\|f_k-f\\|_{WL^\vz}=0$ and $f\in WL^\vz$. This finishes the proof of Theorem \[wlwbx\]. Two boundedness criterions of operators {#s3} ======================================= In this section, we first recall the notion concerning the Musielak-Orlicz Hardy space $H^\vz$ via the non-tangential grand maximal function, and then establish two boundedness criterions of operators from $H^\vz$ to $L^\vz$ or from $H^\vz$ to $WL^\vz$. In what follows, we denote by $\cs$ the [set of all Schwartz functions]{} and by $\cs'$ its [dual space ]{} (namely, the [set of all tempered distributions]{}). For any $m\in\nn$, let $\cs_m$ be the [[set]{}]{} of all $\psi\in\cs$ such that $\\|\psi\\|_{\cs_m}\le1$, where $$\begin{aligned} \\|\psi\\|_{\cs_m}:=\sup_{\az\in\nn^n,\,\|\az\|\le m+1}\sup_{x\in\rn}(1+\|x\|)^{(m+2)(n+1)}\|\partial^\az\psi(x)\|.\end{aligned}$$ Then, for any $m\in\nn$ and $f\in \cs'$, the [[non-tangential grand maximal function]{}]{} $f^\ast_m$ of $f$ is defined by setting, for all $x\in\rn$, $$\begin{aligned} \label{e2.m1} f^\ast_m(x) := \sup_{\psi\in \cs_m}\,\sup_{\|y-x\|<t,\,t\in(0,\,\fz)} \|f\ast\psi_t(y)\|,\end{aligned}$$ where, for any $t\in(0,\,\fz)$, $\psi_t(\cdot):= t^{-n}\psi(\frac{\cdot}{t})$. When $$\begin{aligned} \label{e2.5} m=m(\vz) :=\lf\lfloor n\lf(\frac{q(\vz)}{i(\vz)}-1\r)\r\rfloor,\end{aligned}$$ we denote $f^\ast_m$ simply by $f^\ast$, where $q(\vz)$ and $i(\vz)$ are as in and , respectively. \[d2.5\] [@k14 Definition 2.2] Let $\vz$ be a growth function as in Definition \[d2.3\]. The Musielak-Orlicz Hardy space $H^\vz$ is defined as the set of all $f\in\cs'$ such that $f^\ast\in L^\vz$ endowed with the (quasi-)norm $$\\|f\\|_{H^\vz}:=\\|f^\ast\\|_{L^\vz}.$$ \[r2.111\] Let $\omega$ be a classic Muckenhoupt weight and $\phi$ an Orlicz function. 1. If $\vz(x,\,t):=\omega(x)t^{p}$ for all $(x,\,t)\in \rn\times[0,\,\infty)$ with $p\in(0,\,1]$, then $H^\varphi$ is reduced to weighted Hardy space $H_\omega^p$, and particularly, when $\omega\equiv 1$, the corresponding unweighted space is also obtained. 2. If $\vz(x,\,t):=\omega(x)\phi(t)$ for all $(x,\,t)\in \rn\times[0,\,\infty)$, then $H^\varphi$ is reduced to weighted Orlicz Hardy space $H_\omega^\phi$, and particularly, when $\omega\equiv 1$, the corresponding unweighted space is also obtained. \[d2.11\][[@k14 Definition 2.4]]{} Let $\vz$ be a growth function as in Definition \[d2.3\]. 1. A triplet $(\vz,\,q,\,s)$ is said to be [[admissible]{}]{}, if $q\in (q(\vz),\,\fz]$ and $s \in [m(\vz),\,\fz)\cap\nn$, where $q(\vz)$ and $m(\vz)$ are as in and , respectively. 2. For an admissible triplet $(\vz,\,q,\,s)$, a measurable function $a$ is called a [$(\vz,\,q,\,s)$-atom]{} if there exists some ball $B\subset\rn$ such that the following conditions are satisfied: \(a) $a$ is supported in $B$; \(b) $\\|a\\|_{L^q_\vz(B)}\leq\\|\chi_B\\|^{-1}_{L^\vz}$, where $$\begin{aligned} \\|a\\|_{L_\vz^q(B)}:= \lf\{\begin{array}{ll} \dsup_{t\in(0,\,\fz)}\lf[\frac{1}{\vz(B,\,t)}\int_B\|a(x)\|^q \vz(x,\,t)\,dx\r]^{1/q} ,\,\,\,&q\in[1,\,\fz),\\ \,\\ \\|a\\|_{L^\fz(B)},&q=\fz; \end{array}\r.\end{aligned}$$ \(c) $\int_\rn a(x)x^\az dx=0$ for any $\az\in\nn^n$ with $\|\az\|\leq s$. 3. For an admissible triplet $(\vz,\,q,\,s)$, the [Musielak-Orlicz atomic Hardy space]{} $H^{\vz,\,q,\,s}_{\rm{at}}$ is defined as the set of all $f \in \cs'$ which can be represented as a linear combination of $(\vz,\, q,\, s)$-atoms, that is, $f =\sum_j b_j$ in $\cs'$, where $b_j$ for each $j$ is a multiple of some $(\vz,\, q,\, s)$-atom supported in some ball ${x_j+B_{r_j}}$, with the property $$\sum_{j}\vz\lf({x_j+B_{r_j}},\, \\|b_j\\|_{L^q_\vz({x_j+B_{r_j}})}\r)<\fz.$$ For any given sequence of multiples of $(\vz,\,q,\,s)$-atoms, $\{b_j\}_j$, let $$\Lz_q(\{b_j\}_j):=\inf\lf\{\lz\in(0,\,\fz):\ \sum_j \vz\lf({x_j+B_{r_j}},\,\frac{\\|b_j\\|_{L^q_\vz({x_j+B_{r_j}})}}{\lz}\r)\le 1 \r\}$$ and then the (quasi-)norm of $f\in\cs'$ is defined by $$\\|f\\|_{H^{\vz,\,q,\,s}_{\rm{at}}}:=\inf\lf\{\Lz_q\lf(\{b_j\}_j\r)\r\},$$ where the infimum is taken over all admissible decompositions of $f$ as above. We refer the readers to [@k14] and [@ylk17] for more details on the real-variable theory of Musielak-Orlicz Hardy spaces. \[d2.zh\] Let $X$ and $Y$ be two linear spaces. An operator $T$: $D\subset X\rightarrow Y$ is called a positive sublinear operator if, for any $x\in\mathbb{R}^n$, the following conditions are satisfied: 1. $T(f)(x)\geq0$; <!-- --> 1. $T(\alpha f)(x)\leq\|\alpha\|T(f)(x)$, where $\alpha\in\mathbb{R}$; <!-- --> 1. $T(f+g)(x)\leq T(f)(x)+T(g)(x)$. \[szxz\] Let $X$ and $Y$ be two linear spaces and $T: D\subset X\rightarrow Y$ be a positive sublinear operator as in Definition \[d2.zh\]. Then, for any $f,g\in D$, $$\lf\|T(f)-T(g)\r\|\leq T(f-g).$$ By Definition \[d2.zh\](ii), we obtain that $$T(-f)\leq\|-1\|\,T(f)=T(f)\leq\|-1\|\,T\lf(-f\r)=T\lf(-f\r),$$ therefore, $T(-f)=T(f)$. Moreover, by Definition \[d2.zh\](iii), we know that $$\begin{aligned} T(f)-T(g)=T(f-g+g)-T(g) \leq T(f-g)+T(g)-T(g)=T(f-g)\end{aligned}$$ and $$\begin{aligned} T(g)-T(f)&=T(g-f+f)-T(f) \leq T(g-f)+T(f)-T(f)=T(g-f).\end{aligned}$$ From the above two inequalities and $T(-f)=T(f)$, we deduce that $\lf\|T(f)-T(g)\r\|\leq T(f-g)$. This finishes the proof of Lemma \[szxz\]. The following two lemmas come from [[@k14 Lemma 4.3(i), Theorem 3.1]]{}, respectively, and also can be found in [@ylk17]. \[ky4.3\] Let $\vz$ be a growth function as in Definition \[d2.3\]. For a given positive constant ${\widetilde}{C}$, there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$, $$\int_\rn\vz\lf(x,\,\frac{\|f(x)\|}{\lambda}\r)\,dx\le{\widetilde}{C} \ implies \ that \ \\|f\\|_{L^\vz}\le C \lambda.$$ \[yzfj\] Let $(\vz,\,q,\,s)$ be an admissible triplet as in Definition \[d2.11\]. Then $$H^\vz=H^{\vz,\,q,\,s}_{{\rm{at}}}$$ with equivalent (quasi-)norms. \[cm\][[[@ylk17 Remark 4.1.4]]{}]{} Let $\vz$ be a growth function as in Definition \[d2.3\]. Then $H^\vz\cap L^2$ is dense in $H^\vz$. Recall that a [quasi-Banach space]{} $\mathcal{B}$ is a linear space endowed with a quasi-norm $\\|\cdot\\|_\mathcal{B}$ which is nonnegative, non-degenerate (i.e., $\\|f\\|_\mathcal{B}=0$ if and only if $f={\bf 0}$), homogeneous, and obeys the quasi-triangle inequality, i.e., there exists a constant $K$ no less than 1 such that, for any $f, g\in\mathcal{B}$, $\\|f+g\\|_\mathcal{B}\leq K\lf(\\|f\\|_\mathcal{B}+\\|g\\|_\mathcal{B}\r)$. \[ardl\][[[@s84 Aoki-Rolewicz theorem]]{}]{} Let $\mathcal{B}$ be a quasi-Banach space and $K$ a constant associated with $\cb$ as above. Then, for any $f,\,g\in\mathcal{B}$, $$\\|f+g\\|^\gamma_\mathcal{B}\leq\\|f\\|^\gamma_\mathcal{B}+\\|g\\|^\gamma_\mathcal{B},$$ where $\gamma:=\lf[\log_{2}(2K)\r]^{-1}$. \[nfbds\] Let $\mathcal{B}$ be a quasi-Banach space equipped with the quasi-norm $\\|\cdot\\|_\mathcal{B}$. For any $\{f_k\}_{k\in\zz_+}\subset\mathcal{B}$ and $f\in\mathcal{B}$, if $\mathop{\lim}\limits_{k\rightarrow\fz} \lf\\|f_k-f\r\\|_\mathcal{B}=0$, then $$\lim_{k\rightarrow\fz}\lf\\|f_k\r\\|_\mathcal{B}=\lf\\|f\r\\|_\mathcal{B}.$$ By Lemma \[ardl\], we obtain that, for any $k\in\zz_+$ $$\lf\\|f_k\r\\|^\gamma_\mathcal{B}-\lf\\|f\r\\|^\gamma_\mathcal{B} =\lf\\|f_k-f+f\r\\|^\gamma_\mathcal{B}-\lf\\|f\r\\|^\gamma_\mathcal{B} \leq\lf\\|f_k-f\r\\|^{\gamma}_\mathcal{B},$$ where $\gz$ is a harmless constant as in Lemma \[ardl\]. Similarly, we have $\lf\\|f\r\\|^\gamma_\mathcal{B}-\lf\\|f_k\r\\|^\gamma_\mathcal{B}\leq\lf\\|f-f_k\r\\|^{\gamma}_\mathcal{B},$ which, together with the above inequality, implies that $$\lf\|\lf\\|f_k\r\\|^\gamma_\mathcal{B}-\lf\\|f\r\\|^\gamma_\mathcal{B}\r\|\le\lf\\|f_k-f\r\\|^{\gamma}_\mathcal{B}\rightarrow0 \ {\rm{as}} \ k\rightarrow\fz.$$ This finishes the proof of Lemma \[nfbds\]. The following theorem gives a boundedness criterion of operators from $H^\vz$ to $L^\vz$. \[yt\] Let $\vz$ be a growth function as in Definition \[d2.3\]. Suppose that a linear or a positive sublinear operator $T$ is bounded on $L^2$. If there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$ and multiple of a $(\vz,\,q,\,s)$-atom $b$ associated with some ball $B\subset\rn$, $$\begin{aligned} \label{a1} \int_{\rn}\vz\lf(x,\,\frac{\|T(b)(x)\|}{\lz}\r)\,dx \le C\vz\lf(B,\,\frac{\\|b\\|_{L^q_\vz(B)}}{\lz}\r),\end{aligned}$$ then $T$ extends uniquely to a bounded operator from ${H^\vz}$ to $L^\vz$. We first assume that $f\in H^\vz\cap L^2$. By the well known Calderón reproducing formula (see also [@lfy15 Theorem 2.14]), we know that there exists a sequence of multiples of $(\vz,\,q,\,s)$-atoms $\{b_j\}_{j\in\zz_+}$ associated with balls $\{x_j+B_{r_j}\}_{j\in\zz_+}$ such that $$\begin{aligned} \label{a2} f=\lim_{k\rightarrow\fz}\sum_{j=1}^k b_j =:\lim_{k\rightarrow\fz}f_k \ {\rm{in}} \ \cs' \ {\rm{and \ also \ in}} \ L^2.\end{aligned}$$ From the assumption that the linear or positive sublinear operator $T$ is bounded on $L^2$, Lemma \[szxz\] and , it follows that $$\begin{aligned} \lim_{k\rightarrow\fz}\lf\\|T(f)-T\lf(f_k\r)\r\\|_{L^2} \le \lim_{k\rightarrow\fz}\lf\\|T\lf(f-f_k\r)\r\\|_{L^2}\ls \lim_{k\rightarrow\fz}\lf\\|f-f_k\r\\|_{L^2}\sim0,\end{aligned}$$ which implies that $$\begin{aligned} \label{z14} T(f)=\lim_{k\rightarrow\fz}T\lf(f_k\r) \le\lim_{k\rightarrow\fz}\sum_{j=1}^kT\lf(b_j\r) =\sum_{j=1}^\fz T\lf(b_j\r) \ {\rm{almost \ everywhere}}.\end{aligned}$$ By this, Lemma \[ckj\] and with taking $\lambda=\Lz_q(\{b_j\}_j)$, we obtain $$\begin{aligned} \int_{\rn}\vz\lf(x,\,\frac{\|T(f)(x)\|}{\Lz_q(\{b_j\}_j)}\r)\,dx &\ls \sum_{j=1}^{\fz}\int_{\rn}\vz\lf(x,\,\frac{\|T(b_j)(x)\|}{\Lz_q(\{b_j\}_j)}\r)\,dx \\ &\ls \sum_{j=1}^{\fz}\vz\lf(x_j+B_{r_j},\, \frac{\\|b_j\\|_{L^q_\vz(x_j+B_{r_j})}}{\Lz_q(\{b_j\}_j)}\r)\ls1,\end{aligned}$$ which, together with Lemma \[ky4.3\], further implies that $$\\|T(f)\\|_{L^\vz}\ls \Lz_q(\{b_j\}_j).$$ Taking infimum for all admissible decompositions of $f$ as above and using Lemma \[yzfj\], we obtain that, for any $f\in H^\vz\cap L^2$, $$\begin{aligned} \label{a3} \\|T(f)\\|_{L^\vz} \ls \\|f\\|_{H^{\vz,\,q,\,s}_{{\rm{at}}}} \sim \\|f\\|_{H^\vz}.\end{aligned}$$ Next, suppose $f\in{H^\vz}$. By Lemma \[cm\], we know that there exists a sequence of $\{f_j\}_{j\in\zz_+}\subset H^\vz\cap L^2$ such that $f_j\rightarrow f$ as $j\rightarrow\fz$ in ${H^\vz}$. Therefore, $\{f_j\}_{j\in\zz_+}$ is a Cauchy sequence in $H^\vz$. From this, Lemma \[szxz\] and , we deduce that, for any $j,\,k\in\zz_+$, $$\lf\\|T(f_j)-T(f_k)\r\\|_{L^\vz} \leq \lf\\|T(f_j-f_k)\r\\|_{L^\vz}\ls \lf\\|f_j-f_k\r\\|_{H^\vz}.$$ Thus, by this, we know that $\{T(f_j)\}_{j\in\zz_+}$ is a Cauchy sequence in $L^\vz$. Applying Theorem \[lwbx\], we conclude that there exists some $g\in L^\vz$ such that $T(f_j)\rightarrow g$ as $j\rightarrow\fz$ in $L^\vz$. Consequently, define $T(f):=g$. Below, we claim that $T(f)$ is well defined. Indeed, for any other sequence $\{f'_j\}_{j\in\zz_+}\subset H^\vz\cap L^2$ satisfying $f'_j\rightarrow f$ as $j\rightarrow\fz$ in ${H^\vz}$, by Lemma \[szxz\] and , we have $$\begin{aligned} \lf\\|T(f'_j)-T(f)\r\\|_{L^\vz} &\ls \lf\\|T(f'_j)-T(f_j)\r\\|_{L^\vz}+\lf\\|T(f_j)-g\r\\|_{L^\vz} \\ &\ls \lf\\|f'_j-f_j\r\\|_{H^\vz}+\lf\\|T(f_j)-g\r\\|_{L^\vz} \\ &\ls \lf\\|f'_j-f\r\\|_{H^\vz}+\lf\\|f-f_j\r\\|_{H^\vz}+\lf\\|T(f_j)-g\r\\|_{L^\vz}\rightarrow0 \ {\rm{as}} \ j\rightarrow\fz,\end{aligned}$$ which is wished. From this, Lemma \[nfbds\] and , it follows that, for any $f\in{H^\vz}$, $$\\|T(f)\\|_{L^\vz}=\\|g\\|_{L^\vz}=\lim_{j\rightarrow\fz}\\|T(f_j)\\|_{L^\vz} \ls \lim_{j\rightarrow\fz}\\|f_j\\|_{H^\vz}\thicksim\\|f\\|_{H^\vz}.$$ This finishes the proof of Theorem \[yt\]. To show the boundedness criterion of operators from ${H^\vz}$ to $WL^\vz$, we need the following superposition principle of weak type estimates. \[dj\][[[@bckyy13 Lemma 7.13]]{}]{} Let $\vz$ be a growth function as in Definition \[d2.3\] satisfying $I(\vz)\in(0,\,1)$, where $I(\vz)$ is as in . Assume that $\{f_j\}_{j\in\zz+}$ is a sequence of measurable functions such that, for some $\lz\in(0,\,\fz)$, $$\sum_{j\in\zz+}\sup_{\az\in(0,\,\fz)}\vz\lf(\{\|f_j\|>\az\},\,\frac{\az}{\lz}\r)<\fz.$$ Then there exists a positive constant $C$, depending only on $\vz$, such that, for any $\eta\in(0,\,\fz)$, $$\vz\lf(\lf\{ \sum_{j\in\zz+}\|f_j\|>\eta \r\},\,\frac{\eta}{\lz}\r) \le C\sum_{j\in\zz+}\sup_{\az\in(0,\,\fz)}\vz\lf(\{\|f_j\|>\az\},\,\frac{\az}{\lz}\r).$$ By an argument similar to that used in the proof of [@k14 Lemma 4.3], we easily obtain the following lemma, the details being omitted. \[fs1\] Let $\vz$ be a growth function as in Definition \[d2.3\]. For a given positive constant ${\widetilde}{C}$, there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$, $$\sup_{\alpha\in(0,\,\fz)}\vz\lf(\{\|f\|>\alpha\},\,\frac{\alpha}{\lambda}\r)\le{\widetilde}{C} \ implies \ that \ \\|f\\|_{WL^\vz}\le C \lambda.$$ The following theorem gives a boundedness criterion of operators from ${H^\vz}$ to $WL^\vz$. \[yt2\] Let $\vz$ be a growth function as in Definition \[d2.3\] satisfying $I(\vz)\in(0,\,1)$, where $I(\vz)$ is as in . Suppose that a linear or a positive sublinear operator $T$ is bounded on $L^2$. If there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$ and multiple of a $(\vz,\,q,\,s)$-atom $b$ associated with some ball $B\subset\rn$, $$\begin{aligned} \label{z11} \sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{\|T(b)\|>\alpha\r\},\,\frac{\alpha}{\lambda}\r) \le C\vz\lf(B,\,\frac{{\\|b\\|_{L^q_\vz(B)}}}{\lambda}\r),\end{aligned}$$ then $T$ extends uniquely to a bounded operator from ${H^\vz}$ to $WL^\vz$. Since the proof of Theorem \[yt2\] is similar to that of Theorem \[yt\], we use the same notation as in the proof of Theorem \[yt\]. Here we just give out the necessary modifications. By , Lemma \[dj\] and with taking $\lambda=\Lz_q(\{b_j\}_j)$, we obtain that, for any $\alpha\in(0,\,\fz)$, $$\begin{aligned} \vz\lf(\{\|T(f)\|>\alpha\},\,\frac{\alpha}{\Lz_q(\{b_j\}_j)}\r) &\le \vz\lf(\lf\{\sum_{j=1}^\fz \|T\lf(b_j\r)\|>\alpha\r\},\,\frac{\alpha}{\Lz_q(\{b_j\}_j)}\r) \\ &\ls \sum_{j=1}^\fz\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{\|T\lf(b_j\r)\|>\alpha\r\},\,\frac{\alpha}{\Lz_q(\{b_j\}_j)}\r) \\ &\ls \sum_{j=1}^{\fz}\vz\lf(x_j+B_{r_j},\, \frac{\\|b_j\\|_{L^q_\vz(x_j+B_{r_j})}}{\Lz_q(\{b_j\}_j)}\r)\ls1,\end{aligned}$$ which, together with Lemma \[fs1\], implies that $$\\|T(f)\\|_{WL^\vz}\ls \Lz_q(\{b_j\}_j).$$ Taking infimum for all admissible decompositions of $f$ as above and using Lemma \[yzfj\], we obtain that, for any $f\in H^\vz\cap L^2$, $$\begin{aligned} \label{z13} \\|T(f)\\|_{WL^\vz} \ls \\|f\\|_{H^{\vz,\,q,\,s}_{{\rm{at}}}} \sim \\|f\\|_{H^\vz}.\end{aligned}$$ Next, suppose $f\in{H^\vz}$. By Lemma \[cm\], we know that there exists a sequence of $\{f_j\}_{j\in\zz_+}\subset H^\vz\cap L^2$ such that $f_j\rightarrow f$ as $j\rightarrow\fz$ in ${H^\vz}$. Therefore, $\{f_j\}_{j\in\zz_+}$ is a Cauchy sequence in $H^\vz$. From this, Lemma \[szxz\] and , we deduce that, for any $j,\,k\in\zz_+$, $$\lf\\|T(f_j)-T(f_k)\r\\|_{{WL}^\vz} \leq \lf\\|T(f_j-f_k)\r\\|_{{WL}^\vz}\ls \lf\\|f_j-f_k\r\\|_{H^\vz}.$$ Thus, by this, we know that $\{T(f_j)\}_{j\in\zz_+}$ is a Cauchy sequence in ${WL}^\vz$. Applying Theorem \[wlwbx\], we conclude that there exists some $g\in {WL}^\vz$ such that $T(f_j)\rightarrow g$ as $j\rightarrow\fz$ in ${WL}^\vz$. Consequently, define $T(f):=g$. Below, we claim that $T(f)$ is well defined. Indeed, for any other sequence $\{f'_j\}_{j\in\zz_+}\subset H^\vz\cap L^2$ satisfying $f'_j\rightarrow f$ as $j\rightarrow\fz$ in ${H^\vz}$, by Lemma \[szxz\] and , we have $$\begin{aligned} \lf\\|T(f'_j)-T(f)\r\\|_{WL^\vz} &\ls \lf\\|T(f'_j)-T(f_j)\r\\|_{WL^\vz}+\lf\\|T(f_j)-g\r\\|_{WL^\vz} \\ &\ls \lf\\|f'_j-f_j\r\\|_{H^\vz}+\lf\\|T(f_j)-g\r\\|_{WL^\vz} \\ &\ls \lf\\|f'_j-f\r\\|_{H^\vz}+\lf\\|f-f_j\r\\|_{H^\vz}+\lf\\|T(f_j)-g\r\\|_{WL^\vz}\rightarrow0 \ {\rm{as}} \ j\rightarrow\fz,\end{aligned}$$ which is wished. From this, Lemma \[nfbds\] and , it follows that, for any $f\in{H^\vz}$, $$\\|T(f)\\|_{{WL}^\vz}=\\|g\\|_{{WL}^\vz}=\lim_{j\rightarrow\fz}\\|T(f_j)\\|_{{WL}^\vz} \ls \lim_{j\rightarrow\fz}\\|f_j\\|_{H^\vz}\sim\\|f\\|_{H^\vz}.$$ This finishes the proof of Theorem \[yt2\]. Boundedness of parametric Marcinkiewicz integrals \[s4\] ======================================================== In this section, we first recall the notion concerning the [[$L^q$-Dini type condition of order $\alpha$]{}]{} with $q\in[1,\,\fz]$ and $\az\in(0,\,1]$, and then obtain the boundedness of $\mu^\rho_\boz$ from $H^\vz$ to $L^\vz$ or from $H^\vz$ to $WL^\varphi$. Here and hereafter, we always assume that $\Omega$ is homogeneous of degree zero and satisfies . Recall that, for any $q\in[1,\,\fz)$ and $\alpha\in(0,\,1]$, a function $\Omega\in L^q(S^{n-1})$ is said to satisfy the [[$L^q$-Dini type condition of order $\alpha$]{}]{} (when $\alpha=0$, it is called the [[$L^q$-Dini condition]{}]{}), if $$\int_0^1\frac{\omega_q(\delta)}{\delta^{1+\alpha}}\,d\delta<\fz,$$ where ${\omega_q(\delta)}$ is the integral modulus of continuity of order $q$ of $\Omega$ defined by setting, for any $\delta\in(0,\,1]$, $${\omega_q(\delta)}:=\sup_{\\|\gz\\|<\delta} \lf(\int_{S^{n-1}}\|\Omega(\gz x')-\Omega(x')\|^q\,d\sigma(x')\r)^{1/q}$$ and $\gz$ denotes a rotation on $S^{n-1}$ with $\\|\gz\\|:=\sup_{y'\in S^{n-1}}\|\gz y'-y'\|$. For any $\alpha,\,\beta\in(0,\,1]$ with $\beta<\alpha$, it is easy to see that if $\Omega$ satisfies the $L^q$-Dini type condition of order $\alpha$, then it also satisfies the $L^q$-Dini type condition of order $\beta$. We thus denote by ${\rm{Din}}^q_\alpha(S^{n-1})$ the class of all functions which satisfy the $L^q$-Dini type conditions of all orders $\beta<\alpha$. For any $\alpha\in(0,\,1]$, we define $${\rm{Din}}^\fz_\alpha(S^{n-1}):=\bigcap_{q\ge1}{\rm{Din}}^q_\alpha(S^{n-1}).$$ See [@ll07 pp.89-90] for more properties of ${\rm{Din}}^q_\alpha(S^{n-1})$ with $q\in[1,\,\fz]$ and $\alpha\in(0,\,1]$. The main results of this section are as follows. \[dingli.1\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p\in({n}/{(n+\beta)},\,1)$ therein. Suppose that $\Omega\in L^q(S^{n-1})\cap {\rm{Din}}^1_{\alpha}(S^{n-1})$ with $q\in(1,\,\fz]$. If $q$ and $\varphi$ satisfy one of the following conditions: 1. $q\in(1,\,1/p]$ and $\vz^{q'}\in \mathbb{A}_{\frac{p\beta}{n(1-p)}}$; 2. $q\in(1/p,\,\fz]$ and $\vz^{{1}/{(1-p)}}\in \mathbb{A}_{\frac{p\beta}{n(1-p)}}$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{L^\vz} \leq C\\|f\\|_{H^\vz}.$$ \[dingli.2\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p\in({n}/{(n+\beta)},\,1]$ therein. Suppose that $\Omega\in {\rm{Din}}^q_{\alpha}(S^{n-1})$ with $q\in(1,\,\fz)$. If $\vz^{q'}\in\aa_{(p+\frac{p\beta}{n}-\frac1q)q'}$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{L^\vz} \leq C\\|f\\|_{H^\vz}.$$ \[tuilun.1\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p\in({n}/{(n+\beta)},\,1]$ therein. Suppose that $\Omega\in {\rm{Din}}^\fz_{\alpha}(S^{n-1})$. If $\vz\in\aa_{p(1+\frac{\beta}{n})}$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{L^\vz} \leq C\\|f\\|_{H^\vz}.$$ \[dingli.3\] Let $\rho\in(0,\,\infty)$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p:=1$ therein. For a given positive constant ${\widetilde}{C}$, suppose $\Omega\in{L}^{q}(S^{n-1})$ with $q\in(1,\,\infty)$ such that, for any $y\neq{\mathbf{0}}$, $h\in\rn$ and $t\in[0,\,\infty)$, $$\begin{aligned} \int_{\|x\|\geq2\|y\|}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n}}-\frac{\Omega(x)}{\|x\|^{n}}\r\| \vz(x+h,\,t)\,dx \leq {\widetilde}{C}\vz(x+h,\,t).\end{aligned}$$ If $\varphi^{q'}\in\aa_{1}$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{L^\vz} \leq C\\|f\\|_{H^\vz}.$$ \[dingli.4\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p:={n}/{(n+\beta)}$ therein, and $I(\vz)\in(0,\,1)$, where $I(\vz)$ is as in . Suppose that $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$. If $\varphi\in\aa_{1}$, then there exists a positive constant $C$ independent of $f$ such that $$\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{WL^\vz} \leq C\\|f\\|_{H^\vz}.$$ \[r3\] 1. It is worth noting that Corollary \[tuilun.1\] can be regarded as the limit case of Theorem \[dingli.2\] by letting $q\rightarrow\fz$. 2. Theorem \[dingli.1\], Theorem \[dingli.2\] and Corollary \[tuilun.1\] jointly answer the question: when $\Omega\in{\rm{Din}}^q_{\alpha}(S^{n-1})$ with $q=1$, $q\in(1,\,\infty)$ or $q=\infty$, respectively, what kind of additional conditions on growth function $\varphi$ and $\Omega$ can deduce the boundedness of $\mu^\rho_\Omega$ from $H^\vz$ to $L^\vz$? 3. When $\rho:=1$, Theorem \[dingli.1\], Theorem \[dingli.2\] and Corollary \[tuilun.1\] are reduced to [@lll17 Theorem 2.4, Theorem 2.5 and Corollary 2.6], respectively. 4. Let $\omega$ be a classic Muckenhoupt weight and $\phi$ an Orlicz function. \(a) When $\vz(x,\,t):=\omega(x)\phi(t)$ for all $(x,\,t)\in\rn\times[0,\,\fz)$, we have $H^\vz=H^\phi_\omega$. In this case, Theorem \[dingli.1\], Theorem \[dingli.2\], Corollary \[tuilun.1\] and Theorem \[dingli.3\] hold true for weighted Orlicz Hardy space. Even when $\varphi(x,\,t):=\phi(t)$, the above results are also new. \(b) When $\vz(x,\,t):=\omega(x)\,t^p$ for all $(x,\,t)\in\rn\times[0,\,\fz)$, we have $H^\vz=H^p_\omega$. In this case, if $\rho:=1$, Theorem \[dingli.1\], Theorem \[dingli.2\], Corollary \[tuilun.1\] and Theorem \[dingli.3\] are reduced to [@ll07 Theorem 1.4, Theorem 1.5, Corollary 1.7 and Theorem 1.8], respectively. \(c) When $\vz(x,\,t):=\omega(x)\,t^p$ for all $(x,\,t)\in\rn\times[0,\,\fz)$, we have $H^\vz=H^p_\omega$. In this case, the assumptions of $\rho$ and $\Omega$ in Corollary \[tuilun.1\] are weaker than that in [@w16 Theorem 1.1]. Precisely, in [@w16 Theorem 1.1], $\rho\in(0,\,n)$ and $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$, however, in our case, $\rho\in(0,\,\fz)$ and $\Omega$ satisfies some weaker smoothness conditions, i.e., $\Omega\in{\rm{Din}}^\fz_{\alpha}(S^{n-1})$. \(d) When $\vz(x,\,t):=\omega(x)\,t^p$ for all $(x,\,t)\in\rn\times[0,\,\fz)$, we have $H^\vz=H^p_\omega$. In this case, if $\rho$ is restricted to $(0,\,n)$, Theorem \[dingli.4\] is reduced to [@w16 Theorem 1.2]. To show main results, let us begin with some lemmas. Since $\varphi$ satisfies the uniform Muckenhoupt condition, the proofs of $\rm{(i)}$, $\rm{(ii)}$ and $\rm{(iii)}$ of the following Lemma \[quan\] are identity to that of [[@g09m Exercises 9.1.3, Theorem 9.2.5 and Corollary 9.2.6]]{}, respectively, the details being omitted. \[quan\] Let $q\in[1,\,\fz]$. If $\vz\in \aa_q$, then the following statements hold true: 1. $\vz^\varepsilon\in \aa_q$ for any $\varepsilon\in(0,\,1]$; 2. $\vz^\eta\in \aa_q$ for some $\eta\in(1,\,\fz)$; 3. $\vz\in \aa_d$ for some $d\in(1,\,q)$ with $q\neq1$. The following Lemma \[lemma.1\] is a subtle pointwise estimate for $\mu^\rho_\Omega(b)$, where $b$ is a multiple of a $(\vz,\,\fz,\,s)$ atom. And this lemma plays an important role in the proof of Theorem \[dingli.1\]. \[lemma.1\] Let $\rho\in(0,\,\infty)$ and $b$ be a multiple of a $(\vz,\,\fz,\,s)$-atom associated with some ball $B_r$. Then, for any $x\in B_{2R}\setminus B_{R}$ with $R\in[2r,\,\fz)$, $$\mu^\rho_\Omega(b)(x)\le \\|\Omega\\|_{L^1(S^{n-1})}\\|b\\|_{L^\fz}\frac{1}{\rho}\lf\{\ln{\frac{2R+r}{R-r}}+\frac{[(2R+r)^{\rho}-(R-r)^{\rho}]^2}{2\rho(2R+r)^{2\rho}}\r\}^{1/2}.$$ The key of the proof is to find a subtle segmentation. From $\supp\,b \subset B_r$, we deduced that, for any $y\in B_r$ and $x\in B_{2R}\setminus B_{R}$ with $R\in[2r,\,\fz)$, $$\begin{aligned} \label{eee.1} R-r<\|x-y\|<2R+r.\end{aligned}$$ Therefore, for any $x\in B_{2R}\setminus B_{R}$ with $R\in[2r,\,\fz)$, write $$\begin{aligned} \lf[\mu^\rho_\Omega(b)(x)\r]^2 &=\int_{0}^{\fz}\lf\|\int_{\|x-y\|\leq t} \frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}b(y)\,{dy}\r\|^2\,\frac{dt}{t^{2\rho+1}} \\ &= \int_0^{R-r}\lf\|\int_{\|x-y\|\leq t} \frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}b(y)\,{dy}\r\|^2\,\frac{dt}{t^{2\rho+1}} +\int_{R-r}^{2R+r}\cdot\cdot\cdot+\int_{2R+r}^\fz\cdot\cdot\cdot \\ &=:{\rm{I_1+I_2+I_3}}.\end{aligned}$$ For ${\rm{I_1}}$, from $t\in(0,\,R-r]$ and , it follows that $\{y\in\rn: \ \|x-y\|\le t\}=\emptyset$ and hence ${\rm{I_1}}=0$. For ${\rm{I_2}}$, by the spherical coordinates transform and $\Omega\in L^1(S^{n-1})$ (see ), we obtain $$\begin{aligned} {\rm{I_2}} &\le\\|b\\|_{L^\fz}^2 \int_{R-r}^{2R+r}\lf(\int_{S^{n-1}}\int_0^t \frac{\|\Omega(z')\|}{u^{n-\rho}}u^{n-1}\,{du}\,{d\sigma(z')}\r)^2\,\frac{dt}{t^{2\rho+1}} \\ &=\\|\Omega\\|^2_{L^1(S^{n-1})} \\|b\\|_{L^\fz}^2 \frac{1}{{\rho}^2}\int_{R-r}^{2R+r}\frac1t\,dt =\\|\Omega\\|^2_{L^1(S^{n-1})}\\|b\\|_{L^\fz}^2\frac{1}{{\rho}^2} \ln{\frac{2R+r}{R-r}}.\end{aligned}$$ For ${\rm{I_3}}$, by , the spherical coordinates transform and $\Omega\in L^1(S^{n-1})$ (see ), we have $$\begin{aligned} {\rm{I_3}} &\leq \\|b\\|_{L^\fz}^2 \int_{2R+r}^\fz\lf(\int_{B_{2R+r}\setminus B_{R-r}} \frac{\|\Omega(z)\|}{\|z\|^{n-\rho}}\,{dz}\r)^2\,\frac{dt}{t^{2\rho+1}} \\ &= \\|b\\|_{L^\fz}^2 \int_{2R+r}^\fz\lf(\int_{S^{n-1}}\int_{R-r}^{2R+r} \frac{\|\Omega(z')\|}{u^{n-\rho}}u^{n-1}\,{du}\,{d\sigma(z')}\r)^2\,\frac{dt}{t^{2\rho+1}} \\ &=\\|\Omega\\|^2_{L^1(S^{n-1})}\\|b\\|_{L^\fz}^2\frac{1}{{\rho}^2} [(2R+r)^\rho-(R-r)^\rho]^2 \int_{2R+r}^\fz\frac{1}{t^{2\rho+1}}\,dt\\ &=\\|\Omega\\|^2_{L^1(S^{n-1})} \\|b\\|_{L^\fz}^2\frac{1}{{\rho}^2} \frac{[(2R+r)^{\rho}-(R-r)^{\rho}]^2}{2\rho(2R+r)^{2\rho}}.\end{aligned}$$ Combining the estimates of ${\rm{I_1}}$, ${\rm{I_2}}$ and ${\rm{I_3}}$, we obtain the desired inequality. This finishes the proof of Lemma \[lemma.1\]. \[bcd\][[[@k14 Lemma 4.5]]{}]{} Let $\vz\in\mathbb{A}_q$ with $q\in[1,\,\fz)$. Then there exists a positive constant $C$ such that, for any ball $B\subset\rn$, $\lambda\in(1,\,\fz)$ and $t\in(0,\,\fz)$, $$\vz(\lambda B,\,t)\le C{\lambda}^{nq}\vz(B,\,t).$$ Since $\varphi$ satisfies the uniform Muckenhoupt condition, the proof of Lemma \[fh\] is identity to that of [[@sw85 Corollary 6.2]]{}, the details being omitted. \[fh\] Let $d\in(1,\,\fz)$. Then, $\vz^d\in\mathbb{A}_\fz$ if and only if $\vz\in{\mathbb{RH}_d}$. The proof of the following Lemma \[L3.6\] is motivated by [@kw79 Lemma 5]. \[L3.6\] Suppose that $\rho\in(0,\,\infty)$, $q\in[1,\,\infty)$ and $\Omega$ satisfies the $L^q$-Dini condition. Then there exists a positive constant $C$ such that, for any $R\in(0,\,\infty)$ and $y\in B_{R/2}$, $$\begin{aligned} \lf(\int_{R\leq\|x\|<2R}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\|^q dx\r)^{1/q} \leq CR^{n/q-(n-\rho)}\lf({\frac{\|y\|}{R}}+ \int_{\|y\|/2R}^{\|y\|/R}\frac{\omega_q(\delta)}{\delta}d\delta\r).\end{aligned}$$ Noticing that $\|x\|\geq R$ and $\|y\|<R/2$, we have $\|x-y\|\sim\|x\|$. From this and the mean value theorem, it follows that $$\begin{aligned} \lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\| &\leq \lf\|\frac{\Omega(x)}{\|x\|^{n-\rho}}-\frac{\Omega(x)}{\|x-y\|^{n-\rho}}\r\|+ \lf\|\frac{\Omega(x)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}\r\|\\ &\leq C\lf(\|\Omega(x)\|\frac{\|y\|}{\|x\|^{n-\rho+1}}+\frac{\|\Omega(x-y)-\Omega(x)\|}{\|x\|^{n-\rho}}\r).\end{aligned}$$ We then write $$\begin{aligned} & \lf(\int_{R\leq\|x\|<2R}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\|^q dx\r)^{1/q} \\ &\hs \leq C\lf(\int_{R\leq\|x\|<2R}\|\Omega(x)\|^q\frac{\|y\|^q}{\|x\|^{(n-\rho+1)q}}dx\r)^{1/q} +C\lf(\int_{R\leq\|x\|<2R}\frac{\|\Omega(x-y)-\Omega(x)\|^q}{\|x\|^{(n-\rho)q}}dx\r)^{1/q}\\ &\hs =:C(\mathrm{I_1+I_2}).\end{aligned}$$ For $\mathrm{I_1}$, by the spherical coordinates transform and $\Omega\in{L}^{q}(S^{n-1})$, we know that, for any $y\in B_{R/2}$, $$\begin{aligned} &\mathrm{I_1}=\|y\|\lf(\int_{S^{n-1}}\int_{R}^{2R}\|\Omega(x')\|^q\frac{r^{n-1}}{r^{(n-\rho+1)q}}drd\sigma(x')\r)^{1/q} \\ &\hs\sim \|y\|\lf(\int_{R}^{2R}r^{n-1-(n-\rho+1)q} dr\r)^{1/q} \sim R^{n/q-(n-\rho)}\lf(\frac{\|y\|}{R}\r).\end{aligned}$$ For ${\mathrm{I_2}}$, from the spherical coordinates transform and Fubini’s theorem, it follows that, for any $y\in B_{R/2}$, $$\begin{aligned} & \mathrm{I_2}=\lf(\int_{S^{n-1}}\int_{R}^{2R}\frac{\|\Omega(rx'-y)-\Omega(rx')\|^q}{r^{(n-\rho)q}}r^{n-1}drd\sigma(x')\r)^{1/q} \\ &\hs \sim R^{n/q-(n-\rho)}\lf[\int_{R}^{2R}\lf(\int_{S^{n-1}}\lf\|\Omega\lf(\frac{x'-\alpha}{\|x'-\alpha\|}\r)-\Omega(x')\r\|^q d\sigma(x')\r)\frac{dr}{r}\r]^{1/q},\end{aligned}$$ where $\alpha:=y/r$. Proceeding as in the proof of [@kw79 Lemma 5], $\mathrm{I_2}$ is bounded by a positive constant times $$R^{n/q-(n-\rho)}\lf(\int_{\|y\|/2R}^{\|y\|/R}\omega_{q}(\delta)\frac{d\delta}{\delta}\r).$$ Combining the estimates of ${\rm{I_1}}$ and ${\rm{I_2}}$, we obtain the desired inequality. This finishes the proof of Lemma \[L3.6\]. The following Lemma \[lin\] extends [@ll07 Lemma 4.4] from non-parametric case to the parametric case. \[lin\] For $\alpha\in(0,\,1]$ and $q\in[1,\,\fz)$, suppose that $\Omega$ satisfies the $L^q$-Dini type condition of order $\alpha$. Let $\rho\in(0,\,\infty)$, $\beta:=\min\{\alpha,\,1/2\}$ and $b$ be a multiple of a $(\varphi,\infty,s)$-atom associated with some ball $B_r$. 1. If $q=1$, then there exists a positive constant $C$ independent of $b$ such that, for any $R\in[2r,\,\fz)$, $$\int_{B_{2R}\setminus B_{R}}\mu^\rho_\Omega(b)(x)\,dx \leq C\\|b\\|_{L^\fz}R^n\lf(\frac{r}{R}\r)^{n+\beta}.$$ 2. If $q\in(1,\,\fz)$ and, for any $(x,\,t)\in\rn\times[0,\,\fz)$, $\vz(x,\,t)\ge0$, then there exists a positive constant $C$ independent of $b$ such that, for any $R\in[2r,\,\fz)$ and $t\in[0,\,\fz)$, $$\int_{B_{2R}\setminus B_{R}}\mu^\rho_\Omega(b)(x)\vz(x,\,t)\,dx \leq C\\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q}\lf(\frac{r}{R}\r)^{n+\beta}.$$ We only prove for case (ii), since the proof of case (i) is analogous to that of case (ii) and is left to the readers. For any $R\in[2r,\,\fz)$ and $t\in[0,\,\fz)$, write $$\begin{aligned} & \int_{R\leq\|x\|<2R}\mu^\rho_\Omega(b)(x)\vz(x,\,t)\,dx \\ &\hs \leq\int_{R\leq\|x\|<2R}\lf(\int_{0}^{\|x\|+r}\lf\|\int_{\|x-y\|\leq t} \frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}b(y)\,{dy}\r\|^2\,\frac{dt}{t^{2\rho+1}}\r)^{1/2}\vz(x,\,t)\,dx \\ &\hs\hs+\int_{R\leq\|x\|<2R}\lf(\int_{\|x\|+r}^{\infty}\lf\|\int_{\|x-y\|\leq t} \frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}b(y)\,{dy}\r\|^2\,\frac{dt}{t^{2\rho+1}}\r)^{1/2}\vz(x,\,t)\,dx=:\mathrm{I_1+I_2}.\end{aligned}$$ For $\mathrm{I_1}$, noticing that $y\in B_r$ and $x\in B_{2R}\setminus B_{R}$ with $R\in[2r,\,\fz)$, we know that $$\begin{aligned} \label{eee.2} \|x-y\|\sim\|x\|\sim\|x\|+r\end{aligned}$$ and $$\begin{aligned} \label{eee.3} R/2<\|x-y\|<5R/2.\end{aligned}$$ From and the mean value theorem, it follows that, for any $y\in B_r$ and $x\in B_{2R}\setminus B_{R}$ with $R\in[2r,\,\fz)$, $$\lf\|\frac{1}{\|x-y\|^{2\rho}}-\frac{1}{(\|x\|+r)^{2\rho}}\r\|\ls\frac{r}{\|x-y\|^{2\rho+1}}.$$ By Minkowski’s inequality for integrals, the above inequality and Fubini’s theorem, we obtain that, for any $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $$\begin{aligned} {\mathrm{I_1}} &\leq \int_{R\leq\|x\|<2R}\lf[\int_{B_{r}}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}b(y)\r\| \lf(\int_{\|x-y\|}^{\|x\|+r}\frac{dt}{t^{2\rho+1}}\r)^{1/2}\,dy\r]\vz(x,\,t)\,dx \\ &\ls \\|b\\|_{L^\fz} \int_{R\leq\|x\|<2R}\lf[\int_{B_{r}}\frac{\lf\|\Omega(x-y)\r\|}{\|x-y\|^{n-\rho}} \lf\|\frac{1}{\|x-y\|^{2\rho}}-\frac{1}{(\|x\|+r)^{2\rho}}\r\|^{1/2}\,dy\r]\vz(x,\,t)\,dx \\ &\ls \\|b\\|_{L^\fz}r^{1/2}\int_{B_{r}}\lf(\int_{R\leq\|x\|<2R}\frac{\|\Omega(x-y)\|}{\|x-y\|^{n+{1/2}}}\vz(x,\,t)\,dx\r)dy.\end{aligned}$$ On the other hand, from Hölder’s inequality, , the spherical coordinates transform and $\Omega\in{L}^{q}(S^{n-1})$, we deduced that, for any $y\in {B_{r}}$, $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $$\begin{aligned} &\int_{R\leq\|x\|<2R}\frac{\|\Omega(x-y)\|}{\|x-y\|^{n+{1/2}}}\vz(x,\,t)\,dx\\ &\hs\leq \lf(\int_{R\leq\|x\|<2R}\frac{\|\vz(x,\,t)\|^{q'}}{\|x-y\|^{n+{1/2}}}\,dx\r)^{1/q'} \lf(\int_{R\leq\|x\|<2R}\frac{\|\Omega(x-y)\|^q}{\|x-y\|^{n+{1/2}}}dx\r)^{1/q}\\ &\hs\ls\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'} R^{(-n-1/2)/q'} \lf(\int_{R/2<\|z\|<5R/2}\frac{\|\Omega(z)\|^q}{\|z\|^{n+{1/2}}}dz\r)^{1/q}\\ &\hs\sim \lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{(-n-1/2)/q'}\lf(R^{-n-1/2} \int_{S^{n-1}}\int_{0}^{{5R}/2}\lf\|\Omega(z')\r\|^q u^{n-1} du d\sigma(z')\r)^{1/q}\\ &\hs\sim \lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{-n/q'-1/2}.\end{aligned}$$ Substituting the above inequality into ${\mathrm{I_1}}$ and using the assumption that $\beta=\min\{\alpha,\,1/2\}$, we know that, for any $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $${\mathrm{I_1}}\ls \\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q}\lf(\frac{r}{R}\r)^{n+1/2} \ls \\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q}\lf(\frac{r}{R}\r)^{n+\beta}.$$ For ${\mathrm{I_2}}$, noticing that for $t>\|x\|+r$, it is easy to see that $B_r\subset \lf\{y\in\mathbb{R}^n:\|x-y\|\leq t\r\}$. From this, vanishing moments of $b$, Minkowski’s inequality for integrals and Fubini’s theorem, it follows that, for any $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $$\begin{aligned} {\mathrm{I_2}} &= \int_{R\leq\|x\|<2R}\lf(\int_{\|x\|+r}^{\infty}\lf\|\int_{\|x-y\|\leq t} \lf(\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r)b(y)\,{dy}\r\|^2\,\frac{dt}{t^{2\rho+1}}\r)^{1/2}\vz(x,\,t)\,dx \\ &\leq \int_{R\leq\|x\|<2R}\lf[\int_{B_r}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\|\|b(y)\| \lf(\int_{R}^{\infty}\frac{dt}{t^{2\rho+1}}\r)^{1/2} dy\r]\vz(x,\,t)\,dx \\ &\ls\\|b\\|_{L^\fz}R^{-\rho}\int_{B_r}\lf(\int_{R\leq\|x\|<2R}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\| \vz(x,\,t)\,dx\r)dy.\end{aligned}$$ On the other hand, from Hölder’s inequality and Lemma \[L3.6\] (since $\Omega$ satisfies the $L^q$-Dini type condition of order $\alpha$, it also satisfies the $L^q$-Dini condition), we deduced that, for any $y\in {B_{r}}$, $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $$\begin{aligned} &\int_{R\leq\|x\|<2R}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\| \vz(x,\,t)\,dx\\ &\hs\leq \lf(\int_{R\leq\|x\|<2R}\|\vz(x,\,t)\|^{q'}dx\r)^{1/q'} \lf(\int_{R\leq\|x\|<2R}\lf\|\frac{\Omega(x-y)}{\|x-y\|^{n-\rho}}-\frac{\Omega(x)}{\|x\|^{n-\rho}}\r\|^q dx\r)^{1/q}\\ &\hs\ls \lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q-n+\rho}\lf({\frac{\|y\|}{R}}+ \int_{\|y\|/2R}^{\|y\|/R}\frac{\omega_q(\delta)}{\delta}d\delta\r)\\ &\hs\ls \lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q-n+\rho}\lf[{\frac{\|y\|}{R}}+\lf(\frac{\|y\|}{R}\r)^\alpha \int_{\|y\|/2R}^{\|y\|/R}\frac{\omega_q(\delta)}{\delta^{1+\alpha}}d\delta\r].\end{aligned}$$ Substituting the above inequality into ${\mathrm{I_2}}$ and using the assumptions that $\Omega$ satisfies the $L^q$-Dini type condition of order $\alpha$, and $\beta=\min\{\alpha,\,1/2\}$, we know that, for any $R\in [2r,\,\infty)$ and $t\in[0,\,\fz)$, $$\begin{aligned} {\mathrm{I_2}} &\ls \\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q-n}\int_{B_{r}}\lf[{\frac{r}{R}}+\lf({\frac{r}{R}}\r)^\alpha \int_{0}^{1}\frac{\omega_q(\delta)}{\delta^{1+\alpha}}d\delta\r]dy \\ &\ls \\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q}\lf({\frac{r}{R}}\r)^{n}\lf[{\frac{r}{R}}+\lf({\frac{r}{R}}\r)^\alpha\r] \\ &\ls \\|b\\|_{L^\fz}\lf[\vz^{q'}(B_{2R},\,t)\r]^{1/q'}R^{n/q}\lf({\frac{r}{R}}\r)^{n+\beta}.\end{aligned}$$ Combining the estimates of ${\rm{I_1}}$ and ${\rm{I_2}}$, we obtain the desired inequality. This finishes the proof of Lemma \[lin\]. The following Lemma \[lemma.2\] shows that $\mu_\Oz$ maps all multiple of an atoms into uniformly bounded elements of $L^\vz$. \[lemma.2\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $p\in({n}/{(n+\beta)},\,1)$. Suppose that $\Omega\in L^q(S^{n-1})\cap {\rm{Din}}^1_{\alpha}(S^{n-1})$ with $q\in(1,\,\fz]$. If $q$ and $\varphi$ satisfy one of the following conditions: 1. $q\in(1,\,1/p]$ and $\vz^{q'}\in \mathbb{A}_{\frac{p\beta}{n(1-p)}}$; 2. $q\in(1/p,\,\fz]$ and $\vz^{{1}/{(1-p)}}\in \mathbb{A}_{\frac{p\beta}{n(1-p)}}$, then there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$ and multiple of a $(\vz,\,\fz,\,s)$-atom $b$ associated with some ball $B\subset\rn$, $$\int_{\rn}\vz\lf(x,\,\frac{\mu^\rho_\Omega(b)(x)}{\lz}\r)\,dx \le C\vz\lf(B,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r).$$ We need only consider the case $q\in(1,\,\fz)$, since the case $q=\infty$ can be derived from the case $q=2$. Indeed, when $q=\infty$, a routine computation gives rise to $2>1/p$. If Lemma \[lemma.2\] holds true for $q=2$, by $\Omega\in L^\infty(S^{n-1})\subset L^2(S^{n-1})$, $2>1/p$ and $\vz^{{1}/{(1-p)}}\in \mathbb{A}_{\frac{p\beta}{n(1-p)}}$, we know that Lemma \[lemma.2\] holds true for $q=\fz$. We are now turning to the proof of Lemma \[lemma.2\] under case $q\in(1,\,\fz)$. Without loss of generality, we may assume $b$ is a multiple of a $(\vz,\,\fz,\,s)$-atom associated with a ball $B_r$ for some $r\in(0,\,\fz)$. For the general case, we refer the readers to the method of proof in [@ll07 Theorem 1.4]. First, we claim that, in either case (i) or (ii) of Lemma \[lemma.2\], there exists some $d\in(1,\,{p\beta}/{[n(1-p)]})$ such that $$\begin{aligned} \label{eee.4} \vz^{q'}\in\mathbb{A}_d \ {\rm{and}} \ \vz^{1/(1-p)}\in\mathbb{A}_d.\end{aligned}$$ We only prove under case (ii) since the proof under case (i) is similar. By Lemma \[quan\](iii) with $\vz^{1/{(1-p)}}\in\aa_{\frac{p\beta}{n(1-p)}}$, we see that there exists some $d\in(1,\,{p\beta}/{[n(1-p)]})$ such that $\vz^{1/(1-p)}\in\mathbb{A}_d$. On the other hand, notice that $q'<1/{(1-p)}$, then, by Lemma \[quan\](i), we know $\vz^{q'}\in\mathbb{A}_d$, which is wished. The next thing to do in the proof is to find a subtle segmentation. For any $j\in\zz_+$, let $E_j:=B_{2^{j+1}r}\setminus B_{2^{j}r}$. By Lemma \[lemma.1\], we know that, for any $x\in E_j$, $$\begin{aligned} \label{eee.x} \mu^\rho_\Omega(b)(x) \ls \\|b\\|_{L^\fz}\frac{1}{\rho}\lf\{\ln{\frac{2^{j+1}+1}{2^j-1}}+ \frac{[(2^{j+1}+1)^{\rho}-(2^{j}-1)^{\rho}]^2}{2\rho(2^{j+1}+1)^{2\rho}}\r\}^{1/2}.\end{aligned}$$ Notice that $$\lf\{\ln{\frac{2^{j+1}+1}{2^j-1}}+\frac{[(2^{j+1}+1)^{\rho}-(2^{j}-1)^{\rho}]^2}{2\rho(2^{j+1}+1)^{2\rho}}\r\}^{1/2} \rightarrow\lf[\ln2+\frac{1}{2\rho}\lf(1-\frac1{2^\rho}\r)^2\r]^{1/2}\,\,{\rm{as}}\,\,j\rightarrow\fz,$$ which, together with $$\sup_{\rho\in(0,\,\infty)}\lf[\ln2+\frac{1}{2\rho}\lf(1-\frac1{2^\rho}\r)^2\r]^{1/2}<1,$$ implies that there exists some $J\in\zz_+$ independent of $b$ such that, for any $j\in[J+1,\,\fz)\cap\zz_+$, $$\begin{aligned} \label{eee.5} \lf\{\ln{\frac{2^{j+1}+1}{2^j-1}}+\frac{[(2^{j+1}+1)^{\rho}-(2^{j}-1)^{\rho}]^2}{2\rho(2^{j+1}+1)^{2\rho}}\r\}^{1/2}<1.\end{aligned}$$ Therefore, for any $\lz\in(0,\,\fz)$, write $$\begin{aligned} \int_{\rn}\vz\lf(x,\,\frac{\mu^\rho_\Omega(b)(x)}{\lz}\r)\,dx =\int_{2^JB_{r}}\vz\lf(x,\,\frac{\mu^\rho_\Omega(b)(x)}{\lz}\r)\,dx +\int_{\lf(2^JB_{r}\r)^\complement}\cdot\cdot\cdot =:\mathrm{I_1+I_2}.\end{aligned}$$ Another step in the proof is to estimate $\mathrm{I_1}$ and $\mathrm{I_2}$, respectively. For $\mathrm{I_1}$, by the uniformly upper type 1 property of $\vz$, Theorem A with $\Omega\in L^q(S^{n-1})$ and $\vz^{q'}\in\mathbb{A}_d$, and Lemma \[bcd\] with $\vz\in\mathbb{A}_d$ (which is guaranteed by Lemma \[quan\](i) with ), we know that, for any $\lz\in(0,\,\fz)$, $$\begin{aligned} {\mathrm{I_1}} &\ls \int_{2^JB_{r}}\lf(1+\frac{\mu^\rho_\Omega(b)(x)}{\\|b\\|_{L^\fz}}\r)^d \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \int_{2^JB_{r}}\lf(1+\frac{\lf[\mu^\rho_\Omega(b)(x)\r]^d}{\\|b\\|^d_{L^\fz}}\r) \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(2^JB_{r},\,{\\|b\\|_{L^\fz}}\r)+\frac{1}{\\|b\\|^d_{L^\fz}}\int_\rn \lf[\mu^\rho_\Omega(b)(x)\r]^d \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(2^JB_{r},\,{\\|b\\|_{L^\fz}}\r)+\frac{1}{\\|b\\|^d_{L^\fz}}\int_{B_r} \|b(x)\|^d \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(B_{r},\,\frac{\\|b\\|_{L^\fz}}{\lz}\r),\end{aligned}$$ which is wished. For ${\mathrm{I_2}}$, from the uniformly lower type $p$ properties of $\vz$ with $\frac{\mu^\rho_\Omega(b)(x)}{\\|b\\|_{L^\fz}}\ls1$ (see and ) and Hölder’s inequality, we deduce that, for any $\lz\in(0,\,\fz)$, $$\begin{aligned} {\mathrm{I_2}} &= \sum_{j=J+1}^{\fz}\int_{E_j}\vz\lf(x,\,\frac{\mu^\rho_\Omega(b)(x)}{\lz}\r)\,dx \\ &\ls \frac{1}{\\|b\\|^p_{L^\fz}}\sum_{j=J+1}^{\fz}\int_{E_j}\lf[\mu^\rho_\Omega(b)(x)\r]^p \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \frac{1}{\\|b\\|^p_{L^\fz}}\sum_{j=J+1}^{\fz} \lf(\int_{E_j}\lf[\vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\r]^{1/(1-p)}\,dx\r)^{1-p} \lf(\int_{E_j}\mu^\rho_\Omega(b)(x)\,dx\r)^p.\end{aligned}$$ Notice that $\vz^{1/{(1-p)}}\in\aa_d\subset\aa_\fz\ $(see ). By Lemma \[fh\], we have $\vz\in \mathbb{RH}_{1/(1-p)}$. Thus, from Lemma \[bcd\] with $\vz^{1/{(1-p)}}\in\mathbb{A}_d$, and $\vz\in \mathbb{RH}_{1/(1-p)}$, it follows that, for any $\lz\in(0,\,\fz)$, $$\begin{aligned} \lf(\int_{E_j}\lf[\vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\r]^{1/(1-p)}\,dx\r)^{1-p} &\leq \lf[\vz^{1/(1-p)} \lf(2^{j+1}B_{r},\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\r]^{1-p} \\ &\ls 2^{jnd(1-p)}\lf[\vz^{1/(1-p)}\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\r]^{1-p} \\ &\ls 2^{jnd(1-p)}r^{-np}\vz\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r).\end{aligned}$$ Since $d<{p\beta}/{[n(1-p)]}$, we may choose an ${\widetilde}{\alpha}\in(0,\,\alpha)$ such that $d<{p{\widetilde}{\beta}}/{[n(1-p)]}$, where ${\widetilde}{\beta}:=\min\{{\widetilde}{\alpha},\,1/2\}$. By the assumption $\Omega\in{\rm{Din}}^1_\alpha(S^{n-1})$, $\Omega$ satisfies the $L^1$-Dini type condition of order ${\widetilde}{\alpha}$. Applying Lemma \[lin\](i), we obtain $$\int_{E_j}\mu^\rho_\Omega(b)(x)\,dx \ls \\|b\\|_{L^\fz}\lf(2^jr\r)^n \lf(\frac{r}{2^jr}\r)^{n+{\widetilde}{\beta}} \thicksim \\|b\\|_{L^\fz}r^n 2^{-j{\widetilde}{\beta}}.$$ Substituting the above two inequalities into ${\mathrm{I_2}}$, we know that, for any $\lz\in(0,\,\fz)$, $${\mathrm{I_2}} \ls \vz\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\lf(\sum_{j=J+1}^{\fz}2^{j(nd-ndp-p{\widetilde}{\beta})}\r) \ls \vz\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r),$$ where the last inequality is due to $d<{p{\widetilde}{\beta}}/{[n(1-p)]}$. Finally, combining the estimates of ${\rm{I_1}}$ and ${\rm{I_2}}$, we obtain the desired inequality. This finishes the proof of Lemma \[lemma.2\]. From Theorem A with $\omega\equiv1$, it follows that $\mu^\rho_\Omega$ is bounded on $L^2$. By this, Lemma \[lemma.2\] and the fact that $\mu^\rho_\Omega$ is a positive sublinear operator, applying Theorem \[yt\] with $q=\fz$, we know that $\mu^\rho_\Omega$ extends uniquely to a bounded operator from $H^\vz$ to $L^\vz$. This finishes the proof of Theorem \[dingli.1\]. The proof of Theorem \[dingli.2\] is similar to that of Theorem \[dingli.1\]. We only need to modify the estimate of ${\mathrm{I_2}}$ in the proof of Lemma \[lemma.2\]. And fortunately, the estimate of ${\mathrm{I_2}}$ is nearly identity to that of $J$ in the proof of [@ll07 Theorem 1.5], where [@ll07 Lemma 4.4(a)] is used in that proof, and here Lemma \[lin\](ii) is used instead. We leave the details to the interested readers. By Lemma \[quan\](ii) with $\vz\in\aa_{p(1+\frac{\beta}{n})}$, we see that there exists some $d\in(1,\,\fz)$ such that $\vz^d\in\aa_{p(1+\frac{\beta}{n})}$. For any $q\in(1,\,\fz)$, by $p>n/{(n+\beta)}$, we have $(p+p\beta/n-1/q)q'>p(1+\beta/n)$ and hence $\vz^d\in\aa_{(p+\frac{p\beta}{n}-\frac1q)q'}$. Thus, we may choose $q:=d/{(d-1)}$ such that $$\vz^{q'}=\vz^d\in\aa_{(p+\frac{\beta}{n}-\frac{1}{q})q'}$$ and hence Corollary \[tuilun.1\] follows from Theorem \[dingli.2\]. Observe that, if $\vz$ is of uniformly lower type 1 and of uniformly upper type 1, then, in either $s\in(0,\,1]$ or $s\in[1,\,\infty)$, there exists a positive constant $C$ independent of $s$ such that, for any $x\in\rn$ and $t\in[0,\,\fz)$, $$\begin{aligned} \label{xydy} \vz(x,\,st)\le C s\vz(x,\,t).\end{aligned}$$ On the other hand, we claim that, in either $s\in(0,\,1]$ or $s\in[1,\,\infty)$, there exists a positive constant $C$ independent of $s$ such that, for any $x\in\rn$ and $t\in[0,\,\fz)$, $$\begin{aligned} \label{dydy} \vz(x,\,st)\geq Cs\vz(x,\,t).\end{aligned}$$ In fact, by , we have $$s\vz(x,\,t)=s\vz(x,\,st/s)\ls\vz(x,\,s\,t),$$ which is wished. Combining and , we obtain $\vz(x,\,st)\sim s\vz(x,\,t)$, which implies that $$\int_\rn \vz\lf(x,\, f^\ast(x)\r)\, dx\thicksim\int_\rn f^\ast(x)\vz\lf(x,\,1\r)\, dx$$ and hence, $H^\vz=H^1_{\vz(\cdot\,,\,1)}$. Similarly, $L^\vz=L^1_{\vz(\cdot\,,\,1)}$. Then, by repeating the proof of [@ll07 Theorem 1.8], we know that $\lf\\|\mu_{\Omega}^{\rho}(f)\r\\|_{L^\vz} \ls\\|f\\|_{H^\vz}.$ This finishes the proof of Theorem \[dingli.4\]. \[m11\] Let $\rho\in(0,\,\infty)$, $\alpha\in(0,\,1]$, $\beta:=\min\{\alpha,\,1/2\}$ and $\vz$ be a growth function as in Definition \[d2.3\] with $p:={n}/{(n+\beta)}$ therein. Suppose that $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$. If $\vz\in\aa_1$, then there exists a positive constant $C$ such that, for any $\lz\in(0,\,\fz)$ and multiple of a $(\vz,\,\fz,\,s)$-atom $b$ associated with some ball $B\subset\rn$, $$\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{\mu^\rho_\Omega(b)>\alpha\r\},\,\frac{\alpha}{\lz}\r) \le C\vz\lf(B,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r).$$ We show this lemma by borrowing some ideas from the proof of [@lyj16 Theorem 5.2]. Without loss of generality, we may assume $b$ is a multiple of a $(\vz,\,\fz,\,s)$-atom associated with a ball $B_r$ for some $r\in(0,\,\fz)$. For the general case, we refer the readers to the method of proof in [@ll07 Theorem 1.4]. Proceeding as in the proof of [@w16 Theorem 1.1], we know that, for any $x\in\lf(B_{2r}\r)^{\complement}$, $$\mu^\rho_\Omega(b)(x)\ls \\|b\\|_{L^\fz}\lf(\frac{r^{n+1/2}} {\|x\|^{n+1/2}}+\frac{r^{n+1}}{\|x\|^{n+1}}+\frac{r^{n+\alpha}}{\|x\|^{n+\alpha}}\r),$$ which, together with $\beta:=\min\{\alpha,\,1/2\}$, implies that, for any $x\in\lf(B_{2r}\r)^{\complement}$, $$\begin{aligned} \label{brdtgj} \mu^\rho_\Omega(b)(x)\ls \\|b\\|_{L^\fz}\frac{r^{n+\beta}}{\|x\|^{n+\beta}}\,.\end{aligned}$$ Therefore, for any $\lz\in(0,\,\fz)$, write $$\begin{aligned} &\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{\mu^\rho_\Omega(b)>\alpha\r\},\,\frac{\alpha}{\lz}\r) \\ &\hs\le \sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{x\in B_{2r}: \mu^\rho_\Omega(b)(x)>\alpha\r\},\,\frac{\alpha}{\lz}\r) \\ &\hs\hs+\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{x\in (B_{2r})^\complement: \mu^\rho_\Omega(b)(x)>\alpha\r\},\,\frac{\alpha}{\lz}\r) =:{\rm{I_1}}+{\rm{I_2}}. $$ For $\mathrm{I_1}$, from Lemma \[quan\](ii) with $\vz\in\aa_2$ (since $\vz\in\aa_1$), it follows that $\vz^{q'}\in\mathbb{A}_2$ for some ${q'}\in(1,\,\infty)$. By the uniformly upper type 1 property of $\vz$, Theorem A with $\Omega\in L^q(S^{n-1})$ (since $\Omega\in{\rm{Lip}}_\alpha(S^{n-1})$) and $\vz^{q'}\in\mathbb{A}_2$, and Lemma \[bcd\] with $\vz\in\mathbb{A}_2$, we know that, for any $\lz\in(0,\,\fz)$, $$\begin{aligned} {\mathrm{I_1}} &=\sup_{\alpha\in(0,\,\fz)}\int_{\lf\{x\in B_{2r}: \mu^\rho_\Omega(b)(x)>\alpha\r\}}\vz\lf(x,\,\frac{\alpha}{\lz}\r)\,dx\\ &\le \int_{B_{2r}}\vz\lf(x,\,\frac{\mu^\rho_\Omega(b)(x)}{\lz}\r)\,dx \\ &\ls \int_{B_{2r}}\lf(1+\frac{\mu^\rho_\Omega(b)(x)}{\\|b\\|_{L^\fz}}\r)^2\vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \int_{B_{2r}}\lf(1+\frac{\lf[\mu^\rho_\Omega(b)(x)\r]^2}{\\|b\\|^2_{L^\fz}}\r)\vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(B_{2r},\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)+\frac{1}{\\|b\\|^2_{L^\fz}}\int_\rn \lf[\mu^\rho_\Omega(b)(x)\r]^2 \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(B_{2r},\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)+\frac{1}{\\|b\\|^2_{L^\fz}}\int_{B_r} \|b(x)\|^2 \vz\lf(x,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r)\,dx \\ &\ls \vz\lf(B_{r},\,\frac{\\|b\\|_{L^\fz}}{\lz}\r),\end{aligned}$$ which is wished. For ${\rm{I_2}}$, from , Lemma \[bcd\] with $\vz\in\aa_1$, and the uniformly lower type ${\frac{n}{n+\beta}}$ property of $\vz$, we deduce that, for any $\lz\in(0,\,\fz)$, $$\begin{aligned} {\rm{I_2}} &\ls\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{x\in (B_{2r})^\complement: \ \\|b\\|_{L^\fz}\frac{r^{n+\beta}}{\|x\|^{n+\beta}}>\alpha\r\},\,\frac{\alpha}{\lz}\r) \\ &\sim\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{x\in (B_{2r})^\complement: \ \|x\|^{n+\beta}<\frac{\\|b\\|_{L^\fz}}{\alpha}r^{n+\beta}\r\},\,\frac{\alpha}{\lz}\r) \\ &\sim\sup_{\alpha\in(0,\,\fz)}\vz\lf(\lf\{x\in \rn: \ 2r\le\|x\|<\lf(\frac{\\|b\\|_{L^\fz}}{\alpha}\r)^{\frac{1}{n+\beta}} r \r\},\,\frac{\alpha}{\lz}\r) \\ &\ls\sup_{\alpha\in(0,\,\\|b\\|_{L^\fz})}\vz\lf(\lf\{x\in \rn: \ \|x\|<\lf(\frac{\\|b\\|_{L^\fz}}{\alpha}\r)^{\frac{1}{n+\beta}} r \r\},\,\frac{\alpha}{\lz}\r) \\ &\thicksim\sup_{\alpha\in(0,\,\\|b\\|_{L^\fz})} \vz\lf(\lf[\frac{\\|b\\|_{L^\fz}}{\alpha}\r]^{\frac{1}{n+\beta}} B_r ,\,\frac{\alpha}{\lz}\r) \\ &\ls\sup_{\alpha\in(0,\,\\|b\\|_{L^\fz})} \lf(\frac{\\|b\\|_{L^\fz}}{\alpha}\r)^{\frac{n}{n+\beta}} \vz\lf(B_r ,\,\frac{\alpha}{\lz}\r) \\ &\ls\sup_{\alpha\in(0,\,\\|b\\|_{L^\fz})} \lf(\frac{\\|b\\|_{L^\fz}}{\alpha}\r)^{\frac{n}{n+\beta}} \lf(\frac{\alpha}{\\|b\\|_{L^\fz}}\r)^{\frac{n}{n+\beta}} \vz\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r) \\ &\thicksim \vz\lf(B_r,\,\frac{\\|b\\|_{L^\fz}}{\lz}\r).\end{aligned}$$ Combining the estimates of ${\rm{I_1}}$ and ${\rm{I_2}}$, we obtain the desired inequality. This finishes the proof of Lemma \[m11\]. From Theorem A with $\omega\equiv1$, it follows that $\mu^\rho_\Omega$ is bounded on $L^2$. By this, Lemma \[lemma.2\] and the fact that $\mu^\rho_\Omega$ is a positive sublinear operator, applying Theorem \[yt2\] with $q=\fz$, we know that $\mu^\rho_\Omega$ extends uniquely to a bounded operator from ${H^\vz}$ to $WL^\vz$. This finishes the proof of Theorem \[dingli.4\]. [Competing Interests]{} The authors declare that there is no conflict of interests regarding the publication of this paper. [Authors’ Contributions]{} Li Bo conceived of the study. Liu Xiong, Li Baode, Qiu Xiaoli and Li Bo carried out the main results, participated in the sequence alignment and drafted the manuscript. Moreover, all authors read and approved the final manuscript. [Acknowledgements]{} This work is partially supported by the National Natural Science Foundation of China (Grant Nos. 11461065 & 11661075) and A Cultivate Project for Young Doctor from Xinjiang Uyghur Autonomous Region (No. qn2015bs003). The authors would like to thank the anonymous referees for their constructive comments. [30]{} Akbulut, Ali; Kuzu, Okan, Marcinkiewicz integrals associated with Schrödinger operator on generalized Morrey spaces, J. Math. Inequal. 8 (2014), no. 4, 791-801. Bui, The Anh; Cao, Jun; Ky, Luong Dang; Yang, Dachun; Yang, Sibei, Musielak-Orlicz-Hardy spaces associated with operators satisfying reinforced off-diagonal estimates, Anal. Geom. Metr. Spaces 1 (2013), 69-129. Cao, Jun; Chang, Der-Chen; Yang, Dachun; Yang, Sibei, Riesz transform characterizations of Musielak-Orlicz Hardy spaces, Trans. Amer. Math. Soc. 368 (2016), no. 10, 6979-7018. Diening, Lars, Maximal function on Musielak-Orlicz spaces and generalized Lebesgue spaces, Bull. Sci. Math. 129 (2005), no. 8, 657-700. Diening, Lars; Hästö, Peter A.; Roudenko, Svetlana, Function spaces of variable smoothness and integrability, J. Funct. Anal. 256 (2009), no. 6, 1731-1768. Fan, Xingya; He, Jianxun; Li, Baode; Yang, Dachun, Real-variable characterizations of anisotropic product Musielak-Orlicz Hardy spaces, Sci. China Math. 60 (2017), doi: 10.1007/s11425-016-9024-2, to appear. Guliyev, Vagif S.; Akbulut, Ali; Hamzayev, Vugar H.; Kuzu, Okan, Commutators of Marcinkiewicz integrals associated with Schrödinger operator on generalized weighted Morrey spaces, J. Math. Inequal. 10 (2016), no. 4, 947-970. Grafakos, Loukas, Classical Fourier Analysis, Second edition, Graduate Texts in Mathematics Vol. 249 (Springer, New York, 2009). Grafakos, Loukas, Modern Fourier Analysis, Second edition, Graduate Texts in Mathematics Vol. 250 (Springer, New York, 2009). Hörmander, Lars, Estimates for translation invariant operators in $L^p$ spaces, Acta Math. 104 (1960), 93-140. Hou, Shaoxiong; Yang, Dachun; Yang, Sibei, Lusin area function and molecular characterizations of Musielak-Orlicz Hardy spaces and their applications, Commun. Contemp. Math. 15 (2013), no. 6, 1350029, 37 pp. Janson, Svante, Generalizations of Lipschitz spaces and an application to Hardy spaces and bounded mean oscillation, Duke Math. J. 47 (1980), no. 4, 959-982. Jiang, Renjin; Yang, Dachun, New Orlicz-Hardy spaces associated with divergence form elliptic operators, J. Funct. Anal. 258 (2010), no. 4, 1167-1224. Johnson, Raymond L.; Neugebauer, Christoph Johannes, Homeomorphisms preserving $A_p$, Rev. Mat. Iberoamericana 3 (1987), no. 2, 249-273. Kurtz, Douglas S.; Wheeden, Richard L., Results on weighted norm inequalities for multipliers, Trans. Amer. Math. Soc. 255 (1979), 343-362. Ky, Luong Dang, New Hardy spaces of Musielak-Orlicz type and boundedness of sublinear operators, Integral Equations Operator Theory 78 (2014), no. 1, 115-150. Li, Baode; Fan, Xingya; Yang, Dachun, Littlewood-Paley characterizations of anisotropic Hardy spaces of Musielak-Orlicz type, Taiwanese J. Math. 19 (2015), no. 1, 279-314. Li, Baode; Fan, Xingya; Fu, Zunwei; Yang, Dachun, Molecular characterization of anisotropic Musielak-Orlicz Hardy spaces and their applications, Acta Math. Sin. (Engl. Ser.) 32 (2016), no. 11, 1391-1414. Li, Jinxia; Sun, Ruirui; Li, Baode, Anisotropic interpolation theorems of Musielak-Orlicz type, J. Inequal. Appl. 2016, 2016: 243. Liang, Yiyu; Huang, Jizheng; Yang, Dachun, New real-variable characterizations of Musielak-Orlicz Hardy spaces, J. Math. Anal. Appl. 395 (2012), no. 1, 413-428. Liang, Yiyu; Yang, Dachun, Musielak-Orlicz Campanato spaces and applications, J. Math. Anal. Appl. 406 (2013), no. 1, 307-322. Liang, Yiyu; Yang, Dachun; Jiang, Renjin, Weak Musielak-Orlicz Hardy spaces and applications, Math. Nachr. 289 (2016), no. 5-6, 634-677. Lin, Chin-Cheng; Lin, Ying-Chieh, $H^p_\omega-L^p_\omega$ boundedness of Marcinkiewicz integral, Integral Equations Operator Theory 58 (2007), no. 1, 87-98. Li, Bo; Liao, Minfeng; Li, Baode, Boundedness of Marcinkiewicz integrals with rough kernels on Musielak-Orlicz Hardy spaces, J. Inequal. Appl. 2017, 2017: 228. Marcinkiewicz, J., Sur quelques intégrales du type de Dini, Annales de la Société Polonaise de Mathématiques, 17 (1938), 42-50. Shi, Xinfeng; Jiang, Yinsheng, Weighted boundedness of parametric Marcinkiewicz integral and higher order commutator, Anal. Theory Appl. 25 (2009), no. 1, 25-39. Stefan, Rolewicz, Metric Linear Spaces, Second edition, (PWN-Polish Scientific Publishers, Warsaw, 1984). Stein, Elias M., On the functions of Littlewood-Paley, Lusin, and Marcinkiewicz, Trans. Amer. Math. Soc. 88 (1958), 430-466. Strömberg, Jan-Olov; Torchinsky, Alberto, Weighted Hardy spaces, Lecture Notes in Mathematics Vol. 1381 (Springer-Verlag, Berlin, 1989). Strömberg, Jan-Olov; Wheeden, Richard L., Fractional integrals on weighted $H^p$ and $L^p$ spaces, Trans. Amer. Math. Soc. 287 (1985), no. 1, 293-321. Wang, Hua, Parametric Marcinkiewicz integrals on the weighted Hardy and weak Hardy spaces, J. Math. Inequal. 10 (2016), no. 2, 373-391. Yang, Dachun; Liang, Yiyu; Ky, Luong Dang, Real-Variable Theory of Musielak-Orlicz Hardy spaces, Lecture Notes in Mathematics Vol. 2182 (Springer, Cham, 2017). Liu Xiong, Li Baode, Qiu Xiaoli and Li Bo (Corresponding author) College of Mathematics and System Sciences\ Xinjiang University\ Urumqi 830046\ P. R. China [E-mail ]{}:\ `[email protected]` (Liu Xiong)\ `[email protected]` (Li Baode)\ `[email protected]` (Qiu Xiaoli)\ `[email protected]` (Li Bo)	{ "pile_set_name": "ArXiv" }
--- abstract: 'Consider the situation where a data analyst wishes to carry out an analysis on a given dataset. It is widely recognized that most of the analyst’s time will be taken up with data engineering tasks such as acquiring, understanding, cleaning and preparing the data. In this paper we provide a description and classification of such tasks into high-levels groups, namely data organization, data quality and feature engineering. We also make available four datasets and example analyses that exhibit a wide variety of these problems, to help encourage the development of tools and techniques to help reduce this burden and push forward research towards the automation or semi-automation of the data engineering process.' author: - Alfredo Nazabal - 'Christopher K.I. Williams' - Giovanni Colavizza - Camila Rangel Smith - Angus Williams bibliography: - 'autods\_ckiw.bib' title: \| Data Engineering for Data Analytics:\ A Classification of the Issues, and Case Studies --- Introduction {#sec:introduction} ============ Overview of Case Studies \[sec:case-studies\] ============================================= Classification of Wrangling Challenges \[sec:class\] ==================================================== Related Work \[sec:rel-work\] ============================= Conclusions \[sec:conc\] ======================== In this paper we have identified three high-level groups of data wrangling problems, those related with obtaining a proper representation of the data (data organization), those related to assessing [and improving]{} the quality of the data (data quality), and feature engineering issues, which heavily depend on the task at hand and the model employed to solve it. Furthermore, we have presented the full analysis of four use cases, where we have provided a systematic pipeline for each of the datasets to clean them while identifying and classifying the main problems the data scientists faced during the wrangling steps. We hope that this work helps to further explore and understand the field of data engineering, and to value a part of every data scientist’s work that most of the time goes unnoticed both in research and industry. Additionally, we would like to encourage practitioners to provide their raw data and the scripts necessary to clean it in order to advance the field. In future work we would like to study data engineering workflows across multiple datasets in order to identify (if possible) common structures concerning the ordering of the various wrangling operations. We note that there can be feedback loops in the process, as described e.g. in [@crisp-dm-00]. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported in by The Alan Turing Institute under the EPSRC grant EP/N510129/1 and by funding provided by the UK Government’s Defence & Security Programme in support of the Alan Turing Institute. We would like to thank the AIDA team including Taha Ceritli, Jiaoyan Chen, James Geddes, Zoubin Ghahramani, Ian Horrocks, Ernesto Jimenez-Ruiz, Charles Sutton, Tomas Petrick and Gerrit van den Burg for helpful discussions and comments on the paper.	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper studies the singularities of Cullen-regular functions of one quaternionic variable, as defined in [@advances]. The quaternionic Laurent series prove to be Cullen-regular. The singularities of Cullen-regular functions are thus classified as removable, essential or poles. The quaternionic analogues of meromorphic complex functions, called semiregular functions, turn out to be quotients of Cullen-regular functions with respect to an appropriate division operation. This allows a detailed study of the poles and their distribution.' author: - \| Caterina Stoppato\ Dipartimento di Matematica “Ulisse Dini”, Università di Firenze\ Viale Morgagni 67/A, 50134 Firenze, Italy\ [email protected]\ date: title: 'Poles of regular quaternionic functions' --- Introduction ============ Denote by ${{\mathbb{H}}}$ the skew field of real quaternions. Recall that it is obtained by endowing ${{\mathbb{R}}}^4$ with the multiplication operation defined on the standard basis $1,i,j,k$ by $$i^2 = j^2 = k^2 = -1,$$ $$ij = -ji = k, jk = -kj = i, ki = -ik = j,$$ $$1^2 = 1, 1 i = i 1 = i, 1 j = j 1 = j, 1 k = k 1 = k$$ and extended by distributivity to all quaternions $q = x_0 + x_1 i + x_2 j + x_3 k$. A new theory of quaternion-valued functions of one quaternionic variable has been proposed by G. Gentili and D. C. Struppa in [@cras; @advances]. The theory is based on a definition of regularity for quaternionic functions inspired by C. G. Cullen [@cullen]. Several interesting results are proven in [@advances], including the Cullen-regularity of quaternionic power series and some basic properties of their zeros. The study of the zero-sets has been further developed in [@zeros; @multiplicity]. Let us quickly review the definition of Cullen-regular function and the basic properties of such a function. Denote by ${{\mathbb{S}}}$ the two-dimensional sphere of quaternionic imaginary units: ${{\mathbb{S}}}= \{q \in {{\mathbb{H}}}: q^2 =-1\}$. For all imaginary unit $I \in {{\mathbb{S}}}$, let $L_I = {{\mathbb{R}}}+ I {{\mathbb{R}}}$ be the complex line through $0, 1$ and $I$. \[definition\] Let $\Omega$ be a domain in ${{\mathbb{H}}}$ and let $f : \Omega \to {{\mathbb{H}}}$ be a real differentiable function. $f$ is said to be [Cullen-regular]{.nodecor} if, for all $I \in {{\mathbb{S}}}$, the function $\bar \partial_I f : \Omega \cap L_I \to {{\mathbb{H}}}$ defined by $$\bar \partial_I f (x+Iy) = \frac{1}{2} ( \frac{\partial}{\partial x}+I\frac{\partial}{\partial y} ) f_I (x+Iy)$$ vanishes identically. With the notations $\Omega_I = \Omega \cap L_I$ and $f_I = f_{\|_{\Omega_I}}$, we may refer to the vanishing of $\bar \partial_I f$ saying that the restriction $f_I$ is holomorphic on $\Omega_I$. The following result clarifies the meaning of such a condition. \[splitting\] Let $f : \Omega \to {{\mathbb{H}}}$ be a Cullen-regular function and choose imaginary units $I,J \in {{\mathbb{S}}}$ with $I \perp J$. There exist functions $F,G : \Omega_I \to L_I$ such that $f_I = F + G J$ on $\Omega_I$. With the natural identification between $L_I$ and the complex field ${{\mathbb{C}}}$, the functions $F,G$ are holomorphic, complex-valued functions of one complex variable. From now on we will omit Cullen’s name and refer to these functions just as regular functions. As observed in [@advances], a quaternionic power series $\sum_{n \in {{\mathbb{N}}}} q^n a_n$ with $a_n \in {{\mathbb{H}}}$ defines a regular function in its domain of convergence, which proves to be a ball $B(0,R) = \{q \in {{\mathbb{H}}}: \|q\| <R\}$. In the same paper, it is proven that If $f : B = B(0,R) \to {{\mathbb{H}}}$ is regular then there exist quaternions $a_n \in {{\mathbb{H}}}$ such that $f(q)=\sum_{n \in {{\mathbb{N}}}} q^n a_n$ for all $q \in B$. In particular, $f \in C^{\infty}(B)$. We may thus identify the set of regular functions on $B(0,R)$ with the set ${{\mathcal{D}}}_R$ of quaternionic power series converging in $B(0,R)$. In [@advances] many basic results in complex analysis are extended to regular functions of this type: the identity principle, the maximum modulus principle, the Cauchy representation formula, the Liouville theorem, the Morera theorem and the Schwarz lemma. A version of the open mapping theorem has been recently proven in [@openarxiv; @open]. The peculiar properties of the zeros proven in [@advances; @zeros], which we summarize in section \[preliminarysection\], arouse a new question. Do these functions have point singularities resembling the poles of holomorphic complex functions? In section \[laurentsection\], we give a positive answer to this question: we prove that a quaternionic Laurent series $$\sum_{n \in {{\mathbb{Z}}}} q^n a_n$$ defines a regular function on its domain of convergence, which is a four-dimensional spherical shell $A(0, R_1, R_2) = \{q \in {{\mathbb{H}}}: R_1 < \|q\| < R_2\}$. This allows us to construct functions which are regular on a punctured ball $B(0,R) {{\setminus \{0\}}}$ and have a singularity at $0$. Moreover, we prove that any function which is regular on a spherical shell $A(0, R_1, R_2)$ admits a Laurent series expansion centered at $0$. This result is extended in section \[typesection\] to the following: \[introexpansion\] Let $f$ be a regular function on a domain $\Omega$, let $p \in {{\mathbb{H}}}$ and let $L_I$ be a complex line through $p$. If $\Omega$ contains an annulus $A_I = A(p,R_1,R_2) \cap L_I$ then there exist $\{a_n\}_{n \in {{\mathbb{Z}}}} \subseteq {{\mathbb{H}}}$ such that $f_I(z) = \sum_{n \in {{\mathbb{Z}}}} (z-p)^n a_n$ for all $z \in A_I$. If, moreover, $p \in {{\mathbb{R}}}$ then $f(q) = \sum_{n \in {{\mathbb{Z}}}} (q-p)^n a_n$ for all $q \in A(p,R_1,R_2) \cap \Omega$. This allows us to define quaternionic analogues of the concepts of pole and essential singularity of holomorphic functions of one complex variable. We define the (classical) order $ord_f(p)$ of a pole and call a function $f$ semiregular if it does not have essential singularities or, equivalently, if the restriction $f_I$ is meromorphic for all $I \in {{\mathbb{S}}}$. Note that, by the final statement of theorem \[introexpansion\], real singularities are completely analogous to singularities of holomorphic functions of one complex variable. There is no resemblance to the case of several complex variables: no such result as Hartog’s lemma can hold. As for non-real singularities, we remark that theorem \[introexpansion\] only provides information on the complex line $L_I$ through the point $p$; we apparently cannot predict the behavior of the function in a (four-dimensional) neighborhood of $p$. In order to overcome this difficulty, in section \[divisionsection\] we introduce some new algebraic manipulation. We associate to any couple of regular functions $g,h: B(0,R) \to {{\mathbb{H}}}$ a function $h^{-}g$, called the left regular quotient of $h$ and $g$. We study the basic properties of such a function and prove that it is semiregular on $B(0,R)$. We are thus able to conclude that if $f_I(z) = (z-p)^{-1} g_I(z)$ for $g_I$ holomorphic on a disk $B_I(0,R) = B(0,R) \cap L_I$ containing $p$ then $$f(q) = (q-p)^{-}g(q)$$ for some function $g$ which is regular on $B(0,R)$. As a consequence, in section \[semiregularsection\] we prove the following result. Let $f$ be a semiregular function on $B(0,R_0)$. For all $R<R_0$, $I \in {{\mathbb{S}}}$ there exist a polynomial $P(q)$ having coefficients in $L_I$ and a regular function $g : B(0,R) \to {{\mathbb{H}}}$ such that $f = P^{-}g$ on $B(0,R)$. In particular $f$ is semiregular on $B(0,R_0)$ if and only if $f_{\|_{B(0,R)}}$ is a left regular quotient for all $R<R_0$. This allows the definition of a multiplication operation $$ on the set of semiregular functions on a ball and the proof of the following result (where we denote $h^{n} = h...h = _{j=1}^n h$ the nth power of a regular function $h$ with respect to $$-multiplication). Let $f$ be a semiregular function on $B = B(0,R)$, choose $p \in B$ and let $m = ord_f(p), n = ord_f(\bar p)$. There exists a unique semiregular function $g$ on $B$ such that $$f(q) = [(q-p)^{m}(q-\bar p)^{n}]^{-}* g(q)$$ The function $g$ is regular near $p$ and $\bar p$ and $g(p) \neq 0, g(\bar p) \neq 0$, provided $m>0$ or $n>0$. The previous result allows the study of the structure of the poles: If $f$ is a semiregular function on $B = B(0,R)$ then $f$ extends to a regular function on $B$ minus a union of isolated real points or isolated 2-spheres of the type $x+y{{\mathbb{S}}}= \{x+yI : I \in {{\mathbb{S}}}\}$ with $x,y \in {{\mathbb{R}}}, y \neq 0$. All the poles on each 2-sphere $x+y{{\mathbb{S}}}$ have the same order with the possible exception of one, which must have lesser order. Finally, in section \[multiplicitysection\] we prove the following. Let $f$ be semiregular on $B = B(0,R)$ and suppose $f \not \equiv 0$. For all $x+y{{\mathbb{S}}}\subseteq B$, there exist $m \in {{\mathbb{Z}}}, n \in {{\mathbb{N}}}$, $p_1,...,p_n \in x+y{{\mathbb{S}}}$ with $p_i \neq \bar p_{i+1}$ for all $i,j$ such that $$f(q) = [(q-x)^2+y^2]^m (q-p_1)(q-p_2)...(q-p_n)g(q)$$ for some semiregular function $g$ on $B$ which does not have poles nor zeros in $x+y{{\mathbb{S}}}$. This theorem allows to extend to transcendental functions the concepts of spherical multiplicity and isolated multiplicity of the zeros defined in [@multiplicity] for polynomials. It also leads to analogous definitions for the poles of a semiregular function. Preliminary results {#preliminarysection} =================== We now run through the basic properties of the zero-sets of regular functions. In [@advances] it is proven that \[symmetry\] Let $f : B(0,R) \to {{\mathbb{H}}}$ be a regular function and let $x, y \in {{\mathbb{R}}}$ be such that $x^2 +y^2 <R^2$. If there exist distinct imaginary units $I,J \in {{\mathbb{S}}}$ such that $f(x+yI) = f(x+yJ) = 0$, then $f(x+yK) = 0$ for all $K \in {{\mathbb{S}}}$. In other words, if $f$ has more than one zero on $x+y{{\mathbb{S}}}= \{x+yI : I \in {{\mathbb{S}}}\}$ then it vanishes identically on $x+y{{\mathbb{S}}}$. Note that $x+y{{\mathbb{S}}}$ is a 2-sphere if $y \neq 0$, a real singleton $\{x\}$ if $y= 0$. The polynomial function $f(q) = q^2+1$ vanishes on ${{\mathbb{S}}}$. For all $x,y \in {{\mathbb{R}}}$, the function $g(q) = (q-x)^2 +y^2 = q^2 - q 2x + x^2+y^2$ vanishes on $x+y{{\mathbb{S}}}$. In [@zeros] the zero-set is further characterized as follows. \[structure\] Let $f$ be a regular function on an open ball $B(0,R)$. If $f$ is not identically zero then its zero-set consists of isolated points or isolated 2-spheres of the form $x + y {{\mathbb{S}}}$, for $x,y \in {{\mathbb{R}}}, y\neq 0$. This result and the previous are proven for polynomials in [@shapiro] using quite simple tools. On the contrary, the study of the zero-set conducted in [@zeros] requires the introduction of the following operations on regular functions $f : B(0,R) \to {{\mathbb{H}}}$. \[multiplication\] Let $f, g$ be regular functions on an open ball $B = B(0,R)$ and let $f(q) = \sum_{n \in {{\mathbb{N}}}} q^n a_n, g(q) = \sum_{n \in {{\mathbb{N}}}} q^n b_n$ be their power series expansions. We define the [regular product]{.nodecor} of $f$ and $g$ as the regular function $fg : B \to {{\mathbb{H}}}$ defined by $$fg(q) = \sum_{n \in {{\mathbb{N}}}} q^n c_n, \ c_n = \sum_{k=0}^n a_k b_{n-k}.$$ Moreover, we define the [regular conjugate]{.nodecor} of $f$, $f^c : B \to {{\mathbb{H}}}$, by $f^c(q) = \sum_{n \in {{\mathbb{N}}}} q^n \bar a_n$ and the [symmetrization]{.nodecor} of $f$, as $f^s = f * f^c = f^cf$. The series $fg$ and $f^c$ clearly converge in $B$. Also note that $f^s(q) = \sum_{n \in {{\mathbb{N}}}} q^n r_n$ with $r_n = \sum_{k = 0}^n a_k \bar a_{n-k} \in {{\mathbb{R}}}$. Since no confusion can arise, we may also write $f(q)g(q)$ for $fg(q)$. Fix $R$ with $0<R\leq \infty$ and let ${{\mathcal{D}}}_R$ be the set of regular functions $f:B(0,R)\to {{\mathbb{H}}}$. Then $({{\mathcal{D}}}_R,+,)$ is an associative algebra over ${{\mathbb{R}}}$. As observed in [@zeros], the zeros of regular functions cannot be factored with respect to the standard multiplication of ${{\mathbb{H}}}$. However, a factorization property is proven in [@zeros] in terms of $$-multiplication, extending the results proven for polynomials in [@lam; @serodio]. \[factorization\] Let $f : B = B(0,R) \to {{\mathbb{H}}}$ be a regular function and let $p \in B$. Then $f(p) = 0$ if and only if there exists another regular function $g : B \to {{\mathbb{H}}}$ such that $f(q) = (q-p) * g(q)$. Fix $x,y \in {{\mathbb{R}}}$ and consider $f(q) = (q-x)^2 +y^2$. For all $I \in {{\mathbb{S}}}$, $f$ vanishes at $p = x+yI$ and it can be factored as $f(q) = q^2 - q 2x + x^2+y^2 = q^2 - q (p + \bar p) + p \bar p = (q-p)(q-\bar p)$. If we define the nth regular power of $f$ as $f^{n} = f...f = _{i=1}^n f$, we can give the following definition. \[multiplicity\] Let $f : B = B(0,R) \to {{\mathbb{H}}}$ be a regular function, suppose $f \not \equiv 0$ and let $p \in B$. We define the [(classical) multiplicity]{.nodecor} of $p$ as a zero of $f$ and denote by $m_f(p)$ the largest $n \in {{\mathbb{N}}}$ such that there exists a regular function $g : B \to {{\mathbb{H}}}$ with $f(q)=(q-p)^{n}g(q)$. The classical multiplicity is a consistent generalization of the complex multiplicity; in other words, if $p \in L_I$ then $m_f(p)$ is the largest $n \in {{\mathbb{N}}}$ such that there exists a holomorphic function $g_I$ with $f_I(z) = (z-p)^{n}g_I(z)$. This proves, a posteriori, that definition \[multiplicity\] is well posed: indeed, by the identity principle proven in [@advances], $f_I \not \equiv 0$; thus we cannot factor $z-p$ out of $f_I(z)$ “infinitely many times”. Finally, the zero set of a regular product is completely characterized in terms of the zeros of the two factors by the following result, which is proven in [@zeros] and extends [@lam; @serodio]. \[zerosmultiplication\] Let $f,g$ be regular functions on an open ball $ B = B(0,R)$ and let $p \in B$. If $f(p) = 0$ then $fg(p) = 0$, otherwise $fg(p) = f(p)(f(p)^{-1} p f(p))$. In particular $p$ is a zero of $fg$ if and only if $f(p) = 0$ or $g(f(p)^{-1} p f(p))=0$. We conclude this section recalling that the zero-sets of $f^c$ and $f^s$ are characterized in [@zeros] as follows. \[conjugatezeros\] Let $f$ be a regular function on $B = B(0,R)$. For all $x,y \in {{\mathbb{R}}}$ with $x+y{{\mathbb{S}}}\subseteq B$, the zeros of the regular conjugate $f^c$ on $x+y{{\mathbb{S}}}$ are in one-to-one correspondence with those of $f$. Moreover, the symmetrization $f^s$ vanishes exactly on the sets $x+y{{\mathbb{S}}}$ on which $f$ has a zero. Laurent series and expansion {#laurentsection} ============================ The first step in the study of point singularities is generalizing the theory of Laurent series to the quaternionic case. The domain of convergence of a quaternionic Laurent series $\sum_{n \in {{\mathbb{Z}}}} q^n a_n$ is a four-dimensional spherical shell $A(0, R_1, R_2) = \{q \in {{\mathbb{H}}}: R_1 < \|q\| < R_2\}$. More precisely one can prove, just as in the complex case, the following result. \[convergence\] Let $\{a_n\}_{n \in {{\mathbb{Z}}}} \subseteq {{\mathbb{H}}}$. There exist $R_1, R_2$ with $0 \leq R_i \leq \infty$ such that 1. the series $\sum_{n \in {{\mathbb{N}}}} q^n a_n$ and $ \sum_{n \in {{\mathbb{N}}}} q^{-n} a_{-n}$ both converge absolutely and uniformly on compact subsets of $A = A(0,R_1,R_2)$; 2. for all $q \in {{\mathbb{H}}}\setminus \bar A$ (except possibly $0$ if $A = \emptyset$), either $\sum_{n \in {{\mathbb{N}}}} q^n a_n$ or $ \sum_{n \in {{\mathbb{N}}}} q^{-n} a_{-n}$ diverge. Note that $A=A(0,R_1,R_2)$ is empty if and only if $R_1\geq R_2$. If this is not the case then we define the sum of the series $\sum_{n \in {{\mathbb{Z}}}} q^n a_n$ as $\sum_{n \in {{\mathbb{N}}}} q^n a_n + \sum_{n>0} q^{-n} a_{-n}$ for all $q$ in $A$, which we may call the domain of convergence of the series. \[laurentregularity\] Let $\sum_{n \in {{\mathbb{Z}}}} q^n a_n$ have domain of convergence $A = A(0,R_1,R_2)$ with $R_1 < R_2$. Then $f : A \to {{\mathbb{H}}}\ \ q \mapsto \sum_{n \in {{\mathbb{Z}}}} q^n a_n$ is a regular function. The proof follows by computation from definition \[definition\]. We will now prove that all regular functions $f : A(0,R_1,R_2) \to {{\mathbb{H}}}$ admit Laurent series expansions. In order to prove it, we make use of the identity principle. This result is proven in [@advances] for functions which are regular on a ball $B(0,R)$, but it easily extends to a larger class of domains. \[identity\] Let $\Omega$ be a domain in ${{\mathbb{H}}}$ intersecting the real axis and having connected intersection $\Omega_I = \Omega \cap L_I$ with any complex line $L_I$. If $f,g : \Omega \to {{\mathbb{H}}}$ are regular functions which coincide on $\Omega \cap {{\mathbb{R}}}$ then they coincide on the whole domain $\Omega$. Let $h = f-g$ and let us prove $h \equiv 0$ on $\Omega$. Choose any imaginary unit $I \in {{\mathbb{S}}}$ and consider the restriction $h_I = h_{\|_{\Omega_I}}$. Since $h_I : \Omega_I \to {{\mathbb{H}}}$ is holomorphic and it vanishes on the set $\Omega \cap {{\mathbb{R}}}$, which is not discrete, $h_I$ must vanish identically on $\Omega_I$. We are now ready to prove the following. \[expansion\] Let $A = A(0,R_1,R_2)$ with $0 \leq R_1< R_2$ and let $f : A \to {{\mathbb{H}}}$ be a regular function. There exist $\{a_n\}_{n \in {{\mathbb{Z}}}} \subseteq {{\mathbb{H}}}$ such that $$f(q) = \sum_{n \in {{\mathbb{Z}}}} q^n a_n$$ for all $q \in A$. Choose a complex line $L_I$ and consider the annulus we get by intersecting $L_I$ with the shell $A$: $$A_I = A_I(0,R_1,R_2) = \{z \in L_I : R_1<\|z\| < R_2\}.$$ Consider the restriction $f_I = f_{\|_{A_I}}$ and choose $J \in {{\mathbb{S}}}, J \perp I$. As we saw in lemma \[splitting\], we can find two functions $F, G : A_I \to L_I$ which are holomorphic (with the natural identification between $L_I$ and ${{\mathbb{C}}}$) and such that $f_I = F + G J$. Let $F(z) = \sum_{n \in {{\mathbb{Z}}}} z^n \alpha_n$ and $G(z) = \sum_{n \in {{\mathbb{Z}}}} z^n \beta_n$ be the Laurent series expansions of the functions $F$ and $G$ (which have coefficients $\alpha_n, \beta_n \in L_I$). If we let $a_n = \alpha_n + \beta_n J$ for all $n \in {{\mathbb{Z}}}$, then $$f_I(z) = \sum_{n \in {{\mathbb{Z}}}} z^n a_n$$ for all $z \in A_I$. Now consider the quaternionic Laurent series $\sum_{n \in {{\mathbb{Z}}}} q^n a_n$. By lemma \[convergence\] it converges in $A$. Hence, by theorem \[laurentregularity\], its sum defines a regular function on $A$. This function coincides with $f$ on $A_I$ by construction. We can conclude, using the identity principle \[identity\], that it coincide with $f$ on the whole domain $A$. The above argument also proves that \[existence\] Let $A_I = A_I(0,R_1,R_2)$ with $I \in {{\mathbb{S}}}, 0 \leq R_1< R_2$ and let $f_I : A_I \to {{\mathbb{H}}}$ be holomorphic. There exists exactly one regular function $g : A(0,R_1,R_2) \to {{\mathbb{H}}}$ such that $f_I(z) = g_I(z)$ for all $z \in A_I$. Types of singularities {#typesection} ====================== We now remark that the results proven in the previous section not only work for point $0$. It is easy to prove that for all $r \in {{\mathbb{R}}}$ the series $\sum_{n \in {{\mathbb{Z}}}} (q-r)^n a_n$ converges and defines a regular function on some shell $A(r,R_1,R_2) = \{q \in {{\mathbb{H}}}: R_1<\|q-r\|<R_2 \}$. Thus we can generalize all the above formulae just by substituting $r$ to $0$, $(q-r)^n$ to $q^n$ et cetera. On the contrary, we cannot apply the same procedure to a non-real quaternion $p \in {{\mathbb{H}}}\setminus {{\mathbb{R}}}$. Indeed, a quaternionic Laurent series centered at such a point $p$ has a nice convergence domain, $A(p,R_1,R_2) = \{q \in {{\mathbb{H}}}: R_1<\|q-p\|<R_2 \}$, but its sum does not define, in general, a regular function. Indeed it is easy to check that (due to the non-commutativity of ${{\mathbb{H}}}$) the function $P(q) = (q-p)^n$ is not regular for $n \in {{\mathbb{Z}}}\setminus \{0,1\}$ and $p \in {{\mathbb{H}}}\setminus {{\mathbb{R}}}$. Nevertheless, the first part of the proof of theorem \[expansion\] still works at a non-real point. This leads to the following result: \[weakexpansion\] Let $f$ be a regular function on a domain $\Omega$, let $p \in {{\mathbb{H}}}$ and let $L_I$ be a complex line through $p$. If $\Omega$ contains an annulus $A_I = A(p,R_1,R_2) \cap L_I$ with $0 \leq R_1< R_2$ then there exist $\{a_n\}_{n \in {{\mathbb{Z}}}} \subseteq {{\mathbb{H}}}$ such that $$\label{weakexpansioneq} f_I(z) = \sum_{n \in {{\mathbb{Z}}}} (z-p)^n a_n$$ for all $z \in A_I$. If, moreover, $p =r \in {{\mathbb{R}}}$ then $f(q) = \sum_{n \in {{\mathbb{Z}}}} (q-r)^n a_n$ for all $q \in A(r,R_1,R_2) \cap \Omega$. For all $p$ and all complex lines $L_I$ through $p$, this result allows us to classify the behavior of $f(q)$ when $q$ approaches $p$ along $L_I$. Note that if $p$ does not lie in ${{\mathbb{R}}}$ then there exists exactly one complex line $L_I$ through $p$. If, on the contrary, $p$ is real then it belongs to all complex lines $L_I$; however, in this case the coefficients $a_n$ which appear in equation \[weakexpansioneq\] are the same for all $I \in {{\mathbb{S}}}$, thanks to the final statement of theorem \[weakexpansion\]. We can thus give the following definition. Let $f$, $p$ and $a_n$ be as in theorem \[weakexpansion\]. The point $p$ is called a [pole]{.nodecor} if there exists an $n \in {{\mathbb{N}}}$ such that $a_{-m} = 0$ for all $m>n$; the minimum of such $n \in {{\mathbb{N}}}$ is called the [(classical) order]{.nodecor} of the pole and denoted as $ord_f(p)$. If $p$ is not a pole for $f$ then we call it an [essential singularity]{.nodecor} for $f$. When no confusion can arise, we omit the adjective “classical” for the sake of simplicity. For a real point we derive from theorem \[weakexpansion\] the following classification. \[realclassification\] Let $A = A(r,R_1,R_2)$ with $r \in {{\mathbb{R}}}, 0 \leq R_1< R_2$ and let $f : A \to {{\mathbb{H}}}$ be a regular function. 1. If $r$ is a pole of order $0$ then $f$ extends to $B(r,R_2) = \{q \in {{\mathbb{H}}}: \|q-r\|<R_2\}$ as a regular function. 2. If $r$ is a pole of order $n>0$ then there exists a regular function $g : B = B(r,R_2) \to {{\mathbb{H}}}$ such that $f(q) = (q-r)^{-n}g(q)$ for all $q \in A$. In particular, $f$ extends to $B {{\setminus \{r\}}}$ as a regular function and $\lim_{q \to r} \|f(q)\| = + \infty$. 3. Suppose $r$ to be an essential singularity. If $f$ extends to $B(r,R_2) {{\setminus \{r\}}}$, then the modulus $\|f\|$ is unbounded on $U {{\setminus \{r\}}}$ for all neighborhood $U$ of $r$; moreover, the limit $\lim_{q \to r} f(q)$ is not defined. Note that when $p$ is not real the classification of $p$ as a pole or an essential singularity only depends on the restriction $f_I$ to the complex line through $p$. A priori, it does not predict the behavior of $f$ in a four-dimensional neighborhood of $p$. For instance, if $p$ is a pole of order $m$ then we derive from theorem \[weakexpansion\] the existence of a holomorphic function $g_I : A_I \to {{\mathbb{H}}}$ such that $$f_I(z) = \frac {1}{(z-p)^{m}} g_I(z).$$ However, for $p \in {{\mathbb{H}}}\setminus {{\mathbb{R}}}$ we cannot conclude that $f(q)$ equal $(q-p)^{-m} g(q)$ for some regular $g$ (note that the second expression does not generally define a regular function). In order to prove a result of this type we need an adequate division operation on regular functions, which we will define in section \[divisionsection\]. Remark that even when $m=0$, i.e. $f_I$ extends to a holomorphic function on a neighborhood of $p$ in $L_I$, we may not conclude that $f$ extends to a regular function on a neighborhood of $p$ in ${{\mathbb{H}}}$: theorem \[existence\] was only proven for spherical shells $A(0,R_1,R_2)$ centered at $0$ and it does not immediately generalize to $A(p,R_1,R_2)$ with $p \in {{\mathbb{H}}}\setminus {{\mathbb{R}}}$. This is why we did not call a pole of order $0$ a removable singularity. Our caution will prove correct in section \[semiregularsection\]. We will instead give the following, natural definition. The point $p$ is called a [removable singularity]{.nodecor} if $f$ extends to a neighborhood of $p$ in ${{\mathbb{H}}}$ as a regular function. We conclude this section defining an analogue to the concept of meromorphic function. We will call semiregular a function which does not have essential singularities. More precisely: Let $\Omega$ be a domain in ${{\mathbb{H}}}$, let ${{\mathcal{S}}}\subseteq \Omega$ and suppose the intersection ${{\mathcal{S}}}_I = {{\mathcal{S}}}\cap L_I$ to be a discrete subset of $\Omega_I = \Omega \cap L_I$ for all $I \in {{\mathbb{S}}}$. A regular function $f : \Omega \setminus {{\mathcal{S}}}\to {{\mathbb{H}}}$ is said to be [semiregular]{.nodecor} on $\Omega$ if ${{\mathcal{S}}}$ does not contain essential singularities for $f$. In other words, $f$ is semiregular on $\Omega$ if and only if, for all $I \in {{\mathbb{S}}}$, the restriction $f_I$ is a meromorphic function on $\Omega_I$. Note that we did not ask for the set of singularities ${{\mathcal{S}}}$ to be discrete: in order to classify a point $p \in {{\mathcal{S}}}$ as a pole or an essential singularity for $f$ it is enough for $p$ to be isolated in ${{\mathcal{S}}}_I = {{\mathcal{S}}}\cap L_I$ (see the hypotheses of theorem \[weakexpansion\]). In section \[semiregularsection\] we will present a detailed study of the functions which are semiregular on a ball $B(0,R)$. This study requires the introduction of the above mentioned division operation, which we undertake in the next section. Regular quotients {#divisionsection} ================= As we saw in section \[preliminarysection\], the zeros of a regular function cannot be factored with respect to the multiplication of ${{\mathbb{H}}}$, but they have nice multiplicative properties in terms of the non-standard multiplication $$. Similarly, the apparent difficulties we found in dealing with non-real poles can be solved in terms of a non-standard division operation. Denote by ${{\mathcal{Z}}}_f = \{q \in B(0,R) : f(q) = 0\}$ the zero-set of a function $f : B(0,R) \to {{\mathbb{H}}}$. Let $f,g : B = B(0,R) \to {{\mathbb{H}}}$ be regular functions and let $f^c,f^s$ be the regular conjugate and the symmetrization of $f$. We define the [left regular quotient]{.nodecor} of $f$ and $g$ as the function $f^{-} * g : B \setminus {{\mathcal{Z}}}_{f^s} \to {{\mathbb{H}}}$ such that $$f^{-} g (q) = \frac{1}{f^s(q)} f^c * g(q).$$ The [right regular quotient]{.nodecor} of $g$ and $f$ is the function $gf^{-} : B \setminus {{\mathcal{Z}}}_{f^s} \to {{\mathbb{H}}}$ defined by $gf^{-} =f^{-s}(g f^c)$. Finally, we define the [regular reciprocal]{.nodecor} of $f$ as the function $f^{-} = f^{-} 1 = 1f^{-}$. Since no confusion can arise, we will often write $(f(q))^{-}$ for $f^{-}(q)$. We will also use the shorthand notation $f^{-s}(q)$ for $\frac{1}{f^s(q)}$. Regular quotients are regular on their domains of definition by the following lemma, which can be proven by direct computation. Let $f,g: B = B(0,R) \to {{\mathbb{H}}}$ be regular functions and suppose the power series expansions of $g$ at $0$, $g(q) = \sum_{n \in {{\mathbb{N}}}} q^n r_n$, has real coefficients $r_n \in {{\mathbb{R}}}$. Then the function $h: B \setminus {{\mathcal{Z}}}_g \to {{\mathbb{H}}}$ defined by $h(q) = \frac{1}{g(q)} f(q)$ is regular. Moreover, left and right regular quotients of regular functions on a ball $B(0,R)$ are semiregular on $B(0,R)$: Let $f,g : B = B(0,R) \to {{\mathbb{H}}}$ be regular functions and consider the left quotient $f^{-}g : B \setminus {{\mathcal{Z}}}_{f^s} \to {{\mathbb{H}}}$. Each $p \in {{\mathcal{Z}}}_{f^s}$ is a pole of order $$ord_{f^{-}g}(p) \leq m_{f^s}(p)$$ for $f^{-}g$, where $m_{f^s}(p)$ denotes the classical multiplicity of $f^s$ at $p$. The same holds for the right quotient $g * f^{-}$. In particular $f^{-}g$ and $g f^{-}$ are semiregular on $B$. Let $p = x+yI \in L_I$. If $m_{f^s}(p) = n$ then, as a consequence of theorem \[conjugatezeros\], $m_{f^s}(\bar p) = n$; there exists a holomorphic function $h_I$ with $h_I(p) \neq 0$ such that $f_I^s(z) = (z-p)^n(z-\bar p)^n h_I(z)= \left[(z-x)^2+y^2\right]^n h_I(z)$. We observe that, since $f^s$ is a series with real coefficients, $h_I(z)$ must be have real coefficients, too. As a consequence, $f_I^s(z) = h_I(z) \left[(z-x)^2+y^2\right]^n = h_I(z) (z-\bar p)^n (z-p)^n$ and $$(f^{-}g)_I(z) = f_I^{-s}(z) (f^c g)_I(z) = (z-p)^{-n}(z-\bar p)^{-n} h_I(z)^{-1}(f^c g)_I(z)$$ where $(z-\bar p)^{-n} h_I(z)^{-1}(f^c g)_I(z)$ is holomorphic in a neighborhood of $p$ in $L_I$. Moreover $(gf^{-})_I(z) = (z-p)^{-n}(z-\bar p)^{-n} h_I(z)^{-1}(gf^c)_I(z)$, with $(z-\bar p)^{-n} h_I(z)^{-1}(gf^c)_I(z)$ holomorphic in a neighborhood of $p$ in $L_I$. The regular quotient $f^{-} g(q)$ is related to the quotient $f(q)^{-1} g (q)= \frac{1}{f(q)} g(g)$ by the following result. First remark that, as a consequence of theorem \[conjugatezeros\], ${{\mathcal{Z}}}_{f^c} \subseteq {{\mathcal{Z}}}_{f^s}$, so that $f^c(q) \neq 0$ for all $q \in B \setminus {{\mathcal{Z}}}_{f^s}$. \[quotients\] Let $f,g$ be regular functions on $B=B(0,R)$. If we define $T_f : B \setminus {{\mathcal{Z}}}_{f^s} \to {{\mathbb{H}}}$ as $T_f(q) = f^c(q)^{-1} q f^c(q)$, then $$f^{-}g(q)= \frac{1}{f( T_f(q))} g (T_f(q))$$ for all $q \in B \setminus {{\mathcal{Z}}}_{f^s}$. By theorem \[zerosmultiplication\], $f^c(q) * g(q) = f^c(q) g(T_f(q))$ for all $q \in B \setminus {{\mathcal{Z}}}_{f^s}$. We conclude by computation: $$f^{-}g(q) = f^{-s}(q) f^cg(q)= [f^cf(q)]^{-1} f^cg(q)=$$ $$= [f^c(q) f(T_f(q))]^{-1} f^c(q)g(T_f(q)) = f(T_f(q))^{-1} f^c(q)^{-1} f^c(q) g(T_f(q)) =$$ $$= f(T_f(q))^{-1}g(T_f(q)).$$ Since for all $x,y \in {{\mathbb{R}}}, I \in {{\mathbb{S}}}$ and $p \in {{\mathbb{H}}}{{\setminus \{0\}}}$ we have $p^{-1}(x+yI) p = p^{-1}x p + p^{-1} y I p = x+y J$ with $J = p^{-1} I p \in {{\mathbb{S}}}$, we remark that: For all $x,y \in {{\mathbb{R}}}$ with $x^2+y^2<R^2$, the function $T_f$ maps the 2-sphere (or real singleton) $x+y{{\mathbb{S}}}= \{x+yI : I \in {{\mathbb{S}}}\}$ to itself. In particular, since by theorem \[conjugatezeros\] we have $${{\mathcal{Z}}}_{f^s} = \bigcup_{x+yI \in {{\mathcal{Z}}}_f} (x+y{{\mathbb{S}}}),$$ we conclude $T_f(B \setminus {{\mathcal{Z}}}_{f^s}) \subseteq B \setminus {{\mathcal{Z}}}_{f^s}$. Moreover, the following is proven in [@openarxiv; @open]. Let $f: B=B(0,R) \to {{\mathbb{H}}}$ be a regular function. $T_f$ and $T_{f^c}$ are mutual inverses. In particular $T_f : B \setminus {{\mathcal{Z}}}_{f^s} \to B \setminus {{\mathcal{Z}}}_{f^s}$ is a diffeomorphism. Let us give an example. For any fixed quaternion $p = x+yI \in {{\mathbb{H}}}$, the regular reciprocal of the polynomial $f(q) = q-p$ is $$(q-p)^{-} = \frac{1}{(q-\bar p)(q-p)} (q-\bar p)= \frac{1}{q^2 - q (p + \bar p) + \bar p p} (q-\bar p) = \frac{1}{(q- x)^2 + y^2} (q - \bar p),$$ where the polynomial $(q- x)^2 + y^2$ vanishes exactly on $x+y{{\mathbb{S}}}$. Moreover, by theorem \[quotients\], $$(q-p)^{-} = \frac{1}{T_f(q) -p} = \frac{1}{(q-\bar p)^{-1} q (q-\bar p) -p}$$ In particular, since $T_f(q) = (q-\bar p)^{-1} q (q-\bar p) = (q-\bar p)^{-1} (q-\bar p) q = q$ for $q$ in the complex line $L_I$ through $p$, the function $(q-p)^{-}$ coincides with $(q-p)^{-1} = \frac{1}{q-p}$ on $L_I \setminus \{p,\bar p\}$. We conclude this section explaining the algebraic meaning of regular quotients. In the complex case, the set of quotients $\frac{F}{G}$ of holomorphic functions $F,G$ on a disc $\Delta$ becomes a field when endowed with the usual operations of addition and multiplication. More precisely, it is the field of quotients of the integral domain (i.e. the commutative ring with no zero divisors) obtained by endowing the set of holomorphic functions $F$ on $\Delta$ with the natural addition and multiplication. As explained in [@rowen] (see also [@cohn; @lam2]), the concept of field of quotients of an integral domain can be generalized to the non-commutative case as follows. We define a [left Ore domain]{.nodecor} as a domain (a ring with no zero divisors) $(D,+,\cdot)$ such that $Da \cap Db \neq \{0\}$ for all $a,b \in D {{\setminus \{0\}}}$. If this is the case, then the set of formal quotients $L = \{a^{-1}b : a,b \in D\}$ can be endowed with operations $+, \cdot$ such that: - $D$ is isomorphic to a subring of $L$ (namely $ \{1^{-1}a : a \in D\}$); - $L$ is a skew field, i.e. a ring where every non-zero element has a multiplicative inverse (namely, $(a^{-1}b)^{-1} = b^{-1}a$). The ring $L$ is called the [classical left ring of quotients]{.nodecor} of $D$ and, up to isomorphism, it is the only ring having the properties (i) and (ii). On a right Ore domain* $D$, defined by $aD \cap bD \neq \{0\}$ for all $a,b \in D {{\setminus \{0\}}}$, we can similarly construct the classical right ring of quotients. If $D$ is both a left and a right Ore domain, then (by the uniqueness property) the two rings of quotients are isomorphic and we may speak of the classical ring of quotients of $D$. Fix $R$ with $0<R\leq \infty$. The associative real algebra $({{\mathcal{D}}}_R,+,)$ of regular functions on $B(0,R)$ is a left Ore domain and a right Ore domain. If we endow the set of left regular quotients ${{\mathcal{L}}}_R = \{f^{-}g : f,g \in {{\mathcal{D}}}_R, f \not \equiv 0\}$ with the multiplication $$ defined by $$(f^{-}g)(h^{-}k) = f^{-s}h^{-s} f^cgh^ck$$ then $({{\mathcal{L}}}_R,+,)$ is a division algebra over ${{\mathbb{R}}}$ and it is the classical ring of quotients of $({{\mathcal{D}}}_R,+,)$. The same holds for ${{\mathcal{R}}}_R = \{gf^{-} : f,g \in {{\mathcal{D}}}_R, f \not \equiv 0\}$ with the multiplication defined by $(gf^{-})(kh^{-}) = f^{-s}h^{-s} gf^ckh^c$. The multiplication $$ is well defined on ${{\mathcal{L}}}_R$: $f^{-}g = \tilde f^{-}* \tilde g$ if and only if there exist $l, \tilde l$ such that $l * f = \tilde l * \tilde f$, $lg=\tilde l\tilde g$ and in this case we get by direct computation that $f^{-s}h^{-s} f^cgh^ck = \tilde f^{-s}h^{-s} \tilde f^c\tilde gh^ck$; the same can be done for the second factor $h^{-}k$. Clearly, $({{\mathcal{L}}}_R,+,)$ is an associative algebra over ${{\mathbb{R}}}$. We remark that $f^{-}g$ has inverse element $g^{-}f$ with respect to $$: $$(f^{-}g) * (g^{-}f) = f^{-s}g^{-s} f^cgg^cf = f^{-s}g^{-s} f^cg^sf =$$ $$= f^{-s}g^{-s} g^s f^cf = f^{-s} f^cf = f^{-s} f^s = 1$$ and, switching $f$ and $g$, $(g^{-}f) (f^{-}g) = 1$. Thus $({{\mathcal{L}}}_R,+,)$ is a division algebra. The same holds for $({{\mathcal{R}}}_R,+,)$. The ring ${{\mathcal{D}}}_R$ is a domain, since $fg \equiv 0$ iff $f \equiv 0$ or $g\equiv 0$. Moreover, ${{\mathcal{D}}}_R$ is a left Ore domain: if $f,g \not \equiv 0$ then $({{\mathcal{D}}}_Rf) \cap ({{\mathcal{D}}}_Rg)$ contains the non-zero element $f^s g^s = g^s f^s$, which can be obtained as $(g^s f^c )f$ or as $(f^sg^c)g$. Similarly, ${{\mathcal{D}}}_R$ is a right Ore domain. Thus the classical ring of quotients of ${{\mathcal{D}}}_R$ is well defined. It must be isomorphic to both ${{\mathcal{L}}}_R$ and ${{\mathcal{R}}}_R$ by the uniqueness property: ${{\mathcal{L}}}_R,{{\mathcal{R}}}_R$ are skew fields which have ${{\mathcal{D}}}_R$ as a subring and the inclusions ${{\mathcal{D}}}_R \to {{\mathcal{L}}}_R \ \ f \mapsto f = 1^{-} f$, ${{\mathcal{D}}}_R \to {{\mathcal{R}}}_R \ \ f \mapsto f = f1^{-}$ prove to be ring homomorphisms by direct computation. Poles of semiregular functions {#semiregularsection} ============================== We now prove that all functions which are semiregular in a neighborhood of a bounded ball $B$ can be expressed as left quotients on $B$. \[polesfactorization\] Let $f$ be a semiregular function on $B(0,R_0)$ with $0<R_0\leq \infty$. Let $B = B(0,R)$ with $0<R<R_0$, choose $I \in {{\mathbb{S}}}$ and let $z_1,...,z_n$ be the poles of $f_I$ in $B_I = B \cap L_I$, listed according to their order $ord_f$. There exists a unique regular function $g : B \to {{\mathbb{H}}}$ such that $$\label{polesfactorizationequation} f(q) = \left[(q-z_1)...(q-z_n)\right]^{-} * g(q)$$ for all $q \in B$. Moreover, $g(z_j) \neq 0$ for all $j \in \{1,...,n\}$. As we observed in section \[typesection\], there exists a holomorphic $g_I : B_I \to {{\mathbb{H}}}$, with $g_I(z_j) \neq 0$ for all $j$, such that $$f_I(z) = \frac{1}{(z-z_1)...(z-z_n)}g_I(z)$$ for all $z \in B_I \setminus\{z_1,...,z_n\}$. As a consequence of theorem \[existence\], $g_I$ extends to a regular $g : B \to {{\mathbb{H}}}$. Consider the function $h(q) = \left[(q-z_1)...(q-z_n)\right]^{-} g(q)$: it is regular on its domain of definition, which is $\Omega = B \setminus \bigcup_{j = 1}^n \left(x_j+y_j{{\mathbb{S}}}\right)$ if $z_j=x_j+y_jI_j$. We remark that, as a consequence of theorem \[quotients\], $h_I (z) = [(z-z_1)...(z-z_n)]^{-1} g_I(z)=f_I(z)$ for all $z \in \Omega_I = B_I \setminus\{z_1, \bar z_1,...,z_n, \bar z_n\}$. The identity principle \[identity\] allows us to conclude that $f(q) = h(q)$ for all $q \in \Omega$. Let $0<R_0\leq \infty$. A function $f$ is semiregular on $B(0,R_0)$ if, and only if, $f_{\|_{B(0,R)}} \in {{\mathcal{L}}}_R$ for all $R<R_0$. We use the notation $f_{\|_{B(0,R)}}$ for the sake of simplicity, instead of writing $f _{\|_{\Omega \cap B(0,R)}}$ where $\Omega$ is the domain on which $f$ is regular. Thanks to the previous corollary, we can define a multiplication operation $$ on semiregular functions on $B(0,R_0)$ with $0<R_0\leq \infty$. Consider indeed two such functions $f,g$. On each ball $B(0,R)$ with $0<R<R_0$, the restrictions $f _{\|_{B(0,R)}},g_{\|_{B(0,R)}}$ can be represented as left regular quotients and we may consider their product $f _{\|_{B(0,R)}} g_{\|_{B(0,R)}}$. Moreover, taking $R_2>R_1$ we get that $f _{\|_{B(0,R_2)}}* g_{\|_{B(0,R_2)}}$ equals $f_{\|_{B(0,R_1)}} * g_{\|_{B(0,R_1)}}$ on $B(0,R_1)$. We can thus define: The [(semi)regular product]{.nodecor} of semiregular functions $f,g$ on $B(0,R_0)$ is the semiregular function $fg$ on $B(0,R_0)$ such that $(fg)_{\|_{B(0,R)}} = f_{\|_{B(0,R)}} * g_{\|_{B(0,R)}}$ for all $R<R_0$. We can now remark the following. Recall that we denote by $h^{n} = h...h = _{j=1}^n h$ the nth regular power of $h$. \[polefactorization\] Let $f$ be a semiregular function on $B(0,R_0)$ with $0<R_0\leq\infty$, choose $p = x+yI \in B(0,R_0)$ and let $m = ord_f(p), n = ord_f(\bar p)$. Without loss of generality, $m \leq n$. There exists a unique semiregular function $g$ on $B(0,R_0)$ such that $$f(q) = [(q-p)^{m}(q-\bar p)^{n}]^{-}* g(q) = \left[(q-x)^2+y^2\right]^{-n} (q-p)^{(n-m)} g(q)$$ The function $g$ is regular in a neighborhood of $x+y{{\mathbb{S}}}$ and, if $n>0$, $g(p),g(\bar p) \neq 0$. For all $R<R_0$, the existence of a $g^{(R)} \in {{\mathcal{L}}}_R$ such that $f(q) = [(q-p)^{m}(q-\bar p)^{n}]^{-}* g^{(R)} (q)$ on $B(0,R)$ is an immediate consequence of theorem \[polesfactorization\]. Clearly, if $R_1<R_2$ then $g^{(R_1)}$ equals $g^{(R_2)}$ on $B(0,R_1)$. We can thus define a global $g$, semiregular on $B(0,R_0)$, such that $f(q) = [(q-p)^{m}(q-\bar p)^{n}]^{-}* g(q)$ on $B(0,R_0)$. We conclude by observing that $$[(q-p)^{m}(q-\bar p)^{n}]^{c} = (q-p)^{n}(q-\bar p)^{m} = [(q-x)^2+y^2]^{m} (q-p)^{(n-m)}$$ and $[(q-p)^{m}(q-\bar p)^{n}]^{-s} = [(q-x)^2+y^2]^{-m-n}$, so that $$[(q-p)^{m}(q-\bar p)^{n}]^{-}* g(q) = \left[(q-x)^2+y^2\right]^{-n} (q-p)^{(n-m)} g(q).$$ We will soon use theorem \[polefactorization\] to study the distribution of the poles. Let us first give two significant examples. The regular function $f : {{\mathbb{H}}}\setminus {{\mathbb{S}}}\to {{\mathbb{H}}}$ defined by $$f(q) = (q^2+1)^{-} = \frac{1}{q^2+1}$$ has a pole of order $1$ at any point $I \in {{\mathbb{S}}}$: we indeed have $f_I(z) = \frac{1}{z-I}\frac{1}{z+I}$ for all $z \in L_I \setminus \{I,-I\}$. For any non-real quaternion $p = x+yI \in {{\mathbb{H}}}\setminus {{\mathbb{R}}}$, the function $f : {{\mathbb{H}}}\setminus (x+y{{\mathbb{S}}}) \to {{\mathbb{H}}}$ defined by $$f(q) = (q-p)^{-} = \left[(q-x)^2+y^2\right]^{-1} (q - \bar p)$$ has poles of order $1$ at all points of $x+y{{\mathbb{S}}}$ except $\bar p$, which has order $0$. Indeed, $f_I(z) = \frac{1}{(z- p)(z-\bar p)}(z-\bar p) = \frac{1}{z -p}$ for all $z \in L_I \setminus \{p,\bar p\}$, while for $p' = x+yJ$ with $J \in {{\mathbb{S}}}\setminus \{I, -I\}$ we have $f_J(z) = \frac{1}{(z- p')(z-\bar p')}(z-\bar p)$ for all $z \in L_J \setminus \{p',\bar p'\}$, where $p'-\bar p \neq 0$ and $\bar p' - \bar p \neq 0$. The previous example proves that A pole of order $0$ is not always a removable singularity. This is because $f(q) = (q-p)^{-} $ does not extend as a regular function to a neighborhood of $\bar p$. Indeed any such neighborhood $U$ contains poles of order $1$ for $f$ and we conclude that $\|f\|$ is unbounded on $U$. We now study the distribution of the poles of a generic function which is semiregular on a ball. \[polestructure\] Let $f$ be a semiregular function on $B = B(0,R)$ with $0<R\leq\infty$. Then $f$ extends to a regular function on $B \setminus {{\mathcal{S}}}$ with ${{\mathcal{S}}}$ consisting of isolated 2-spheres (or real singletons) of the form $x+y{{\mathbb{S}}}$. All the poles on each sphere $x+y{{\mathbb{S}}}$ have the same order with the possible exception of one, which must have lesser order. Take $x+y{{\mathbb{S}}}\subseteq B$ and suppose that there exists $I \in {{\mathbb{S}}}$ such that $p = x+yI$ and $\bar p = x-yI$ have orders $m$ and $n$ with $m>0$ or $n>0$. By possibly substituting $-I$ to $I$, we may suppose $m \leq n$. By theorem \[polefactorization\], there exists a semiregular function on $B$ which is regular in a neighborhood $U$ of $x+y{{\mathbb{S}}}$ such that $f(q) = [(q-x)^2+y^2]^{-n} (q-p)^{(n-m)}* g(q)$. Observing that the last expression is regular on $U \setminus (x+y{{\mathbb{S}}})$ proves the first statement of the theorem. We now prove the second statement. If we set $\tilde f(q) = (q-p)^{(n-m)} g(q)$ then $$f(q) = [(q-x)^2+y^2]^{-n} \tilde f(q),$$ $$f_J(z) = [z-(x+yJ)]^{-n} [z-(x-yJ)]^{-n} \tilde f_J(z)$$ for all $J \in {{\mathbb{S}}}$. If $m<n$ then $\tilde f(x+yI) = 0$ and $\tilde f (x+yJ) \neq 0$ for all $J \in {{\mathbb{S}}}\setminus \{I\}$. The previous equation allows us to conclude $ord_f(x+yJ) = n$ for all $J \in {{\mathbb{S}}}\setminus \{I\}$. Since we know by hypothesis that $ord_f(x+yI)= ord_f(p) = m<n$, the thesis holds. If $m = n$ then $\tilde f(x+yI) \neq 0$. If $\tilde f$ does not have zeros in $x +y {{\mathbb{S}}}$ then we conclude $ord_f(x+yJ) = n$ for all $J \in {{\mathbb{S}}}$. If, on the contrary, $\tilde f(x + yK)=0$ for some $K \in {{\mathbb{S}}}$ then we can factor $z-(x+yK)$ out of $\tilde f_K(z)$ and conclude that $ord_f(x + yK)<n$ while $ord_f(x + yJ)=n$ for all $J \in {{\mathbb{S}}}\setminus \{K\}$, as desired. A different approach to multiplicity and order {#multiplicitysection} ============================================== The following peculiar property of the zeros of a quaternionic polynomial is shown in [@zeros]. Recall that $m_f(p)$ denotes the classical multiplicity of $p$ as a zero of a regular function $f$, defined in \[multiplicity\] as the largest $n \in {{\mathbb{N}}}$ such that $f(q) = (q-p)^{n}g(q)$ for some regular $g$. Let $P(q)$ be a regular quaternionic polynomial of degree $d$ which does not have spherical zeros. Then $d \geq \sum_{q \in {{\mathcal{Z}}}_P} m_P(q)$ and the inequality can be strict. If $P$ has a spherical zero $x+y {{\mathbb{S}}}$ then the situation is even more peculiar: clearly $m_P(q) > 0$ for all the (infinite) points $q \in x+y{{\mathbb{S}}}$. In order to overcome these apparent difficulties, an alternative approach has been recently introduced in [@multiplicity]. We may rephrase the definition given in [@multiplicity] as follows. \[newmultiplicity\] Let $P(q)$ be a regular quaternionic polynomial and let $x,y \in {{\mathbb{R}}}$. We say that $P$ has [spherical multiplicity]{.nodecor} $2m$ at $x+y{{\mathbb{S}}}$ if $m$ is the largest natural number such that $P(q) = [(q-x)^2+y^2]^m \tilde P(q)$ for some other polynomial $\tilde P$. If $\tilde P$ has a zero $p_1 \in x+y{{\mathbb{S}}}$ then we say that $P$ has [isolated multiplicity]{.nodecor} $n$ at $p_1$, where $n$ is the largest natural number such that there exist $p_2,...,p_n \in x+y{{\mathbb{S}}}$ and a polynomial $R(q)$ with $\tilde P(q) = (q-p_1)(q-p_2)...(q-p_n)R(q)$. As observed in [@multiplicity], the previous definition yields: If $P(q)$ is a regular quaternionic polynomial of degree $d$, then the sum of the spherical multiplicities and the isolated multiplicities of $P$ is $d$. The classical multiplicity $m_P$ of a polynomial $P$ is related to the spherical and isolated multiplicities of $P$ as follows: \[relationmult\] Let $P(q)$ be a regular quaternionic polynomial and let $p = x+yI \in {{\mathbb{H}}}$. Then $P$ has spherical multiplicity $2 \min \{ m_P(p),m_P(\bar p)\}$ at $x+y{{\mathbb{S}}}$. Moreover, if $m_P(p)>m_P(\bar p)$, then $P$ has isolated multiplicity $n \geq m_P(p)-m_P(\bar p)$ at $p$. Let us give some examples to clarify the previous remark. The polynomial $P(q) = q^2+1$ vanishes on ${{\mathbb{S}}}$. For all $I \in {{\mathbb{S}}}$, $P(q) = (q-I)(q+I)$ has classical multiplicity $m_P(I) =1$ at $I$. Moreover, $P$ has spherical multiplicity $2$ at ${{\mathbb{S}}}$. If $I \in {{\mathbb{S}}}$, then the polynomial $P(q) = (q-I) (q-I) = (q-I)^{2}$ only vanishes at $I$. $P$ has classical multiplicity $m_P(I) =2$ at $I$, it has spherical multiplicity $0$ at ${{\mathbb{S}}}$ and isolated multiplicity $2$ at $I$. If $I,J \in {{\mathbb{S}}}, I\neq J,I \neq -J$, then the polynomial $P(q) = (q-I) (q-J) = q^2 - q (I+J) +IJ$ only vanishes at $I$, where it has classical multiplicity $m_P(I) = 1$ (see [@zeros] for details). Note that this is an example of polynomial having degree 2 greater that the sum $\sum_{q \in {{\mathcal{Z}}}_P} m_P(q) = 1$ of the classical multiplicities of its zeros. According to definition \[newmultiplicity\], $P$ has spherical multiplicity $0$ at ${{\mathbb{S}}}$ and isolated multiplicity $2$ at $I$. We note that it is possible to combine the three cases presented above to build new examples. Fix $I \in {{\mathbb{S}}}$. We can easily construct a quaternionic polynomial $P$ of degree $d$ having classical multiplicities $m_P(I) = M, m_P(-I) = m$ at $I, -I$, spherical multiplicity $2m$ at ${{\mathbb{S}}}$ and isolated multiplicity $n$ at $I$ whenever $d,m,M,n \in {{\mathbb{N}}}$ are such that $m \leq M$ and $M-m \leq n \leq d-m$. Thus remark \[relationmult\] is sharp. Definition \[newmultiplicity\] does not immediately extend to transcendental functions: a priori, there may exist a regular function $f : B \to {{\mathbb{H}}}$ and a 2-sphere $x+y{{\mathbb{S}}}\subseteq B$ such that we can factor out $q-p_j$ for “infinitely many” $p_j \in x+y{{\mathbb{S}}}$. We now prove that this is not the case. It is convenient to take care, at the same time, of the case of an $f$ having poles on $x+y{{\mathbb{S}}}$. \[transfactorization\] Let $f$ be a semiregular function on $B = B(0,R)$, suppose $f \not \equiv 0$ and let $x+y{{\mathbb{S}}}\subseteq B$. There exist $m \in {{\mathbb{Z}}}, n \in {{\mathbb{N}}}$, $p_1,...,p_n \in x+y{{\mathbb{S}}}$ with $p_i \neq \bar p_{i+1}$ for all $i$ such that $$\label{eqpoles} f(q) = [(q-x)^2+y^2]^m (q-p_1)(q-p_2)...(q-p_n)g(q)$$ for some semiregular function $g$ on $B$ which does not have poles nor zeros in $x+y{{\mathbb{S}}}$. As we saw in the proof of theorem \[polefactorization\], if $x+y{{\mathbb{S}}}$ contains poles then there exists a $j> 0$ such that $f(q) = [(q-x)^2+y^2]^{-j} \tilde f(q)$ where $\tilde f$ does not have poles in $x+y{{\mathbb{S}}}$. Hence it is enough to prove the theorem for functions which are regular around $x+y{{\mathbb{S}}}$. This proof requires two steps: 1. If $f$ is regular around $x+y{{\mathbb{S}}}$ and $f \not \equiv 0$ on $B$, then there exists an $m \in {{\mathbb{N}}}$ such that $$f(q) = [(q-x)^2+y^2]^m h(q)$$ for some $h$ which does not vanish identically on $x+y{{\mathbb{S}}}$. Suppose indeed it were possible to find, for all $k \in {{\mathbb{N}}}$, a function $h^{(k)}(q)$ such that $f(q) = [(q-x)^2+y^2]^k h^{(k)}(q)$. Then, choosing an $I \in {{\mathbb{S}}}$, the meromorphic function $f_I$ would have the factorization $f_I(z) = [(z-x)^2+y^2]^k h^{(k)}_I(z) = [z-(x+yI)]^k[z-(x-yI)]^k h^{(k)}_I(z)$ for all $k \in {{\mathbb{N}}}$. This would imply $f_I \equiv 0$ and, by the identity principle \[identity\], $f \equiv 0$. 2. Let $h$ be a semiregular function on $B$ which does not have poles in $x+y{{\mathbb{S}}}$ nor vanishes identically on $x+y{{\mathbb{S}}}$. By theorem \[symmetry\], $g^{(0)}=h$ has at most one zero $p_1 \in x+y{{\mathbb{S}}}$. If this is the case then $h(q) = (q-p_1) * g^{(1)}(q)$ for some function $g^{(1)}$ which does not vanish identically on $x+y{{\mathbb{S}}}$. If for all $k \in {{\mathbb{N}}}$ there existed a $p_{k+1} \in x+y{{\mathbb{S}}}$ and a $g^{(k+1)}$ such that $g^{(k)}(q) = (q-p_{k+1})g^{(k+1)}$ then we would have $$h(q) = (q-p_1)...(q-p_k)g^{(k)}(q)$$for all $k \in {{\mathbb{N}}}$. This would imply, for the symmetrization $h^s$ of $h$, $$h^s(q) = [(q-x)^2+y^2]^k (g^{(k)})^s(q)$$ for all $k \in {{\mathbb{N}}}$. By point (i), this would imply $h^s \equiv 0$. We could then conclude, applying theorems \[conjugatezeros\] and \[structure\], that $h \equiv 0$, a contradiction. Thus there exists an $n \in {{\mathbb{N}}}$ such that $g^{(n)}$ does not have zeroes in $x+y{{\mathbb{S}}}$ and, setting $g = g^{(n)}$, we conclude. The function $f$ in equation (\[eqpoles\]) extends to a regular function in a neighborhood of $x+y{{\mathbb{S}}}$ if and only if $m \geq 0$. In this case, we can extend the definitions of spherical and isolated multiplicity to transcendental functions. If $m<0$, then we can make use of theorem \[transfactorization\] to give an analogous definition for the poles of $f$. Let $f, m, n, p_1$ be as in theorem \[transfactorization\]. If $m\geq0$ then we say that $f$ has [spherical multiplicity]{.nodecor} $2m$ at $x+y{{\mathbb{S}}}$. If, on the contrary, $m<0$ then we say that $f$ has [spherical order]{.nodecor} $-2m$ at $x+y{{\mathbb{S}}}$. In both cases, if $n>0$ then we say that $f$ has [isolated multiplicity]{.nodecor} $n$ at $p_1$. We extend definition \[multiplicity\] to semiregular functions as: Let $f$ be a semiregular function on $B=B(0,R)$ and let $p \in B$. The [(classical) multiplicity]{.nodecor} of $f$ at $p$, denoted $m_f(p)$, is the largest $k \in {{\mathbb{N}}}$ such that $f(q) = (q-p)^{k}g(q)$ for some semiregular function $g$ on $B$ with $ord_g(p) = 0$, if such a $k$ exists. If not, then $m_f(p) = 0$. Remark \[relationmult\] immediately extends to all semiregular functions $f$ on a ball $B(0,R)$ and all points $p$ in the domain where $f$ is regular. Similarly, in the case of poles the spherical order and isolated multiplicity of $f$ are related to the (classical) order $ord_f$ as follows. Let $f$ be a semiregular function on $B(0,R)$ which is not regular at $p = x+yI \in B(0,R)$. Then $f$ has spherical order $2 \max \{ord_f(p),ord_f(\bar p) \}$ at $x+y{{\mathbb{S}}}$. If moreover $ord_f(p)>ord_f(\bar p)$, then $f$ has isolated multiplicity $n \geq ord_f(p)-ord_f(\bar p)$ at $\bar p$. Acknowledgements {#acknowledgements .unnumbered} ================ The author acknowledges support by the Unione Matematica Italiana, the MIT-Italy Program and the Department of Mathematics of the Massachusetts Institute of Technology while writing this paper. [20]{} P. M. Cohn, Skew fields. Theory of general division rings. Encyclopedia of Mathematics and its applications, 57. Cambridge University Press, Cambridge, 1995. 14–18. C. G. Cullen, An integral theorem for analytic intrinsic functions on quaternions. Duke Math. J. [32]{} (1965), 139–148. G. Gentili, C. Stoppato, The open mapping theorem for quaternionic regular functions, E-print. arXiv:0802.3861v1 \[math.CV\] G. Gentili, C. Stoppato, The open mapping theorem for quaternionic regular functions, Preprint. Dipartimento di Matematica “U. Dini”, Università di Firenze, n. 2 (2008). G. Gentili, C. Stoppato, Zeros of regular functions and polynomials of a quaternionic variable, Preprint. Dipartimento di Matematica “U. Dini”, Università di Firenze, n. 1 (2007). G. Gentili, D. C. Struppa, A new approach to Cullen-regular functions of a quaternionic variable. C. R. Math. Acad. Sci. Paris [342]{} (2006), 741–744. G. Gentili, D. C. Struppa, A new theory of regular functions of a quaternionic variable. Adv. Math. [216]{} (2007), 279–301. G. Gentili, D. C. Struppa, On the multiplicity of the zeroes of polynomials with quaternionic coefficients, Preprint. Dipartimento di Matematica “U. Dini”, Università di Firenze, n. 11 (2007). T. Y. Lam, A first course in noncommutative rings. Graduate Texts in Mathematics, 123. Springer-Verlag, New York, 1991. 261–263. T. Y. Lam, Lectures on modules and rings. Graduate Texts in Mathematics, 189. Springer-Verlag, New York, 1999. 299–303. L. H. Rowen, Ring theory. Student edition. Academic press, San Diego, 1991. 272–279. A. Pogorui, M. V. Shapiro, On the structure of the set of zeros of quaternionic polynomials. Complex Variables Theory Appl. [49]{} (2004), no. 6, 379–389. R. Serôdio, L. S. Siu, Zeros of quaternion polynomials. Appl. Math. Letters [14]{} (2001), 237–239.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Spicules have been proposed as significant contributors to the mass and energy balance of the corona. While previous observations have provided a glimpse of short-lived transient brightenings in the corona that are associated with spicules, these observations have been contested and are the subject of a vigorous debate both on the modeling and the observational side. Therefore, it remains unclear whether plasma is heated to coronal temperatures in association with spicules. We use high-resolution observations of the chromosphere and transition region with the Interface Region Imaging Spectrograph ([[IRIS]{}]{}) and of the corona with the Atmospheric Imaging Assembly (AIA) onboard the Solar Dynamics Observatory (SDO) to show evidence of the formation of coronal structures associated with spicular mass ejections and heating of plasma to transition region and coronal temperatures. Our observations suggest that a significant fraction of the highly dynamic loop fan environment associated with plage regions may be the result of the formation of such new coronal strands, a process that previously had been interpreted as the propagation of transient propagating coronal disturbances (PCD)s. Our observations are supported by 2.5D radiative MHD simulations that show heating to coronal temperatures in association with spicules. Our results suggest that heating and strong flows play an important role in maintaining the substructure of loop fans, in addition to the waves that permeate this low coronal environment.' author: - 'B. De Pontieu & I. De Moortel & J. Martinez-Sykora & S. W. McIntosh' title: Observations and numerical models of solar coronal heating associated with spicules --- Introduction {#sec:intro} ============ Chromospheric spicules are dynamic jet-like features that dominate the solar limb and appear to penetrate the million-degree corona before falling back to the surface. Their nature has remained mysterious with many explanations proposed for their origin [for reviews, see @Sterling2000; @Tsiropoula2012]. They have long been considered as a plausible mechanism to provide plasma to the corona [@Beckers1968; @Pneuman1978; @Athay1982]. The discussion about their role in the outer atmosphere was recently revived, with the advent of Hinode [@Kosugi2007], in particular the Solar Optical Telescope [SOT @Tsuneta2008] and the EUV Imaging Spectrometer [EIS @Culhane2007], the Atmospheric Imaging Assembly (AIA) onboard Solar Dynamics Observatory [@Lemen2012], and the Interface Region Imaging Spectrograph [[[IRIS]{}]{} @De-Pontieu2014], as well as advanced 3D radiative MHD simulations [e.g., @Martinez2011; @Martinez2013]. Observations with these spacecraft have provided a new view on spicules, revealing the presence of: 1) relatively slow ($< 40$ km s$^{-1}$), longer-lived (5-10 min) spicules that do not appear to show signficant heating to transition region (TR) or coronal temperatures, falling back towards the surface as chromospheric features [type I spicules, or dynamic fibrils when seen on the disk, @Hansteen2006]; 2) fast (40-100 km s$^{-1}$) (type II) spicules that are only briefly visible (1-2 min) in chromospheric observables such as Ca II H 3968Å [@De-Pontieu2007a; @Pereira2012] and in which a fraction of the plasma appears to be heated to at least TR temperatures [@Pereira2014; @Skogsrud2015; @Rouppe2015] before returning to the surface. We focus here on the impact of type II spicules on the coronal mass and energy balance. There have been several suggestions that these spicules play a significant role in heating the corona. @De-Pontieu2009a suggested that strong upflows of (multi)million-degree plasma seen at the footpoints of coronal loops as spectral line asymmetries [@Hara2008] were associated with upper chromospheric activity [e.g., @McIntosh2009]. @De-Pontieu2011 suggested that the disk counterparts of spicules [rapid blueshifted events or RBEs, see @Rouppe2009] are associated with brightenings in coronal SDO/AIA lines in active regions [for quiet Sun, see @Henriques2016]. These studies have been the subject of significant debate, both from an observational and a theoretical point of view. @Madjarska2011 studied coronal hole spicules at the limb and found no evidence for coronal counterparts using lower resolution and lower signal-to-noise observations from SOHO/SUMER [@Wilhelm1995]. @Klimchuk2012 used simplified theoretical considerations to reject a significant role of spicules in the coronal heating issue, while @Tripathi2013 and @Patsourakos2014 studied spectral line asymmetries from EIS and argued that while spicules may play a role in the coronal mass and energy balance, it is likely not a dominant one. Lacking a theoretical model that captures the complexity of the spicular environment, these studies are based on simplifying assumptions about the physical scenario focusing on single-field-line approaches that underestimate the complexity of the spicular environment. The latter is observationally challenging to capture: they rapidly fade in and out of various passbands, necessitating a multi-instrument approach, and are so dynamic and finely structured that many instruments do not resolve their spatio-temporal evolution. Here we attempt to address both issues. We exploit the [[IRIS]{}]{} discovery of the TR counterparts of spicules, in particular the fact that these are longer-lived than the chromospheric spicules, allowing us to more easily track their evolution. We use SDO/AIA to focus on the impact of spicules on coronal loops that are connected to plage or enhanced network regions. These loops have long been known to be permeated by propagating coronal disturbances (PCDs): rapid ($\sim 100$ km s$^{-1}$) intensity disturbances whose exact nature remains unknown. While it is clear that PCDs that originate from sunspot umbrae are caused by sound waves, it is less clear whether waves or flows cause these PCDs in plage regions, with both receiving observational and theoretical support [e.g. @De-Moortel2002a; @De-Moortel2002b; @De-Pontieu2005; @De-Moortel2009b; @De-Pontieu2010; @Verwichte2010; @De-Moortel2012; @Ofman2012; @Tian2012; @Wang2013; @Petralia2014; @De-Moortel2015; @Samanta2015; @Bryans2016]. We also take advantage of recent developments in spicule modeling [@Martinez2016; @Martinez2017] that appear to show coronal heating associated with spicules. Observations {#sec:obs} ============ We use [[IRIS]{}]{} slit-jaw observations of AR 12171 at x$_{cen}$=464, y$_{cen}$=-476, taken on 26-Sep-2014 from 00:34-01:37 UTC using OBS-ID 3820107266. The [[IRIS]{}]{} level 2 data was corrected for dark current, flat-field, geometry and co-aligned as described in @De-Pontieu2014. To boost signal-to-noise, the [[IRIS]{}]{} data was summed onboard 2x2, so that spatial pixels are 0.33x 0.33. Both 1330Å and 1400Å passbands were used which are dominated by far-ultraviolet continuum (formed in the low chromosphere) and, respectively, 1335/1336Å lines [formed at upper chromospheric and low TR temperatures, from 15,000-40,000 K @Rathore2015] and 1394/1403Å lines [formed at TR temperatures from 20,000-300,000 K @Olluri2015]. Co-temporal [[SDO]{}]{}/AIA observations in the 1600Å, 171Å and 193Åpassbands were prepped, coaligned, and normalized using the SolarSoft [aia\_prep]{} routine. The AIA data were interpolated in time and space to match the [[IRIS]{}]{}/SJI temporal cadence (10.4 s) and spatial resolution (0.33). We focus on a decayed plage region that is associated with “plume”-like coronal structures that emanate towards the south (Fig. \[f1\]). Results {#sec:res} ======= Observations {#observations} ------------ Movies of the [[IRIS]{}]{} 1400Å passband show that the footpoints of these coronal loops are dominated by a multitude of spicule-like features, which, given their appearance, are most likely caused by TR emission (Si$^{3+}$ ions). These spicules originate from magnetic flux concentrations, shooting away from the weak plage at apparent speeds of 50-200 km s$^{-1}$. One can often see the spicules retract after reaching a maximum extent, although not always. They are often not clearly visible along their whole length and fade as they extend away from the plage, suggesting a complex thermal environment and evolution, involving heating and/or cooling. The coronal loops rooted in the same region similarly show a lot of complexity, which traditionally has been associated with “propagating coronal disturbances” along pre-existing coronal loops. However, closer inspection of the timeseries associated with Fig. \[f1\] shows that much of this activity is actually caused by a variety of coronal strands appearing and disappearing, typically starting from the bottom of the loops, followed by “propagation” away from the footpoint. It is also clear that the PCDs that are observed here are not as cleanly periodic as those associated with sunspot umbrae [see also @De-Pontieu2010]). This is illustrated in Figs. \[f2\], \[f3\] and accompanying movies. ![image](f1.eps){width="0.95\hsize"} The line-of-sight superposition and the multitude of events overlapping in space and time means it is not straightforward to disentangle individual events. However, analysis of the evolution of several of the larger spicules that stand out individually show an intriguing connection between the and spicules in [[IRIS]{}]{} slit-jaw movies and the formation of strands in the coronal loop system (AIA). This is illustrated in Fig. \[f2\] and the accompanying movies which show the temporal evolution of the event shown in Fig. \[f2\] as well as a second event. In both cases we show the images (top row) and the running difference of the 171ÅAIA channel – calculated by differencing the current image with that taken 62 s earlier, a commonly used method to enhance the visibility of the intensity disturbances. For both cases, we see the spicule form with apparent velocities in the plane-of-the-sky of $\sim 50$ km s$^{-1}$, accompanied by a brightening in 171Å that initially grows with the same apparent speed as the spicule. Towards the time of maximum extension of the spicule ($t=2396$ s), the spicule seems to stay roughly constant in length and we see (red rectangles in Fig. \[f2\] and movies) that the coronal counterpart grows rapidly to cover 30in 50 s, suggesting an apparent speed of order 400 km s$^{-1}$. A similar evolution can be seen in the second movie accompanying Fig. \[f2\]: that spicule grows with an apparent speed of order 40 km s$^{-1}$ until $t\sim1980~s$, appears to stay at its maximum extent for a while, and then the coronal counterpart grows another 30 in $\sim40$ s, suggesting an apparent speed for this phase of $\sim 500$ km s$^{-1}$. ![image](f2.eps){width="0.9\hsize"} Figure \[f3\] shows how the coronal loop strand that is associated with the spicule of Fig. \[f2\] becomes visible in both AIA 171Å and 193Å passbands when using a different color table that accentuates small intensity differences. The loop strand is more clearly visible in 171Å than in 193Å suggesting that it reaches temperatures closer to the formation temperature of (log $T = 5.9$) rather than that of (log $T = 6.2$). This figure does not show the running difference, but rather the original AIA intensity. It illustrates how this particular feature is not a disturbance on top of a pre-existing coronal loop structure, but the formation of a completely new coronal strand. Detailed inspection of the AIA timeseries shows that the formation of such strands is not a rare event, but a common occurrence throughout this loop fan structure. ![image](f3a.eps){width="0.9\hsize"}\ ![image](f3b.eps){width="0.9\hsize"} However, the formation of such a strand can also, deceptively, look like a propagating coronal disturbance (PCD) as illustrated by Fig. \[f4\]. The example in the top row shows the event that was highlighted in Figs. \[f2\] and \[f3\]. As shown in Fig. \[f3\], this event involved the formation of a new loop strand by $t=2200$ s, which briefly appears as a propagating coronal disturbance in both 171 and 193Å channels. Because of the nature of the running difference, the longevity of the strand is not clear at all from panels (A) and (B). This becomes much clearer in panels (C) and (D) which do not show running difference and instead of the original 171 and 193Å intensity, the intensity after unsharp masking to enhance the small spatial scales of the coronal loop strands. We show here the spatio-temporal evolution for a cut (see upper horizontal line in Fig. \[f1\]) across the loop strand that we highlighted earlier. We see that this particular strand forms when the PCD hits this location. While the PCD continues to propagate away from this location, and the running difference plot suggests that a “wave” (perturbation) just passed through a background structure, panels (C) and (D) indicate that the strand continues to exist long after the PCD has left this region. This particular strand in fact undergoes repeated activity with many PCDs passing through, and the strand being strengthened every time such a PCD passes by. A more isolated case of strand formation associated with the passage of a PCD is shown in the bottom of Fig. \[f4\] which reveals a short-lived passage of a PCD (at $t=1700$ s) that leads to the formation of a strand, initially brightest in 211 Å (not shown), followed by a brief event in 193Å (panel H) and then a prolonged presence in 171Å. This event occurs where the right hand track and bottom horizontal line cross in Fig. \[f1\]. This sequence of events strongly supports a scenario in which the PCD appears to be associated not only with a spicule, but also with heating of plasma to $\sim 1.5$ MK followed by apparently relatively rapid cooling and subsequent fading from the 171Åpassband after 10-15 minutes. Many more examples can be found in the data that support this scenario of several phenomena associated with spicules: triggering of PCDs, heating to coronal temperatures, and the formation of loop strands. ![image](f4a.eps){width="0.95\hsize"} ![image](f4b.eps){width="0.95\hsize"} Simulations ----------- These results fit well with synthetic 171Åand 193Å observations from a 2.5D radiative MHD simulation using Bifrost [@Gudiksen2011], which captures many physical processes important for the dynamics and energetics of the solar atmosphere. This simulation covers a domain from the top of the convection zone into the corona, including self-consistent chromospheric and coronal heating [@Martinez2016; @Martinez2017]. We included the effects of interactions between ions and neutrals, or ambipolar diffusion. Ambipolar diffusion plays a key role in the formation of features that closely resemble type II spicules, through a complex mechanism outlined in @Martinez2017. In summary, the interaction between weak, granular-scale fields and strong flux concentrations leads to strong magnetic tension, which can emerge into the chromosphere through ambipolar diffusion and leads to a violent release of tension when the low plasma $\beta$ regime is reached in the middle to upper chromosphere. This violent release leads to strong upward acceleration of plasma, the formation of fast spicules and the generation of strong transverse waves. In addition, currents, created through several mechanisms including wave-mode coupling, gradients in ambipolar diffusion and the interaction between emerging flux and pre-existing ambient field, are in part dissipated by ambipolar diffusion in the spicule (leading to heating to TR temperatures), and in part propagated into the corona at Alfvén speeds where they lead to significant heating. The coronal impact of this scenario is illustrated in Figure \[f5\], which includes running difference of and intensities along a track that covers a spicule. While the response is relatively simple with the apparent propagation of a “PCD” at speeds of $\sim150$ km s$^{-1}$, the running difference shows two PCDs, one that is similar in slope as the , and another at much higher speeds ($\sim1600$ km s$^{-1}$). Comparison with the original intensity of both lines and various physical variables paints a complicated picture and indicates that the interpretation of PCDs in terms of physical mechanisms is not straightforward, as explained below. Both PCDs are causally linked to the launch of a fast spicule. The chromospheric part of this spicule is visible as a parabolic path in the temperature (e) which starts at the same time and location as the PCD. The cause of the PCD is, at low heights, a mix of flows associated with the TR and coronal counterparts of the spicule (visible as a parabolic path that reaches distances of 10 Mm in the original intensity of (a) and in the density (f)), and the coronal remnant of the shock wave that was involved in the spicule acceleration. These flows are caused by the acceleration, compression and heating of plasma associated with the spicule eruption. At greater heights ($>10$ Mm), the PCD is mostly determined by the shock-wave related disturbance [similar to @De-Pontieu2005; @Petralia2014] since the coronal counterpart of the spicule fades beyond distances of 12 Mm. The PCDs are even more complex: the slower PCD is a mix of spicular flows and shock-wave related disturbances. However, the average slope that is drawn through the slow PCD in this wavelength ignores the fact that the different physical mechanisms lead to different slopes in the space-time plots, with the apparent speed increasing with distance from the spicule footpoints. The running difference (b) also reveals a faster “PCD” that is caused by the rapid formation of a loop strand. This loop strand is formed because of the heating associated with the arrival and dissipation in the corona of the current that also heats the spicular plasma, as well as thermal conduction that spreads the released heat. This faster PCD at Alfvénic speeds could well be responsible for the rapid propagation (400 km s$^{-1}$) we see in our observations. The slope of the PCD with modest speed can also be affected by the heating from the currents (g). The movie accompanying Fig. \[f5\] shows how the currents that are created during the spicule formation propagate rapidly into the corona (at Alfvénic speeds) and appear to “meander” through the coronal volume, similar to what is seen in observations of coronal loop strands that often appear to “move” perpendicularly to their own axis. In our simulation, the spicules are thus the harbingers of significant coronal heating both through heating of spicular plasma and heating from current dissipation associated with the spicule. The current dissipation in the corona of our model occurs because of numerical resistivity [@Gudiksen2011]. In the solar atmosphere the dissipation of this energy could be because of current dissipation on small scales or dissipation of the Alfvén waves that are triggered when the spicule is formed [@Martinez2017]. ![image](f5.eps){width="0.85\hsize"} Discussion {#sec:dis} ========== Our results support a scenario in which “PCDs” along loops originating from plage or strong network regions are not necessarily only a signature of magneto-acoustic waves, but often caused by a complex sequence of events that involves generation of spicular flows and associated shock waves that propagate into the corona, as well as plasma heating through dissipation of electrical currents and magnetic waves. These currents are a key component of the spicule formation which critically depends on ambipolar diffusion caused by the interaction between ions and neutrals. Our observations provide a detailed view of how spicules, heated from chromospheric to TR temperatures, set off PCDs, but also lead to the formation of new coronal loop strands, thus locally contributing to the mass and energy balance of the corona. Our results suggest that analysis of PCDs through running differencing misses the fact that plage-related loop strands are continuously formed and persist after PCDs have “passed”. Our results provide a natural explanation for the often confusing reports of apparent speeds, which in our simulations are caused by a mixture of real mass motions of coronal plasma in response to spicular flows and heating, remnants of shock waves generated during the spicule formation, heating through spicule-associated currents, and subsequent thermal conduction. Our simulations also show that idealized “single-field line” approaches [e.g. @Klimchuk2012] to spicule-associated coronal heating are bound to fail: the spicular environment is highly complex, takes place on many neighboring “field lines” some of which carry accelerated plasma that is heated by ambipolar dissipation of electrical currents, and others carry plasma that is heated by spicule-associated currents that rapidly propagate into the corona. The complexity of physical mechanisms in these simulations highlights why it is so difficult to determine the “spicule contribution to coronal heating”. Such an endeavor is driven by an approach that is based on observational phenomena (“chromospheric” spicules) which our simulated (and likely solar) reality defies: spicules are neither chromospheric, TR or coronal phenomena; they are all of the above, and their dynamics and energetics are intimately tied to that of the corona. Our results indicate that the currents and waves associated with spicule formation should not be ignored in future studies of coronal heating. [[[IRIS]{}]{} is a NASA small explorer developed and operated by LMSAL with major contributions to downlink communications by ESA and Norwegian Space Centre. This work is supported by NASA (NNG09FA40C, [[IRIS]{}]{}) and the UK Science and Technology Facilities Council and EU Horizon 2020 research programme (grant No. 647214). The simulations were run on Pleiades (project s1061).]{} [44]{} natexlab\#1[\#1]{} , R. G. & [Holzer]{}, T. E. 1982, , 255, 743 , J. M. 1968, , 3, 367 , P., [McIntosh]{}, S. W., [De Moortel]{}, I., & [De Pontieu]{}, B. 2016, , 829, L18 , J. L., [Harra]{}, L. K., [James]{}, A. M., [Al-Janabi]{}, K., [Bradley]{}, L. J., [Chaudry]{}, R. A., [Rees]{}, K., [Tandy]{}, J. A., [Thomas]{}, P., [Whillock]{}, M. C. R., [Winter]{}, B., [Doschek]{}, G. A., [Korendyke]{}, C. M., [Brown]{}, C. M., [Myers]{}, S., [Mariska]{}, J., [Seely]{}, J., [Lang]{}, J., [Kent]{}, B. J., [Shaughnessy]{}, B. M., [Young]{}, P. R., [Simnett]{}, G. M., [Castelli]{}, C. M., [Mahmoud]{}, S., [Mapson-Menard]{}, H., [Probyn]{}, B. J., [Thomas]{}, R. J., [Davila]{}, J., [Dere]{}, K., [Windt]{}, D., [Shea]{}, J., [Hagood]{}, R., [Moye]{}, R., [Hara]{}, H., [Watanabe]{}, T., [Matsuzaki]{}, K., [Kosugi]{}, T., [Hansteen]{}, V., & [Wikstol]{}, [Ø]{}. 2007, , 243, 19 , I. 2009, Space Science Reviews, 38 , I., [Antolin]{}, P., & [Van Doorsselaere]{}, T. 2015, , 290, 399 , I., [Hood]{}, A. W., [Ireland]{}, J., & [Walsh]{}, R. W. 2002, , 209, 89 , I., [Ireland]{}, J., [Walsh]{}, R. W., & [Hood]{}, A. W. 2002, , 209, 61 , I. & [Nakariakov]{}, V. M. 2012, Philosophical Transactions of the Royal Society of London Series A, 370, 3193 , B., [Erd[é]{}lyi]{}, R., & [De Moortel]{}, I. 2005, , 624, L61 , B., [McIntosh]{}, S., [Hansteen]{}, V. H., [Carlsson]{}, M., [Schrijver]{}, C. J., [Tarbell]{}, T. D., [Title]{}, A. M., [Shine]{}, R. A., [Suematsu]{}, Y., [Tsuneta]{}, S., [Katsukawa]{}, Y., [Ichimoto]{}, K., [Shimizu]{}, T., & [Nagata]{}, S. 2007, , 59, 655 De Pontieu, B. & McIntosh, S. W. 2010, The Astrophysical Journal, 722, 1013 , B., [McIntosh]{}, S. W., [Carlsson]{}, M., [Hansteen]{}, V. H., [Tarbell]{}, T. D., [Boerner]{}, P., [Martinez-Sykora]{}, J., [Schrijver]{}, C. J., & [Title]{}, A. M. 2011, Science, 331, 55 , B., [McIntosh]{}, S. W., [Hansteen]{}, V. H., & [Schrijver]{}, C. J. 2009, , 701, L1 , B., [Title]{}, A. M., [Lemen]{}, J. R., [Kushner]{}, G. D., [Akin]{}, D. J., [Allard]{}, B., [Berger]{}, T., [Boerner]{}, P., [Cheung]{}, M., [Chou]{}, C., [Drake]{}, J. F., [Duncan]{}, D. W., [Freeland]{}, S., [Heyman]{}, G. F., [Hoffman]{}, C., [Hurlburt]{}, N. E., [Lindgren]{}, R. W., [Mathur]{}, D., [Rehse]{}, R., [Sabolish]{}, D., [Seguin]{}, R., [Schrijver]{}, C. J., [Tarbell]{}, T. D., [W[ü]{}lser]{}, J.-P., [Wolfson]{}, C. J., [Yanari]{}, C., [Mudge]{}, J., [Nguyen-Phuc]{}, N., [Timmons]{}, R., [van Bezooijen]{}, R., [Weingrod]{}, I., [Brookner]{}, R., [Butcher]{}, G., [Dougherty]{}, B., [Eder]{}, J., [Knagenhjelm]{}, V., [Larsen]{}, S., [Mansir]{}, D., [Phan]{}, L., [Boyle]{}, P., [Cheimets]{}, P. N., [DeLuca]{}, E. E., [Golub]{}, L., [Gates]{}, R., [Hertz]{}, E., [McKillop]{}, S., [Park]{}, S., [Perry]{}, T., [Podgorski]{}, W. A., [Reeves]{}, K., [Saar]{}, S., [Testa]{}, P., [Tian]{}, H., [Weber]{}, M., [Dunn]{}, C., [Eccles]{}, S., [Jaeggli]{}, S. A., [Kankelborg]{}, C. C., [Mashburn]{}, K., [Pust]{}, N., [Springer]{}, L., [Carvalho]{}, R., [Kleint]{}, L., [Marmie]{}, J., [Mazmanian]{}, E., [Pereira]{}, T. M. D., [Sawyer]{}, S., [Strong]{}, J., [Worden]{}, S. P., [Carlsson]{}, M., [Hansteen]{}, V. H., [Leenaarts]{}, J., [Wiesmann]{}, M., [Aloise]{}, J., [Chu]{}, K.-C., [Bush]{}, R. I., [Scherrer]{}, P. H., [Brekke]{}, P., [Martinez-Sykora]{}, J., [Lites]{}, B. W., [McIntosh]{}, S. W., [Uitenbroek]{}, H., [Okamoto]{}, T. J., [Gummin]{}, M. A., [Auker]{}, G., [Jerram]{}, P., [Pool]{}, P., & [Waltham]{}, N. 2014, , 289, 2733 , B. V., [Carlsson]{}, M., [Hansteen]{}, V. H., [Hayek]{}, W., [Leenaarts]{}, J., & [Mart[í]{}nez-Sykora]{}, J. 2011, , 531, A154 , V. H., [De Pontieu]{}, B., [Rouppe van der Voort]{}, L., [van Noort]{}, M., & [Carlsson]{}, M. 2006, , 647, L73 , H., [Watanabe]{}, T., [Harra]{}, L. K., [Culhane]{}, J. L., [Young]{}, P. R., [Mariska]{}, J. T., & [Doschek]{}, G. A. 2008, , 678, L67 , V. M. J., [Kuridze]{}, D., [Mathioudakis]{}, M., & [Keenan]{}, F. P. 2016, , 820, 124 , J. A. 2012, Journal of Geophysical Research (Space Physics), 117, A12102 , T., [Matsuzaki]{}, K., [Sakao]{}, T., [Shimizu]{}, T., [Sone]{}, Y., [Tachikawa]{}, S., [Hashimoto]{}, T., [Minesugi]{}, K., [Ohnishi]{}, A., [Yamada]{}, T., [Tsuneta]{}, S., [Hara]{}, H., [Ichimoto]{}, K., [Suematsu]{}, Y., [Shimojo]{}, M., [Watanabe]{}, T., [Shimada]{}, S., [Davis]{}, J. M., [Hill]{}, L. D., [Owens]{}, J. K., [Title]{}, A. M., [Culhane]{}, J. L., [Harra]{}, L. K., [Doschek]{}, G. A., & [Golub]{}, L. 2007, , 243, 3 , J. R., [Title]{}, A. M., [Akin]{}, D. J., [Boerner]{}, P. F., [Chou]{}, C., [Drake]{}, J. F., [Duncan]{}, D. W., [Edwards]{}, C. G., [Friedlaender]{}, F. M., [Heyman]{}, G. F., [Hurlburt]{}, N. E., [Katz]{}, N. L., [Kushner]{}, G. D., [Levay]{}, M., [Lindgren]{}, R. W., [Mathur]{}, D. P., [McFeaters]{}, E. L., [Mitchell]{}, S., [Rehse]{}, R. A., [Schrijver]{}, C. J., [Springer]{}, L. A., [Stern]{}, R. A., [Tarbell]{}, T. D., [Wuelser]{}, J.-P., [Wolfson]{}, C. J., [Yanari]{}, C., [Bookbinder]{}, J. A., [Cheimets]{}, P. N., [Caldwell]{}, D., [Deluca]{}, E. E., [Gates]{}, R., [Golub]{}, L., [Park]{}, S., [Podgorski]{}, W. A., [Bush]{}, R. I., [Scherrer]{}, P. H., [Gummin]{}, M. A., [Smith]{}, P., [Auker]{}, G., [Jerram]{}, P., [Pool]{}, P., [Soufli]{}, R., [Windt]{}, D. L., [Beardsley]{}, S., [Clapp]{}, M., [Lang]{}, J., & [Waltham]{}, N. 2012, , 275, 17 , M. S., [Vanninathan]{}, K., & [Doyle]{}, J. G. 2011, , 532, L1 , J., [De Pontieu]{}, B., & [Hansteen]{}, V. 2012, , 753, 161 , J., [Hansteen]{}, V., & [Moreno-Insertis]{}, F. 2011, , 736, 9 Mart[í]{}nez-Sykora, J., De Pontieu, B., Leenaarts, J., et al. 2013, , 771, 66 Mart[í]{}nez-Sykora, J., De Pontieu, B., Carlsson, M., & Hansteen, V. 2016, , 831, L1 Mart[í]{}nez-Sykora, J., De Pontieu, B., Hansteen, V., Rouppe van der Voort, L., Carlsson, M., & Pereira, T. 2017, accepted for publication in Science, 10.1126science.aah5412 , S. W. and [De Pontieu]{}, B., 2009, , 706, L80 , L., [Wang]{}, T. J., & [Davila]{}, J. M. 2012, , 754, 111 , K., [Gudiksen]{}, B. V., [Hansteen]{}, V. H., & [De Pontieu]{}, B. 2015, , 802, 5 , S., [Klimchuk]{}, J. A., & [Young]{}, P. R. 2014, , 781, 58 , T. M. D., [De Pontieu]{}, B., & [Carlsson]{}, M. 2012, , 759, 18 , T. M. D., [De Pontieu]{}, B., [Carlsson]{}, M., [Hansteen]{}, V., [Tarbell]{}, T. D., [Lemen]{}, J., [Title]{}, A., [Boerner]{}, P., [Hurlburt]{}, N., [W[ü]{}lser]{}, J. P., [Mart[í]{}nez-Sykora]{}, J., [Kleint]{}, L., [Golub]{}, L., [McKillop]{}, S., [Reeves]{}, K. K., [Saar]{}, S., [Testa]{}, P., [Tian]{}, H., [Jaeggli]{}, S., & [Kankelborg]{}, C. 2014, , 792, L15 Petralia, A., Reale, F., Orlando, S., & Klimchuk, J. A. 2014, , 567, A70 , G. W. & [Kopp]{}, R. A. 1978, , 57, 49 , B., [Carlsson]{}, M., [Leenaarts]{}, J., & [De Pontieu]{}, B. 2015, , 811, 81 , L., [De Pontieu]{}, B., [Pereira]{}, T. M. D., [Carlsson]{}, M., & [Hansteen]{}, V. 2015, , 799, L3 , L., [Leenaarts]{}, J., [de Pontieu]{}, B., [Carlsson]{}, M., & [Vissers]{}, G. 2009, , 705, 272 , T., [Pant]{}, V., & [Banerjee]{}, D. 2015, , 815, L16 , H., [Rouppe van der Voort]{}, L., [De Pontieu]{}, B., & [Pereira]{}, T. M. D. 2015, , 806, 170 , A. C. 2000, , 196, 79 , H., [McIntosh]{}, S. W., [Wang]{}, T., [Ofman]{}, L., [De Pontieu]{}, B., [Innes]{}, D. E., & [Peter]{}, H. 2012, , 759, 144 , D. & [Klimchuk]{}, J. A. 2013, , 779, 1 , G., [Tziotziou]{}, K., [Kontogiannis]{}, I., [Madjarska]{}, M. S., [Doyle]{}, J. G., & [Suematsu]{}, Y. 2012, , 169, 181 , S., [Ichimoto]{}, K., [Katsukawa]{}, Y., [Nagata]{}, S., [Otsubo]{}, M., [Shimizu]{}, T., [Suematsu]{}, Y., [Nakagiri]{}, M., [Noguchi]{}, M., [Tarbell]{}, T., [Title]{}, A., [Shine]{}, R., [Rosenberg]{}, W., [Hoffmann]{}, C., [Jurcevich]{}, B., [Kushner]{}, G., [Levay]{}, M., [Lites]{}, B., [Elmore]{}, D., [Matsushita]{}, T., [Kawaguchi]{}, N., [Saito]{}, H., [Mikami]{}, I., [Hill]{}, L. D., & [Owens]{}, J. K. 2008, , 249, 167 , E., [Marsh]{}, M., [Foullon]{}, C., [Van Doorsselaere]{}, T., [De Moortel]{}, I., [Hood]{}, A. W., & [Nakariakov]{}, V. M. 2010, , 724, L194 , T., [Ofman]{}, L., & [Davila]{}, J. M. 2013, , 775, L23 , K., [Curdt]{}, W., [Marsch]{}, E., [Sch[ü]{}hle]{}, U., [Lemaire]{}, P., [Gabriel]{}, A., [Vial]{}, J.-C., [Grewing]{}, M., [Huber]{}, M. C. E., [Jordan]{}, S. D., [Poland]{}, A. I., [Thomas]{}, R. J., [K[ü]{}hne]{}, M., [Timothy]{}, J. G., [Hassler]{}, D. M., & [Siegmund]{}, O. H. W. 1995, , 162, 189	{ "pile_set_name": "ArXiv" }
--- abstract: 'We show how the partition function of a network of parallel superconducting wires weakly coupled together by the proximity effect, subjected a vector potential along the wires can be mapped onto N-distinguishable two dimensional quantum-mechanics problem with a perpendicular imaginary magnetic field. Then, we show, using a mean field approximation, that, for a given coupling, there is a critical temperature for onset of inter-wire phase coherence. The transition temperature $T_c$ is plotted on both cases for non-magnetic and a magnetic field perpendicular to the wires.' author: - Kohjiro Kobayashi and David Stroud title: Theory of Fluctuations in a Network of Parallel Superconducting Wires --- Introduction ============ There has been considerable recent interest in thin wires that undergo transitions into an ordered state, such as superconducting or ferromagnetic. For example, a recent experiment [@Tang2001] has suggested that single-walled carbon nanotubes (which have diameters of only about 4 Å) are superconducting up to temperatures as high as 20 K. Because these tubes are so thin, they behave very much like one-dimensional superconductors. It was therefore proposed [@Tang2001] that they could be described by a complex order parameter $\psi(z)$ which varies only in one dimension, say the z direction, i.e. along the tube. $\psi(z)$ might represent the complex energy gap, or, in a different normalization, it could represent the condensate wave function in a BCS superconductor. Moreover, there have been many experiments for investigating superconductivity on nanowires. Ropes of carbon nanotubes between superconducting electrodes can show superconductivity due to the proximity effect of the electrodes [@Kasumov1999; @Morpurgo1999; @Gonzalez2001]. Furthermore, superconductivity on carbon nanowires connected to normal contacts, has been observed [@Kociak2001; @Kasumov2003]. On the other hand, superconductivity of nanowires of Zn or Sn has been investigated [@Tian2005a; @Tian2005b]. Fluctuations are, of course, especially important in one-dimensional systems. It was shown many years ago by Scalapino [et al]{} [@Scalapino1972] that [classical]{} fluctuations in one dimension could be treated [exactly]{}, within the context of a Ginzburg-Landau (GL) free energy functional. Their treatment involved mapping the GL functional onto a single-particle quantum mechanics problem, using an exact connection between the classical partition function and a path integral treatment of the quantum mechanics problem. These authors showed that classical fluctuations could give rise to a non-zero order parameter even above the GL transition temperature. This mapping was extended to treat Josephson-coupled thin wires [@Stoeckly1975; @Scalapino1975]. However, in the mapping, the effect of a magnetic field was ignored. In the case of a non-zero perpendicular magnetic field, we show that the partition function for the wires maps onto a certain zero-temperature quantum mechanics problem in two dimensions with an effective imaginary perpendicular magnetic field, which brings to a non-Hermitian quantum mechanics problem. The non-Hermitian problem in physics has not been new recently. Nonequilibrium processes can be described by non-Hermitian Liouville operators [@Kadanoff1968; @Fogedby1995; @Kim1995]. The non-Hermitian quantum mechanics are well studied in order to study the pinning of magnetic flux lines in high temperature superconductors [@Nelson1993; @Hatano1996; @Hatano1997]. The remainder of this paper is organized as follows. In Section \[sec:one\_formalism\], we describe our formalism and mapping. In Section \[sec:one\_result\], we give our numerical results including phase diagrams. This is followed by a concluding discussion and an outline of possible future research. Formalism \[sec:one\_formalism\] ================================ Mapping to a quantum mechanics problem for interacting superconducting wires when $ \vec{B} $ is perpendicular to the wires --------------------------------------------------------------------------------------------------------------------------- Let us consider a network of N parallel superconducting wires in a non-zero vector potential. We assume, for convenience, that these wires all have the same GL parameters, though the formalism can easily be generalized to the case when the parameters are different. Then the partition function can be written as a functional integral over the $N$ complex order parameters $\psi_1(z_1),...\psi_N(z_N)$: $$Z = \int{\cal D}\psi_1(z_1)...{\cal D}\psi_N(z_N)\exp\{-\beta F[\psi_1(z_1),...\psi_N(z_N)]\}.$$ We assume that the free energy functional is the sum of two parts: a single-wire term $F_s$ and a term describing inter-wire interactions, which we denote $F_{int}$. The single-wire term will just be the sum of general GL equation for each wire: $$F_s = \sum_{i=1}^NF_{GL}[\psi_i(z_i)].$$ Here, $$F_{GL}[\psi_i(z_i)] = \int_0^{z_{max}}[\frac{1}{2m^}\|\left(\frac{\hbar}{i}\nabla-\frac{e^\vec{A} }{c} \right)\psi (z)\|^2+\alpha \|\psi(z)\|^2+\gamma \|\psi(z)\|^4+\frac{HB}{8\pi}\Sigma]dz,$$ where $\alpha$, $\gamma$, and $m^$ are material-dependent (and possibly temperature-dependent) coefficients. Commonly, it is assumed that $\gamma$ is positive and that $\alpha = \alpha^\prime(T - T_c)$, where $T$ is the temperature, $T_c$ is the critical temperature, and $\alpha^\prime$ is greater than zero. Also, $\Sigma$ is the cross-sectional area of the sample, but for one-dimensional wire we may ignore this term. For the interaction term, we assume a form similar to that used by Lawrence and Doniach for interacting superconducting layers [@Lawrence], namely $$F_{int} = \sum_{\langle ij \rangle}\int_0^{z_{max}}K_{ij}\|\psi_i(z) - \psi_j(z)\|^2.$$ where $z_{max}$ is the length of the wires. Basically, we are assuming that there is a Josephson coupling of strength $K_{ij}$ between different wires, but at the same point along the length, $z$. We choose a gauge such that the vector potential is parallel to the superconducting wires, has only $z$ component and independent of $z$. When a wire is a loop, a vector potential is related to the total flux $\Phi$ through the loop, $A_z = \Phi /z_{max}$. In this case, using $\psi_i(z)=\psi_{iR}(z)+i\psi_{iI}(z)$, $F_s$ and $F_{int}$ take the forms $$F_s \!=\! \sum_i\!\!\int_0^{z_{max}}\!\!\!\!\![\frac{\hbar^2}{2m^}\|\psi_i'\|^2 -\frac{e^\hbar A_z}{m^c}(\psi_{iR}\psi_{iI}'-\psi_{iR}'\psi_{iI}) +\left\{\!\alpha+\frac{1}{2m^}\left(\frac{e^}{c}\right)^2A_z^2\!\right\}\!\!\|\psi_i\| ^2 + \gamma \|\psi_i\|^4]dz, \label{eq:sheng1}$$ and $$F_{int} = \sum_{\langle ij \rangle}\int_0^{z_{max}}K_{ij} (\|\psi_i(z)\|^2+\|\psi_j(z)\|^2-2(\psi_{iR}\psi_{jR}+\psi_{iI}\psi_{jI})), \label{eq:sheng2}$$ where $\psi '(z)=d\psi(z)/dz$. Finally, the partition function takes the form $$Z = \int\sum_{i}{\cal D}\psi_{iR}{\cal D}\psi_{iI}\exp(-\beta F[\psi_{iR},\psi_{iI} ]), \label{eq:sheng3}$$ where we use $\psi_{i}=\psi_{iR}+i\psi_{iI}$. We now show that eqs. (\[eq:sheng1\]), (\[eq:sheng2\]) and (\[eq:sheng3\]) for $Z$ are actually equivalent to a [quantum mechanical]{} problem of a N distinguishable particles in N distinct quantum wells in two dimensions in the presence of a perpendicular magnetic field. In order to simplify our argument, we consider the case of single particle with mass, $m$ and a charge $e^$ subjected to a 2D potential, $V(x,y)$. The density matrix of a two-dimensional system, using $\psi^I$ and $\psi^F$ are boundary condition at initial and final time, can be written [@Feynman1965] $$\langle \psi^F\|e^{-S/\hbar}\|\psi^I\rangle = \int_{\psi^I}^{\psi^F}{\cal D}x(\tau){\cal D}y(\tau)\exp{\{-\frac{1}{\hbar}S[x(\tau),y(\tau)]\}},$$ where $$S = \int_{0}^{\beta_{eff}\hbar}[\frac{m}{2}(x'^2+y'^2)+V(x,y)-i\frac{e^}{c}\vec {A}\cdot\vec{v}]d\tau .$$ For the given $\vec{B}=B_{eff}\hat{z}$ with the gage $$\vec{A}_{eff} = \frac{B_{eff}}{2}(x\hat{y}-y\hat{x}),$$ this $S$ becomes $$S = \int_{0}^{\beta_{eff}\hbar}d\tau[\frac{m}{2}(x'^2+y'^2)+V(x,y)-i\frac{e^B_{eff}} {2c}(xy'-yx')] .$$ This is a similar equation to the partition function of the superconducting wires. In order to simplify this mapping, we use the suitable dimensionless form. $\tau = \frac{\beta\hbar}{\xi_0}z, \tilde{\psi}_{ix} = \xi_0^{3/2} \psi_{iR}$, and $\tilde{\psi}_{iy} = \xi_0^{3/2} \psi_{iI}$. Then we can make the identifications of Table \[tab:table1\]. ------------------------------------------------------------------------------------------------------------------------------------ Q.M. S.C. ---------------------------------- ------------------------------------------------------------------------------------------------- $\tau$ $z$ $\vec{\rho_i}=\{x_i(u),y_i(u)\}$ $\vec{\psi_i}=\{\tilde{\psi}_{ix}(z),\tilde{\psi}_{iy}(z)\}$ $E$ $F\frac{\xi_0}{z_{max}}$ $\beta_{eff}$ $\beta z_{max}/\xi_0$ $V_i(x_i,y_i)$ $\left(\frac{\alpha}{\xi_0^2}+\frac{1}{2m^\xi_0^2}(\frac{e^A_z}{c})^2\right)\|\tilde{\psi}_i\|^ 2+\frac{\gamma}{\xi_0^5}\|\tilde{\psi}_i\|^4$ $m$ $\frac{\hbar^4\beta^2}{m^\xi_0^4}$ $B_{eff}$ $-i\frac{2\hbar^2A_z\beta }{m^\xi_0^3}$ $J_{ij}$ $\frac{K_{ij}}{\xi_0^2}$ ------------------------------------------------------------------------------------------------------------------------------------ : \[tab:table1\]Correspondence on the mapping between Q.M. and S.C. (in each case, the left-hand variable corresponds to the parameters on the quantum mechanics problem and the right hand variable corresponds to the parameters on the superconductor wires) We find that the magnetic field has two effects: (i) it determines an effective perpendicular magnetic field in which the equivalent quantum-mechanical particle moves; and (ii) it changes the quadratic part of the effective potential. The Hamiltonian for the analogous quantum problem is $$H = \sum_{i=1}^N[\frac{1}{2m}\left(p_{ix}+\frac{e^B_{eff}}{2c}y\right)^2+\frac{1}{2m}\left(p_{iy}-\frac{e^B_{eff}}{2c}x\right)^2 +V_i(\vec{\rho}_i)]+\sum_{<ij>}2J_{ij}\|\vec{\rho}_i-\vec{\rho}_j\|^2 ,\label{eq:effh}$$ where $p_{ix}$ and $p_{iy}$ are momentum operators of x and y components of $i$th particle, respectively. Probability distribution of the order parameter ----------------------------------------------- We consider the probability distribution of the order parameter, which corresponds to the probability distribution of particles in quantum mechanics. In order to simplify our discussion, we consider single wire case. The probability distribution function of the order parameter can be defined as $$P(\vec{\rho}(\tau))=\frac{1}{Z}\langle\psi^F\|e^{-\frac{H}{\hbar}(L_{\tau}-\tau)}\|\vec{\rho}(\tau)\rangle\langle\vec{\rho}(\tau)\|e^{-\frac{H}{\hbar}\tau}\|\psi^I\rangle ,$$ where $Z=\langle\psi^F\|e^{-\frac{H}{\hbar}L_{\tau}}\|\psi^I\rangle$ and $\|\psi^I\rangle$ represents the boundary condition at $\tau=0$ and $\langle\psi^F\|$ represents the boundary condition at $\tau=L_{\tau}$. Using the eigenstates of the Hamiltonian, $H\|n\rangle=E_n\|n\rangle$, the probability can be written as $$P(\vec{\rho}(\tau))=\frac{1}{Z}\sum_{m,n}\langle\psi^F\|m\rangle\langle m\|\vec{\rho}(\tau)\rangle\langle\vec{\rho}(\tau)\|n\rangle\langle n\|\psi^I\rangle e^{-\frac{E_m}{\hbar}(L_{\tau}-\tau)}e^{-\frac{E_n}{\hbar}\tau}$$ with $$Z=\sum_n\langle\psi^F\|n\rangle\langle n\|\psi^I\rangle e^{-\frac{E_n}{\hbar}\tau} .$$ Explicitly, the expectation value of operator, $\rho$ at the distance $\tau$ from the bottom of the wires is given by $$\langle\hat{\rho}\rangle_{\tau}=\frac{1}{Z}\langle\psi^F\|e^{-\frac{H}{\hbar}(L_{\tau}-\tau)}\int d\vec{\rho}(\tau)\|\vec{\rho}(\tau)\rangle\rho\langle\vec{\rho}(\tau)\|e^{-\frac{H}{\hbar}\tau}\|\psi^I\rangle ,$$ where $\hat{\rho}\|\vec{\rho}\rangle=\rho\|\vec{\rho}\rangle$. The case of periodic boundary condition, our problem can be simplified. If $\psi^F$ corresponds to $\psi^I$ and summed over all possible this configuration, the density matrix can be written as. $$\begin{aligned} P(\vec{\rho}(\tau))&=&\frac{1}{Z}\sum_{m,n}\sum_I\langle\psi^I\|m\rangle\langle m\|\vec{\rho}(\tau)\rangle\langle\vec{\rho}(\tau)\|n\rangle\langle n\|\psi^I\rangle e^{-\frac{E_m}{\hbar}(L_{\tau}-\tau)}e^{-\frac{E_n}{\hbar}\tau} \\ &=&\frac{1}{Z}\sum_{n}\langle n\|\vec{\rho}(\tau)\rangle\langle\vec{\rho}(\tau)\|n\rangle e^{-\frac{E_n}{\hbar}L_{\tau}} ,\end{aligned}$$ where $Z=\sum_n\sum_I\langle\psi^I\|n\rangle\langle n\|\psi^I\rangle e^{-\frac{E_n}{\hbar}L_{\tau}}=\sum_n e^{-\frac{E_n}{\hbar}L_{\tau}}$. So, if the wire is actually in the form of a loop, which means the boundary conditions $\psi(0)=\psi(z_{max})$, our problem corresponds to this statistical mechanics. Of course, in the limit of a very long wire, the periodic boundary condition imposed by the loop should become unimportant. In the case of the periodic boundary condition for single wire, we can see qualitative behavior of order parameter. The average gap in the GL problem (denoted $\tilde{\Delta}(t)$) corresponds to the mean distance $\langle \rho \rangle$ in the quantum-mechanical problem, i.e. $$\begin{aligned} \langle\rho\rangle \leftrightarrow \tilde{\Delta}(t) ,\end{aligned}$$ where $\Delta (t) = \tilde{\Delta}(t)/\xi_0^{3/2}$. At much lower temperature than the critical temperature $T_c^0$, the mean distance from the origin of the particle approaches the value predicted for the quantum problem in the limit of infinite mass, i.e. the value of $\rho$ for which the quartic potential is a minimum although the magnitude of these gaps at $T=0$ are different. The function $\sqrt{1-t}$, is the classical solution, i.e., in the case when thermal fluctuations in the GL case are negligible. These fluctuations do indeed become very small when $T\rightarrow 0$, because in this regime, the effective potential rises steeply above its minimum, and the $\langle\rho\rangle$ becomes very close to the value that minimizes the GL free energy. When $\langle\rho\rangle$ has this value, the corresponding value for $\tilde{\Delta}(t)$ is $$\tilde{\Delta}(t) = \sqrt{\tilde{\psi_R}^2+\tilde{\psi_I}^2} = \sqrt{\frac{\alpha_0 T^0_c\xi_0^3}{2\gamma}}\sqrt{1-t}=\tilde{\Delta}(0)g(t), \label{eq:rho_one}$$ where $\tilde{\Delta}(0)$ is the gap at $T=0$. These considerations may suggest that we can approximate $g(t)=\tilde{\Delta}(t)/\tilde{\Delta}(0)=\sqrt{1-t}$. Phase only model and mean-field approximation --------------------------------------------- This system will undergo a phase transition into a phase-ordered state below a critical temperature $T_c$ which is distinct from (and lower than) the single wire mean-field transition temperature $T_{c}^0$. To do this, we consider a simplified, “phase-only” version of this Schrödinger equation (\[eq:effh\]). We assume that the [magnitudes]{} $\rho_i$ of the variables ${\bf x}_i$ are fixed at the values which minimize the single-wire GL free energy, i.e. $\rho_i \equiv \rho_0$ (\[eq:rho\_one\]). All terms in the Hamiltonian involving $\partial/\partial\rho_i$ can be ignored in this phase-only model. The effective Hamiltonian (\[eq:effh\]) then becomes $$H = -\sum_i\frac{\hbar^2}{2m\rho_0^2}\frac{\partial^2}{\partial\phi_i^2} -\sum_i\frac{e^B_{eff}}{2mc}\frac{\hbar}{i}\frac{\partial}{\partial\phi_i} +2\sum_{\langle ij\rangle}J_{ij}\rho_0^2(1-\cos{(\phi_i-\phi_j)}), \label{eq:phaseH}$$ where the sum runs over distinct nearest neighbor pairs. This is the well-known [quantum XY model]{}, which exhibits a quantum phase transition at a critical value. The mean field approximation can be applied to this Hamiltonian, assuming that $J_{ij}=J$ for only nearest neighbors, by replacing the second term according to the prescription $$\cos{(\phi_i-\phi_j)}=2\cos{\phi_i}\langle\cos{\phi}\rangle-\langle\cos{\phi}\rangle^2 ,$$ where we are supposing $\langle\sin{\phi}\rangle=0$ because of the symmetry. Thus, $$\begin{aligned} 2\sum_{\langle ij\rangle}J\rho_0^2 (1-\cos{(\phi_i-\phi_j)}) &=&2\sum_{\langle ij\rangle}J\rho_0^2\left(1-2\cos{\phi_i}\langle\cos{\phi}\rangle+\langle\cos{\phi}\rangle^2\right) \\ &=&-4z_nJ\rho_0^2\langle\cos{\phi}\rangle\sum_i\cos{\phi_i}+2\sum_{\langle ij\rangle}J\rho_0^2(1+\langle\cos{\phi}\rangle^2) \\ &=&-4z_nJ\rho_0^2\langle\cos{\phi}\rangle\sum_i\cos{\phi_i}+2z_nNJ\rho_0^2(1+\langle\cos{\phi}\rangle^2) ,\end{aligned}$$ where $z_n$ is the number of nearest neighbors in the lattice. Thus, the effective Hamiltonian corresponding to eq. (\[eq:phaseH\]) becomes a following Schrödinger equation: $$\left\{\!\!\frac{-\hbar^2}{2m\rho_0^2}\frac{\partial^2}{\partial\phi_i^2} \!\!-\!\frac{e^B_{eff}}{2mc}\frac{\hbar}{i}\frac{\partial}{\partial\phi_i} \!\!-\!4z_n\rho_0^2J\langle\cos{\phi}\rangle\!\!\cos{\phi_i}\!\!+\!2z_nJ\rho_0^2(1\!\!+\!\langle\cos{\phi}\rangle^2) \!\!\right\} \!\!\psi_n(\phi_i)\! = \! E_n\psi_n(\phi_i). \label{eq:phaseonlyh}$$ We consider the self consistent equation for $\cos{\phi}$ on the periodic boundary condition. The mean field theory is defined by the self-consistency requirement on $\langle\cos{\phi}\rangle$: $$\langle\cos{\phi}\rangle =\frac{\sum_n e^{-\beta_{eff}E_n}\langle\psi_n(\phi_i)\|\cos{\phi_i}\|\psi_n(\phi_i)\rangle} {\sum_n e^{-\beta_{eff}E_n}}. \label{eq:condition}$$ For example, when the wires are sufficiently long where only the ground state contribution may be important, the self-consistent condition becomes $$\langle\cos{\phi}\rangle=\langle\psi_0(\phi_i)\|\cos{\phi_i}\|\psi_0(\phi_i)\rangle . \label{eq:gcondition}$$ These equations may be solved for $\langle\cos{\phi}\rangle$ and $T_c$, where the critical temperature can be determined by $\langle\cos{\phi}\rangle\rightarrow 0$. Results and Discussion \[sec:one\_result\] ========================================== We have considered long-range phase coherence among wires in the bundle in order to see whether the phases on the wires are coherent and the bundle as a whole is superconducting or not. The self-consistent equation gives rise to a phase diagram exhibiting superconductivity, which can be defined as the greatest temperature and field such that $\cos{\theta}$ takes on a non-zero value [@Simanek1979]. Here, we assume that the Josephson coupling is independent of a temperature. We consider the temperature dependence, $\sqrt{1-t}$ for $\rho$. In order to simplify our calculations, we consider the case of the periodic boundary condition. No magnetic field ----------------- We consider the following self-consistent equation, substituting $B_{eff}=0$ for the differential eq. (\[eq:phaseonlyh\]), $$\left(-\frac{\hbar^2}{2m\rho_0^2}\frac{\partial^2}{\partial\phi_i^2} -4z_n\rho_0^2J\langle\cos{\phi}\rangle\cos{\phi_i}+2z_nJ\rho_0^2(1+\langle\cos{\phi}\rangle^2)\right)\psi_n(\phi_i) = E_n\psi_n(\phi_i) . \label{eq:pH}$$ This equation can be reduced to the standard Mathieu equation [@Mathieu], using $v=\phi/2$, $y(v)=\psi_n(\phi_i/2)$, $$\frac{d^2y_n(v)}{dv^2}+(a_n-2q\cos{2v})y_n(v)=0 , \label{eq:mathieu_one}$$ where the characteristic value of the Mathieu equation and $q$ are written as $$\begin{aligned} a_n&=&4(E_n-2z_nJ\rho_0^2(1+\langle\cos{\phi}\rangle^2))\frac{2m\rho_0^2}{\hbar^2}=\frac{E_n-B(1+\langle\cos{\phi}\rangle^2)}{A} ,\\ q&=&-8z_n\rho_0^2J\langle\cos{\phi}\rangle\frac{2m\rho_0^2}{\hbar^2}=-\frac{B}{A}\langle\cos{\phi}\rangle ,\end{aligned}$$ where we define $A=\frac{\hbar^2}{8m\rho_0^2}$ and $B=2z_nJ\rho_0^2$. The eigenvalues are explicitly written as $$E_n=Aa_n+B(1+\langle\cos{\phi}\rangle^2) .$$ The allowed eigenfunctions are determined by the condition that the wave functions be single-valued, i.e., that $\psi_n(\phi+2\pi)=\psi_n(\phi)$, or equivalently, that $y_n(v+\pi)=y_n(v)$. The allowed three lowest solutions, up to the order of $q^2$, are [@Mathieu] $$\begin{aligned} y_0(v,q)&=&\frac{1}{\sqrt{\pi}}\left[1-\frac{q}{2}\cos{2v}+q^2\left(\frac{\cos{4v}}{32}-\frac{1}{16}\right)\right], \quad a_0=-\frac{q^2}{2}, \\ y_2(v,q)&=&\frac{2}{\sqrt{\pi}}\left[\cos{2v}\!-\!q\left(\frac{\cos{4v}}{12}-\!\frac{1}{4}\right)\!+\!q^2\left(\frac{\cos{6v}}{384}-\frac{19\cos{2v}}{288}\right)\right],\quad\!\! a_2\!=\!4\!+\!\frac{5q^2}{12}, \\ y_{-2}(v,q)&=&\frac{2}{\sqrt{\pi}}\left[\sin{2v}-q\frac{\sin{4v}}{12}+q^2\left(\frac{\sin{6v}}{384}-\frac{\sin{2v}}{288}\right)\right], \quad a_{-2}=4-\frac{q^2}{12} ,\end{aligned}$$ where these are normalized like $\int_0^{2\pi}\psi_n(\phi)d\phi=1$. Thus, the matrix elements for $\cos{\theta}$ on the corresponding bases, $n=0,2,$ and $-2$, are $$\langle\cos{\phi}\rangle= \left( \begin{array}{ccc} -\frac{q}{2} & \frac{1}{\sqrt{2}} & 0 \\ \frac{1}{\sqrt{2}} & \frac{5q}{12} & 0 \\ 0 & 0 & -\frac{q}{12} \end{array} \right).$$ From the mapping, we can get the self-consistent condition for the critical temperature of the phase ordering in terms of the parameters of the GL equation for sufficient or infinite long wires eq. (\[eq:gcondition\]), which corresponds to the only consideration of the ground state $(n=0)$ in the quantum mechanics problem, and it takes the following form. $$\langle\cos{\phi}\rangle=-\frac{A}{B}q = -\frac{q}{2}. \label{eq:conditionatt0}$$ The temperature dependence of order parameter obtained by eq. (\[eq:conditionatt0\]) for infinite long wires is shown in Fig. \[fig:order1\]. ![\[fig:order1\] Temperature dependence of the phase-ordering for infinite long wires. ](order1.eps){height="8cm" width="9cm"} This figure clearly shows that there is a second order phase transition at $t=0.5$ because the order parameter continuously becomes zero at the critical point. As expected, the critical temperature of the whole wires is lower than the critical temperature of a single wire. The transition temperature of phase ordering can be calculated by finding the temperature where $\langle\cos{\phi}\rangle$ becomes zero. Thus, because with $ A\rightarrow\frac{m^\xi_0^4}{8\hbar^2\beta^2\tilde{\Delta}^2(t)}$ and $B\rightarrow\frac{2z_nK\tilde{\Delta}^2(t)}{\xi_0^2}$, $$\frac{B}{A}\rightarrow\frac{16z_n\hbar^2\beta^2K\tilde{\Delta}^4(t)}{m^\xi_0^6}=2\alpha\frac{(1-t)^2}{t^2},$$ where $\alpha=\frac{8z_n\hbar^2\tilde{\Delta}^4(0)K}{m^\xi_0^6(k_BT_c^0)^2}$, the condition becomes $$\langle\cos{\phi}\rangle = 0 \rightarrow t_c=\sqrt{\alpha}g^2(t_c) .$$ Therefore, using $g(t)=\sqrt{1-t}$, this critical temperature becomes $$t_c=\frac{\sqrt{\alpha}}{1+\sqrt{\alpha}} .$$ On the other hand, for finite length wires, contributions from excited states in the quantum mechanics problem need to be considered because the effective temperature is not zero. Using up to the order $\|n\|\le 2$ for the solution of Mathieu’s equation, using eq. (\[eq:condition\]), the following self-consistent condition can be obtained, $$-\frac{A}{B}q=\frac{-\frac{q}{2}e^{-\beta_{eff}E_0}+\frac{5q}{12}e^{-\beta_{eff}E_2}-\frac{q}{12}e^{-\beta_{eff}E_{-2}}}{e^{-\beta_{eff}E_0}+e^{-\beta_{eff}E_2}+e^{-\beta_{eff}E_{-2}}} ,$$ where $\beta_{eff}E_n=\beta_{eff}(Aa_n+B(1+\langle\cos{\phi}\rangle))$, but the second term can be canceled. Therefore, with the mapping, we can get $$\left(\frac{t}{1-t}\right)^2=\alpha \frac{1-\frac{2}{3}e^{-4x\frac{t}{1-t}}}{1+2e^{-4x\frac{t}{1-t}}},$$ where we use the following mapping $$\beta_{eff}A\rightarrow x\frac{t}{1-t}=x_0\frac{z_{max}}{\xi_0}\frac{t}{1-t}=\frac{m^\xi_0^2k_BT_c^0}{8\hbar^2}\frac{\xi_0^2}{\tilde{\Delta}^2(0)}\frac{z_{max}}{\xi_0}\frac{t}{1-t} .$$ Using the numerical values according to Tang [et al]{} [@Tang2001], $x_0\approx 1.4\times 10^{-4}$. A plot of $T_c$ versus $\alpha$ for several lengths ($100\xi_0$, $1000\xi_0$, $2000\xi_0$, and $5000\xi_0$) and infinite length are given in Fig. \[fig:phase1\]. ![\[fig:phase1\] Phase diagram of $t_c=T_c/T_c^0$ as a function of $\alpha$ for several values of length of the wires, $100\xi_0$, $1000\xi_0$, $2000\xi_0$, $5000\xi_0$, and $\infty$ where $\xi_0=42$Å.](phase1.eps){height="8cm" width="9cm"} This figure shows that as the length of the wires has increased, the phase critical temperature has increased. Perpendicular magnetic field ---------------------------- The critical temperature for the presence of a magnetic field on the wires can be obtained by solving the non-Hermitian eq. (\[eq:phaseonlyh\]). $$\left\{\!\!\frac{-\hbar^2}{2m\rho_0^2}\frac{\partial^2}{\partial\phi_i^2} \!\!-\!\frac{e^B_{eff}}{2mc}\frac{\hbar}{i}\frac{\partial}{\partial\phi_i} \!\!-\!4z_n\rho_0^2J\langle\cos{\phi}\rangle\cos{\phi_i}\!\!+\!2z_nJ\rho_0^2(1\!\!+\!\langle\cos{\phi}\rangle^2)\!\right\} \!\psi_n(\phi_i) \!\!=\!E_n\psi_n(\phi_i). \label{eq:pAH}$$ Using $\psi_n(\phi)=e^{pv}F(v)$ with and $v=\phi/2$ and $$p=i\frac{2e^\rho_0^2B_{eff}}{c\hbar},$$ again this equation reduces to the standard Mathieu equation: $$\frac{d^2F(v)}{dv^2}-(2q\cos{2v})F(v)=-a_{\nu}F(v),$$ where $$\begin{aligned} a_{\nu}-p^2&=&4(E_n-2z_nJ\rho_0^2(1+\langle\cos{\phi}\rangle^2))\frac{2m\rho_0^2}{\hbar^2}=\frac{E_n-B(1+\langle\cos{\phi}\rangle^2)}{A}, \\ q&=&-8z_n\rho_0^2J\langle\cos{\phi}\rangle\frac{2m\rho_0^2}{\hbar^2}=-\frac{B}{A}\langle\cos{\phi}\rangle .\end{aligned}$$ The allowed eigenvalues are determined by the boundary condition that $\psi_n(\phi+2\pi)=\psi_n(\phi)$, or equivalently $F(v+\pi)=\exp(-p\pi)F(v)$. Thus we are interested only in the Floquet solutions of the Mathieu equation with Floquet exponent $\nu= 2n+ip$, where $n=0,\pm1,\pm2,....$. These solutions are explicitly written as [@Mathieu] $$F_\nu(v)=c_0e^{i\nu v}\left[1-q\left(\frac{e^{2iv}}{4(\nu+1)}-\frac{e^{-2iv}}{4(\nu-1)}\right)\right],$$ where $c_0$ is a normalization constant. The eigenvalues are, using $q=-\frac{B}{A}\cos{\theta}$, $$a_{\nu}=\nu^2+\frac{q^2}{2(\nu^2-1)}.$$ The allowed three lowest solutions, up to the order of $q^2$, are [@Mathieu] $$\begin{aligned} \psi_{ip}(v)&\!\!\!\!\!=\!\!\!\!\!& \sqrt{\frac{1}{\pi}}\left(1-q\frac{\cos{2v}+p\sin{2v}}{2(1+p^2)}\right), \quad\!\!\! a_{ip}\!=\!-\frac{q^2}{2(1+p^2)}, \\ \psi_{2+ip}(v)&\!\!\!\!\!=\!\!\!\!\!& \sqrt{\frac{1}{\pi}}\left(e^{2iv}-\frac{q}{4}\left(\frac{e^{4iv}}{3+ip}-\frac{1}{1+ip}\right)\right), \quad\!\!\! a_{2+ip}\!=\!4(1+ip)+\frac{q^2}{2(-p^2+4ip+3)}, \\ \psi_{-2+ip}(v)&\!\!\!\!\!=\!\!\!\!\!& \sqrt{\frac{1}{\pi}}\left(e^{-2iv}+\frac{q}{4}\left(\frac{e^{-4iv}}{ip-3}-\frac{1}{ip-1}\right)\right), \quad\!\!\! a_{-2+ip}\!=\!4(1-ip)+\frac{q^2}{2(3-p^2-4ip)}.\end{aligned}$$ Left wave functions can be obtained from right wave function with $\psi_n^L(v,p)=\psi_n^R(v,-p)^$. The self-consistent condition for long wires becomes the following form, $$\langle \cos{\phi}\rangle=-\frac{A}{B}q = -\frac{q}{2(1+p^2)},$$ because the matrix elements for $\cos{\theta}$ corresponding to $n=0,2,$ and $-2$ are, using $q=-\frac{B}{A}\langle\cos{\theta}\rangle$, $$\langle\cos{\phi}\rangle=\frac{1}{2} \left( \begin{array}{ccc} -\frac{q}{1+p^2} & 1 & 1 \\ 1 & \frac{q}{3+4ip-p^2} & \frac{1}{2(1+p^2)} \\ 1 & \frac{1}{2(1+p^2)} & \frac{q}{3-4ip-p^2} \end{array} \right) .$$ Again, we can determine the transition temperature of the phase ordering, where $\langle\cos{\phi}\rangle$ becomes zero. $$\langle\cos{\phi}\rangle = 0 \rightarrow t_c=\sqrt{\frac{\alpha}{1+p^2(t_c)}}g^2(t_c).$$ The approximation $g(t)=\sqrt{1-t}$ is again used for this case and then we can get $$t_c=\frac{\sqrt{\alpha-f^2}}{1+\sqrt{\alpha-f^2}},$$ where we define $f$ as $$p\rightarrow\frac{fg^2(t)}{t}=f_0\frac{A_z\xi_0}{\Phi_0}\frac{g^2(t)}{t}=\frac{8\pi\hbar^2 }{k_BT_c^0m^\xi_0^2}\frac{ \tilde{\Delta}^2(0)}{\xi_0^2}\frac{A_z\xi_0}{\Phi_0}\frac{1-t}{t},$$ where $\Phi_0 = hc/e^$. When $f=0$, this solution corresponds to the previous case. A plot of $T_c$ versus $\alpha$ for $f=0, 0.2, 0.4, 0.6$, and $0.8$ is given in Fig. \[fig:phase2\]. ![\[fig:phase2\] Phase diagram of $t_c=T_c/T_c^0$ as a function of $\alpha$ for several values of magnetic field strength, $f=0.2$, $f=0.4$, $f=0.6$, and $f=0.8$ for infinity length wires](phase2.eps){height="8cm" width="9cm"} This figure shows that the critical temperatures have the minimum values for the interaction between the wires. These values can be calculated by $$\alpha\geq f^2 \rightarrow z_nK \geq \frac{8\pi^2\hbar^2}{m^\xi_0^2}\left(\frac{A_z\xi_0}{\Phi_0}\right)^2.$$ The critical $f_c$, which is related to the maximum flux in the wires can be obtained, $$f_c =\frac{\sqrt{\alpha (1-t)^2-t^2}}{1-t}.$$ Near the critical temperature of phase ordering, using $t=\frac{\sqrt{\alpha}}{1+\sqrt{\alpha}}-\delta t$, this can be written $$f_c\approx\sqrt{2}(1+\sqrt{\alpha})\alpha^{1/4}\sqrt{\delta t}.$$ Fig. \[fig:phase3\] shows that this critical magnetic field $f_c$ for $\alpha=0.2, 0.4, 0.6, 0.8$, and $1$ as a function of a temperature. ![\[fig:phase3\] Temperature dependence of the critical field strength, $f_c$ for $\alpha=0.2, 0.4, 0.6, 0.8$, and $1$ for infinity long wires.](phase3.eps){height="8cm" width="9cm"} Summary ======= We have presented a mapping between a one-dimensional GL problem in the presence of a vector potential along wires and a two-dimensional quantum mechanics problem with a perpendicular magnetic field. Moreover, in the case of weak links between wires, we have obtained, using the mean-field approximation, the phase diagrams for the presence of a magnetic field and absence of it. Next, we discuss the parameters used in this paper. Using the numerical values of the various parameters appropriate to those of a single-walled carbon nanotube, which according to Tang [et al]{} [@Tang2001], where superconducting with a relatively high transition temperature $T^0_c = 15 \mathrm{K}$, $k_BT^0_c = 1.3 \mathrm{meV}$, $\alpha_0T^0_c = 6 \mathrm{meV}$, $\gamma = 1.3 \mathrm{meV}$Å, $m^ = 0.36 m_{\mathrm{e}}$, and $\xi_0 = \frac{\hbar}{\sqrt{2 m^* \alpha_0 T^0_c}}= 42$Å, we can obtain the following values for $\alpha$ and $f$, $\alpha=\frac{z_nK}{8.6\times 10^{-6}\mathrm{meV}}$ and $f=1.7\times 10^4\frac{A_z\xi_0}{\Phi_0}$. The Josephson coupling energy $K$ is approximated by $\frac{\hbar^2}{2m_cs^2}$ where $s$ is the distance between nearest wires. If we use $m_c=m^$, $K$ can be written as $\frac{\xi_0^2\alpha_0 T^0_c}{s^2}$. Thus, supposing $s\approx 5$Å, $K$ is order of $100\sim 1000 [\mathrm{meV}]$. Therefore, our values used in the figures are well suited for describing real systems. We discuss about the use of the GL free energy functional. In principle, this free energy functional is applicable only near the critical temperature, $T-T_c^0\ll T_c^0$. Besides near the critical temperature $T_c^0$, the qualitative description of this functional may not be reasonable, although we can employ higher order expansions of the order parameter in the G.L equation. We want to comment the effect on the interaction term by a magnetic field. When there is a magnetic field, the phase difference needs to be replaced by $\phi_i-\phi_{i+1}-\frac{2\pi}{\Phi_0}\int\vec{A}\cdot d\vec{l}$ where the integration is between different wires. However, because the direction of vector potential is taken in the direction of the wires, $z$, there is no contribution from the integral on the phase difference. In this paper, we only consider the periodic boundary condition for simplification. When wires are sufficient long, the effect of the boundary conditions may not change the physical properties of the system. However, these boundary conditions may affect the properties of the system because of finite length of wires. Moreover, our theory neglects the effects of disorder, which plays an important role on balk superconductors. With these degrees of freedom, the properties of the system may be changed. Thus, it might be an interest to consider these cases for our future research. [99]{} Z. K. Tang, L. Zhang, N. Wang, X. X. Zhang, G. H. Wen, G. D. Li, J. N. Wang, C. T. Chan, and Ping Sheng, Science [292]{}, 2462 (2001). A.Yu. Kasumov, R. Deblock, M. Kociak, B. Reulet, H. Bouchiat, I.I. Khodos, Yu.B. Gorbatov, V.T. Volkov, C. Journet, and M. Burghard, Science [284]{}, 1508 (1999). A.F. Morpurgo, J. Kong, C.M. Marcus, and H. Dai, Science [286]{} 263 (1999). J. González, Phys. Rev. Lett. [87]{} 136401 (2001). M. Kociak, A. Yu. Kasumov, S. Guéron, B. Reulet, I.I. Khodos, Yu. B. Gorbatov, V.T. Volkov, L. Vaccarini, and H. Bouchiat, Phys. Rev. Lett. [86]{}, 2416 (2001). A. Kasumov, M. Kociak, M. Ferrier, R. Deblock, S. Guéron, B. Reulet, I. Khodos, O. Stéphan, and H. Bouchiat, Phys. Rev. B [68]{}, 214521 (2003). M. Tian, N. Kumar, S. Xu, J.S. Kurtz, and M.H.W. Chan, Phys. Rev. B [*]{}, (2005). M.L. Tian, J.G. Wang, J.S. Kurtz, Y. Liu, T.S. Mayer, T.E. Mallouk, and M.H.W. Chan, Phys. Rev. B []{}, (2005). D. J. Scalapino, M. Sears, and R. A. Ferrell, Phys. Rev. B [6]{}, 3409 (1972). B. Stoeckly and D. J. Scalapino, Phys. Rev. B [11]{}, 205 (1975). D. J. Scalapino, Y. Imry, and P. Pincus, Phys. Rev. B [11]{}, 2042 (1975). L.P. Kadanoff and J. Swift, Phys. Rev. [165]{}, 310 (1968). H.D. Fogedby, A.B. Eriksson, and L.V. Mikheev, Phys. Rev. Lett. [75]{}, 1883 (1995). D. Kim, Phys. Rev. E [52]{}, 3512 (1995). D. R. Nelson and V. Vinokur, Phys. Rev. B [48]{}, 13060 (1993). N. Hatano, and D. R. Nelson, Phys. Rev. Lett [77]{}, 570 (1996). N. Hatano, and D. R. Nelson, Phys. Rev. B [56*]{}, 8651 (1997). W.E. Lawrence and S. Doniach, in E.Danda (ed.) Proc. 12th Int. Conf. Low Temp. Phys. (Kyoto, 1970; Keigaku, Tokyo, 1971), p.361. R. P. Feynman and A. R. Hibbs, [Quantum Mechanics and Path Integrals]{} (McGraw-Hill, New York, 1965). E. Simanek, Solid State Commun. [31]{}, 419 (1979). See, for example, [Handbook of Mathematical Functions*]{}, edited by M. Abramowitz and I.A. Stegun (New York, Dover, 1964), p. 721.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A stochastic ordering approach is applied with Stein’s method for approximation by the equilibrium distribution of a birth–death process. The usual stochastic order and the more general $s$-convex orders are discussed. Attention is focused on Poisson and translated Poisson approximation of a sum of dependent Bernoulli random variables, for example $k$–runs in i.i.d. Bernoulli trials. Other applications include approximation by polynomial birth–death distributions.' title: 'Stein’s method and stochastic orderings' --- Introduction {#intro} ============ Stein’s method has proved to be an effective tool in probability approximation, and has the advantage of being applicable in the presence of dependence. See, for example, Stein (1986), and Barbour and Chen (2005) for more recent developments. It is well–known that error bounds obtained via Stein’s method may be simplified under some assumptions on the dependence present. For example, in the presence of negative or positive relation, Stein’s method gives simple error bounds in the Poisson approximation of a sum of indicator random variables. This is exploited throughout the work of Barbour et al. (1992), and will be returned to in our Section \[pois\]. In this work, we consider the more general situation of approximation by the equilibrium distribution of a birth–death process, and examine the situations in which Stein’s method leads to simple, easily calculable error bounds. These error bounds will typically be differences of moments of our random variables. As we will see, the assumptions under which we can obtain such error bounds are naturally phrased in terms of stochastic orderings. Consider a birth-death process on (some subset of) $\mathbb{Z}^+$ with birth rates $\alpha_j$ and death rates $\beta_j$ for $j\geq0$. Suppose $\beta_0=0$. Let $\pi$ be the stationary distribution of such a process, with $\pi_j=P(\pi=j)$, $j\geq 0$. In this work we combine Stein’s method with a stochastic ordering construction to consider the approximation by $\pi$ of some random variable $W$ on $\mathbb{Z}^+$. Our random variable $\pi$ satisfies the identity $E[Ag(\pi)]=0$ for any bounded function $g:\mathbb{Z}^+\mapsto\mathbb{R}$, where $A$ is the linear operator defined by $$\label{1} Ag(j)=\alpha_j g(j+1)-\beta_j g(j), \quad j\geq 0.$$ $A$ is a characterising operator for $\pi$, in the sense that a random variable $Z=_d\pi$ if and only if $E[Ag(Z)]=0$ for all $g$ bounded. The construction of such a characterising operator is the basis of Stein’s method for probability approximation. See the books by Stein (1986), Barbour et al. (1992), Barbour and Chen (2005) and references therein. For Stein’s method applied to birth-death processes, see Brown and Xia (2001) and Holmes (2004). Given some test function $h$, the so-called Stein equation is defined by $$\label{2} h(j) - E[h(\pi)] =Af(j), \quad j\geq 0.$$ Its solution is denoted $f=f_h=Sh$. We call $S$ the Stein operator. Bounds on $S$ are an essential ingredient of Stein’s method. Note that the solution $f$ of the Stein equation depends on the chosen test function $h$. However, for notational convenience in much of the work that follows we will write $f$ rather than $f_h$ or $Sh$. We will often choose $h(j)=I_{(j\in B)}$ for some $B\subseteq\mathbb{Z}^+$, in which case the solution $f$ will depend on the chosen set $B$. There are several common distributions $\pi$ covered by this framework. For each of the examples below, bounds are available on the corresponding Stein operator $S$. Theorem 2.10 of Brown and Xia (2001) may also be applied to give bounds on $S$ in many cases. - If $\alpha_j=\lambda$ and $\beta_j=j$, then $\pi \sim \textrm{Po}(\lambda)$, the Poisson distribution with mean $\lambda$. See Barbour et al. (1992) and references therein. - If $\alpha_j=q(r+j)$ and $\beta_j=j$, then $\pi \sim \textrm{NB}(r,1-q)$ has a negative binomial distribution. See Brown and Phillips (1999). - If $\alpha_j=(n-j)p$ and $\beta_j=(1-p)j$, then $\pi \sim \textrm{Bin}(n,p)$. See Ehm (1991). - In the geometric case, we may, of course, use the negative binomial operator above. Alternatively we may choose $\alpha_j=q$ and $\beta_j=I_{(j\geq1)}$, so that $\pi \sim \textrm{Geom}(1-q)$. See Peköz (1996). The present work is organized as follows. In Section \[sect2\], we will derive abstract error bounds using Stein’s method combined with some stochastic ordering assumptions in the setting of approximation by the equilibrium distribution of a birth–death process. In Section \[simple\], a simple sufficient condition under which these stochastic ordering assumptions hold is considered, and some applications are given. Section \[pois\] discusses Poisson approximation for a sum of dependent indicators. We will see how concepts of negative and positive relation relate to our stochastic ordering assumptions, and present generalizations of error bounds derived by Barbour et al. (1992). Based on this work we move on, in Section \[tp\], to consider translated Poisson approximation. Applications here will include approximation of the number of $k$–runs in i.i.d. Bernoulli trials. Finally, in Section \[bd3\], we give another abstract approximation theorem, and consider its application to a sum of independent indicator random variables. An abstract approximation theorem {#sect2} ================================= Consider Stein’s method for approximating the equilibrium distribution of a birth-death process. Our purpose in this section is to derive abstract error bounds under some stochastic ordering assumptions. A first-order bound {#bd1} ------------------- Suppose that $W$ is a random variable supported on (some subset of) $\mathbb{Z}^+$ with $\mu_j=P(W=j)$, $j\geq 0$. Set $\mu_{-1}=0$. Our concern is the approximation of such a variable $W$ by $\pi$, specifically by estimating the difference $\|Eh(W)-Eh(\pi)\|$, i.e. $\|E[Af(W)]\|$. For this, a simple representation of this difference will be applied with some stochastic ordering assumptions to yield bounds using Stein’s method. We may then bound, for example, the total variation distance between $\mathcal{L}(W)$ and $\mathcal{L}(\pi)$, defined by $$d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi))=\sup_{B\subseteq\mathbb{Z}^+}\|P(W\in B)-P(\pi\in B)\|.$$ Although we are mainly concerned with approximation in total variation distance, the results we derive may also be used with other probability metrics. Let $\Delta$ be the forward difference operator. Since, with the operator (\[1\]), the choice of $f(0)$ is arbitrary, we follow Brown and Xia (2001) and choose $f(0)=0$. Writing $f(j)=\Delta f(0)+\cdots+\Delta f(j-1)$, we thus obtain the representation $$\label{3} Eh(W)-Eh(\pi) = \sum_{k=0}^\infty\Delta f(k) \sum_{j=k+1}^\infty (\alpha_{j-1}\mu_{j-1}-\beta_{j}\mu_{j}).$$ In the next subsection, we will extend (\[3\]) to include the $l$th forward differences of $f(\cdot)$, for all $l\geq 1$. We now consider how this representation may be applied in conjunction with the usual stochastic ordering, denoted $\succeq_{st}$. Define two random variables $W_\alpha$ and $W_\beta$ by $$\label{4} P(W_\alpha=j)=\frac{\alpha_{j-1} \mu_{j-1}}{E\alpha_W}, \; \mbox{ and } \; P(W_\beta=j)=\frac{\beta_{j}\mu_{j}}{E\beta_W}, \quad j\geq 1.$$ If $W_\alpha\succeq_{st}W_\beta$ and $E\alpha_W\geq E\beta_W$, we have that $\sum_{j=i}^\infty\alpha_{j-1}\mu_{j-1}\geq\sum_{j=i}^\infty\beta_{j}\mu_{j}$ for all $i\geq1$. In this case, (\[3\]) may be bounded to obtain $$\|Eh(W)-Eh(\pi)\| \;\leq\; \lVert\Delta f\rVert_\infty \, E[\alpha_W (W+1) - \beta_W W].$$ A similar argument holds if we instead assume that $W_\beta\succeq_{st}W_\alpha$ and $E\beta_W\geq E\alpha_W$. We thus obtain the following result. \[res1\] Assume that one of the two following conditions holds: $$\label{5} \mbox {either (i)} \; W_\alpha\succeq_{st} W_\beta \; \mbox { with } E\alpha_W\geq E\beta_W, \mbox { or (ii)} \; W_\beta\succeq_{st}W_\alpha\; \mbox { with } \; E\beta_W\geq E\alpha_W.$$ Then, $$\label{6} \|Eh(W)-Eh(\pi)\| \;\leq\; \lVert\Delta Sh\rVert_\infty\, \left\|E[\alpha_W (W+1) - \beta_W W]\right\|.$$ A $s$-order bound {#bd2} ----------------- We will now establish our main abstract result. For that, we will have recourse to the concept of discrete $s$-convex stochastic ordering, denoted $\succeq_{s-cx}$, for any integer $s\geq 1$. See, for example, Lefèvre and Utev (1996) for this notion. Briefly, given any two non-negative integer-valued random variables $X$ and $Y$, one says that $X \succeq_{s-cx} Y$ when $$E[f(X)]\leq E[f(Y)] \; \mbox{ for all }\mbox{$s$-convex functions }f,$$ that is, for all functions $f$ satisfying $\Delta^s f(j) \geq 0, \; j\geq 0$. Note that this ordering implies that $X$ and $Y$ have the same first $s-1$ moments. To begin with, we introduce a Bernoulli random variable $v_p$ with $$P(v_p=1) = p = 1-P(v_p=0),$$ independently of all other entries. We write $\alpha =E\alpha_W$, $\beta =E\beta_W$, and in an analogous way to (\[4\]), we define the random variables $W_\alpha$ and $W_\beta$ by $$\label{7} P(W_\alpha\in B) = \alpha^{-1} E[\alpha_W I_{(W+1\in B)}],\; \mbox { and } \; P(W_\beta\in B) = \beta^{-1} E[\beta_W I_{(W\in B)}],$$ for any Borel set $B$. For notational convenience, we choose to write $C^k_n = {n\choose k}$. The key theorem and an immediate corollary will be first stated, the proof of the theorem being given after. \[res3\] Assume that there exists a random variable $Y$ on $\mathbb{Z}^+$ such that $W_\beta-Y\geq0$ a.s. and $$\label{8} W_\alpha \succeq_{s-cx} v_p(W_\beta -Y).$$ Then, $$\begin{gathered} \label{9} \|Eh(W)-Eh(\pi)\| \leq \sum_{t=0}^{s-1}\;\|\Delta^t Sh(0)\|\; \|E(\alpha_W C_{W+1}^{t}) - E(\beta_W C_{W}^{t})\|\\ + \\|\Delta^s Sh\\|_\infty \; (\alpha E[C_{W_\alpha}^{s}] -2\alpha p E[C_{W_\beta-Y}^{s}] + (\alpha p+\|\alpha p-\beta\|) E[C_{W_\beta}^{s}]).\end{gathered}$$ Consider the special case of (\[8\]) when $p=1$ and $Y=0$ a.s. When $\alpha = \beta$ and under the condition (\[11\]) below, one has that $$E[\alpha_W (W+1)^{t}] = E[\beta_W W^{t}], \quad t=0, \ldots, s-1,$$ so that the inequality (\[9\]) reduces to (\[12\]). \[res4\] Assume that $\alpha = \beta$, and one of the two following conditions holds: $$\label{11} \mbox { either (i) } \; W_\alpha\succeq_{s-cx} W_\beta, \; \mbox { or (ii) } \; W_\beta\succeq_{s-cx}W_\alpha.$$ Then, $$\label{12} \|Eh(W)-Eh(\pi)\| \;\leq\; \\|\Delta^s Sh\\|_\infty \, \|E[\alpha_W C_{W+1}^{s}] - E[\beta_W C_{W}^{s}]\|.$$ We note that Proposition \[res1\] does not follow as a special case of Corollary \[res4\], since this latter result requires the condition $\alpha=\beta$ not needed in Proposition \[res1\]. In the first step we derive a representation of $E[Af(W)]$ that generalizes the representation (\[3\]). Observe that (\[1\]) and (\[7\]) give $$E[Af(W)] =E[\alpha_W f(W+1)] - E[\beta_W f(W)] = \alpha E[f(W_\alpha)] - \beta E[f(W_\beta)].$$ Expanding the function $f$ by the discrete Taylor formula, we obtain, for any $s=1,2,\ldots$, $$f(x) = f(0)+\sum_{k=0}^\infty \Delta f(k)\; I_{(x>k)} =\sum_{t=0}^{s-1}\Delta^t f(0) \; C_{x}^{t} \;+\; \sum_{k=0}^\infty \Delta^s f(k) \; C_{x-k-1}^{s-1};$$ see Lefèvre and Utev (1996). Thus, we find that $$\begin{aligned} \label{13} E[Af(W)]&=&\sum_{t=0}^{s-1}\Delta^t f(0) \; E[AC_{W}^{t}] + \sum_{k=0}^\infty \Delta^s f(k) \; E[A C_{W-k-1}^{s-1}]\nonumber\\ &=&\sum_{t=0}^{s-1}\Delta^t f(0)\; (\alpha E[C_{W_\alpha}^{t}] - \beta E[C_{W_\beta}^{t}])\nonumber\\ && \; + \; \sum_{k=0}^\infty \Delta^s f(k)\; (\alpha E[C_{W_\alpha-k-1}^{s-1}]- \beta E[C_{W_\beta-k-1}^{s-1}]).\end{aligned}$$ Our next step is to derive an abstract metrics-ordering relationship result, which is stated below as a separate lemma. Using the bound (\[15\]) in the representation (\[13\]) then leads to the announced bound (\[12\]). \[res5\] Let $X$, $Y$ and $Z$ be random variables on $\mathbb{Z}^+$ such that $$\label{14} Z-Y \geq 0 \; \mbox { a.s., } \; \mbox { and } \; X\succeq_{s-cx} v_p(Z-Y).$$ Then, for all $a,b \in \mathbb{R^+}$, $$\label{15} \sum_{k=0}^\infty \|a E[C^{s-1}_{X-k-1}] - b E[C^{s-1}_{Z-k-1}]\| \;\leq\; aE[C^s_{X}] - 2ap E[C^s_{Z-Y}] + (ap +\|ap-b\|) E[C^s_{Z}].$$ Letting $$w_k^{(s)}(x) =w_{k}(x) = C_{x-k-1}^{s-1},$$ we get that $$\begin{gathered} \label{16} \sum_{k=0}^\infty \|a E(C^{s-1}_{X-k-1}) - b E(C^{s-1}_{Z-k-1})\| \;\; = \;\; \sum_{k=0}^\infty \|a E[w_k(X)] - b E[w_k(Z)]\| \\ \leq a \sum_{k=0}^\infty \|E[w_k(X)]-E[w_k(v_p(Z-Y))]\| + a \sum_{k=0}^\infty \|E[w_k(v_p Z)]-E[w_k(v_p(Z-Y))]\| \\ + \sum_{k=0}^\infty \|a E[w_k(v_p Z)]-b E[w_k(Z)]\| \;\; = \;\; S_1+S_2+S_3.\end{gathered}$$ Let us examine the three sums in (\[16\]). First, we easily check that $$\label{17} \sum_{k=0}^\infty E[w_k(Z)]=E[C_{Z}^{s}].$$ Using (\[17\]), we successively find that $$S_3 \;=\; \|ap-b\| \sum_{k=0}^\infty E[w_k(Z)] \;=\; \|ap-b\| E[C_{Z}^{s}];$$ since $Z-Y\geq 0$ and $Z\succeq_{st} Z-Y$, $$S_2 \;=\; ap \sum_{k=0}^\infty (E[w_k(Z)]-E[w_k(Z-Y)]) \;=\; ap (E[C_{Z}^{s}]-E[C_{Z-Y}^{s}]);$$ finally, by the assumption (\[14\]) and a standard property of the order $\succeq_{s-cx}$, $$S_1 \;=\; a \sum_{k=0}^\infty [Ew_k(X)- p Ew_k(Z-Y)] \;=\; a (E[C_{X}^{s}] - pE[C_{Z-Y}^{s}]).$$ Inserting these three terms in (\[16\]), we then deduce the bound (\[15\]). For $s=p=1$ and $a=b=1$, Lemma \[res5\] states that if $X\succeq_{st} Z-Y \geq 0$, then an upper bound for the Wasserstein distance between $\mathcal{L}(X)$ and $\mathcal{L}(Z)$ is $$\label{18} d_{W}(\mathcal{L}(X),\mathcal{L}(Z)) \;=\; \sum_{k=0}^\infty \|P(X>k)-P(Z>k)\| \;\leq\; 2EY + EX -EZ.$$ This bound is of interest in the stochastic ordering context investigated by Kamae et al. (1977), with random variables on $\mathbb{Z}^+$ here. Note that by choosing the optimal coupling $X$, $Z$ and $Y=(Z-X)_+$, (\[18\]) gives the exact bound since $$d_{W}(\mathcal{L}(X),\mathcal{L}(Z)) \;\leq\; 2 E(Z-X)_+ + EX - EZ \;=\; E\|Z-X\| \;=\; d_{W}(\mathcal{L}(X),\mathcal{L}(Z)).$$ It is worth indicating that an analogous argument allows us to show that the same bound (\[18\]) holds under the single condition $X+Y\succeq_{st} Z$. A priori, this result seems to be preferable, since the extra condition $Z-Y\geq 0$ is not required. One can see, however, that $X\succeq_{st} Z-Y$ does not imply $X+Y\succeq_{st} Z$ in general. As an example, choose $X=U$, $Y=U$ and $Z=n$ a.s., where $n$ is any fixed positive integer and $U$ is discrete uniform on the set $\{0,1,\ldots,n\}$. Then, $X=U =_d n-U=Z-Y$ so that $X\succeq_{st} Z-Y$, but $X+Y=2U$ is not $\succeq_{st}$ than $n=Z$. A simple sufficient condition and examples {#simple} ========================================== In practice, it may be difficult to check directly such conditions as stochastic ordering between $W_\alpha$ and $W_\beta$, as required by (\[5\]) and (\[11\]). It is thus useful to have available a simple sufficient condition which we may then apply. Throughout this subsection, we assume that $\alpha=\beta$ and $W_\alpha$ and $W_\beta$ have equal moments of order $t=1,\ldots, s-1$. That is, we assume $$\mbox{condition } (A_s):\quad\quad\quad E[\alpha_W (W+1)^t]=E[\beta_W W^t],\hspace{20pt} t=0,\ldots, s-1.$$ A well-known Karlin-Novikoff sufficient condition to guarantee the $s$-convex ordering in (\[11\]) under $(A_s)$ is that our sequence $\{\alpha_{j-1}\mu_{j-1} - \beta_j\mu_j\}$ has at most $s$ changes of sign. \[prk\_n\] Suppose that the condition $(A_s)$ is satisfied and that the sequence $\{\alpha_{j-1}\mu_{j-1} - \beta_j\mu_j\}$ has at most $s$ changes of sign. Then (\[12\]) holds. As a consequence, we obtain the following corollary, which extends Proposition A.1 of Barbour and Pugliese (2000) to birth-death processes. \[res2\] Suppose that $E\alpha_W=E\beta_W$. If the sequence $\{\alpha_{j-1}\mu_{j-1} - \beta_j\mu_j\}$ is monotone, then $W_\alpha$ and $W_\beta$ are stochastically ordered, so that the inequality (\[6\]) may be applied. We illustrate these results with the following examples. Our first example is motivated by Phillips and Weinberg (2000). Let $W$ have a Bose-Einstein occupancy distribution. That is, given $m, d \geq 1$, $$\mu_j=P(W=j)=\binom{d+m-j-2}{m-j}\binom{d+m-1}{m}^{-1}, \quad 0\leq j\leq m.$$ We wish to approximate $W$ by $\pi \sim \textrm{Geom}(p)$ where $p=(d-1)/(d+m-1)$. Let $q=1-p$. To obtain our geometric law, we choose $\alpha_j=q$ and $\beta_j=I_{(j>0)}$, $j\geq 0$ as birth and death rates. Firstly, one can easily check that in this case, $E\alpha_W=E\beta_W$ and the sequence $\{q\mu_{j-1}-\mu_j\}$ is non-decreasing, so that $W_\alpha\succeq_{st}W_\beta$. Using Corollary \[res2\], the bound (\[6\]) then becomes $$\label{19} \|Eh(W)-Eh(\pi)\| \;\leq\; p\, \lVert\Delta Sh\rVert_\infty\; \|EW-E\pi\|.$$ Moreover, it is known (see Peköz (1996, Section 2)) that the Stein operator $S$ admits here the representation $$Sh(j)=-\sum_{i=j}^\infty\left[h(i)-Eh(\pi)\right]q^{i-j}.$$ From this, we find that $\Delta Sh(k)=-\sum_{i=k}^\infty\Delta h(i)q^{i-k}$, which leads to the bound $$\lVert\Delta Sh\rVert_\infty\leq\ p^{-1} \lVert\Delta h\rVert_\infty.$$ Inserting this bound in (\[19\]) yields the following. \[cor\_ex1\] With $W$ and $\pi$ as above, $$\|Eh(W)-Eh(\pi)\| \;\leq\; \lVert\Delta h\rVert_\infty\, \frac{m}{d(d-1)}.$$ In particular, $d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi))\leq m/d(d-1)$. Our next examples centre around approximation by so–called polynomial birth–death distributions, defined by Brown and Xia (2001) as the equilibrium distribution of a birth–death process with birth and death rates $\alpha_j$ and $\beta_j$ which are polynomial in $j$. With such choices, we will write $\pi \sim \mbox{PBD}(\alpha_j,\beta_j)$. Suppose that $W$ satisfies $\mu_j=(a+bj^{-1})\mu_{j-1}$ for some $a,b\in\mathbb{R}$. That is, $W$ belongs to the Katz (or Panjer) family of distributions (see Johnson et al. (1992, Section 2.3.1)). It is well known that in this case $W$ must have either either a binomial, Poisson or negative binomial distribution. We fix some $l\geq1$ and consider the approximation of $W$ by the polynomial birth-death distribution $\pi \sim \mbox{PBD}(\alpha,jQ_{l-1}(j))$. Here we have chosen a constant birth rate $\alpha$ and a death rate $\beta_j=jQ_{l-1}(j)$, where $Q_{l-1}(j)$ is a non–decreasing, monic polynomial in $j$ of degree $l-1$. This gives us $l$ parameters needed to specify the distribution of $\pi$. We choose these parameters in such a way that the condition ($A_l$) is satisfied. With our choice of birth and death rates we have that $$\alpha\mu_{j-1}-\beta_{j}\mu_j \;=\; \alpha\mu_{j-1}-jQ_{l-1}(j)(a+bj^{-1})\mu_{j-1} \;=\; \mu_{j-1}[\alpha-ajQ_{l-1}(j)-bQ_{l-1}(j)].$$ Noting that $\alpha-ajQ_{l-1}(j)-bQ_{l-1}(j)$ is a polynomial of degree $l$ in $j$, and therefore has at most $l$ real roots, we have that the sequence $\{\alpha_{j-1}\mu_{j-1} - \beta_j\mu_j\}$ has at most $l$ changes of sign, so that either $W_\alpha\succeq_{l-cx}W_\beta$ or $W_\beta\succeq_{l-cx}W_\alpha$. Theorem 2.10 of Brown and Xia (2001) gives us that $$\sup \lbrace \\|\Delta Sh\\|_\infty : h(j)=I_{(j\in B)} , B\subseteq{\mathbb{Z}^+} \rbrace \leq \alpha^{-1}.$$ Hence, with $h(j)=I_{(j\in B)}$ for some $B\subseteq\mathbb{Z}^+$, $$\\|\Delta^lSh\\|_\infty \;\leq\; 2^{l-1}\\|\Delta Sh\\|_\infty \;\leq\; 2^{l-1}\alpha^{-1}.$$ From Corollary \[res4\] we thus obtain Corollary \[cor\_ex2\_gen\]. \[cor\_ex2\_gen\] With $W$ and $\pi$ as above, $$\label{21} d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi)) \;\leq\; 2^{l-1}\alpha^{-1} \left\| E\left[\alpha{W+1\choose l}-WQ_{l-1}(W){W\choose l}\right] \right\|.$$ For example, consider the case where $W\sim\mbox{Bin}(n,p)$ and $\pi\sim\mbox{PBD}(\alpha, \gamma j+j(j-1))$, so that $l=2$. Choosing our constants $\alpha$ and $\gamma$ according to the prescription above, straightforward calculations give us that $$\alpha=n(n-1)p(1-p), \hspace{12pt}\mbox{ and }\hspace{12pt} \gamma=(n-1)(1-2p).$$ Furthermore, $$\begin{gathered} E[W(W+1)]=np(np+2-p),\hspace{15pt}E[W^2(W-1)]=n(n-1)p^2(np+2-2p),\hspace{15pt}\mbox{and}\\ E[W^2(W-1)^2]=n(n-1)p^2(n^2p^2+4np-5np^2-8p+6p^2+2).\end{gathered}$$ Evaluating the bound (\[21\]) then gives \[cor\_ex2\] Assume that $W\sim\mbox{Bin}(n,p)$ and $\pi\sim\mbox{PBD}(\alpha, \gamma j+j(j-1))$. Then, $$\label{22} d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi)) \leq 2p^2.$$ We note that (\[21\]) does not necessarily give a bound of the optimal order. In the case covered by (\[22\]), Theorem 3.1 of Brown and Xia (2001) gives a bound on total variation distance of order $O(p^2/\sqrt{\lambda})$, where $\lambda=E[W]=np$. This disparity is due to our rather crude use of the supremum norm in obtaining bounds such as (\[21\]). In Sections \[tp\] and \[bd3\], we will consider more refined ways to bound the terms of our Stein equation in some particular cases when we have two parameters to choose in our approximating distribution $\pi$. Despite this disadvantage, we nevertheless note that (\[21\]) gives an explicit bound which may be applied in many contexts. Our final example of this section focuses on mixture distributions of the polynomial birth–death type. Suppose that $\pi\sim\mbox{PBD}(\alpha,\beta_j)$ and $W\sim\mbox{PBD}(\xi,\beta_j)$, for some constant birth rate $\alpha$, polynomial death rate $\beta_j$ and random variable $\xi$ on $\mathbb{R}^+$. In this case we have that $$\label{23} \mu_j = \frac{E\left[\mu_0(\xi)\xi^j\right]}{\prod_{k=1}^j\beta_k}, \quad j\geq 0.$$ We choose $\alpha$ such that $\alpha=E\beta_W$, that is, $$\alpha \;=\; E\sum_{j=0}^\infty\beta_j\mu_j \;=\; E\sum_{j=0}^\infty\xi\mu_{j-1} \;=\; E\xi.$$ Using (\[23\]), we obtain $$\begin{aligned} \alpha\mu_{j}-\beta_{j+1}\mu_{j+1} &=& E\left[\frac{\mu_0(\xi)\alpha^{j+1}}{\prod_{k=1}^j\beta_j}\left\{\left(\frac{\xi}{\alpha}\right)^j-\left(\frac{\xi}{\alpha}\right)^{j+1}\right\}\right]\\ &=&E\left[\frac{\alpha\mu_0(\xi)}{\mu_0(\alpha)} \left(1-\frac{\xi}{\alpha}\right)\pi_j\left(\frac{\xi}{\alpha}\right)^j\right].\end{aligned}$$ From this, we can see that the sequence $\{\alpha\mu_{j}-\beta_{j+1}\mu_{j+1}\}$ is monotone. Hence, Corollary \[res2\] gives us the following. \[cor\_ex3\] With $W$ and $\pi$ as above, $$\label{24} \|Eh(W)-Eh(\pi)\| \;\leq\; \lVert\Delta Sh\rVert_\infty \left\|E[\alpha (W+1) - \beta_W W]\right\|.$$ For example, if $\beta_j=j$ then $W\sim\textrm{Po}(\xi)$ and we take $\pi\sim\textrm{Po}(\lambda)$, where $\lambda=E\xi$. Using the well–known bound on the Stein operator $S$ in this case, namely $$\label{25} \\|\Delta Sh\\|_\infty \leq \; \lambda^{-1} (1-e^{-\lambda}) \; \\|h\\|_\infty,$$ evaluating (\[24\]) gives, after some straightforward calculation, $$d_{TV}(\mathcal{L}(W),\mbox{Po}(\lambda)) \;\leq\; \lambda^{-1} (1-e^{-\lambda})\mbox{Var}(\xi),$$ a bound that has also been obtained by Barbour et al. (1992, Theorem 1.C). Poisson approximation for a sum of indicators {#pois} ============================================= Throughout this section, the random variable $W$ of interest is a sum of indicators: $$W=X_1+\cdots+X_n,$$ where the $X_i$ are Bernoulli variables, possibly dependent, with $$p_i=P(X_i=1)=1-P(X_i=0), \quad 1\leq i \leq n.$$ Using Propositions \[res1\] and \[res3\], we are going to investigate the approximation of the sum $W$ by a Poisson random variable $\pi \sim \textrm{Po}(\lambda)$. Recall that our Poisson variable is derived from (\[1\]) when $\alpha_j=\lambda$ and $\beta_j=j$, so that by (\[7\]), $$\label{26} W_{\alpha}=W+1, \; \mbox { and } \; P(W_{\beta} \in B)=\frac{E[WI_{(W\in B)}]}{EW},$$ for any Borel set $B$. In the analysis, an important role will be played by the variables $$W_i=W-X_i, \quad 1\leq i \leq n.$$ Total dependence {#total} ---------------- Firstly, we consider the case where the indicators $X_i$ are totally negatively dependent in the sense of Papadatos and Papathanasiou (2002). Let us recall that $n$ random variables $X_i,\; 1\leq i \leq n$, are totally negatively dependent (TND) if $$\label{28} \textrm{Cov}[g_1(X_i),g_2(W_i)]\leq 0, \quad 1\leq i \leq n,$$ for all non-decreasing functions $g_1$, $g_2$ such that the covariance exists. Papadatos and Papathanasiou (2002, Theorem 3.1) show that the class of TND indicators includes the standard class of negatively related indicators. Stein’s method for Poisson approximation of a sum of negatively related indicators is discussed by, for example, Barbour et al. (1992) and Erhardsson (2005). Recall that indicator random variables $X_1,\ldots,X_n$ are said to be negatively related if $$\begin{gathered} \label{29} E[g(X_1,\ldots,X_{i-1},X_{i+1},\ldots,X_n)\|X_i=1] \;\leq\; E[g(X_1,\ldots,X_{i-1},X_{i+1},\ldots,X_n)],\\ 1\leq i \leq n,\end{gathered}$$ for all non-decreasing functions $g:\lbrace0,1\rbrace^{n-1} \mapsto \lbrace0,1\rbrace$. We wish to bound the total variation distance between $\mathcal{L}(W)$ and $\textrm{Po}(\lambda)$. For that, we will apply Proposition \[res1\]. By (\[26\]), we have that, for any function $g:\mathbb{Z}^+\mapsto\mathbb{R}$, $$Eg(W_\alpha)=Eg(W+1), \; \mbox { and } \; Eg(W_\beta)=\frac{E[Wg(W)]}{EW}.$$ Thus, to show that $W_\alpha \succeq_{st} W_\beta$, we must prove that if $g$ is non-decreasing, then $EW Eg(W+1)\geq E[Wg(W)]$. In fact, this was established by Papadatos and Papathanasiou (2002, Lemma 3.1). Using the bound (\[25\]) on the Stein operator in the Poisson case, (\[5\]) and (\[6\]) provide the following result. \[res6\] If the indicators $\{X_i: 1\leq i\leq n\}$ are TND, then $W_\alpha \succeq_{st}W_\beta$. If, in addition, $EW\geq\lambda$, then $$d_{TV}(\mathcal{L}(W),\textrm{Po}(\lambda)) \leq \frac{1-e^{-\lambda}}{\lambda}\; ([\lambda +1]EW - E[W^2]).$$ Further results on, and examples of, TND indicator random variables can be found in Papadatos and Papathanasiou (2002). Let us now consider the case where the indicators $X_i$ are positively dependent in a certain sense. We adapt the definition (\[28\]) and say that $n$ random variables $X_1\ldots,X_n$, are totally positively dependent (TPD) if $$\textrm{Cov}[g_1(X_i),g_2(W_i)]\geq 0, \quad 1\leq i \leq n,$$ for all non-decreasing functions $g_1$, $g_2$ such that the covariance exists. Association or positive relation is sufficient for TPD. This can be established analogously to the proof of Theorem 3.1 of Papadatos and Papathanasiou (2002). Recall that our indicator random variables are said to be positively related if (\[29\]) holds with the inequality reversed for all non-decreasing functions $g:\lbrace0,1\rbrace^{n-1} \mapsto \lbrace0,1\rbrace$. This standard property is used with Stein’s method by, for example, Barbour et al. (1992) and Erhardsson (2005). In the sequel, it is assumed that $EW=\lambda$. To get a bound for the total variation distance, we will apply Proposition \[res3\], using the lemma stated below. To begin with, we introduce a random variable $X_V$, a mixing of our $n$ indicators, in which the index $V$ is a random variable of law $$\label{32} P(V=i)=\frac{EX_i}{\lambda}, \quad 1\leq i \leq n.$$ \[res7\] If $EW=\lambda$ and the indicators $\{X_i : 1\leq i\leq n\}$ are TPD, then $$\label{33} W_\beta\succeq_{st}W_\alpha -X_V,$$ where $W_\alpha -X_V\geq 0$ a.s. As seen in (\[26\]), $W_\alpha = W+1$ and thus, $W_\alpha -X_V\geq 0$ a.s. Moreover, $W_\beta$ has the so-called $W$-size-biased distribution: see, for example, Goldstein and Rinott (1996). $W$ being a sum of indicators, it is then known that $W_\beta$ admits the representation $$\label{34} W_\beta=\sum_{i\not=V}\hat{X}_i+1,$$ where $V$ is a random variable of law (\[32\]), and if $V=v$, $$\hat{X}_i =_d (X_i\|X_v=1), \quad i\neq v.$$ Thus, by (\[34\]), the ordering (\[33\]) is equivalent to $\sum_{i\not=V}\hat{X}_i \succeq_{st}W-X_V$. To establish this, it is enough to prove that $$\sum_{i\not=v}\hat{X}_i \succeq_{st}W-X_v, \quad 1\leq v\leq n;$$ see Shaked and Shanthikumar (2007). Now, by (\[34\]) and the TPD assumption, we get, for any real $a \geq 0$, $$\begin{aligned} P(\sum_{i\not=v}\hat{X}_i>a) &=& P(\sum_{i\not=v}X_i>a\|X_v=1)\\ &\geq& P(\sum_{i\not=v}X_i>a) \;=\; P(W-X_v>a),\end{aligned}$$ which is the desired result. Thanks to Lemma \[res7\], we may apply Proposition \[res3\] with $s=p=1$. Noting that by (\[32\]), $$EX_V \;=\; \sum_{i=1}^n p_i\,P(V=i) \;=\; \frac{1}{\lambda}\,\sum_{i=1}^n p_i^2 ,$$ we then get the following result. \[res8\] If $EW=\lambda$ and the indicators $\{X_i : 1\leq i\leq n\}$ are TPD, then $$d_{TV}(\mathcal{L}(W),\textrm{Po}(\lambda)) \leq \frac{1-e^{-\lambda}}{\lambda}\; \left\{E[W^2] +2\sum_{i=1}^np_i^2 - \lambda(\lambda +1)\right\}.$$ This bound is obtained (and applied) by Barbour et al. (1992, Corollary 2.C.4) under the condition of positive relation. See also Erhardsson (2005). Local dependence {#local} ---------------- Our goal in this part is to combine the previous $s$-convex ordering approach with a more flexible property of dependence. More precisely, we first introduce a concept of local dependence between a set of $n$ indicators $X_1,\ldots,X_n$. Let $\mathcal{F}_s $ be the class of all functions $g: \lbrace0,1\rbrace^{n-1} \mapsto \mathbb{R^+}$ that are non-decreasing and $s$-convex with $g(0)=0$. We say that the $n$ indicators $X_1\ldots,X_n,$ are $(s,{\bf \delta})$[-locally negatively dependent]{} ($(s,{\bf \delta})$-LND) if there exist $n$ non-negative reals $\delta_1\ldots,\delta_n$ (of sum $>0$) such that $$\label{37} E[X_ig(W_i)] \;\leq\; \delta_i\, E[g(W_i)] \; \mbox { for all functions } g\in \mathcal{F}_s, \quad 1\leq i \leq n.$$ Similarly, $X_1\ldots,X_n,$ are said to be $(s,{\bf \delta})$-locally positively dependent ($(s,{\bf \delta})$-LPD) if $$\label{38} E[X_ig(W_i)] \;\geq\; \delta_i \, E[g(W_i)] \; \mbox { for all functions } g\in \mathcal{F}_s, \quad 1\leq i \leq n.$$ Let $\delta:=\delta_1+\ldots+ \delta_n$, and denote $$v_i= \delta^{-1}\delta_i, \quad 1\leq i \leq n, \; \mbox{ and }\; p=EW/\delta \wedge \delta/EW.$$ We then adopt the notation $v_p$ and $X_V$ of Sections \[intro\] and \[total\]. \[res9\] If the indicators $\{X_i : 1\leq i\leq n\}$ are $(s,{\bf \delta})$-LND, then $$\label{40} W_{\alpha} \succeq_{s-cx} v_p W_{\beta},$$ while if the indicators $\{X_i : 1 \leq i\leq n\}$ are $(s,{\bf \delta})$-LPD, then $$\label{41} W_{\beta} \succeq_{s-cx} v_p (W_{\alpha} -X_V).$$ The method of proof is built on ideas in Barbour at al. (1992), Goldstein and Rinott (1996), Papadatos and Papathanasiou (2002) and Reinert (2005). Let $g$ be any function belonging to $\mathcal{F}_s$. As a preliminary, we observe that $W\leq W_i+1\leq W+1$ a.s. for each $i=1,\ldots,n$. Now, consider the case of $(s,{\bf \delta})$-LND. Using (\[37\]) and the assumption that $g$ is non-decreasing, we obtain that $$\begin{aligned} E[Wg(W)] &=&\sum_{i=1}^n E[X_ig(W)] \;\;=\;\; \sum_{i=1}^n E[X_ig(W_i+1)] \;\;\leq\;\; \sum_{i=1}^n\delta_i E[g(W_i+1)]\nonumber\\ &\leq& \sum_{i=1}^n\delta_i E[g(W+1)] \;\;=\;\; \delta E[g(W_{\alpha})].\end{aligned}$$ As $g(0)=0$, and $EW/\delta \geq p\in (0,1]$, we find from (\[43\]) that $$\begin{aligned} E[g(W_{\alpha})] &\geq& \frac{E[Wg(W)]}{EW}\; \frac{EW}{\delta}\\ &\geq& pE[g(W_{\beta})] \;\;=\;\; E[g(v_p W_{\beta})],\end{aligned}$$ hence the ordering (\[40\]). The case of $(s,{\bf \delta})$-LPD is treated similarly. By (\[38\]) and since $g$ is non-decreasing, we get $$\begin{aligned} \label{43} E[Wg(W)] &=&\sum_{i=1}^n E[X_i g(W_i+1)] \;\;\geq\;\; \sum_{i=1}^n\delta_i E[g(W_i+1)]\nonumber\\ &=& \delta\sum_{i=1}^nP(V=i) E[g(W+1-X_i)] \;\;=\;\; \delta E[g(W_{\alpha}-X_V)].\end{aligned}$$ As before, we then deduce from (\[43\]) that $$\begin{aligned} E[g(W_{\beta})]&=& \frac{E[Wg(W)]}{EW} \;\;\geq\;\; E[g(W_{\alpha}-X_V)]\;\frac{\delta}{EW}\\ &\geq& p E[g(W_{\alpha}-X_V)] \;\;=\;\; E[g(v_p(W_{\alpha}-X_V))],\end{aligned}$$ proving the ordering (\[41\]). Combining Proposition \[res3\] and Lemma \[res9\] would then allow us to derive an upper bound for the total variation distance. Approximate local dependence {#aplocal} ---------------------------- Approximate local dependence is becoming a rather popular topic in probability. For works related to this idea, see for example Chen (1975), Barbour et al. (1992) and Chatterjee et al. (2005). We wish now to to derive an abstract Poisson approximation theorem by combining stochastic ordering with such an approach. We say that the $n$ indicators $X_1,\ldots,X_n$ are approximately locally negatively dependent (ALND) if there exist $n$ non-negative reals $\delta_1,\ldots,\delta_n$ (of sum $\delta>0$), and $n$ random variables $Y_1,\ldots,Y_n$ on $\mathbb{Z}^+$ such that $$\label{44} E[X_ig(W_i-Y_i)] \;\leq\; \delta_i \, E[g(W_i-Y_i)], \quad 1\leq i \leq n,$$ for all non-negative, non-decreasing functions $g$. Similarly, $X_1,\ldots,X_n$ are said to be approximately locally positively dependent (ALPD) if $$\label{45} E[X_ig(W_i-Y_i)] \;\geq\; \delta_i \, E[g(W_i-Y_i)], \quad 1\leq i \leq n,$$ for all non-negative, non-decreasing functions $g$. Define $$\varepsilon = \sum_{i=1}^n E[X_iY_i], \; \mbox { and } \; \varepsilon_\ast = \varepsilon + \sum_{i=1}^n\delta_i E[X_i+Y_i],$$ and let $$c_\lambda = (\lambda+1)(1-e^{-\lambda})/\lambda + 2d_\lambda, \; \mbox { with } \; d_\lambda = 1 \wedge \sqrt{2/e\lambda}.$$ \[res11\] If $EW=\lambda$ and the indicators $\{X_i : 1\leq i\leq n\}$ are ALND, then $$\label{47} d_{TV}(\mathcal{L}(W),\textrm{Po}(\lambda)) \leq \frac{1-e^{-\lambda}}{\lambda}\; \left(\|\mbox{Var}(W)-\lambda\|+ 2\varepsilon\right) + c_\lambda\; \|\delta-\lambda\|,$$ while if the indicators $\{X_i : 1\leq i\leq n\}$ are ALPD, then $$d_{TV}(\mathcal{L}(W),\textrm{Po}(\lambda)) \leq \frac{1-e^{-\lambda}}{\lambda} \; (\|\mbox{Var}(W)-\lambda\|+2\varepsilon_\ast)+ c_\lambda \;\|\delta-\lambda\|.$$ Before proving Theorem \[res11\], we give an example of its application. We examine a variation of the classical birthday problem; see also Barbour et al. (1992). Suppose we independently colour $N\geq2$ points with one of $m$ colours, each colour being chosen equiprobably. Let $\Gamma$ be the set of all subsets $i\subseteq\{1,\ldots,N\}$ of size $2$. For $i\in \Gamma$, let $Z_i$ be the indicator that the points indexed by $i$ have the same colour. Moreover, suppose we choose uniformly $r$ of the $\|\Gamma\|={N\choose 2}$ pairs of points, independently of the colourings chosen. For $i\in\Gamma$, we let $\xi_i=0$ if the pair of points indexed by $i$ is chosen, and otherwise set $\xi_i=1$. Set $W=\sum_{i\in\Gamma}Z_i\xi_i$. This counts the number of pairs of points with the same colour, excluding those $r$ pairs of points we have chosen. In the case where $r=0$, this corresponds to the classical birthday problem. A bound in the Poisson approximation of $W$ in this case is given by Arratia et al. (1989, Example 2). We observe that for all $i,j\in\Gamma$, $E\left[Z_i\right]=m^{-1}$ and $E\left[Z_iZ_j\right]=m^{-2}$. Furthermore, $$E\left[\xi_i\right]=\frac{{N\choose 2}-r}{{N\choose 2}},\hspace{12pt} \mbox{ and }\hspace{12pt} E\left[\xi_i\xi_j\right]= \frac{{N\choose 2}-r}{{N\choose 2}} \left(\frac{{N\choose 2}-r-1}{{N\choose 2}-1}\right), \quad i\not=j.$$ Straightforward calculations then give $$\lambda=E[W]=\frac{{N\choose 2}-r}{m},\hspace{12pt} \mbox{ and }\hspace{12pt} \lambda-\mbox{Var}(W)=\frac{{N\choose 2}-r}{m^2}.$$ Now, we write $W_i=W-Z_i\xi_i\;$ and choose $$Y_i=\sum_{j\not=i}Z_j\xi_jI_{(i\cap j\not=\emptyset)}, \;\;\mbox{ and }\;\; \delta_i=E\left[Z_i\xi_i\right].$$ The condition (\[44\]) holds true with these choices. Indeed, $W_i-Y_i$ is independent of $Z_i$ and the $\xi_i$ are negatively related by construction. Thus, for all non–decreasing functions $g$, we have $$E\left[Z_i\xi_ig(W_i-Y_i)\right] \;\;=\;\; E\left[Z_i\xi_i\right]E\left[g(W_i-Y_i)\|\xi_i=1\right] \;\;\leq\;\; E\left[Z_i\xi_i\right]E\left[g(W_i-Y_i)\right],$$ as required. We further see that $$\begin{aligned} \varepsilon \;=\; \sum_{i\in\Gamma}E[Z_i\xi_iY_i] &=& \sum_{i\in\Gamma}\sum_{j\not=i}E\left[Z_iZ_j\right] E\left[\xi_i\xi_j\right]I_{(i\cap j\not=\emptyset)}\\ &=& \frac{2(N-1)\left\{{N\choose 2}-r\right\} \left\{{N\choose 2}-r-1\right\}}{m^2\left\{{N\choose 2}-1\right\}}.\end{aligned}$$ Evaluating (\[47\]) then gives the following bound. \[cor\_ex4\] With $W$ as above, $$d_{TV}(\mathcal{L}(W),\textrm{Po}(\lambda)) \leq \frac{1-e^{-\lambda}}{m}\left\{1+4(N-1)\left(\frac{{N\choose 2}-r-1}{{N\choose 2}-1}\right)\right\}.$$ In the case $r=0$, a bound of the same order was established by Arratia et al. (1989, Example 2). Proof of Theorem \[res11\] {#proof1} -------------------------- \(i) Consider the ALND case. We suppose first that $f$ is any non-negative, non-decreasing function. Arguing as for Lemma \[res9\], we have $$\begin{aligned} E[Wf(W)] &=&\sum_{i=1}^n E[X_if(W)] \;\;=\;\; \sum_{i=1}^n E[X_if(W_i+1)]\\ &=& \sum_{i=1}^n E[X_if(W_i-Y_i+1)] +\sum_{i=1}^n E\{X_i[f(W_i+1)-f(W_i-Y_i+1)]\},\end{aligned}$$ which we denote by $T_1+T_2$. We bound the sum $T_2$ by noting that $$\|f(x)-f(y)\| \;\leq\; \\|\Delta f\\|_\infty\; \|x-y\|,$$ which yields $$T_2 \;\leq\; \\|\Delta f\\|_{\infty} \sum_{i=1}^n E(X_i Y_i) \;=\; \\|\Delta f\\|_{\infty} \; \varepsilon.$$ For the sum $T_1$, by (\[44\]) and since $f$ is non-decreasing, we get $$T_1 \;\leq\; \sum_{i=1}^n\delta_i E[f(W_i-Y_i+1)] \;\leq\; \sum_{i=1}^n \delta_i E[f(W+1)] \;=\; \delta E[f(W+1)].$$ Inserting these two bounds, we find that $$\begin{aligned} \label{50} E[Af(W)]&=&\lambda E[f(W+1)] - E[Wf(W)] \nonumber\\ &\geq & -(\delta-\lambda) E[f(W+1)] - \\|\Delta f\\|_\infty \; \varepsilon.\end{aligned}$$ To get an upper bound, we define a function $\tilde{f}$ on $\{0, 1, \ldots, n-1\}$ by $$\label{51} \tilde{f}(x)=\\|f\\|_\infty + \\|\Delta f\\|_\infty\; x - f(x).$$ Note that $\tilde{f}$ is, as $f$, a non-negative, non-decreasing function. By assumption, $EW=\lambda$ so that $E[A1]=0$; observe also that $E[AW]=\lambda E[W+1]- E[W^2] = -[\mbox{Var}(W)-\lambda]$. Thus, $$\begin{aligned} E[A\tilde{f}(W)]&=&\\|g\\|_\infty \; E[A1] + \\|\Delta f\\|_\infty \; E[AW]- E[Af(W)]\\ &=& -\\|\Delta f\\|_\infty \; [\mbox{Var}(W)-\lambda]- E[Af(W)].\end{aligned}$$ On the other hand, (\[50\]) is applicable to the function $\tilde{f}$, so that $$E[A \tilde{f}(W)] \geq -(\delta-\lambda) E[\tilde{f}(W+1)] -\\|\Delta \tilde{f}\\|_\infty \; \varepsilon.$$ From these two formulas, we deduce that $$\label{52} E[Af(W)] \leq \\|\Delta \tilde{f}\\|_\infty \; \varepsilon + (\delta-\lambda) E[\tilde{f}(W+1)] + \\|\Delta f\\|_\infty \; \|\mbox{Var}(W)-\lambda\|.$$ Now, let $f$ be an arbitrary function. We start with the standard decomposition $f=f_+ - f_-$, where $f_+$ and $f_-$ are non-negative, non-decreasing functions with, of course, $$\label{53} \\|\Delta^j f_+\\|_\infty\leq \\|\Delta^j f\\|_\infty, \; \mbox { and }\; \\|\Delta^j f_-\\|_\infty\leq \\|\Delta^j f\\|_\infty, \quad j=0,1.$$ By (\[50\]) and (\[52\]), we obtain an upper bound $$\begin{aligned} E[Af(W)]&=& E[Af_+(W)] - E[Af_-(W)]\\ &\leq& \\|\Delta \tilde{f_+}\\|_{\infty} \;\varepsilon + (\delta-\lambda) E[\tilde{f_+}(W+1)] + \\|\Delta f_+\\|_\infty \; \|\mbox{Var}(W)-\lambda\|\\ &&\qquad +(\delta-\lambda) E[f_-(W+1)] +\\|\Delta f_-\\|_\infty \; \varepsilon\\ &=& \\|\Delta f_+\\|_\infty \; \|\mbox{Var}(W)-\lambda\| + (\\|\Delta \tilde{f_+}\\|_{\infty}+\\|\Delta f_-\\|_\infty) \;\varepsilon\\ &&\qquad + (\delta-\lambda) \; \{\\|f_+\\|_\infty + \\|\Delta f_+\\|_\infty \;(\lambda +1) -E[f(W+1)]\},\end{aligned}$$ using (\[51\]) and $EW=\lambda$ for the last equality. By a similar method, we find as a lower bound $$\begin{aligned} E[Af(W)] &\geq& - (\delta-\lambda) E[{f_+}(W+1)] - \\|\Delta f_+\\|_\infty \; \varepsilon\\ && \qquad - \\|\Delta \tilde{f_-}\\|_\infty\; \varepsilon -(\delta-\lambda) E[\tilde{f_-}(W+1)] -\\|\Delta f_-\\|_\infty \; \|\mbox{Var}(W)-\lambda\|\\ &=& -\\|\Delta f_-\\|_\infty \; \|\mbox{Var}(W)-\lambda\| - (\\|\Delta f_+\\|_\infty + \\|\Delta \tilde{f_-}\\|_\infty)\; \varepsilon\\ &&\qquad -(\delta-\lambda) \; \{\\|f_-\\|_\infty + \\|\Delta f_-\\|_\infty \;(\lambda +1) +E[f(W+1)]\}.\end{aligned}$$ By (\[53\]) and since $\\|\Delta \tilde{f}\\|_\infty \leq \\|\Delta f\\|_\infty$, combining the two previous bounds then yields $$\label{54} \|E[Af(W)]\|\leq \\|\Delta f\\|_\infty \; (\|\mbox{Var}(W)-\lambda\| +2 \varepsilon) + \|\delta-\lambda\| \;[2 \\|f\\|_\infty + \\|\Delta f\\|_\infty \;(\lambda +1)].$$ With $f=Sh$, it now suffices to apply in (\[54\]) the standard bounds $$\\|\Delta Sh\\|_\infty \leq \lambda^{-1}(1-e^{-\lambda}) \\|h\\|_\infty, \mbox { and } \; \\|Sh\\|_\infty\leq d_\lambda \\|h\\|_\infty,$$ which gives (\[47\]). \(ii) The ALPD case is dealt with analogously. For $f$ non-negative, non-decreasing, we first write that $$\begin{aligned} E[Wf(W)] & = &\sum_{i=1}^n E[X_if(W_i-Y_i+1)] +\sum_{i=1}^n E[X_i\{f(W_i+1)-f(W_i-Y_i+1)\}]\\ &\geq &\sum_{i=1}^n E[X_if(W_i-Y_i+1)] - \\|\Delta f\\|_\infty \; \varepsilon.\end{aligned}$$ By (\[45\]), we then get that $$\begin{aligned} E[Wf(W)] &\geq& \sum_{i=1}^n\delta_i E[f(W_i-Y_i+1)]- \\|\Delta f\\|_\infty \;\varepsilon\\ &=& \delta E[f(W+1)] - \sum_{i=1}^n\delta_i E[f(W+1)-f(W_i-Y_i+1)] - \\|\Delta f\\|_\infty \;\varepsilon\\ &\geq& \delta E[f(W+1)] - \\|\Delta f\\|_\infty \; \sum_{i=1}^n\delta_i E(X_i+Y_i) - \\|\Delta f\\|_\infty \; \varepsilon \\ &=& \delta E[f(W+1)] - \\|\Delta f\\|_\infty \;\varepsilon_\ast.\end{aligned}$$ Overall, we find that $$E[Af(W)] = \lambda E[f(W+1)] - E[Wf(W)] \geq -(\delta-\lambda) E[f(W+1)] +\\|\Delta f\\|_\infty\; \varepsilon_\ast.$$ The rest of the proof follows as in the ALND case. Translated Poisson approximation {#tp} ================================ We assume, as in Section \[pois\], that $W=X_1+\cdots+X_n$ is a sum of (possibly dependent) indicator random variables, with $p_i=P(X_i=1)$. Denote $$\lambda_k=\sum_{i=1}^np_i^k, \quad \lambda=\lambda_1=E[W], \quad \mbox {and}\quad \sigma^2=\mbox{Var}(W).$$ We are going to discuss the approximation of $W$ by a translated Poisson distribution. Main results ------------ A random variable $Z$ has a translated Poisson distribution $\textrm{TP}(\lambda,\sigma^2)$ if $Z$ is distributed as $Z^\prime+\rho$, where $Z^\prime\sim\mbox{Po}(\sigma^2+\gamma)$ with $$\rho=\lambda-\sigma^2-\gamma, \;\;\mbox{ and }\;\; \gamma=\langle\lambda-\sigma^2\rangle\in[0,1),$$ $\langle x\rangle=x-\lfloor x\rfloor$ denoting the fractional part of $x$. We note that $E[Z]=\lambda$ and $\sigma^2\leq\mbox{Var}(Z)= \sigma^2+\gamma<\sigma^2+1$, so that our approximating translated Poisson distribution has a mean equal to, and variance close to, that of $W$. We would thus expect a closer approximation than could be obtained by simply using the one–parameter Poisson distribution. The variances of $W$ and $Z$ cannot necessarily be made to match exactly, as we must shift our Poisson distribution by an integer. However, the error term arising from this mismatch does not adversely affect the order of the bounds we obtain, as we shall see below. The following results give us bounds in translated Poisson approximation for $W$ under some stochastic ordering assumptions. We defer the proofs of Theorems \[tpneg\] and \[tppos\] until Section \[tpproofs\], giving first some examples of their application, in Section \[runs\]. Our bounds demonstrate convergence to a translated Poisson distribution if $\sigma\rightarrow\infty$ as $n\rightarrow\infty$. Bounds on the total variation distance between $\mathcal{L}(W)$ and a translated Poisson random variable may still be found if this is not the case, but require a different analysis of the error terms. For example, in proving Theorems \[tpneg\] and \[tppos\], we write $P(W-\rho<0)\leq\sigma^{-2}$. This error term may be reduced, or even omitted altogether depending on the problem at hand, with a more careful analysis. This could give us good bounds in cases where $\sigma\rightarrow\sigma_\infty<\infty$ as $n\rightarrow\infty$. In the sequel, we let $W^s$ be a random variable having the $W$–size–biased distribution, and $v_q$ be an indicator random variable, independent of all else, with $P(v_q=1)=q$. As before, we write $W_i=W-X_i$, $1\leq i\leq n$, and for any random index $V$ we let $W_V=W-X_V$. \[tpneg\] Suppose that $X_1,\ldots,X_n$ are negatively related, and there is $q\in[0,1]$ and $l\in\mathbb{Z}^+$ such that $$\label{56} (W+1\|X_k=0)\preceq_{st}(W+l+v_q\|X_k=1),\quad 1\leq k\leq n.$$ Then, $$\begin{gathered} \label{57} d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq \frac{2}{\sigma^2} + \frac{\lambda_2+(l+q)(\lambda-\lambda_2)}{\lambda\sigma} \\ + \frac{l(l+2q-1)(\lambda-\lambda_2)}{\sigma^2}\,d_{TV}(\mathcal{L} (W^s),\mathcal{L}(W^s+1)).\end{gathered}$$ \[tppos\] Suppose that $X_1,\ldots,X_n$ are positively related, and there is $q\in[0,1]$ and $l\in\mathbb{Z}^+$ such that $$\label{58} (W+1\|X_k=0)\succeq_{st}(W-l-v_q\|X_k=1),\quad 1\leq k\leq n.$$ Then, $$\begin{gathered} \label{59} d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq \frac{2}{\sigma^2} + \frac{\lambda_2+(l+q)(\lambda-\lambda_2)}{\lambda\sigma} \\ + \frac{(l+1)(l+2q)(\lambda-\lambda_2)}{\sigma^2}\,d_{TV}(\mathcal{L}(W^s), \mathcal{L}(W^s+1)).\end{gathered}$$ Consider the stochastic ordering assumptions (\[56\]) and (\[58\]). We note that the choice of $l$ and $q$ is not unique, in that choosing $l=m$, $q=1$ gives the same assumption as choosing $l=m+1$, $q=0$. It is easily checked, however, that each of these choices gives rise to the same bounds in (\[57\]) and (\[59\]). In the examples below, we will verify the validity of such stochastic orderings by using an appropriate coupling argument. Applications {#runs} ------------ Suppose that $X_1,\ldots,X_n$ are independent. Thus, they are also negatively related. Moreover, the condition (\[56\]) is true for $q=l=0$. Therefore, (\[57\]) is applicable and yields the following. \[cor\_ex5\] With $W$ as above, $$d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq \frac{\lambda_2}{\lambda\sigma} + \frac{2}{\sigma^2}.$$ This bound is of the order we would expect: see also Čekanavičius and Vaǐtkus (2001). Suppose that $m$ balls are placed into $N$ urns, in such a way that no urn contains more than one ball and all arrangements are equally likely. Let $W$ be the number of balls in the first $n$ urns. Thus, $W$ has a hypergeometric distribution with $$\lambda=\frac{mn}{N},\;\;\mbox{ and }\;\;\sigma^2=\frac{mn(N-m)(N-n)}{(N-1)N^2}.$$ We set $X_i$ to be the indicator that the $i$th urn contains a ball, so that $W=X_1+\cdots+X_n$. By construction, these indicators are negatively related. The condition (\[56\]) holds for $q=1$ and $l=0$. To see this, we construct $(W+1\|X_k=0)$ by considering the $N$ urns and excluding the $k$th. Distribute the $m$ balls in these $N-1$ urns, such that all arrangements are equally likely, and count the number of the first $n$ urns that are occupied. Adding one to this count gives us our random variable. We then choose (uniformly and independently of what has gone before) one of the occupied urns. Take the ball from this urn and place it in urn $k$. This gives us $(W+1\|X_k=1)$. If the ball chosen is from one of the first $n$ urns, the number of occupied urns is the same as before. Otherwise, we have increased the number of occupied urns within the first $n$. Evaluating the bound (\[57\]) then gives Corollary \[cor\_ex6\]. \[cor\_ex6\] For $W$ having our hypergeometric distribution, $$d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \;\leq\; \frac{1}{\sigma}+\frac{2}{\sigma^2} \;=\; \sqrt{\frac{N^2(N-1)}{mn(N-m)(N-n)}} + \frac{2N^2(N-1)}{mn(N-m)(N-n)}.$$ Röllin (2007, Section 4.1) has considered translated Poisson approximation for the hypergeometric distribution, and shows that if $m=O(n)$ and $N=O(n)$, then one gets a bound in total variation distance of order $O(1/\sqrt{n})$. This order is also reflected in our result. Suppose $\xi_1,\ldots,\xi_n$ are i.i.d. Bernoulli random variables with $$p=P(\xi_i=1)=1-P(\xi_i=0), \quad 1\leq i \leq n.$$ Fix an integer $k\geq2$, and define $$X_i=\xi_{i}\xi_{i+1}\cdots\xi_{i+k-1}, \quad \mbox {and} \quad W=\sum_{i=1}^nX_i,$$ in which, to avoid edge effects, all indices are treated modulo $n$. Thus, $W$ counts the number of $k$–runs in our Bernoulli trials. Observe that $$\lambda= np^k, \quad \lambda_2=np^{2k}, \quad \mbox{and}\quad \sigma^2=\frac{np^k}{1-p}\;(1+p-p^k[2+(2k-1)(1-p)]).$$ Translated Poisson approximation for $k$–runs was treated by Röllin (2005, Section 3.2), who gives a bound in total variation distance of the form $K/\sqrt{n}$, for some constant $K=K(k,p)$ independent of $n$. Barbour and Xia (1999, Section 5) also give a bound of this order for 2–runs. We shall use our Theorem \[tppos\] to give an explicit bound with this same order. It is easily seen that the variables $X_1,\ldots,X_n$ are positively related. The condition (\[58\]) holds by choosing $q=1$ and $l=2k-3$. To see that, consider the following construction. Given the Bernoulli random variables $\xi_1,\ldots,\xi_n$, fix some $m\leq n$ and set $\xi_{m}=\xi_{m+1}=\cdots=\xi_{m+k-1}=1$, while the others remain independent Bernoulli random variables with parameter $p$. Counting the number of $k$–runs in these $n$ Bernoulli trials gives us $(W\|X_m=1)$. Suppose now we resample the random variables $\xi_m,\ldots,\xi_{m+k-1}$, conditional on at least one of these being zero. Counting the number of $k$–runs now gives us $(W\|X_m=0)$. In this resampling procedure, one can remove at most $2k-1$ of the $k$–runs that were originally present. Thus, our construction implies that $(W\|X_m=0) + 2k-1 \geq (W\|X_m=1)$, or, equivalently, $(W+1\|X_m=0)\geq(W-2k+2\|X_m=1)$, hence the announced values of $q$ and $l$. Following the work of Section \[pois\], to construct $W^s$ we choose an index $V$ uniformly from $\{1,\ldots,n\}$, and set $\xi_V=\xi_{V+1}=\cdots=\xi_{V+k-1}=1$, while the other $\xi_i$ remain independent Bernoulli random variables with parameter $p$. Lemma 2.1 of Wang and Xia (2008) thus gives us that $$d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1)) \leq 1 \wedge \frac{2.3}{\sqrt{(n-k-1)p^k(1-p)^3}}.$$ Using this, Theorem \[tppos\] yields the following. \[runsprop\] Let $W$ count the number of $k$–runs in $n$ independent Bernoulli trials, each with success probability $p$. Then, $$\begin{gathered} \label{60} d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq \frac{2}{\sigma^2} + \frac{p^k+(2k-2)(1-p^k)}{\sigma}\\ + \frac{(2k-2)(2k-1)np^k(1-p^k)}{\sigma^2} \left(1\wedge\frac{2.3}{\sqrt{(n-k-1)p^k(1-p)^3}}\right).\end{gathered}$$ Our bound (\[60\]) has the same order as that of Röllin (2005, Theorem 5) and Barbour and Xia (1999, Theorem 5.2) (this latter result applying only to the 2–runs case). Numerical comparison of the bounds shows that ours generally performs well compared to these other bounds, often giving a better result. Table 1 gives some illustrations, with values for comparison taken from Röllin (2005). $p=0.10$ $p=0.25$ $p=0.50$ $p=0.75$ $p=0.90$ ------------- ----- ---------- ---------- ---------- ---------- ---------- (a) 0.1553 0.0675 0.0500 0.0814 0.2512 $n=10^6$ (b) 0.4463 0.2334 0.1747 0.5528 $>1$ (c) 0.0304 – 0.1251 0.6014 – (a) 0.0155 0.0067 0.0050 0.0081 0.0251 $n=10^8$ (b) 0.0445 0.0233 0.0175 0.0553 0.2554 (c) 0.0030 – 0.0125 0.0601 – (a) 0.0016 0.0007 0.0005 0.0008 0.0025 $n=10^{10}$ (b) 0.0045 0.0023 0.0017 0.0055 0.0255 (c) 0.0003 – 0.0013 0.0060 – : Numerical comparisons for 2–runs. Upper bounds on total variation distance from (a) our result (\[60\]), (b) Röllin (2005) and (c) Barbour and Xia (1999). Missing values are due to restrictions on choice of parameters. Proof of Theorems \[tpneg\] and \[tppos\] {#tpproofs} ----------------------------------------- Our proof is based on that of Propositions \[res1\] and \[res3\], using the characterising operator for the Poisson distribution. We find representations of our Stein equation in conjunction with which our dependence and stochastic ordering assumptions may be applied. Throughout this section we let $f=Sh$ be the solution to the Stein equation (\[2\]) with the choices $\alpha_j=\sigma^2+\gamma$ and $\beta_j=j$, corresponding to the Poisson distribution with mean $\sigma^2+\gamma$. We suppose the test function $h$ has the form $h(j)=I_{(j\in B)}$ for some $B\subseteq\mathbb{Z}^+$. We write $g_B(j)=f(j-\rho)$. We note that $g_B$ depends on the choice of set $B$, though for notational convenience we will often write simply $g$ for $g_B$. We note further that bounds on the supremum norm of $f$ also apply to $g$, so that in particular $\\|\Delta g_B\\|_\infty\leq\sigma^{-2}$ for each $B\subseteq\mathbb{Z}^+$. Following Röllin (2007, Section 3), we obtain from the Stein equation that $$\label{61} d_{TV}(\mathcal{L}(W),TP(\lambda,\sigma^2))\leq \sup_{B\subseteq\mathbb{Z}^+} \|E[(\sigma^2+\gamma)g_B(W+1)-(W-\rho)g_B(W)]\| + P(W-\rho<0).$$ One may bound $P(W-\rho<0)\leq\sigma^{-2}$ using Chebyshev’s inequality. So, we now concentrate on the first term on the right–hand side of (\[61\]). Throughout our proof, we will make use of the following equalities in distribution: $$\label{tp1} (W\|X_V=1)=_dW^s,\hspace{15pt}\mbox{ and }\hspace{15pt}(W_V\|X_V=0)=_d(W\|X_V=0).$$ . For this part of the proof, we will consider separately the cases where $\sigma^2\leq\lambda$ and $\sigma^2\geq\lambda$. We begin by assuming $\sigma^2\leq\lambda$, so that $\rho\geq0$. Recall that $$\label{62} E[Wg(W)]=\lambda E[g(W^s)].$$ Using (\[62\]), we can then write that $$\label{63} E[(\sigma^2+\gamma)g(W+1)-(W-\rho)g(W)] = \lambda E[g(\widetilde{W})-g(W^s)],$$ where $$P(\widetilde{W}=j)= \lambda^{-1}\left\{(\sigma^2+\gamma)P(W+1=j)+\rho P(W=j)\right\}, \quad j\geq0.$$ That is, $\widetilde{W}=W+v_r$ where $v_r$ is a Bernoulli variable with success probability $r=\lambda^{-1}(\sigma^2+\gamma)$. Note that $r\leq1$ by assumption. We rewrite (\[63\]) as $$\label{64} \lambda E[g(\widetilde{W})-g(W^s)] = \lambda E[g(\widetilde{W})-g(\overline{W})] + \lambda E[g(\overline{W})-g(W^s)],$$ by defining $\overline{W}=W_V+1$, where $V$ is a random index chosen according to (\[32\]). For the first term in (\[64\]) we note that, by conditioning on $v_r$, $$\label{tp2} \lambda Eg(\widetilde{W}) \;=\; \lambda Eg(W+v_r) \;=\; (\sigma^2+\gamma)E\Delta g(W) + \lambda Eg(W).$$ Furthermore, by conditioning on $X_V$ and using the equalities (\[tp1\]), $$\label{tp3} \lambda Eg(\overline{W}) \;=\;\lambda Eg(W_V+1) \;=\;\lambda_2 Eg(W^s) + (\lambda-\lambda_2) E[g(W)\|X_V=0],$$ since $P(X_V=1)=\lambda^{-1}\lambda_2$. Again by considering conditioning on $X_V$ and using (\[tp1\]), we have that $$\label{tp4} (\lambda-\lambda_2) E[g(W)\|X_V=0] \;=\; \lambda Eg(W+1) - \lambda_2 Eg(W^s+1).$$ Combining (\[tp2\]), (\[tp3\]) and (\[tp4\]) we obtain the following. $$\begin{aligned} \nonumber \lambda E[g(\widetilde{W})-g(\overline{W})] &=& (\sigma^2+\gamma-\lambda)E\Delta g(W) + \lambda_2E\Delta g(W^s)\\ \nonumber &=& \lambda_2E[\Delta g(W^s)-\Delta g(W)] + \gamma E\Delta g(W) \\ \label{66} &&\qquad\qquad+ (\sigma^2-\lambda+\lambda_2)E\Delta g(W).\end{aligned}$$ Now consider the second term of (\[64\]). Let us combine it with the final term of (\[66\]). Since $$E[\overline{W}-W^s]=-\lambda^{-1}(\sigma^2-\lambda+\lambda_2),$$ and proceeding as we did in deriving (\[3\]), we get that $$\begin{gathered} \label{tp5} \lambda E[g(\overline{W})-g(W^s)] + (\sigma^2-\lambda+\lambda_2)E\Delta g(W)\\ = \lambda E\sum_{j=0}^\infty\left(\Delta g(j) - \Delta g(W)\right) \left[P(\overline{W}> j)-P(W^s> j)\right].\end{gathered}$$ Using the definition of $\overline{W}$, conditioning on $X_V$ and employing (\[tp1\]), we have that $$\begin{gathered} \label{tp6} \lambda\big[P(\overline{W}> j)-P(W^s> j)\big] \\ =(\lambda-\lambda_2)\big[P(W_V+1> j\|X_V=0)-P(W_V+1> j\|X_V=1)\big].\end{gathered}$$ Hence, the right–hand side of (\[tp5\]) becomes $$\label{67} (\lambda-\lambda_2) E\sum_{j=0}^\infty\left(\Delta g(j) - \Delta g(W)\right)\big[P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1)\big].$$ Let us now insert the representations (\[66\]) and (\[67\]) into (\[63\]) and then (\[61\]). We obtain $$\begin{gathered} d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq (\lambda-\lambda_2)\sup_{B\subseteq\mathbb{Z}^+}\big\{\Lambda_B\big\} +\lambda_2\sup_{B\subseteq\mathbb{Z}^+}\big\|E[\Delta g_B(W^s)-\Delta g_B(W)]\big\|\\ + \gamma\sup_{B\subseteq\mathbb{Z}^+}\big\|E\Delta g_B(W)\big\| + P(W-\rho<0),\end{gathered}$$ where $$\Lambda_B = E\sum_{j=0}^\infty\|\Delta g_B(j)-\Delta g_B(W)\|\|P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1)\|.$$ Recalling that $P(W-\rho<0)\leq\sigma^{-2}$, $\gamma\leq1$ and $\\|\Delta g_B\\|_\infty\leq\sigma^{-2}$, we have that $$\gamma\sup_{B\subseteq\mathbb{Z}^+}\big\|E\Delta g_B(W)\big\| + P(W-\rho<0) \;\leq\; 2\sigma^{-2}.$$ Furthermore, the random variable $W^s$ having the $W$–size–biased distribution satisfies $$P(W^s=j) \;=\; \lambda^{-1}jP(W=j),\quad 0\leq j\leq n,$$ and so, $$\label{71} 2d_{TV}(\mathcal{L}(W),\mathcal{L}(W^s)) \;=\; \sum_{j=0}^\infty\|P(W=j)- P(W^s=j)\| \;=\; E\|1-\lambda^{-1}W\| \;\leq\; \lambda^{-1}\sigma.$$ We thus have that $$\lambda_2\big\|E[\Delta g_B(W^s)-\Delta g_B(W)]\big\| \;\leq\; 2\lambda_2\\|\Delta g_B\\|_\infty d_{TV}(\mathcal{L}(W),\mathcal{L}(W^s)) \;\leq\; \frac{\lambda_2}{\lambda\sigma}.$$ Combining the above bounds, we obtain $$\label{68} d_{TV}(\mathcal{L}(W),\mbox{TP}(\lambda,\sigma^2)) \leq (\lambda-\lambda_2)\sup_{B\subseteq\mathbb{Z}^+}\big\{\Lambda_B\big\} + \frac{\lambda_2}{\lambda\sigma} + \frac{2}{\sigma^2}.$$ In the second step of the proof, we consider how $\Lambda_B$ may be bounded. Before doing this, we show that if $\sigma^2\geq\lambda$ then the bound (\[68\]) continues to hold. Consider now the case where $\sigma^2\geq\lambda$, so that $\rho\leq0$. We will use an analogous argument to show that the bound (\[68\]) continues to hold. In place of (\[64\]), we this time write $$\begin{gathered} \label{69} E[(\sigma^2+\gamma)g(W+1)-(W-\rho)g(W)] = (\sigma^2+\gamma) E[g(W+1)-g(\widehat{W})]\\ + (\sigma^2+\gamma) E[g(\widehat{W})-g(W^\star)],\end{gathered}$$ where $\widehat{W}=W+v_t(1-X_V)$, $W^\star=v_tW^s+(1-v_t)W$ and $t=\lambda(\sigma^2+\gamma)^{-1}$. Consider the first term on the right–hand side of (\[69\]). For this term, we argue as we did to derive (\[66\]). Conditioning on $v_t$ and $X_V$ and employing the equalities (\[tp1\]), we find, as for (\[66\]), that $$\begin{gathered} (\sigma^2+\gamma)E[g(W+1)-g(\widehat{W})]\\ =\lambda_2E[\Delta g(W^s)-\Delta g(W)] + \gamma E\Delta g(W) + (\sigma^2-\lambda+\lambda_2)E\Delta g(W).\end{gathered}$$ As we have that $$E[\widehat{W}-W^\star]=-(\sigma^2+\gamma)^{-1}(\sigma^2-\lambda+\lambda_2),$$ we then write $$\begin{gathered} \label{70} (\sigma^2+\gamma) E[g(\widehat{W})-g(W^\star)] + (\sigma^2-\lambda+\lambda_2)E\Delta g(W)\\ = (\sigma^2+\gamma) E\sum_{j=0}^\infty\left(\Delta g(j) - \Delta g(W)\right)[P(\widehat{W} > j)-P(W^\star > j)].\end{gathered}$$ Using the definitions of $\widehat{W}$ and $W^\star$, and conditioning on $v_t$, we find that $$P(\widehat{W} > j)-P(W^\star > j) = t\left[P(\overline{W} > j)-P(W^s > j)\right].$$ Comparing this with (\[tp5\]), recalling the definition of $t$ and using (\[tp6\]), we find that (\[67\]) also gives us a representation of (\[70\]). Continuing the argument as before, the bound (\[68\]) holds too in the present case. . In this part of the proof, we bound $\Lambda_B$, and thus obtain the bounds of our theorems. In doing so, we will use our stochastic ordering and dependence assumptions. The cases where $X_1,\ldots,X_n$ are positively and negatively related will be discussed separately. In the positive related case, the argument of Lemma \[res7\] shows that $$P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1)\leq0, \quad j\geq 0.$$ Noting that $(W_V+1\|X_V=1)=_dW^s$, we fix some $l\in\mathbb{Z}^+$ and write $$\begin{gathered} \label{72} P(W_V+1>j\|X_V=1)-P(W_V+1>j\|X_V=0) \\ = P(W_V+1>j+l\|X_V=1)-P(W_V+1>j\|X_V=0) + \sum_{i=1}^lP(W^s=j+i). \end{gathered}$$ Suppose now that there is some $q\in[0,1]$ such that for each $j\geq0$ $$\begin{aligned} \nonumber && P(W_V+1>j+l\|X_V=1)-P(W_V+1>j\|X_V=0) \\ \label{73} &&\qquad\qquad\qquad\qquad\qquad \leq q\,P(W_V=j+l\|X_V=1)\\ \label{tp10} &&\qquad\qquad\qquad\qquad\qquad =q\,P(W^s=j+l+1).\end{aligned}$$ We will show presently that this is implied by the stochastic ordering assumption (\[58\]). Using (\[72\]) and (\[tp10\]), we find that $$\begin{gathered} \label{tp7} \Lambda_B \leq qE\|\Delta g_B(W^s-l-1)-\Delta g_B(W)\| + \sum_{i=1}^lE\|\Delta g_B(W^s-i)-\Delta g_B(W)\|\\ \leq 2q\\|\Delta g_B\\|_\infty d_{TV}(\mathcal{L}(W),\mathcal{L}(W^s-l-1)) + 2\\|\Delta g_B\\|_\infty \sum_{i=1}^ld_{TV}(\mathcal{L}(W),\mathcal{L}(W^s-i)).\end{gathered}$$ Using our bound on $\\|\Delta g_B\\|_\infty$ and the triangle inequality for total variation distance, the first term of (\[tp7\]) is bounded by $$\begin{gathered} \label{tp8} 2q\sigma^{-2}\big\{d_{TV}(\mathcal{L}(W), \mathcal{L}(W^s))+(l+1)d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1))\big\}\\ \leq 2q\sigma^{-2}\left\{\frac{\sigma}{2\lambda}+(l+1)d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1))\right\},\end{gathered}$$ where this last inequality uses (\[71\]). Similarly, the second term of (\[tp7\]) may be bounded by $$\begin{gathered} \label{tp9} 2\sigma^{-2}\sum_{i=1}^l\big\{d_{TV}(\mathcal{L}(W),\mathcal{L}(W^s))+i\,d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1))\big\}\\ \leq \sigma^{-2}\left\{\frac{l\sigma}{\lambda}+l(l+1)d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1))\right\}.\end{gathered}$$ Combining (\[tp7\]), (\[tp8\]) and (\[tp9\]) with the bound (\[68\]) yields the desired inequality (\[59\]). So, the proof of Theorem \[tppos\] is completed upon showing that the stochastic ordering condition (\[58\]) implies the inequality (\[73\]). Writing $$P(W_V=j+l\|X_V=1)=P(W_V+1>j+l\|X_V=1)-P(W_V>j+l\|X_V=1),$$ for $0\leq j\leq n$, it can be seen that (\[73\]) is equivalent to $$P(W_V+1>j\|X_V=0) \geq (1-q)P(W_V+1-l>j\|X_V=1) + qP(W_V-l>j\|X_V=1),$$ for $j\geq 0$. This, in turn, is equivalent to the stochastic ordering $$\label{75} (W+1\|X_V=0) \succeq_{st} (1-v_q)(W-l\|X_V=1) + v_q(W-l-1\|X_V=1),$$ which can be seen using (\[tp1\]). Some rearranging shows that the stochastic ordering assumption (\[58\]) implies the stochastic ordering (\[75\]), hence the result of Theorem \[tppos\]. We turn our attention now to the case of negative relation, and complete the proof of Theorem \[tpneg\]. When $X_1,\ldots,X_n$ are negatively related, one can use a similar argument to the above. We have here that $$P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1)\geq0, \quad 0\leq j\leq n.$$ Analogously to the positively related case, we write, for some fixed $l\in\mathbb{Z}^+$, $$\begin{gathered} P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1) \\ = P(W_V+1>j\|X_V=0) - P(W_V+1>j-l\|X_V=1) + \sum_{i=0}^{l-1}P(W^s=j-i).\end{gathered}$$ This time, we suppose that there is $q\in[0,1]$ such that $$\label{76} P(W_V+1>j\|X_V=0)-P(W_V+1>j\|X_V=1) \leq qP(W_V+1+l=j\|X_V=1).$$ Following a similar argument to that used in the case of positive relation, we find that $$\Lambda_B \leq \frac{l+q}{\lambda\sigma} + \frac{l(l+2q-1)}{\sigma^2}\, d_{TV}(\mathcal{L}(W^s),\mathcal{L}(W^s+1)).$$ Combining this with (\[68\]) gives us the desired inequality (\[57\]). It remains to show that the stochastic ordering assumption (\[56\]) implies the inequality (\[76\]), which can be done as above. Another abstract approximation theorem {#bd3} ====================================== Our aim hereafter is to consider an alternative approximation theorem which may be found within the present framework. For concreteness, we suppose that the birth rates $\alpha_j$ and death rates $\beta_j$ are such that the random variable $\pi$ has two parameters available to choose. This will be the case in the application presented afterwards. Let us return to the basic representation (\[13\]). To choose the two parameters of $\pi$, it seems natural, in our context, to consider $s=2$ and introduce the two conditions $\alpha=\beta$ and $EW_\alpha=EW_\beta$ (i.e., $E[\alpha_W(W+1)]=E[\beta_WW]$). With these choices, the representation (\[13\]) becomes $$\label{78} Eh(W)-Eh(\pi) = \alpha \sum_{i=0}^\infty \Delta^2f(i)\, E[(W_\alpha-i-1)_+-(W_\beta-i-1)_+].$$ Moreover, suppose that one can construct $W_\alpha$ and $W_\beta$ on the same probability space in such a way that $W_\beta=W_\alpha+Y$ for some random variable $Y$ which takes values in the set $\{-1,0,1\}$. Under this assumption, $E[W_\alpha]=E[W_\beta]=E[W_\alpha+Y]$, which implies $E[Y]=0$. It is easily seen that the representation (\[78\]) can be rewritten as $$\begin{aligned} \label{79} \nonumber Eh(W)-Eh(\pi) &=& -\alpha \sum_{i=0}^\infty \Delta^2f(i) \; E[ YI_{(W_\alpha-1 \geq i+1)} + Y_+I_{(W_\alpha-1=i)} ]\nonumber\\ &=& -\alpha \; E[ I_{(Y=1)}\Delta^2f(W_\alpha-1) + Y\Delta f(W_\alpha-1) ].\end{aligned}$$ Noting that $$\begin{aligned} \|E[ I_{(Y=1)}\Delta^2f(W_\alpha-1) ]\| &\leq& 2 \\|\Delta f\\|_\infty \; d_{TV}(\mathcal{L}(W_\alpha),\mathcal{L} (W_\alpha+1)) \; \sup_W\{P(Y=1\|W_\alpha)\},\\ \|E[ Y\Delta f(W_\alpha-1) ]\| \; &\leq& \; \\|\Delta f\\|_\infty E\|E[Y\|W_\alpha]\| \;\; \leq \;\; \\|\Delta f\\|_\infty \sqrt{\textrm{Var}(E[Y\|W_\alpha])},\end{aligned}$$ we can immediately bound the right-hand side of (\[79\]) to obtain the following. \[pbd1\] Suppose that $\alpha=\beta$ and $EW_\alpha=EW_\beta$. If $W_\alpha$ and $W_\beta$ can be constructed on the same probability space such that $$\label{80} W_\beta=W_\alpha+Y \quad \mbox {for some random variable $Y$ valued in $\{-1,0,1\}$},$$ then, $$\begin{aligned} \label{81} \|Eh(W)-Eh(\pi)\| \leq \;\; 2\alpha \\|\Delta Sh\\|_\infty\, d_{TV}(\mathcal{L} (W_\alpha),\mathcal{L}(W_\alpha+1))\, \sup_W\{P(Y=1\|W_\alpha)\} \nonumber\\ +\; \alpha \\|\Delta Sh\\|_\infty \sqrt{\mbox{Var}(E[Y\|W_\alpha])}.\end{aligned}$$ Clearly, if such a random variable $Y$ takes values on a bounded set other than $\{-1,0,1\}$, a representation analogous to (\[79\]) may still be found, and a result analogous to Proposition \[pbd1\] is available. We now apply our Proposition \[pbd1\] to approximate a sum of independent indicator random variables. Suppose that $W=X_1+\cdots+X_n$ is the sum of independent Bernoulli random variables with success probabilities $p_i$, $1\leq i\leq n$. Brown and Xia (2001, Section 3) showed that in this case, one can improve on Poisson or binomial approximation for $W$ by using a so–called polynomial birth–death distribution, with the choices $\alpha_j=\alpha$ and $\beta_j=\gamma j+j(j-1)$ for some constants $\alpha$ and $\gamma$. We will follow that approach and choose here $\alpha$ and $\gamma$ such that $\alpha=\beta$ and $E[\alpha_W(W+1)]=E[\beta_WW]$. Straightforward computations then give us expressions for these parameters: $$\label{82} \gamma = \lambda^2\lambda_2^{-1}-1-2\lambda+2\lambda_3\lambda_2^{-1}, \quad \mbox {and} \quad \alpha = \gamma\lambda+\lambda^2-\lambda_2,$$ where $\lambda_k=\sum_{i=1}^np_i^k$ and $\lambda=\lambda_1=E[W]$ (as in Section 5). Note that the parameter choices (\[82\]) are the same as those employed by Brown and Xia (2001), who based their selection on minimising the error bound obtained in their result. To begin with, let us prove that the condition (\[80\]) is satisfied. Since the birth rate is constant (as in the Poisson case), we again have that $W_\alpha=W+1$. Let us turn our attention to $W_\beta$. We let $W_i=W-X_i$, and $W_{i,j}=W-X_i-X_j$, $0\leq i,j \leq n$ and observe that $W(W-1)=\sum_{1\leq i\not=j\leq n}X_iX_j$. By the definition of $W_\beta$, we get that $$\begin{aligned} P(W_\beta=k) &=& \alpha^{-1}E\left\{[\gamma W+W(W-1)]I_{(W=k)}\right\}\\ &=& \alpha^{-1}[\gamma\sum_{i=1}^np_iP(W_i+1=k)+ \sum_{1\leq i\not=j\leq n}p_ip_jP(W_{i,j}+2=k)],\end{aligned}$$ for $1\leq k\leq n$. In the spirit of the size-biasing construction of Section \[pois\], we define now two random indices $T,U\in\{1,\ldots,n\}$ chosen according to the distribution $$P(T=i,U=j)=\frac{p_ip_j}{\lambda^2-\lambda_2},\quad i\neq j, \quad\quad \mbox{and} \quad\quad P(T=U=i)=0.$$ Recall also the definition (\[32\]) of the random index $V$. Combining these definitions with the above, we may write $$P(W_\beta=k) \;=\; \alpha^{-1}\gamma\lambda P(W+1-X_V=k) + \alpha^{-1} (\lambda^2-\lambda_2)P(W+2-X_T-X_U=k),$$ for $1\leq k\leq n$. Let $q=\alpha^{-1}\gamma\lambda$; note from (\[82\]) that $0\leq q\leq 1$ whenever $\gamma\geq0$. In the sequel we will assume that this is indeed the case. Introduce a Bernoulli random variable $v_q$ with success probability $q$, independent of all other entries. We may then write $$W_\beta = v_q(W+1-X_V)+(1-v_q)(W+2-X_T-X_U) = W+1+Y = W_\alpha+Y,$$ where $$\label{84} Y=(1-v_q)(1-X_T-X_U)-v_qX_V,$$ $Y$ being valued in $\{-1,0,1\}$ with $E[Y]=0$, as desired. Now, let us evaluate the bound (\[81\]). First, we need a bound on the solution $f$ of the Stein equation in this situation. By Theorem 2.10 of Brown and Xia (2001), one knows that $$\label{85} \sup \lbrace \\|\Delta Sh\\|_\infty : h(j)=I_{(j\in B)},\; B\subseteq{\mathbb{Z}^+} \rbrace \leq \alpha^{-1}.$$ Further, $W$ being a sum of independent indicators, one has (from Barbour and Jensen (1989, Lemma 1)) $$\label{86} d_{TV}(\mathcal{L}(W),\mathcal{L}(W+1)) \;\leq\; \frac{1}{2\sqrt{\sum_{i=1}^n p_i(1-p_i)}}.$$ Finally, consider the two conditional terms in (\[81\]). Note from (\[84\]) that $Y=1$ if and only if $v_q=X_T=X_U=0$, so that $$\begin{aligned} P(Y=1\|W) &=& (1-q)P(X_T=X_U=0\|W) \;\;=\;\; (1-q)E[(1-X_T)(1-X_U)\|W]\\ &=& \alpha^{-1} \!\! \sum_{1\leq i\not=j\leq n}p_ip_jE[(1-X_i)(1-X_j)\|W].\end{aligned}$$ This probability takes its greatest value when $W=0$, with $E[(1-X_i)(1-X_j)\|W=0]=1$ for all $i$ and $j$. Hence, $$\label{88} \sup_W \{P(Y=1\|W)\} = \alpha^{-1} \!\! \sum_{1\leq i\not=j\leq n}p_ip_j = \alpha^{-1}(\lambda^2-\lambda_2).$$ Now, let $\\|Z\\|=(E[Z^2])^{1/2}$ be the $L_2$ norm for any random variable $Z$. Since $T=_d U$ and $E[Y]=0$, we write $$E[Y\|W]= -q(E[X_V\|W] -E[X_V]) - 2(1-q)(E[X_T\|W] -E[X_T]),$$ and thus $$\begin{aligned} \sqrt{\mbox{Var}(E[Y\|W])} &=& \\|E[Y\|W]\\|\\ &\leq & q\sum_{j=1}^n \\|E[X_j\|W] -E[X_j]\\| P(V=j)\\ &&\qquad + 2(1-q) \sum_{j=1}^n \\|E[X_j\|W] -E[X_j]\\| P(T=j) \\ &\leq& (q+2(1-q)) \max_{1\leq j\leq n} \sqrt{\mbox{Var}(E[X_j\|W])}.\end{aligned}$$ When $p_j=p$ for $j=1,\ldots,n$, $E[X_j\|W]=W/n$ and so the bound becomes the equality $$\label{87} \sqrt{\mbox{Var}(E[Y\|W])} = (2-q) \sqrt{\mbox{Var}(W/n)}.$$ Inserting (\[85\]), (\[86\]), (\[88\]) and (\[87\]) in (\[81\]) then provides the following bound: $$d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi)) \;\leq\; \frac{p}{(1-p)\sigma} +\frac{(2-q)\sigma}{n} =O(p/\sqrt{\lambda}),$$ where $\sigma^2=\mbox{Var}(W)=np(1-p)$. By exploring the explicit structure of the auxiliary variable $Y$, it is possible to derive better bounds. Throughout this part we let $\bar{a}=1-a$ for any $a\in\mathbb{R}$ and $\sigma_k=\sqrt{\sum_{i=k+1}^n\rho_i}$, where $\rho_i$ is the $i$th largest number of $p_1(1-p_1),\ldots,p_n(1-p_n)$. From Barbour and Jensen (1989, Lemma 1) we have that for all $i,j=1,\ldots,n$ and $i\not=j$, $$2d_{TV}(\mathcal{L}(W_i),\mathcal{L}(W_i+1))\leq\sigma_1^{-1} \quad\mbox{and}\quad 2d_{TV}(\mathcal{L}(W_{i,j}),\mathcal{L}(W_{i,j}+1))\leq\sigma_2^{-1}.$$ Notice that, from representation (\[84\]), $$\label{90} I_{(Y=1)}=\bar{v}_q \bar{X}_T\bar{X_U}\;,\; I_{(Y=-1)}=v_q X_V+\bar{v}_q X_T X_U.$$ The derivations below are based on the conditional independence of $X_T$ and $W_T$, given $T$ and similarly $X_U$ and $W_U$, given $U$ and $X_V$ and $W_V$, given $V$. By substituting (\[90\]) in (\[79\]), integrating with respect to $v_q$, separating linear and quadratic terms and noticing that $T=_d U$, we derive, after some simple calculations, $$\begin{aligned} I&=& Eh(W)-Eh(\pi)\\ &=& -\alpha E[\bar{v}_q \bar{X}_T\bar{X}_U \Delta f(W+1)] + \alpha E[(v_q X_V + \bar{v}_q X_T X_U)\Delta f(W)] \\ &=&-\alpha \bar{q} E[X_T X_U \Delta^2 f(W)] \\ &&\quad\quad +2\alpha \bar{q} E[X_T\Delta^2 f(W)]\\ &&\quad\quad - \alpha( \bar{q} E[\Delta f(W+1)] - E[(2\bar{q}X_T + q X_V)\Delta f(W)])\\ &=&I_1+I_2+I_3.\end{aligned}$$ Using the conditional independence of $W_{T,U}$ and $X_T, X_U$ given $T$ and $U$, the first term $I_1$ is bounded by $$\begin{aligned} \|I_1\|&=& \alpha\bar{q}\big\|EE[X_T X_U\|T,U] E[ \Delta^2 f(W_{T,U}+2)]\big\|\\ &\leq& 2\alpha\bar{q}\\|\Delta f\\|_\infty E[X_T X_U] \max_{i\neq j}\big\{d_{TV}(\mathcal{L}(W_{i,j}),\mathcal{L}(W_{i,j}+1))\big\} \;\leq\; \frac{\lambda_2^2-\lambda_4}{\alpha \sigma_2}.\end{aligned}$$ By conditioning on $T$, $$\begin{aligned} \|I_2\|&=& 2\alpha \bar{q}\big\|EE[X_T\|T] E[ \Delta^2 f(W_{T}+1)]\big\|\\ &\leq& 4\alpha\bar{q}\\|\Delta f\\|_\infty E[X_T] \max_{i}\big\{d_{TV}(\mathcal{L}(W_{i}),\mathcal{L}(W_{i}+1))\big\} \;\leq\; \frac{2(\lambda \lambda_2-\lambda_3)}{\alpha \sigma_1}.\end{aligned}$$ To bound $I_3$, we first notice that since $E[Y]=0$, $$\bar{q}=2\bar{q}E[X_T] + q E[X_V].$$ Thus, $$\begin{aligned} \|I_3\| &=& \big\|2\alpha\bar{q} E\big\{X_T \big(E[\Delta f(W_T+1)\|T] - E[\Delta f(W_T+X_T+1)]\big)\big\}\\ &&\quad + \alpha q E\big\{X_V \big(E[\Delta f(W_V+1)\|V] - E[\Delta f(W_V+X_V+1)]\big)\big\}\big\|\\ &\leq& 2\alpha \big\{2\bar{q} E[\{E(X_T\|T)\}^2] \\ &&\qquad\qquad\qquad + qE[\{E(X_V\|V)\}^2]\big\} \\|\Delta f\\|_\infty \max_{i}\big\{d_{TV}(\mathcal{L}(W_{i}),\mathcal{L}(W_{i}+1))\big\}\\ &\leq&\frac{2(\lambda\lambda_3-\lambda_4)}{\alpha\sigma_1}+\frac{\gamma\lambda_3}{\alpha\sigma_1}.\end{aligned}$$ By combining the bounds on $I_1,I_2$ and $I_3$ we derive the following. \[cor\_ex8\] With $W$ and $\pi$ as above, $$\label{newbound} d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi)) \;\leq\; \frac{\lambda_2^2-\lambda_4}{\alpha \sigma_2} +\frac{2(\lambda \lambda_2-\lambda_3)}{\alpha \sigma_1} +\frac{2(\lambda\lambda_3-\lambda_4)}{\alpha\sigma_1}+\frac{\gamma\lambda_3}{\alpha\sigma_1}.$$ Let us conclude by comparing our result with that of Brown and Xia (2001, Theorem 3.1), who obtain $$\label{bxbound} d_{TV}(\mathcal{L}(W),\mathcal{L}(\pi)) \;\leq\; \frac{\gamma\lambda_3}{\alpha\sigma_1} + \frac{2\lambda\lambda_2}{\alpha\sigma_2}.$$ When $p_i= p\to 0$ for each $i$ and $\lambda\to \infty$, both the bounds (\[newbound\]) and (\[bxbound\]) are asymptotically equivalent to $3p^2/\sqrt{\lambda}$. Fraser Daly gratefully acknowledges the financial support of the Schweizerische Nationalfonds, and thanks the Belgian Fonds National de la Recherche Scientifique for support during a visit to the Université Libre de Bruxelles. Thanks are also due to Andrew Barbour for several useful discussions. [99]{} (1989). Two moments suffice for Poisson approximations: the Chen–Stein method. Ann. Prob. [17]{}, 9–25. , editors (2005). An Introduction to Stein’s Method. Lect. Notes Ser., Inst. Math. Sci. 4, Natl. Univ. Singap., Singapore. (1992). Poisson Approximation. Oxford Univ. Press, Oxford. (1989). Local and tail approximations near the Poisson limit. Scand. J. Statist. [16]{}, 75–87. (2000). On the variance-to-mean ratio in models of parasite distributions. Adv. Appl. Prob. [32]{}, 701–719. (1999). Poisson perturbations. ESAIM Prob. Stat. [3]{}, 131–150. (1999). Negative binomial approximation with Stein’s method. Methodol. Comput. Appl. Prob. [1]{}, 407–421. (2001). Stein’s method and birth-death processes. Ann. Prob. [29]{}, 1373–1403. (2001). Centred Poisson approximation via Stein’s method. Lith. Math. J. [41]{}, 319–329. (2005). Exchangeable pairs and Poisson approximation. Probab. Surv. [2]{}, 64–106. (1975). Poisson approximation for dependent trials. Ann. Prob. [3]{}, 534–545. (1991). Binomial approximation to the Poisson binomial distribution. Stat. Prob. Lett. [11]{}, 7–16. (2005). Stein’s method for Poisson and compound Poisson approximation. In [Barbour, A. D. and Chen, L. H. Y.]{}, editors (2005). An Introduction to Stein’s Method. Lect. Notes Ser., Inst. Math. Sci. 4, Natl. Univ. Singap., Singapore, 61–113. (1996). Multivariate normal approximations by Stein’s method and size bias couplings. J. Appl. Prob. [33]{}, 1–17. (2004). Stein’s method for birth and death chains. In [Diaconis, P. and Holmes, S.]{}, editors (2004). Stein’s Method: Expository Lectures and Applications. IMS Lect. Notes 46, Beachwood, Ohio, 45–67. (1992). Univariate Discrete Distributions. Wiley, New York. (1977). Stochastic inequalities on partially ordered spaces. Ann. Prob. [5]{}, 899–912. (1996). Comparing sums of exchangeable Bernoulli random variables. J. Appl. Prob. [33]{}, 285–310. (2002). Approximation for a sum of dependent indicators: an alternative approach. Adv. Appl. Prob. [34]{}, 609–625. (1996). Stein’s method for geometric approximation. J. Appl. Prob. [33]{}, 707–713. (2000). Non-uniform bounds for geometric approximation. Stat. Prob. Lett. [49]{}, 305–311. (2005). Three general approaches to Stein’s method. In [Barbour, A. D. and Chen, L. H. Y.]{}, editors (2005). An Introduction to Stein’s Method. Lect. Notes Ser., Inst. Math. Sci. 4, Natl. Univ. Singap., Singapore, 183–221. (2005). Approximation of sums of conditionally independent variables by the translated Poisson distribution. Bernoulli [11]{}, 1115–1128. (2007). Translated Poisson approximation using exchangeable pair couplings. Ann. Appl. Prob. [17]{}, 1596–1614. (2007). Stochastic Orders. Springer, New York. (1986). Approximate Computation of Expectations. IMS Lect. Notes, Monogr. Ser. 7, Hayward, California. (2008). On negative binomial approximation to $k$-runs. J. Appl. Prob. [45]{}, 456–471.	{ "pile_set_name": "ArXiv" }
--- abstract: 'For $m,n>0$ or $mn<0$ we estimate the sums $$\sum_{c \leq x} \frac{S(m,n,c,\chi)}{c},$$ where the $S(m,n,c,\chi)$ are Kloosterman sums attached to a multiplier $\chi$ of weight $1/2$ on the full modular group. Our estimates are uniform in $m, n$ and $x$ in analogy with the bounds for the case $mn<0$ due to Ahlgren–Andersen, and those of Sarnak–Tsimerman for the trivial multiplier when $m,n>0$. In the case $mn<0$, our estimates are stronger in the $mn$-aspect than those of Ahlgren–Andersen. We also obtain a refinement whose quality depends on the factorization of $24m-23$ and $24n-23$ as well as the best known exponent for the Ramanujan–Petersson conjecture.' address: 'Department of Mathematics, University of Illinois, 1409 West Green Street, Urbana, IL 61801, USA' author: - Alexander Dunn title: Uniform bounds for sums of Kloosterman sums of half integral weight --- Introduction and statement of results ====================================== The classical Kloosterman sum $$S(m,n,c):=\sum_{d \pmod c} e \Big( \frac{m d+n \overline{d}}{c} \Big), \quad e(x):=\text{exp}(2 \pi i x)$$ plays a central part in analytic number theory. For applications, see [@HB; @S] for example. In this paper we study generalised Kloosterman sums $S(m,n,c,\chi)$ attached to the Dedekind eta multiplier $\chi$ of weight $1/2$. These are given by $$\label{genklo} S(m,n,c,\chi):=\sum_{\substack{0 \leq a,d<c \\ ( \begin{smallmatrix} a & b \\ c & d \end{smallmatrix} ) \in \text{SL}_2(\mathbb{Z})}} \overline{\chi} \begin{pmatrix} a & b \\ c & d \end{pmatrix} e \Big( \frac{\tilde{m}a +\tilde{n}d }{c} \Big),$$ where $$\tilde{m}:=m-\frac{23}{24}, \quad m \in \mathbb{Z}.$$ The Kloosterman sums defined in appear in Rademacher’s formula for the partition function $p(n)$ [@R1; @R2]. Cancellation amongst such sums plays a central role in establishing power saving error terms when one truncates of the formula for $p(n)$. This can be found in the work of Ahlgren and Andersen [@AA]. The study of closely related Kloosterman sums has applications to the coefficients of Ramanujan’s well known mock theta function $f(q)$ as well. This can be found in the work of Ahlgren and the author [@AD]. Kloosterman sums with general multipliers have been studied by Bruggeman [@Br], Goldfeld–Sarnak [@GS] and Pribitkin [@Pr], amongst many others. For the ordinary Kloosterman sums $S(m,n,c)$, Linnik [@Li] and Selberg [@Se] conjectured that there should be considerable cancellation in the sums $$\label{classickl} \sum_{c \leq x} \frac{S(m,n,c)}{c}.$$ Sarnak and Tsimerman [@ST] proposed a modified version of Linnik’s and Selberg’s conjecture with an $\varepsilon$-“safety valve" in $m$ and $n$. In particular, the refined conjecture for is $$\sum_{c \leq x} \frac{S(m,n,c)}{c} \ll_{\varepsilon} (\|mn\| x)^{\varepsilon}.$$ One obtains the “trivial bound" for by applying Weil’s bound [@W] $$\|S(m,n,c)\| \leq \tau(c) (m,n,c)^{\frac{1}{2}} \sqrt{c},$$ where $\tau(c)$ is the number of divisors of $c$. This yields $$\sum_{c \leq x} \frac{S(m,n,c)}{c} \ll \tau((m,n)) x^{\frac{1}{2}} \log x.$$ Still the best known bound in the $x$ aspect was obtained by Kuznetsov [@Ku], who proved for $m,n>0$ that $$\label{Ku} \sum_{c \leq x} \frac{S(m,n,c)}{c} \ll_{m,n} x^{\frac{1}{6}} (\log x)^{\frac{1}{3}}.$$ Sarnak and Tsimerman [@ST] refined Kuznetsov’s method and made the dependence on $m$ and $n$ explicit. They proved that for $m,n>0$ we have $$\label{STb} \sum_{c \leq x} \frac{S(m,n,c)}{c} \ll \big(x^{\frac{1}{6}}+(mn)^{\frac{1}{6}}+(m+n)^{\frac{1}{8}} (mn)^{\frac{\theta}{2}} \big)(xmn)^{\varepsilon},$$ where $\theta$ is any admissible exponent in the Ramanujan–Petersson conjecture for the coefficients of weight zero Maass cusp forms. By work of Kim and Sarnak [@Ki Appendix 2], the exponent $\theta=7/64$ is available. Ganguly and Sengupta [@GSe] have generalised the results of Sarnak and Tsimerman to sums over $c$ that are divisible by a fixed integer $q$. Kiral [@Kir] obtained estimates in the case $mn<0$ using the opposite sign Kloosterman zeta function. He obtained the bound $$\sum_{c \leq x} \frac{S(m,n,c)}{c} \ll x^{\frac{1}{6}+\varepsilon} \big( (m,n)^{\varepsilon}+(mn)^{\theta} \big)+x^{\varepsilon} (mn)^{\frac{1}{4}+\varepsilon},$$ where $\theta$ is as above. For the Kloosterman sum $S(m,n,c,\chi)$, Ahlgren and Andersen [@AA Theorem 1.3] proved that for $mn<0$ we have $$\label{AAeta} \sum_{c \leq x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \big(x^{\frac{1}{6}}+\|mn\|^{\frac{1}{4}} \big) \|mn\|^{\varepsilon} \log x.$$ They obtained a stronger result for sums of Kloosterman sums $S(1,n,c,\chi)$ when $n<0$ [@AA Theorem 9.1]. This leads to an improvement in error term [@AA Theorem 1.1] when one truncates Rademacher’s formula [@R1; @R2] for the partition function $p(n)$. Our first Theorem improves the $mn$-aspect in . \[mainthm\] Let $m>0$ and $n<0$ be integers such that $24n-23$ is not divisible by $5^4$ or $7^4$. Then we have $$\sum_{c \leq x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \Big(x^{\frac{1}{6}}+\frac{m^{\frac{1}{4}}+\|n\|^{\frac{1}{4}}}{\|mn\|^{\frac{1}{308}}}+\|mn\|^{\frac{19}{77}} \Big) \|mn\|^{\varepsilon} \log^3 x.$$ We also consider the case when $m,n>0$. Let $\mathcal{P}:=\big \{\frac{k(3k \pm 1)}{2} : k \in \mathbb{Z} \big \}$ be the set of generalized pentagonal numbers. \[mainthm2\] Let $m,n>0$ be integers be such that $m-1 \not \in \mathcal{P}$ or $n-1 \not \in \mathcal{P}$. Then we have $$\sum_{c \leq x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \big(x^{\frac{1}{6}}+(mn)^{\frac{1}{4}} \big) (mn)^{\varepsilon} \log^2 x.$$ When both $m,n>0$ are such that $m-1 \in \mathcal{P}$ and $n-1 \in \mathcal{P}$, we have the asymptotic formula $$\sum_{c \leq x} \frac{S(m,n,c,\chi)}{c}=C(m,n) x^{\frac{1}{2}}+O_{m,n} (x^{\frac{1}{6}}),$$ for some constant $C(m,n)$. See [@AA2 Theorem 8]. We also obtain refined bounds which recognise the arithmetic of $24m-23$ and $24n-23$. In analogy with the result of Sarnak and Tsimerman, these depend on progress toward the Ramanujan–Petersson conjecture for weight zero Maass forms on $\Gamma_0(N)$ (cf. the $H_{\theta}$-hypothesis in Section \[heckemass\]). \[auxthm2\] Let $m,n>0$ be integers such that $m-1 \not \in \mathcal{P}$ or $n-1 \not \in \mathcal{P}$. Suppose $m_0,n_0$ are integers such that $24m-23=m_0^2 s$ and $24n-23=n_0^2 t$ with $s$ and $t$ square-free integers. Then $$\sum_{c \leq x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \Big( x^{\frac{1}{6}}+ (st)^{\frac{1}{4}}+(st)^{\frac{1}{12}} (mn)^{\frac{1}{6}}+ (mnst)^{\frac{1}{8}+\frac{\theta}{4}} \Big) (mn)^{\varepsilon} \log^3 x,$$ where $\theta$ is any admissible exponent toward the Ramanujan–Petersson conjecture. The proofs of Theorems \[mainthm\]–\[auxthm2\] depend on generalisations of Kuznetsov’s trace formula due to Proskurin [@P] and Ahlgren–Andersen [@AA]. These are given in Sections \[proskurin\] and \[AAkuzsec\]. This formula transfers the task at hand to that of establishing bounds for sums involving the coefficients of half integral weight holomorphic and Maass cusp forms. To bound the contribution from holomorphic forms we appeal to Petersson’s formula. We also use the Shimura lift for half integer weight forms and Deligne’s bound to obtain bounds in terms of the factorisation of $24m-23$ and $24n-23$. Details can be found in Sections \[heckeholom\] and \[holbd\]. To bound the contribution from the Maass cusp forms we modify a dyadic argument in the spectral parameter that appears in [@AA; @ST]. Our new treatment involves estimating an initial segment for the spectral parameter using an averaged bound of Duke [@D]. We will also make use of a mean value estimate for the coefficients of Maass cusp forms due to Andersen and Duke [@And]. These bounds can be found in Section \[massbd\]. A Shimura lift for half integral weight Maass cusp forms developed by Ahlgren and Andersen is applied to obtain the dependence on $\theta$ in Theorem \[auxthm2\]. This lift appears in Section \[heckemass\]. The proofs of Theorems \[mainthm\]–\[auxthm2\] can be found in Sections \[mainsec\]–\[auxsec\] respectively. We follow the methods and presentation in [@AA]. Preliminaries {#pre} ============= We give only a concise background related to congruence subgroups. More details can be found in [@AA] and [@DFI] for example. See also [@AD Section 2]. Let $\mathbb{H}$ denote the upper-half plane. We have the usual action of $\text{SL}_2(\mathbb{R})$ on $\mathbb{H}$ given by $$\gamma \tau =\frac{a \tau +b}{c \tau +d}, \quad \text{for} \quad \tau \in \mathbb{H} \quad \text{and} \quad \gamma= \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \text{SL}_2(\mathbb{R}).$$ For $\gamma \in \text{SL}_2(\mathbb{R})$ we define the weight $k$ slash operator for real analytic forms by $$f \vert_k \gamma:=j(\gamma,z)^{-k} f(\gamma z), \quad j(\gamma,z):=\frac{cz+d}{\|cz+d\|}=e^{i \text{arg}(cz+d)},$$ where the argument is always chosen in $(-\pi,\pi]$. The weight $k$ Laplacian is defined by $$\Delta_k:=y^2 \Big(\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2} \Big)-iky \frac{\partial }{\partial x}.$$ For simplicity we will work only with the groups $\Gamma_0(N)$ for $N \in {\mathbb{N}}$ and with weights $k \in \frac{1}{2} \mathbb{Z}$, although much of what is said here holds in more generality. Let $\Gamma$ denote such a group. We say that $\nu: \Gamma \rightarrow {\mathbb{C}}^{\times}$ is a multiplier system of weight $k$ if - $\|\nu\|=1$ - $\nu(-I)=e^{-\pi i k}$, and - $\nu(\gamma_1 \gamma_2) j(\gamma_1 \gamma_2,\tau)^k=\nu(\gamma_1) \nu(\gamma_2) j(\gamma_2,\tau)^k j(\gamma_1, \gamma_2 \tau)^k$ for all $\gamma_1,\gamma_2 \in \Gamma$. Given a cusp $\mathfrak{a}$, let $\Gamma_{\mathfrak{a}}:=\{\gamma \in \Gamma: \gamma \mathfrak{a}=\mathfrak{a} \}$ denote the stabilizer in $\Gamma$ and let $\sigma_{\mathfrak{a}}$ denote the unique (up to translation on the right) matrix in $\text{SL}_2(\mathbb{R})$ satisfying $\sigma_{\mathfrak{a}} \infty=\mathfrak{a}$ and $\sigma_{\mathfrak{a}}^{-1} \Gamma_{\mathfrak{a}} \sigma_{\mathfrak{a}}=\Gamma_{\infty}$. Define $\alpha_{\nu,\mathfrak{a}} \in [0,1)$ by the condition $$\nu \left( \sigma_{\mathfrak{a}} \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \sigma_{\mathfrak{a}}^{-1} \right)=e \left({-\alpha_{\nu,\mathfrak{a}}} \right).$$ The cusp $\mathfrak{a}$ is singular with respect to $\nu$ if $\alpha_{\nu,\mathfrak{a}}=0$. When $\mathfrak a=\infty$ we suppress the subscript. If $\nu$ is multiplier of weight $k$, then it is a multiplier in any weight $k^{\prime} \equiv k \pmod{2}$, and $\bar \nu$ is a multiplier of weight $-k$. If $\alpha_{\nu}=0$ then $\alpha_{\bar{\nu}}=0$, while if $\alpha_{\nu}>0$ then $\alpha_{\bar{\nu}}=1-\alpha_{\nu}$. For $n \in {\mathbb{Z}}$ we define $$n_\nu:=n-\alpha_{\nu};$$ then we have $$\label{eq:n_nu_conj} n_{\bar{\nu}}=\begin{cases} -(1-n)_\nu\quad& \text{if $\alpha_\nu \neq 0$},\\ n\quad&\text{if $\alpha_\nu= 0$}. \end{cases}$$ With this notation we define the generalized Kloosterman sum (at the cusp $\infty$) by $$\label{eq:kloos_def} S(m,n,c,\nu) := \sum_{\substack{0\leq a,d<c \\ \gamma= \left(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix} \right) \in \Gamma}} \bar\nu(\gamma) e \left(\frac{m_\nu a+n_\nu d}{c} \right).$$ We have the relationships $$\label{eq:nuconj} \overline{S(m,n,c,\nu)}= \begin{cases} S(1-m, 1-n, c, \overline\nu)&\quad\text{if $\alpha_{\nu}>0$,}\\ S(-m, -n, c, \overline\nu)&\quad \text{if $\alpha_{\nu}=0$.} \end{cases}$$ A function $f:\mathbb{H} \rightarrow \mathbb{C}$ is automorphic of weight $k$ and multiplier $\nu$ for $\Gamma_0(N)$ if $$f \vert_k \gamma =\nu(\gamma) f, \quad \text{for all} \quad \gamma \in \Gamma_0(N).$$ Let $\mathcal{A}_k(N,\nu)$ denote the space of such functions. If $f \in \mathcal{A}_k(N,\nu)$ is a smooth eigenfunction of $\Delta_k$ which satisfies the growth condition $$f(\tau) \ll y^{\sigma}+y^{1-\sigma},$$ for some $\sigma$ and all $\tau \in \mathbb{H}$, then it is called a Maass form. Let $$\mathcal{L}_k (N,\nu):=\{f \in \mathcal{A}_k(N,\nu): \\| f\\| < \infty \},$$ where the norm is induced by the Petersson inner product $$\langle f,g \rangle:=\int_{\mathbb{H} \backslash \Gamma_0(N)} f(\tau) \overline{g(\tau)} d \mu, \quad d \mu:=\frac{dx dy}{y^2}.$$ Let $\mathcal{B}_k(N, \nu)$ denote the subspace of $\mathcal{L}_k(N, \nu)$ consisting of smooth functions $f$ such that $f$ and $\Delta_k f$ are bounded on $\mathbb{H}$. For all $f,g \in \mathcal{B}_k(N,\nu)$ we have $$\langle \Delta_k f,g \rangle=\langle f, \Delta_k g \rangle.$$ Furthermore, for any $f \in \mathcal{B}_k(N,\nu)$ we have $$\langle f,-\Delta_k f \rangle \geq \frac{\|k\|}{2} \left(1-\frac{\|k\|}{2} \right) \geq 0.$$ Thus by a theorem of Friedrichs, the operator $-\Delta_k$ has a unique self-adjoint extension to $\mathcal{L}_k(N, \nu)$ (which we also denote $-\Delta_k$). Then by a theorem of von Neumann, the space $\mathcal{L}_k(N,\nu)$ has a complete spectral resolution with respect $-\Delta_k$, which we describe in detail now. There is both a continuous and discrete spectrum. For each singular cusp $\mathfrak a$ (and only at such cusps) there is an Eisenstein series $E_{\mathfrak a}(z, s)$. These provide the continuous spectrum on the line ${\operatorname{Re}}s=\frac{1}{2}$, which covers $[1/4, \infty)$. The reminder of the spectrum is discrete. It is countable and of finite multiplicity (with $\infty$ being the only limit point). We denote it by $$\lambda_0\leq \lambda_1\leq \dots$$ where we have $$\lambda_0\geq \frac{\|k\|}{2} \left(1-\frac{\|k\|}{2} \right).$$ One component of the discrete spectrum is provided by residues of the Eisenstein series $E_{\mathfrak a}(z,s)$ at possible simple poles $s$ with $\frac12<s\leq 1$; the corresponding eigenvalues have $\lambda<\frac{1}{4}$. The remainder of the discrete spectrum arises from Maass cusp forms. We give more details about the discrete spectrum in what follows. Denote by $\tilde{\mathcal{L}}_k(N, \nu)$ the subspace of $\mathcal{L}_k(N, \nu)$ spanned by the eigenfunctions of $\Delta_k$. If $f \in \tilde{\mathcal{L}}_k(N, \nu)$ has Laplace eigenvalue $\lambda$, then we write (here $0 \leq k<2$) $$\lambda=\frac{1}{4}+r^2,\qquad r\in i \Big(0, \sqrt{1/4-(\|k\|/2) \big(1-\|k\|/2 \big)} \Big] \cup [0, \infty),$$ and refer to $r$ as the spectral parameter of $f$. Denote by $\tilde{\mathcal{L}}_k(N, \nu, r)$ the subspace of such functions. Let $W_{\kappa, \mu}$ denote the usual $W$-Whittaker function (cf. [@DL Section 13.14]). Then each $f \in \tilde{\mathcal{L}}_k(N, \nu, r)$ has a Fourier expansion of the form $$\label{eq:f_fourier} f(\tau)=c_{0}(y) + \sum_{n_\nu\neq 0} \rho(n) W_{\frac{k \text{sgn}(n_{\nu})}{2}, ir}(4\pi \|n_\nu\| y)e(n_\nu x),$$ where $$c_{0}(y)=\begin{cases} 0\quad&\text{if $\alpha_\nu\neq 0$},\\ 0\quad&\text{if $\alpha_\nu=0$ and $r\geq 0$,}\\ \rho(0)y^{\frac12+i r}&\text{if $\alpha_\nu=0$ and $r\in i(0, 1/4]$,} \end{cases}$$ with coefficients $\rho(n)$. Note that in the last case, we have $\rho(0) \neq 0$ only when $f$ arises as a residue. Let $\mathcal{S}_k(N,\nu) \subseteq \tilde{\mathcal{L}}_k(N, \nu)$ be the subspace spanned by the Maass cusp forms (i.e. $c_0(y)=0$). Two important multipliers of weight $\frac{1}{2}$ are the eta-multiplier $\chi$ on $\text{SL}_2(\mathbb{Z})$, given by $$\label{eq:etamult} \eta(\gamma\tau)=\chi(\gamma)\sqrt{c\tau+d}\,\eta(\tau), \qquad \gamma=\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \text{SL}_2({\mathbb{Z}}),$$ and the theta-multiplier $\nu_\theta$ on $\Gamma_0(4)$, given by $$\label{eq:thetamult} \theta(\gamma\tau)=\nu_\theta(\gamma)\sqrt{c\tau+d}\,\theta(\tau), \qquad \gamma=\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \Gamma_0(4).$$ Here $\eta(\tau)$ and $\theta(\tau)$ are the two fundamental theta functions $$\begin{aligned} \eta(\tau)&:=q^\frac1{24}\prod_{n=1}^\infty(1-q^n),\\ \theta(\tau)&:=\sum_{n=-\infty}^\infty q^{n^2}, \end{aligned}$$ where we use the standard notation $$q:=e(\tau)=e^{2\pi i\tau}.$$ For $\nu_\theta$ we have the formula $$\label{eq:def-theta-mult} \nu_\theta \begin{pmatrix} a & b \\ c & d \end{pmatrix} = \left( \frac{c}{d} \right) \epsilon_d^{-1},$$ where $\left(\frac{\bullet}{\bullet} \right)$ is the extended Kronecker symbol and $$\epsilon_d = \begin{cases} 1 & \text{ if }d\equiv 1\pmod{4}, \\ i & \text{ if }d\equiv 3 \pmod{4}. \end{cases}$$ From this we obtain $$\label{eq:thetaconj} \bar{\nu_\theta}(\gamma)=\left( \frac{-1}{d} \right) \nu_\theta(\gamma), \qquad \gamma=\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \Gamma_0(4).$$ For $c>0$ and $\gamma=\left(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix} \right) \in \text{SL}_2(\mathbb{Z})$, we have another convenient formula [@Kno Section 4.1] $$\label{kron} \chi(\gamma) = \begin{dcases} \left( \frac{d}{c} \right) e \left( \frac{1}{24} \left[(a+d)c-bd(c^2-1)-3c\right] \right) & \text{ if $c$ is odd}, \\ \left(\frac{c}{d} \right) e \left(\frac {1}{24} \left[(a+d)c-bd(c^2-1)+3d-3-3cd\right] \right) & \text{ if $c$ is even.} \end{dcases}$$ We have $\chi \left( \begin{smallmatrix} 1 & b \\ 0 & 1 \end{smallmatrix} \right)=e(\frac{b}{24})$. Finally, if $c>0$ we have $\chi(-\gamma)=i \chi(\gamma)$ (this follows since $\gamma$ and $-\gamma$ act the same way on $\mathbb{H}$). When $(k,\nu)=(1/2,\chi)$ and we work on the full modular group $\Gamma$, there is neither continuous spectrum nor discrete spectrum arising from Eisenstein series because the only cusp is non-singular. This follows from the evaluation $$\chi \left( \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \right)=e \left( \frac{1}{24} \right).$$ Thus $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ is spanned by Maass cusp forms. We fix an orthonormal basis $\{u_j\}$ with corresponding spectral parameters $r_j$ and with Fourier series given by $$\label{fouriercusp} u_j(\tau)=\sum_{n \neq 0} \rho_j(n) W_{\frac{\text{sgn}(n)}{4},ir_j}(4 \pi \| \tilde{n}\| y) e(\tilde{n} x),$$ where $$\tilde{n}:=n_{\chi}=n-\frac{23}{24}.$$ Hecke theory for holomorphic cusp forms {#heckeholom} ======================================= We briefly review Hecke theory for holomorphic cusp forms of half integral weight. For $N \in \mathbb{N}$ and $k \in 2 \mathbb{N}$, let $S_{\frac{1}{2}+k}(4N,\nu)$ denote the space of holomorphic cusp forms of weight $1/2+k$ on $\Gamma_0(4N)$ with a multiplier $\nu$ of weight $\frac{1}{2}$. Let $\Psi$ denote an even Dirichlet character mod $4N$. For all primes $p \nmid 4N$, the action of the Hecke operator $T_{p^2}$ on $$f:=\sum_{r=1}^{\infty} a(r) e(r \tau)\in S_{\frac{1}{2}+k}(4N,\Psi \nu_{\theta})$$ is given in [@Sh Theorem 1.7] by $$\label{coeffrel} T_{p^2}(f)(z):=\sum_{n=1}^{\infty} \Big( a_f(p^2 n)+\Psi^(p) \Big( \frac{n}{p} \Big) p^{k-1} a_f(n)+ \Psi(p^2) p^{2k-1} a_f \big(n/p^2 \big) \Big) e(n \tau) \in S_{\frac{1}{2}+k}(4N,\Psi \nu_{\theta}),$$ where $\Psi^$ is the character modulo $4N$ defined by $\Psi^(m)=\Psi(m) \left( \frac{-1}{m} \right)^k$ and $a(n/p^2)=0$ if $p^2 \nmid n$. There are Hecke operators $T_{n^2}$ for all integers $n$ such that $(n,4N)=1$. For $v \in \mathbb{N}$, the operators $T_{p^{2v}}$ are polynomials in the $T_{p^2}$. If $(nm,4N)=1$ and $(n,m)=1$ then $$\label{mult} T_{n^2} T_{m^2}=T_{n^2 m^2}.$$ Recall that $\chi$ is the Dedekind eta multiplier defined in . Letting $\chi_{12}:=\Big(\frac{12}{\bullet} \Big)$, we have the map $$\label{image} L: S_{\frac{1}{2}+k}(1,\chi) \rightarrow S_{\frac{1}{2}+k}(576,\chi_{12} \nu_{\theta}), \quad L(f)(\tau):=f(24 \tau).$$ We justify the target space in with the following computation. Let $\Gamma_0(M,N)$ denote the subgroup of $\Gamma_0(N)$ consisting of matrices whose upper right entry is divisible by $M$. Equation and the remark following it show that $$\chi(\gamma)=\left( \frac{c}{d} \right) e \left( \frac{d-1}{8} \right) \quad \text{for} \quad \gamma \in \Gamma_0(24,24).$$ This implies that $$\chi \left( \begin{pmatrix} a & 24b \\ c/24 &d \end{pmatrix} \right)=\left( \frac{12}{d} \right) \left( \frac{c}{d} \right) \epsilon_d^{-1} \quad \text{for} \quad \gamma=\begin{pmatrix} a & b \\ c &d \end{pmatrix} \in \Gamma_0(576).$$ For primes $p \nmid 6$, we can define Hecke operators $\tilde{T}_{p^2}$ on $S_{\frac{1}{2}+k}(1,\chi)$. Let $$f(\tau):=\sum_{n=1}^{\infty} a(n) e \left( \bigg(n-\frac{23}{24} \bigg) \tau \right) \in S_{\frac{1}{2}+k}(1,\chi).$$ Then we can define the action of $\tilde{T}_{p^2}$ on $S_{\frac{1}{2}+k}(1,\chi)$ by $$\tilde{T}_{p^2} f=\sum_{n=1}^{\infty} \Big( a_f(p^2 n)+\Psi^(p) \Big( \frac{n}{p} \Big) p^{k-1} a_f(n)+\Psi(p^2) p^{2k-1} a_f \big(n/p^2 \big) \Big) e \left( \bigg(n-\frac{23}{24} \bigg) \tau \right).$$ Observe that $$\label{commute} L(\tilde{T}_{p^2} f )=T_{p^2}(Lf).$$ We recall the Shimura correspondence for half-integral weight holomorphic cusp forms. [@Sh Main Theorem] and [@Cip Proposition 5.1] \[shim\] Let $N,k \in \mathbb{N}$ and $\Psi$ be a character modulo $4N$. Suppose that $g(\tau):=\sum_{n=1}^{\infty} a(n) e(n \tau) \in S_{k+\frac{1}{2}}(4N,\Psi \nu_{\theta})$. Let $t$ be a positive square-free integer, and define the Dirichlet character $\Psi_t$ by $\Psi_t(n):=\Psi(n) \big(\frac{-1}{n} \big)^{k} \big( \frac{t}{n} \big)$. Define $b_t(n) \in \mathbb{C}$ by $$\sum_{n=1}^{\infty} \frac{b_t(n)}{n^s}:=L(s-k+1,\Psi_t) \sum_{n=1}^{\infty} \frac{a(tn^2)}{n^s}.$$ Then $$\emph{Sh}_t(g):=\sum_{n=1}^{\infty} b_t(n) e(n \tau) \in M_{2 k}(2N,\Psi^2).$$ Moreover, if $k \geq 2$, then $\emph{Sh}_t(g)$ is a cusp form. For all primes $p \nmid 4N$ and squarefree $t$, we have $$\emph{Sh}_t \big(T_{p^2} g \big)=T_p \big( \emph{Sh}_t(g) \big),$$ where $T_p$ denotes the usual Hecke operator on $M_{2 k}(2N,\Psi^2)$. Hecke theory for Maass cusp forms {#heckemass} ================================= We discuss Hecke theory for the spaces $\mathcal{S}_{0}(N,\mathbf{1})$ and $\mathcal{S}_{\frac{1}{2}}(1,\chi)$. For $(n,N)=1$, the Hecke operator on $\mathcal{S}_{0}(N,\mathbf{1})$ can be defined as $$(\mathcal{T}_n f)(\tau)=\frac{1}{\sqrt{n}} \sum_{ad=n} \hspace{0.1cm} \sum_{b \hspace{-0.2cm} \mod{d}} f \left( \frac{a \tau+b}{d} \right) \in \mathcal{S}_0(N,\mathbf{1}).$$ We have $$\mathcal{T}_m \mathcal{T}_n=\sum_{d \mid (m,n)} \mathcal{T}_{mnd^{-2}}.$$ We now record the explicit action of the Hecke operators on Fourier coefficients. If $f \in \mathcal{S}_{0}(N,\mathbf{1}) \cap \tilde{\mathcal{L}}_{0}(N,\mathbf{1},r)$ is a Maass cusp form with Fourier expansion $$f(\tau)=\sum_{n \neq 0} \rho(n) W_{0, ir}(4\pi \|n \| y)e(n x),$$ then $$\mathcal{T}_p f=\sum_{n \neq 0} \left (p^{\frac{1}{2}} \rho(pn)+p^{-\frac{1}{2}} \rho \left(\frac{n}{p} \right) \right) W_{0, ir}(4\pi \|n \| y)e(n x).$$ Note that $\mathcal{T}_n$ commutes with $\Delta_0$, so $\mathcal{T}_n$ is an endomorphism of $\mathcal{S}_0(N,\mathbf{1}) \cap \tilde{\mathcal{L}}_0(N,\mathbf{1},r)$. Furthermore, for all $(n,N)=1$ we have $$\langle \mathcal{T}_n f, g \rangle= \langle f, \mathcal{T}_n g \rangle.$$ Thus we can produce an orthonormal basis $\{v_j\}$ (each having spectral parameter $r_j$) for $\mathcal{S}_0(N,\mathbf{1})$ that consists of Hecke eigenforms for all $\mathcal{T}_n$ with $n$ coprime to $N$. Suppose that each $v_j$ has Fourier coefficients $\tilde{\rho}_j(n)$. The $H_{\theta}$–hypothesis asserts that $$\lambda_j(n) \ll_{\varepsilon} n^{\theta+\varepsilon},$$ where the $\lambda_j(n)$ are the Hecke–Maass eigenvalues defined by $$\mathcal{T}_n v_j= \lambda_j(n) v_j.$$ The Ramanujan–Petersson conjecture asserts that $H_0$ is true. The best known result is due to Kim and Sarnak [@Ki Appendix 2], who showed that the exponent $\theta=7/64$ is available. Applying $\mathcal{T}_{n}$ to the Fourier expansion of $v_j$ we see that [@DFI (6.14),(6.15)] $$\label{coeff} \tilde{\rho}_j(n)=\lambda_j(\|n\|) \tilde{\rho}_j \big( \text{sgn}(n) \big) \|n\|^{-1/2}.$$ The Hecke operators $\mathcal{T}_{p^2}$ for $p \nmid 6$ are defined on $\mathcal{S}_{\frac{1}{2}}(1,\chi)$ [@AA Section 2.6] by $$T_{p^2} f=\frac{1}{p} \bigg[ \sum_{b \mod p^2} e \Big( \frac{-b}{24} \Big) f \vert_{\frac{1}{2}} \begin{pmatrix} \frac{1}{p} & \frac{b}{p} \\ 0 & p \end{pmatrix} + e \Big( \frac{p-1}{8} \Big) \sum_{h=1}^{p-1} e \Big(\frac{-hp}{24} \Big) \Big( \frac{h}{p} \Big) f \vert_{\frac{1}{2}} \begin{pmatrix} 1 & \frac{h}{p} \\ 0 & 1 \end{pmatrix} +f \vert_{\frac{1}{2}} \begin{pmatrix} p & 0 \\ 0 & \frac{1}{p} \end{pmatrix} \bigg].$$ Each $\mathcal{T}_{p^2}$ commutes with $\Delta_{\frac{1}{2}}$, so $\mathcal{T}_{p^2}$ is an endomorphism of $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi,r)$. The analogous discussion above guarantees that there exists an orthonormal basis of $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ consisting of Hecke eigenforms. Ahlgren and Andersen [@AA] developed a Shimura type correspondence between Maass cusp forms of weight $1/2$ on $\Gamma_0(N)$ with the eta multiplier twisted by a Dirichlet character and Maass cusp forms of weight $0$. Here we provide details only in the simplest case. In this setting it is most convenient to write the expansion of $f \in \tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi,r)$ in the form $$\label{normalise} f(\tau)=\sum_{n \neq 0} a(n) W_{\frac{\text{sgn}(n)}{4},ir} \Big( \frac{\pi \|n\| y}{6} \Big) e \Big( \frac{nx}{24} \Big).$$ [@AA Theorem 5.1] \[thelif\] Suppose that $G \in \tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi,r)$ with $r \neq i/4$ and Fourier expansion given by . Let $t \equiv 1 \pmod{24}$ be a square-free positive integer and define $b_t(n) \in \mathbb{C}$ by the relation $$\label{shimrel} \sum_{n=1}^{\infty} \frac{b_t(n)}{n^s}=L\Big (s+1, \Big( \frac{t}{\bullet} \Big) \Big) \sum_{n=1}^{\infty} \Big( \frac{12}{n} \Big) \frac{a(tn^2)}{n^{s-\frac{1}{2}}}.$$ Then the function $S_t(G)$ defined by $$(S_t G)(\tau):=\sum_{n=1}^{\infty} b_t(n) W_{0,2ir}(4 \pi n y) \cos( 2 \pi n x)$$ is a Maass cusp form in $\tilde{\mathcal{L}}_0(6,\mathbf{1},2r)$. For any prime $p \geq 5$ we have $$\mathcal{T}_p S_t(G)= \Big(\frac{12}{p} \Big) S_t \big( \mathcal{T}_{p^2} G \big).$$ \[specremark\] Using Theorem \[thelif\], Ahlgren and Andersen rule out the existence of exceptional eigenvalues in $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$. If $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi,r) \neq \{0\}$, then either $r=i/4$ or $r>1.9$. Note that $r_0=i/4$ corresponds to the minimal eigenvalue $\lambda_0=\frac{3}{16}$. This is achieved by the unique normalised cusp form $$u_0(\tau):=\sqrt{\frac{3}{\pi}} (6y)^{\frac{1}{4}} \eta(\tau).$$ The Fourier coefficients $\rho_0(m)$ of $u_0$ are non-zero only when $m-1 \in \mathcal{P}$. See [@AA2 pg. 435]. Kuznetsov–Proskurin formula {#proskurin} =========================== Here we develop some tools for the case $m,n>0$. Let $\phi \in C^{4} \big([0,\infty) \big)$ be such that $$\label{phicond} \phi(0)=\phi^{\prime}(0), \quad \phi(t) \ll_{\varepsilon} t^{-1-\varepsilon} \quad \text{and} \quad \phi^{(j)} \ll_{\varepsilon} t^{-2-\varepsilon} \quad \text{for} \quad j=1,2,3,4,$$ as $t \rightarrow \infty$ for some fixed $\varepsilon>0$. We define the auxiliary integrals $$\label{orig} \check{\phi}(r):=\int_{0}^{\infty} J_{r-1}(y) \phi(y) \frac{dy}{y},$$ and $$\label{hat} \hat{\phi}(r):= \pi^2 e^{3 \pi i/4} \frac{\int_{0}^{\infty} \Big( \cos \pi \big(\frac{1}{4} +ir \big) J_{2ir}(y)-\cos \pi \big( \frac{1}{4}-ir \big) J_{-2 i r}(y) \Big) \phi(y) \frac{dy}{y} }{\text{sh}(\pi r) \text{ch} (2 \pi r) \Gamma \big(\frac{1}{4}+ir \big) \Gamma \big(\frac{1}{4}-ir \big)},$$ where $J_{\nu}$ for $\nu \in \mathbb{C}$ denotes the $J$–Bessel function [@DL Section 10.2]. Using a trigonometric identity, we write the integrand occurring in $\hat{\phi}(r)$ in the more convenient form $$\label{con} \frac{1}{\sqrt{2}} \frac{\phi(y)}{y} \Big( \cos( \pi i r) \big(J_{2ir}(y)-J_{-2ir}(y) \big)-\sin( \pi i r) \big(J_{2ir}(y)+J_{-2ir}(y) \big) \Big).$$ Endow $S_{\frac{1}{2}+2l}(1,\chi)$ with the usual inner-product [@DFI2 pg 2514]. For each integer $l \geq 1$, let $B_l$ denote an orthonormal basis for $S_{\frac{1}{2}+2l}(1,\chi)$ and $$\mathcal{S}:=\bigcup_{l=1}^{\infty} B_l.$$ Suppose each $f \in \mathcal{S}$ has Fourier expansion given by $$f(\tau):=\sum_{n=1}^{\infty} a_f(n) e \big (\tilde{n} \tau \big),$$ and weight denoted by $w(f)$. Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ with Fourier expansion given by . For $m,n>0$, Proskurin’s formula [@P p. 3888] asserts that $$\label{PrKu} \sum_{c \geq 1} \frac{S(m,n,c,\chi)}{c} \phi \Big(\frac{4 \pi \sqrt{\tilde{m} \tilde{n}}}{c} \Big)=\mathcal{U}+\mathcal{V},$$ where $$\begin{aligned} \mathcal{U}&:=\sum_{f \in \mathcal{S}} \frac{4 \Gamma \big(w(f) \big) e^{\pi i w(f) / 2}}{(4 \pi)^{w(f)} (\tilde{m} \tilde{n})^{(w(f)-1)/2} } \overline{a_f(m)} a_f(n) \check{\phi} \big(w(f) \big) \label{U}, \\ \mathcal{V}&:=4 \sqrt{\tilde{m} \tilde{n}} \sum_{j \geq 0} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch} (\pi r_j)} \hat{\phi}(r_j) \label{V}.\end{aligned}$$ Given $a,x>0$, choose a parameter $T>0$ such that $$T \leq x/3, \quad T \asymp x^{1-\delta} \quad \text{with} \quad 0<\delta<1/2.$$ Now we choose a smooth $\phi=\phi_{a,x,T}: [0,\infty) \rightarrow [0,1]$ satisfying - $\phi(t)=1$ for $\frac{a}{2x} \leq t \leq \frac{a}{x}$ - $\phi(t)=0$ for $t \leq \frac{a}{2x+2T}$ and $t \geq \frac{a}{x-T}$ - $\phi^{\prime}(t) \ll \big(\frac{a}{x-T}-\frac{a}{x} \big)^{-1} \ll \frac{x^2}{aT}$ - $\phi$ and $\phi^{\prime}$ are piecewise monotone on a fixed number of intervals. Here we provide bounds for some useful expressions involving $\check{\phi}$ and $\hat{\phi}$. \[int2\] Let $\phi=\phi_{a,x,T}$ be as above. For $a:=4 \pi \sqrt{ \tilde{m} \tilde{n}}$ we have $$\label{trans2} \sum_{l=1}^{\infty} \Big({-\frac{1}{2}+2l }\Big) \Big \| \check{\phi} \Big(\frac{1}{2}+2l \Big) \Big \| \ll 1+\frac{\sqrt{mn}}{x}.$$ For the reader’s convenience we sketch the argument that appears on [@ST pp. 630–632], indicating what differs in our situation. When $x \geq 4 \pi \sqrt{\tilde{m} \tilde{n}}$, the support of $\phi$ is contained in $[0,3/2]$, and it is immediate from the decay of the Bessel function [@DL (10.14.4)] that $$\check{\phi} \Big(\frac{1}{2}+2l \Big) \ll \frac{1}{\Gamma(\frac{1}{2}+2l)}.$$ Thus the left hand side of is bounded by $O(1)$. Now consider the case when $x \leq 4 \pi \sqrt{\tilde{m} \tilde{n}}$. In what follows we write $k:=1/2+2l$ for convenience. We treat each integral occurring in the summand of according to the transitional ranges of the $J$-Bessel function. In the range $0 \leq y \leq k-k^{\frac{1}{3}}$, $J_k(y)$ is exponentially small and the contribution is $O(1)$. We now consider the transitional range $k-k^{\frac{1}{3}} \leq y \leq k +k^{\frac{1}{3}}$. A computation using [@DL (10.20.2)–(10.20.4),(9.6.1),(9.6.2),(9.6.6) and (9.6.7)] establishes the asymptotics for $J_k(y)$ asserted in [@ST (22) and (23)] when $k$ is half-integral and positive. Using these asymptotics and the fact that the support of $\phi$ is contained in the interval $\big [0, 6 \pi \sqrt{mn} /x \big]$, we see that there are at most $O(\sqrt{m n} /x)$ choices of $k$ for which the transitional range is present on the left hand side of . Each $(k-1) \check{\phi}(k)$ is $O(1)$ for all $k$ in the transitional range using the asymptotics for $J_{\pm 1/3}$ and $K_{1/3}$ in [@DL (10.7.3) and (10.30.2)] and hence the total contribution from all such $k$ to is $O(\sqrt{mn}/x)$. We are now left to bound the contribution for the range $y \geq k+k^{\frac{1}{3}}$. For this one can follow the argument in [@ST pp. 631–632] starting with the asymptotic in [@ST Eqn (52)]. The contribution in this last case is $O(\sqrt{mn} /x)$. \[int\] Suppose that $a,x,T$ are as above and that $\phi=\phi_{a,x,T}$. Then we have $$\hat{\phi}(r) \ll \begin{cases} \min \big ( r^{-1}, r^{-2} \frac{x}{T} \big ) & \quad \text{if} \quad r \geq \max \big(\frac{a}{x},1 \big) \\ r^{-1} & \quad \text{if} \quad r \geq 1. \end{cases}$$ In view of and , it is sufficient to bound $\check{\phi}(2ir+1)$. Sarnak and Tsimerman [@ST pp 629–630] prove that $$\frac{\text{ch}(\pi r )}{\text{sh}(2 \pi r)} \big \| \check{\phi}(2ir+1) \big \| \ll \begin{cases} r^{-\frac{3}{2}} & \text{for} \quad r \geq 1 \\ \min \big ( r^{-\frac{3}{2}}, r^{-\frac{5}{2}} \frac{x}{T} \big ) & \text{for} \quad r \geq \max \big(a/x,1 \big). \end{cases}$$ The result follows by recalling the definition of $\hat{\phi}(r)$ in and observing that $$\frac{1}{\big \| \Gamma( \frac{1}{4}+ir) \big \|^2} \sim \frac{\sqrt{r}}{2 \pi} e^{\pi r} \quad \text{as} \quad r \rightarrow \infty,$$ by Stirling’s formula. Variant of Proskurin–Kuznetsov formula {#AAkuzsec} ====================================== We introduce the tools for the mixed sign case $m>0$ and $n<0$. Let $\phi$ be as in Section \[proskurin\]. Define $$\check{\Phi}(r):=\cosh(\pi r) \int_{0}^{\infty} K_{2ir}(y) \phi(y) \frac{dy}{y},$$ where $K_{\nu}$ for $\nu \in \mathbb{C}$ denotes the $K$–Bessel function [@DL Section 10.25]. Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ with Fourier expansions given by . Then for $m>0$ and $n<0$, [@AA Theorem 4.1] asserts that $$\label{AAkuz} \mathcal{W}:=\sum_{c \geq 1} \frac{S(m,n,c,\chi)}{c} \phi \Big(\frac{4 \pi \sqrt{\tilde{m} \|\tilde{n}\|}}{c} \Big)=8 \sqrt{i} \sqrt{\tilde{m} \|\tilde{n}\|} \sum_{j \geq 0} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch} (\pi r_j)} \check{\Phi}(r_j).$$ [@AA Theorem 6.1] \[int3\] Let a,x,T be as above and let $\phi=\phi_{a,x,T}$. Then $$\check{\Phi}(r) \ll \begin{cases} r^{-\frac{3}{2}} e^{-\frac{r}{2}} & \emph{for} \quad 1 \leq r \leq \frac{a}{8x} \\ r^{-1} & \emph{for} \quad \max \big(1,\frac{a}{8x} \big) \leq r \leq \frac{a}{x} \\ \min \big(r^{-\frac{3}{2}}, r^{-\frac{5}{2}} \frac{x}{T} \big ) & \emph{for} \quad r \geq \max \big( \frac{a}{x},1 \big). \end{cases}$$ Bound for holomorphic forms {#holbd} =========================== In this section we bound the $\mathcal{U}$ term in uniformly in $m,n$ and $x$. To obtain bounds in terms of the square-free parts of $24m-23$ and $24n-23$, we exploit the Shimura correspondence and Deligne’s bound. For our purposes, it is sufficient to consider $\Psi$ an even Dirichlet character mod $4N$ and $k \in 2 \mathbb{N}$. Endow $S_{\frac{1}{2}+k}(4N,\Psi \nu_{\theta})$ with the usual inner product [@DFI2 pg 2514]. We first recall the half–integral weight Petersson formula [@Blo Lemma 4] that will apply in this setting. Let $\big \{ \psi_j:=\sum_{r \geq 1} a_j(n) e(n \tau) \big \}_{j=1}^J$ be an orthonormal basis for $S_{\frac{1}{2}+k} (4N,\Psi \nu_{\theta})$. Then for $k \geq 2$ we have $$\label{pet} \frac{\Gamma(k-\frac{1}{2})}{(4 \pi n)^{k-\frac{1}{2}}} \sum_{j=1}^J \|a_j(n)\|^2=1+2 \pi i^{-\frac{1}{2}-k} \sum_{4N \mid c} c^{-1} J_{k-\frac{1}{2}} \Big(\frac{4 \pi n}{c} \Big) K_{\Psi}(n,n,c),$$ where $$K_{\Psi}(m,n,c):=\sum_{d \pmod c} \varepsilon_d \Big( \frac{c}{d} \Big) \Psi(d) e \Big( \frac{m d+n \overline{d}}{c} \Big),$$ is a twisted Kloosterman sum. \[Ubd\] Suppose $m,n>0$ are integers and $\phi$ is as above with $x \geq 1$ and $a:=4 \pi \sqrt{\tilde{m} \tilde{n}}$. Then for any $\varepsilon>0$ we have $$\label{normalholo} \mathcal{U} \ll_{\varepsilon} (mn)^{\varepsilon} \Big( (mn)^{\frac{1}{4}} + \frac{(m n)^{\frac{3}{4}}}{x} \Big).$$ Furthermore, if $24 m-23=m_0^2 s$ and $ 24n-23=n_0^2 t$ with $s$ and $t$ square-free, then $$\label{holohecke} \mathcal{U} \ll_{\varepsilon} \|m_0 n_0\|^{\varepsilon} (st)^{\frac{1}{4}+\varepsilon} \Big(1+\frac{(mn)^{\frac{1}{2}}}{x} \Big).$$ Let $\{f_{jl} \}_{1 \leq j \leq \text{dim} S_{\tiny{\frac{1}{2}}+2l}(1,\chi)}$ be an orthonormal Hecke eigenbasis with respect to $\tilde{T}_{p^2}$ for all primes $p \nmid 6$. Suppose that $$f_{jl}(\tau)=\sum_{r=1}^{\infty} a_{jl}(r) e \Big(\Big(r-\frac{23}{24} \Big) \tau \Big)$$ and let $$g_{jl}(\tau):=L(f_{jl})(\tau)=\sum_{m=1}^{\infty} c_{jl}(m) e (m \tau) \in S_{\frac{1}{2}+2l}(576,\chi_{12} \nu_{\theta}).$$ For fixed $l$, the set $\{f_{jl}\}$ injects into its image $\{g_{jl}\} \subset S_{\frac{1}{2}+2l}(576,\chi_{12} \nu_{\theta})$ and $\{g_{jl}\}$ is an orthogonal set consisting of Hecke eigenforms for all $T_{p^2}$ with $p \nmid 6$. For each $l$, a computation shows that $$\label{rescale} \{(24)^{l+\frac{1}{4}} [\Gamma:\Gamma_0(24,24)]^{-\frac{1}{2}} g_{jl} \}$$ is an orthonormal set of Hecke eigenforms. Since $a_{jl}(n)=c_{jl}(24n-23)$, we trivially have $$\label{ineq} \sum_{j} \|a_{jl}(n)\|^2=\sum_{j} \|c_{jl}(24n-23)\|^2.$$ Applying the triangle and Cauchy–Schwarz inequalities to the right side of and using we obtain $$\label{CR} \mathcal{U} \ll \sum_{l=1}^{\infty} \frac{ \Gamma(\frac{1}{2}+2l) \| \check{\phi}(\frac{1}{2}+2l) \| }{(4 \pi)^{\frac{1}{2}+2l} (\tilde{m} \tilde{n})^{-\frac{1}{4}+l}} \Big( \sum_j \| c_{jl}(24m-23)\|^2 \Big)^{1/2} \Big( \sum_j \| c_{jl}(24n-23) \|^2 \Big)^{1/2}.$$ The set in can be extended to an orthonormal basis of $S_{\frac{1}{2}+2l}(576,\chi_{12} \nu_{\theta} )$. Applying and the triangle inequality we obtain $$\begin{gathered} \label{petersson} \sum_{j} \|c_{jl}(24n-23)\|^2 \ll \frac{(24)^{-2l} \Big(4 \pi (24n-23) \Big)^{-\frac{1}{2}+2l}}{\Gamma(-\frac{1}{2}+2l)} \\ \times \Bigg(1+2 \pi \sum_{c \equiv 0 \pmod{576}} \frac{\big\| K_{\chi_{12}}(24n-23,24n-23,c) \big\|}{c} \Big\| J_{-\frac{1}{2}+2l} \Big(\frac{4 \pi (24 n-23)}{c} \Big) \Big\| \Bigg).\end{gathered}$$ Let $\delta>0$ be fixed and small. To bound the right hand side of we consider the cases $c \leq n^{1+\delta}$ and $c> n^{1+\delta}$. Here we stress that the estimates obtained are uniform in $l$. When $c \leq n^{1+\delta}$, we use [@Wai Lemma 4] $$\label{kloost} \|K_{\chi_{12}}(24n-23,24n-23,c) \| \leq (24n-23,c)^{\frac{1}{2}} \tau(c) c^{\frac{1}{2}},$$ together with [@DL (10.14.1)] to obtain $$\label{initseg} \sum_{\substack{1 \leq c \leq n^{1+\delta} \\ c \equiv 0 \pmod{576}}} \frac{\big\| K_{\chi_{12}}(24n-23,24n-23,c) \big\|}{c} \Big \| J_{-\frac{1}{2}+2l} \Big(\frac{4 \pi (24n-23)}{c} \Big) \Big \| \ll_{\varepsilon} n^{\frac{1}{2}+\frac{\delta}{2}+\varepsilon}.$$ In the case $c \geq n^{1+\delta}$ we apply [@DL (10.14.4)] and in the following computation $$\begin{aligned} \sum_{\substack{c>n^{1+\delta} \\ c \equiv 0 \pmod{576}}} & \frac{\big\|K_{\chi_{12}}(24n-23,24n-23,c) \big\|}{c} \Big\| J_{-\frac{1}{2}+2l} \Big(\frac{4 \pi (24n-23)}{c} \Big) \Big \| \nonumber \\ & \ll \frac{1}{\Gamma(-\frac{1}{2}+2l)} \sum_{c>n^{1+\delta}} \frac{\|K_{\chi_{12}}(24n-23,24n-23,c)\|}{c} \Big( \frac{4 \pi (24n-23)}{2c} \Big)^{-\frac{1}{2}+2l} \nonumber \\ & \ll_{\varepsilon} \frac{1}{ \Gamma(-\frac{1}{2}+2l)} \sum_{c>n^{1+\delta}} \frac{1}{c^{\frac{1}{2}-\varepsilon}} (24n-23,c)^{\frac{1}{2}} \Big( \frac{4 \pi (24n-23)}{2c} \Big) \Big( \frac{4 \pi (24n-23)}{2c} \Big)^{-\frac{3}{2}+2l} \nonumber \\ & \ll_{\varepsilon} \frac{(48 \pi)^{2l} n^{1-\delta(-\frac{3}{2}+2l)} }{ \Gamma(-\frac{1}{2}+2l)} \sum_{d \mid 24n-23} \frac{1}{d^{1-\varepsilon}} \sum_{c^{\prime}>n^{1+\delta}/d} \frac{1}{(c^{\prime})^{\frac{3}{2}-\varepsilon}} \nonumber \\ & \ll_{\varepsilon} n^{\frac{1}{2}+\varepsilon} \label{case2}. \end{aligned}$$ Combining – we obtain $$\label{trans} \mathcal{U} \ll_{\varepsilon} (mn)^{\frac{1}{4}+\varepsilon} \sum_{l=1}^{\infty} \Big({-\frac{1}{2}+2l} \Big) \Big \| \check{\phi} \Big(\frac{1}{2}+2l \Big) \Big \|.$$ Thus follows from Lemma \[int2\]. We now prove . Since $g_{jl} \in S_{\frac{1}{2}+2l}(576,\chi_{12} \nu_{\theta})$ is an eigenform under the action of $T_{n^2}$ for all $n$ coprime to $6$, we know that $\text{Sh}_t(g_{jl}) \in S_{4l}(288,\mathbf{1})$ is an eigenform under the action of $T_{n}$ with the same eigenvalue. Denote this eigenvalue by $\lambda_{jl}(n)$. For each $l$ and $j$ define $b_t(n) \in \mathbb{C}$ by $$\text{Sh}_{t}(g_{jl}) (\tau)=\sum_{n=1}^{\infty} b_t(n) e(n \tau) \in S_{4l}(288,\mathbf{1}).$$ We also define the arithmetic functions $$\begin{aligned} g(u):=c_{jl}(t u^2) \quad \text{and} \quad h(u):=u^{2l-1} \Big( \frac{12t}{u} \Big).\end{aligned}$$ The equality in Lemma \[shim\] implies that $b_t=g * h$. Observe that $h$ is multiplicative and $h(1)=1$, so $h$ has a multiplicative Dirichlet inverse. We have $h^{-1}(1)=1$ and a computation for $ p$ prime and $\alpha \in \mathbb{N}$ yields $$h^{-1}(p^{\alpha})=\begin{cases} -\big(\frac{12t}{p} \big) p^{2l-1} & \quad \text{if} \quad \alpha=1 \\ 0 & \quad \text{if} \quad \alpha \geq 2. \end{cases}$$ Thus $$\label{invh} \|h^{-1}(u)\| \leq u^{2l-1} \quad \text{for all} \quad u \in \mathbb{N}.$$ We now use the fact that the $\text{Sh}_t(g_{jl})$ are Hecke eigenforms for the $T_n$ with $(n,6)=1$ and are of integral weight. We have $b_t(1)=c_{jl}(t)$, so for $(d,6)=1$ we have $$b_t (d)=\lambda_{jl}(d) b_t(1)=\lambda_{jl} (d) c_{jl}(t).$$ Thus $$\label{cjl} c_{jl}(tn_0^2)=\sum_{d \mid n_0} b_t(d) h^{-1} \Big( \frac{n_0}{d} \Big)=c_{jl}(t) \sum_{d \mid n_0} \lambda_{jl}(d) h^{-1} \Big( \frac{n_0}{d} \Big).$$ By Deligne’s bound [@De] we have $$\label{delbd} \|\lambda_{jl}(d)\| \ll_{\varepsilon} d^{2l-\frac{1}{2}+\varepsilon}.$$ Using and , becomes $$\|c_{jl}(24n-23)\|=\|c_{jl}(tn_0^2)\| \ll_{\varepsilon} \|c_{jl}(t)\| n_0^{2l-\frac{1}{2}+\varepsilon}.$$ We may replace each summand on the right hand side of with $\|c_{jl}(t) \| n_0^{2l-\frac{1}{2}+\varepsilon}$. Performing similar computations to those occuring in –, we see that becomes $$\mathcal{U} \ll_{\varepsilon} (m_0 n_0)^{\varepsilon} (st)^{\frac{1}{4}+\varepsilon} \sum_{l=1}^{\infty} \Big({-\frac{1}{2}+2l} \Big) \Big \| \check{\phi} \Big(\frac{1}{2}+2l \Big) \Big \|.$$ Thus Lemma \[int2\] implies . Estimates for the coefficients of Maass cusp forms {#massbd} ================================================== We bound the quantities in and in the proofs of the main theorems by modifying the dyadic arguments of [@ST]. Here we collect the main inputs required for this argument. [@AA Theorem 1.5] \[AAmve\] Suppose $\{u_j\}$ is an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ with spectral parameters $r_j$ and Fourier expansion given by . If $n<0$ then we have $$\|\tilde{n}\| \sum_{0<r_j \leq x} \frac{\|\rho_j(n)\|^2}{\emph{ch}(\pi r_j)}=\frac{x^{\frac{5}{2}}}{5 \pi^2}+ O_{\varepsilon} \big( x^{\frac{3}{2}} \log x+ \|n\|^{\frac{1}{2}+\varepsilon} x^{\frac{1}{2}} \big).$$ Andersen and Duke improve the mean value estimate of [@AA Theorem 1.5] when $n>0$. This improvement will be important in the proofs of our main theorems. We state only a special case of their result. [@And Theorem 4.1] \[andersen\] Suppose $\{u_j\}$ is an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ with spectral parameters $r_j$ and Fourier expansion given by . If $n>0$ then we have $$\tilde{n} \sum_{x \leq r_j \leq 2x} \frac{\|\rho_j(n)\|^2}{\emph{ch}(\pi r_j)} \ll_{\varepsilon} x^{\frac{3}{2}}+x^{-\frac{1}{2}} (\log x)^{\frac{1}{2}+\varepsilon} n^{\frac{1}{2}+\varepsilon}.$$ One can relate the positively (resp. negatively) indexed coefficients of a Maass cusp form of weight $\frac{1}{2}$ to negatively (resp. positively) indexed coefficients of a Maass cusp form of weight $-\frac{1}{2}$ via the conjugation map (cf. [@AD (2.6)]). This accounts for difference in the order of magnitude for the $x$ parameter in the estimates occurring in Propositions \[AAmve\] and \[andersen\]. The second main idea is the application of an averaged form of a pointwise bound due to Duke [@D] for the Fourier coefficients of Maass cusp forms of half integral weight with multiplier $\big(\frac{12}{\bullet} \big) \nu_{\theta}$. This average bound was established by Ahlgren and Andersen [@AA Theorem 8.1] using a modified version of Duke’s argument. When $n>0$, we remove a hypothesis in their Theorem which restricts divisibility of $n$ by arbitrarily high powers of $5$ and $7$. \[hyperem\] Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ with spectral parameters $r_j$ and Fourier expansion given by . If all the $u_j$ are Hecke eigenforms of $\mathcal{T}_{p^2}$ with $p \nmid 6$, then for $x \geq 1$ and $n>0$ we have, $$n \Big \| \sum_{0<r_j \leq x} \frac{\|\rho_j(n)\|^2}{\emph{ch}(\pi r_j)} \Big \| \ll_{\varepsilon} n^{\frac{3}{7}+\varepsilon} x^{\frac{9}{2}}.$$ Suppose $24n-23=n_0^2 t$ with $t$ square-free. We first prove $$\label{replace} \|\rho_j(n) \| \ll_{\varepsilon} \Big \| \rho_j \Big(\frac{t+23}{24} \Big) \Big \| n_0^{-1+\theta+\varepsilon}.$$ Recalling the normalisations and we define $$\label{fourierrel} a_j(n)=\rho_j \Big(\frac{n+23}{24} \Big).$$ We write $$S_t(u_j):=\sum_{n=1}^{\infty} b_t(n,j) W_{0,2ir}(4 \pi n y) \cos( 2 \pi n x) \in \tilde{\mathcal{L}}_0(6,\mathbf{1},2r_j),$$ where $S_t$ denotes the lift for cusp forms in Theorem \[thelif\]. The $S_t(u_j)$ are also eigenforms under the action of $\mathcal{T}_n$ with $(n,6)=1$. Let the eigenvalue be denoted by $\lambda_j(n)$. We define the arithmetic functions $$g(u):=a_{j}(t u^2) u^{\frac{1}{2}} \Big( \frac{12}{u} \Big) \quad \text{and} \quad h(u):=u^{-1} \Big( \frac{t}{u} \Big).$$ The equality in Theorem \[thelif\] implies that $b_t(\cdot,j)=g * h$. Observe that $h(1)=1$, so arguing as in Section \[holbd\] gives $$h^{-1}(p^{\alpha})=\begin{cases} -\big(\frac{t}{p} \big) p^{-1} & \quad \text{if} \quad \alpha=1 \\ 0 & \quad \text{otherwise}. \end{cases}$$ Thus $$\label{invh2} \|h^{-1}(u)\| \leq u^{-1} \quad \text{for all} \quad u \in \mathbb{N}.$$ Using and the relation $b_t(1,j)=a_{j}(t)$, we have $$b_t (d,j)= d^{-\frac{1}{2}} \lambda_{j}(d) b_t(1,j)=d^{-\frac{1}{2}} \lambda_{j} (d) a_{j}(t).$$ Thus $$g_j(tn_0^2)=\sum_{d \mid n_0} b_t(d,j) h^{-1} \Big( \frac{n_0}{d} \Big)=a_{j}(t) \sum_{d \mid n_0} d^{-\frac{1}{2}} \lambda_{j}(d) h^{-1} \Big( \frac{n_0}{d} \Big).$$ Applying the $H_{\theta}$–hypothesis and we obtain $$\|g_j(tn_0^2)\| \ll_{\varepsilon} \|a_j(t)\| n_0^{-\frac{1}{2}+\theta+\varepsilon}.$$ By the definition of $g$ and , we conclude that holds. Under the $H_{\theta}$ hypothesis we can apply , followed by [@AA Theorem 8.1] to obtain $$n \Big \| \sum_{0<r_j \leq x} \frac{\|\rho_j(n)\|^2}{\text{ch}(\pi r_j)} \Big \| \ll_{\varepsilon} n_0^{2 \theta+\varepsilon} t \Big \| \sum_{0<r_j \leq x} \frac{\|\rho_j \big(\frac{t+23}{24} \big)\|^2}{\text{ch}(\pi r_j)} \Big \| \ll_{\varepsilon} n_0^{2 \theta+\varepsilon} t^{\frac{3}{7}+\varepsilon} x^{\frac{9}{2}}.$$ Since $\theta=7/64$ is an acceptable exponent [@Ki Appendix 2] we see that $$n_0^{2 \theta+\varepsilon} t^{\frac{3}{7}+\varepsilon} x^{\frac{9}{2}} \ll (n_0^2)^{\frac{7}{64}+\varepsilon} t^{\frac{3}{7}+\varepsilon} x^{\frac{9}{2}} \ll_{\varepsilon} n^{\frac{3}{7}+\varepsilon} x^{\frac{9}{2}}.$$ This averaged bound allows us to optimise the dyadic argument in the spectral parameter, and is ultimately responsible for the improved $mn$-aspect occurring in Theorem \[mainthm\]. Proof of Theorem \[mainthm\] {#mainsec} ============================ \[mainprop\] Let $m>0$ and $n<0$ be integers such that $24n-23$ is not divisible by $5^4$ or $7^4$. Then for $x \geq \|\tilde{m} \tilde{n}\|^{\frac{38}{77}}$ we have $$\sum_{x \leq c \leq 2x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \Big( x^{\frac{1}{6}}+ \frac{m^{\frac{1}{4}}+\|n\|^{\frac{1}{4}}}{\|mn\|^{\frac{1}{308}}} +\|mn\|^{\frac{19}{77}} \Big) \|mn\|^{\varepsilon} \log^2 x.$$ We show that Proposition \[mainprop\] implies Theorem \[mainthm\]. Considering the sum $$\label{demsum} \sum_{1 \leq c \leq X} \frac{S(m,n,c,\chi)}{c},$$ the initial segment $1 \leq c \leq \|\tilde{m} \tilde{n}\|^{\frac{38}{77}}$ contributes $O_{\varepsilon}\big(\|\tilde{m} \tilde{n}\|^{\frac{19}{77}+\varepsilon} \big)$ by [@AA (2.30)]. One then breaks the interval $\|\tilde{m} \tilde{n}\|^{\frac{38}{77}} \leq c \leq X$ into $O(\log X)$ dyadic intervals $x \leq c \leq 2x$ with $\|\tilde{m} \tilde{n}\|^{\frac{38}{77}} \leq x \leq X/2$, and then applies Proposition \[mainprop\]. Let $\phi$ be a smooth test function with the properties listed in Section \[proskurin\]. Fix $a:=4 \pi \sqrt{\tilde{m} \|\tilde{n}\|}$ and let $T>0$, $0<\beta<1/2$ both be chosen later. Suppose $$x \geq (\tilde{m} \|\tilde{n}\|)^{\frac{1}{2}-\beta}.$$ Using the Weil bound [@AA Proposition 2.1] and the mean value bound for the divisor function we have $$\begin{gathered} \label{approx} \Bigg \| \sum_{c=1}^{\infty} \frac{S(m,n,c,\chi)}{c} \phi \big(\frac{a}{c} \big)-\sum_{x \leq c \leq 2x} \frac{S(m,n,c,\chi)}{c} \Bigg \| \leq \sum_{\substack{x-T \leq c \leq x \\ 2x \leq c \leq 2x+2T}} \frac{\|S(m,n,c,\chi)\|}{c} \\ \ll_{\varepsilon} \frac{T \log x}{\sqrt{x}} \|mn\|^{\varepsilon}.\end{gathered}$$ Recall from that we have $$\label{newdyad2} \sum_{c=1}^{\infty} \frac{S(m,n,c,\chi)}{c} \phi \Big( \frac{a}{c} \Big)=\mathcal{W}.$$ Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ that are also a Hecke eigenbasis with respect to the $\mathcal{T}_{p^2}$ for all primes $p \nmid 6$. Note that there is no contribution from $r_0=i/4$ since $n$ is negative. For the initial segment $0<r_j \leq 4 \pi(\tilde{m} \|\tilde{n}\|)^{\beta}$, we apply Proposition \[hyperem\] (to the $m>0$ variable) and [@AA Theorem 8.1] (to the $n<0$ variable) to obtain $$\label{avgduke3} m \|n\| \Bigg \| \sum_{0<r_j \leq 4 \pi(\tilde{m} \|\tilde{n}\|)^{\beta}} \frac{\|\rho_j(m)\|^2}{\text{ch}(\pi r_j)} \Bigg \| \cdot \Bigg \| \sum_{0<r_j \leq 4 \pi(\tilde{m} \|\tilde{n}\|)^{\beta}} \frac{\|\rho_j(n)\|^2}{\text{ch}(\pi r_j)} \Bigg \| \ll_{\varepsilon} \|mn\|^{\frac{3}{7}+10 \beta+\varepsilon}.$$ Then applying the Cauchy–Schwarz inequality, , Lemma \[int3\] and the fact that $r_j>1.9$ [@AA Corollary 5.3] yields the estimate $$\label{avgduke2} \sqrt{\tilde{m} \|\tilde{n}}\| \Bigg \| \sum_{0<r_j \leq 4 \pi(\tilde{m} \|\tilde{n}\|)^{\beta} } \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \check{\Phi}(r_j) \Bigg \| \ll_{\varepsilon} \|mn\|^{\frac{3}{14}+5 \beta+\varepsilon}.$$ We now estimate the contribution to $\mathcal{W}$ from the dyadic intervals $A \leq r_j \leq 2A$ with $A \geq 4 \pi (\tilde{m} \|\tilde{n}\|)^{\beta}$. Since we are assuming $x \geq (\tilde{m} \|\tilde{n}\|)^{\frac{1}{2}-\beta}$, we have $A \geq \max \big(a/x,1 \big)$. Using Lemma \[int3\], Proposition \[AAmve\] and Proposition \[andersen\] we have $$\begin{aligned} \sqrt{\tilde{m} \|\tilde{n}\|} & \Big \| \sum_{A \leq r_j \leq 2A} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \check{\Phi}(r_j) \Big \| \nonumber \\ & \ll \min \big (A^{-\frac{3}{2}},A^{-\frac{5}{2}} \frac{x}{T} \big) \Big( m \sum_{A \leq r_j \leq 2A} \frac{\|\rho_j(m)\|^2}{\text{ch}(\pi r_j)} \Big)^{\frac{1}{2}} \Big( \|n\| \sum_{A \leq r_j \leq 2A} \frac{\|\rho_j(n)\|^2}{\text{ch}(\pi r_j)} \Big)^{\frac{1}{2}} \label{CR2} \\ & \ll_{\varepsilon} \min \big (A^{-\frac{3}{2}},A^{-\frac{5}{2}} \frac{x}{T} \big) \big(A^{\frac{3}{2}}+m^{\frac{1}{2}+\varepsilon} A^{-\frac{1}{2}} \|\log A\|^{\frac{1}{2}+\varepsilon} \big)^{\frac{1}{2}} \big(A^{\frac{5}{2}}+\|n\|^{\frac{1}{2}+\varepsilon} A^{\frac{1}{2}} \big)^{\frac{1}{2}} \nonumber \\ & \ll_{\varepsilon} \min \big (\sqrt{A},\frac{x}{T \sqrt{A}} \big) \big(1+A^{-1} \|\log A\| (m^{\frac{1}{4}+\varepsilon}+\|n\|^{\frac{1}{4}+\varepsilon})+A^{-2} \|\log A\| \|mn\|^{\frac{1}{4}+\varepsilon} \big) \label{compbd} \\ & \ll_{\varepsilon} \min \big (\sqrt{A},\frac{x}{T \sqrt{A}} \big) \bigg(1+A^{-\frac{1}{2}} \|\log A\| \frac{(m^{\frac{1}{4}+\varepsilon}+\|n\|^{\frac{1}{4}+\varepsilon})}{\|mn\|^{\frac{\beta}{2}}}+A^{-\frac{1}{2}} \|\log A\| \|mn\|^{\frac{1}{4}-\frac{3}{2} \beta+\varepsilon} \bigg) \nonumber.\end{aligned}$$ Summing over dyadic intervals $[2^{j} \cdot 4 \pi (\tilde{m} \|\tilde{n}\|)^{\beta},2^{j+1} \cdot 4 \pi (\tilde{m} \|\tilde{n}\|)^{\beta} ]$ with $j=0,1,2,\ldots$ we see that when $x \geq (\tilde{m} \|\tilde{n}\|)^{\frac{1}{2}-\beta}$, the following holds: $$\label{fullcusp} \sqrt{\tilde{m} \tilde{n}} \Big \| \sum_{r_j \geq 4 \pi (\tilde{m} \|\tilde{n}\|)^{\beta} } \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \check{\Phi}(r_j) \Big \| \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+\frac{m^{\frac{1}{4}}+\|n\|^{\frac{1}{4}}}{\|mn\|^{\frac{\beta}{2}}}+\|mn\|^{\frac{1}{4}-\frac{3}{2} \beta} \Big) \|mn\|^{\varepsilon} \log^2 x.$$ Combining and we obtain $$\label{fullcusp2} \mathcal{W} \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+\|mn\|^{\frac{3}{14}+5 \beta+\varepsilon }+ \frac{m^{\frac{1}{4}}+\|n\|^{\frac{1}{4}}}{\|mn\|^{\frac{\beta}{2}}}+\|mn\|^{\frac{1}{4}-\frac{3}{2} \beta} \Big) \|mn\|^{\varepsilon} \log^2 x.$$ To balance the $x$-aspect of and we choose $T:=x^{\frac{2}{3}}$. To balance the $mn$-aspect of and the contribution from the initial segment for $c$, we set $\beta=\frac{1}{154}$. Combining and leads to the result. . Proof of Theorem \[mainthm2\] {#mainsec2} ============================= \[mainprop2\] Let $m,n>0$ be integers such that $m-1 \not \in \mathcal{P}$ or $n-1 \not \in \mathcal{P}$. Then for $x \geq 4 \pi \sqrt{\tilde{m} \tilde{n}}$ we have $$\sum_{x \leq c \leq 2x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \Big( x^{\frac{1}{6}}+(mn)^{\frac{1}{4}} \Big) (mn)^{\varepsilon} \log x.$$ Proposition \[mainprop2\] implies Theorem \[mainthm2\]. The initial segment $1 \leq c \leq 4 \pi \sqrt{\tilde{m} \tilde{n}}$ contributes $O_{\varepsilon} \big((mn)^{\frac{1}{4}+\varepsilon} \big)$ by [@AA (2.30)] to . One then breaks the interval $4 \pi \sqrt{\tilde{m} \tilde{n}} \leq c \leq X$ into $O(\log X)$ dyadic intervals $x \leq c \leq 2x$ with $4 \pi \sqrt{\tilde{m} \tilde{n}} \leq x \leq X/2$, and applies Proposition \[mainprop2\]. Let $\phi$ be a smooth test function with the properties listed in Section \[proskurin\]. Fix $a:=4 \pi \sqrt{\tilde{m} \tilde{n}}$ and let $T>0$ be chosen later. Recall that holds (for $m,n>0$ as well) and that by we have $$\label{newdyad} \sum_{c=1}^{\infty} \frac{S(m,n,c,\chi)}{c} \phi \Big( \frac{a}{c} \Big)=\mathcal{U}+\mathcal{V}.$$ Note that there is no contribution from $r_0=i/4$ by Remark \[specremark\] and the hypothesis on $m$ and $n$. When $x \geq 4 \pi \sqrt{\tilde{m} \tilde{n}}$, Lemma \[Ubd\] guarantees $$\label{Ubd2} \mathcal{U} \ll_{\varepsilon} (mn)^{\frac{1}{4}+\varepsilon}.$$ Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$. Applying Lemma \[int\] and Proposition \[andersen\] we see that for $A \geq 1$ (this is sufficient by Remark \[specremark\]) we have $$\begin{aligned} \sqrt{\tilde{m} \tilde{n}} & \Big \| \sum_{A \leq r_j \leq 2A} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \hat{\phi}(r_j) \Big \| \nonumber \\ & \ll \min \big (A^{-1},A^{-2} \frac{x}{T} \big) \Big( m \sum_{j \geq 1} \frac{\|\rho_j(m)\|^2}{\text{ch}(\pi r_j)} \Big)^{\frac{1}{2}} \Big( n \sum_{j \geq 1} \frac{\|\rho_j(n)\|^2}{\text{ch}(\pi r_j)} \Big)^{\frac{1}{2}} \label{startpoint2} \\ & \ll_{\varepsilon} \min \big (A^{-1},A^{-2} \frac{x}{T} \big) \big(A^{\frac{3}{2}}+m^{\frac{1}{2}+\varepsilon} A^{-\frac{1}{2}} \|\log A\|^{\frac{1}{2}+\varepsilon} \big)^{\frac{1}{2}} \big(A^{\frac{3}{2}}+n^{\frac{1}{2}+\varepsilon} A^{-\frac{1}{2}} \|\log A\|^{\frac{1}{2}+\varepsilon} \big)^{\frac{1}{2}} \nonumber \\ & \ll_{\varepsilon} \min \big (\sqrt{A},\frac{x}{T \sqrt{A}} \big) \big(1+A^{-1} \|\log A\| (m^{\frac{1}{4}+\varepsilon}+n^{\frac{1}{4}+\varepsilon})+A^{-2} \|\log A\| (mn)^{\frac{1}{4}+\varepsilon} \big). \label{startpoint}\end{aligned}$$ Summing over dyadic intervals $[A,2A]$ for $A \geq \max \big(a/x,1 \big)=1$ we obtain $$\label{fullcusp4} \mathcal{V} \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+(mn)^{\frac{1}{4}+\varepsilon} \Big) \log x.$$ To balance the $x$-aspect of and we choose $T:=x^{\frac{2}{3}}$. Combining , , and leads to the result. . Proof of Theorem \[auxthm2\] {#auxsec} ============================ The heart of the argument is to use Theorem \[thelif\] as a means to access the $H_{\theta}$–hypothesis. \[auxprop2\] Let $m,n>0$ be such that $m-1 \not \in \mathcal{P}$ or $n-1 \not \in \mathcal{P}$. Suppose $m_0,n_0$ be integers such that $24m-23=m_0^2 s$ and $24n-23=n_0^2 t$ with $s$ and $t$ square-free. If $x \geq (st)^{\frac{1}{6}} (mn)^{\frac{1}{3}}$, then under the $H_{\theta}$–hypothesis we have $$\sum_{x \leq c \leq 2x} \frac{S(m,n,c,\chi)}{c} \ll_{\varepsilon} \Big( x^{\frac{1}{6}}+ (st)^{\frac{1}{4}}+(st)^{\frac{1}{12}} (mn)^{\frac{1}{6}}+ (mnst)^{\frac{1}{8}+\frac{\theta}{4}} \Big) (mn)^{\varepsilon} \log^2 x.$$ As before, Proposition \[auxprop2\] implies Theorem \[auxthm2\] with the choice of initial segment $1 \leq c \leq (st)^{\frac{1}{6}} (mn)^{\frac{1}{3}}$. Let $\phi$ be a smooth test function with the properties listed in Section \[proskurin\]. Fix $a:=4 \pi \sqrt{\tilde{m} \tilde{n}}$ and let $T>0$ be chosen later. Let $\{u_j\}$ be an orthonormal basis for $\tilde{\mathcal{L}}_{\frac{1}{2}}(1,\chi)$ consisting of Hecke eigenforms of $\mathcal{T}_{p^2}$ for all primes $p \nmid 6$. We follow the proof of Proposition \[mainprop2\] to (note that holds for $m,n>0$ as well). To bound $\mathcal{V}$, we will treat the spectral parameter separately on different ranges. Note that it is sufficient to consider $r_j \geq 1$ by Remark \[specremark\]. In view of Lemma \[int\], we consider the ranges $$\begin{aligned} 1 \leq & r_j \leq a/x, \\ & r_j \geq \max \big(a/x,1 \big). \end{aligned}$$ We first consider the case $r_j \geq \max \big( a/x,1 \big)$. For $A \geq \max \big( a/x,1 \big)$, we first prove $$\begin{gathered} \label{bound2} \sqrt{\tilde{m} \|\tilde{n}\|} \Bigg \| \sum_{A \leq r_j \leq 2A} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \hat{\phi}(r_j) \Bigg \| \ll_{\varepsilon} (m_0 n_0)^{\theta+\varepsilon} \min \big \{\sqrt{A},\frac{x}{T \sqrt{A}} \big \} \\ \times \big(1+A^{-1} \|\log A\| (s^{\frac{1}{4}+\varepsilon}+t^{\frac{1}{4}+\varepsilon})+A^{-2} \|\log A\| (st)^{\frac{1}{4}+\varepsilon} \big). \end{gathered}$$ To prove we start with . Using in and then applying Proposition \[andersen\], we establish . Combining and , for $A \geq \max \big( a/x,1 \big)$ we have $$\begin{gathered} \label{mincomp} \sqrt{\tilde{m} \tilde{n}} \Bigg \| \sum_{A \leq r_j \leq 2A} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \hat{\phi}(r_j) \Bigg \| \ll_{\varepsilon} \min \Big (\sqrt{A},\frac{x}{T \sqrt{A}} \Big) \times \\ \min \Bigg(1+ A^{-1} \|\log A\| (m^{\frac{1}{4}+\varepsilon}+n^{\frac{1}{4}+\varepsilon})+A^{-2} \|\log A\| (mn)^{\frac{1}{4}},\\ (m_0 n_0)^{\theta+\varepsilon} \Big[1+A^{-1} \| \log A \| (s^{\frac{1}{4}+\varepsilon}+t^{\frac{1}{4}+\varepsilon})+A^{-2} \|\log A\| (st)^{\frac{1}{4}+\varepsilon} \Big] \Bigg).\end{gathered}$$ Using the facts that for positive $B,C$ and $D$ we have $$\min \big( B+C,D \big) \leq \min \big( B,D \big)+ \min \big( C,D \big) \quad \text{and} \quad \min \big( B,C \big) \leq \sqrt{BC},$$ we simplify . This right side of is $$\begin{gathered} \label{mincompup} \ll_{\varepsilon} \min \Big (\sqrt{A},\frac{x}{T \sqrt{A}} \Big) \Bigg( 1+ (m_0 n_0)^{\frac{\theta}{2}+\varepsilon} \|\log A\| \Big[ A^{-1} (m^{\frac{1}{8}+\varepsilon}+n^{\frac{1}{8}+\varepsilon})(s^{\frac{1}{8}+\varepsilon}+t^{\frac{1}{8}+\varepsilon}) \\+A^{-\frac{1}{2}} (m^{\frac{1}{8}+\varepsilon}+n^{\frac{1}{8}+\varepsilon}) +A^{-\frac{3}{2}} (m^{\frac{1}{8}+\varepsilon}+n^{\frac{1}{8}+\varepsilon})(st)^{\frac{1}{8}+\varepsilon} +A^{-1} (mn)^{\frac{1}{8}+\varepsilon} \\ +A^{-\frac{3}{2}} (mn)^{\frac{1}{8}+\varepsilon} (s^{\frac{1}{8}+\varepsilon}+t^{\frac{1}{8}+\varepsilon} \big)+A^{-2} (mnst)^{\frac{1}{8}+\varepsilon} \Big] \Bigg).\end{gathered}$$ Summing over all dyadic intervals $[A,2A]$ with $A \geq \max \big( a/x,1 \big)$ and ignoring the smallest terms we obtain $$\begin{aligned} \sqrt{\tilde{m} \tilde{n}} \Big \| \sum_{r_j \geq \max (a/x,1)} \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \hat{\Phi}(r_j) \Big \| & \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+(mnst)^{\frac{1}{8}+\varepsilon} (m_0 n_0)^{\frac{\theta}{2}+\varepsilon} \Big) \log^2 x \nonumber \\ & \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+(mnst)^{\frac{1}{8}+\frac{\theta}{4}+\varepsilon} \Big) \log^2 x \label{Vapprox2}.\end{aligned}$$ Since $x \geq (mn)^{\frac{1}{3}}$, we have $a/x \leq 4 \pi (mn)^{\frac{1}{6}}$. When $r_j \leq a/x$, Lemma \[int\] and the fact that $r_j>1.9$ guarantees that we can replace $\min \big(\sqrt{A},x/T \sqrt{A} \big)$ in and with $\sqrt{A}$. Thus $$\label{ineq2} \sqrt{\tilde{m} \tilde{n}} \Big \| \sum_{0<r_j \leq a/x } \frac{\overline{\rho_j(m)} \rho_j(n)}{\text{ch}(\pi r_j)} \hat{\phi}(r_j) \Big \| \ll_{\varepsilon} (mnst)^{\frac{1}{8}+\frac{\theta}{4}+\varepsilon}.$$ Combining and we have $$\label{ineq1} \mathcal{V} \ll_{\varepsilon} \Big( \sqrt{\frac{x}{T}}+ (mnst)^{\frac{1}{8}+\frac{\theta}{4}} \Big) (mn)^{\varepsilon} \log^2 x.$$ We choose $T:=x^{\frac{2}{3}}$ to balance (with $m,n>0$) and . When $x \geq (st)^{\frac{1}{6}} (mn)^{\frac{1}{3}}$, Lemma \[Ubd\] gives $$\label{finalU} \mathcal{U} \ll_{\varepsilon} (st)^{\frac{1}{4}+\varepsilon} +(st)^{\frac{1}{12}} (mn)^{\frac{1}{6}+\varepsilon}.$$ Combining , and finishes the proof. Acknowledgements ================ The author thanks Professor Scott Ahlgren for his careful reading of the manuscript and both of the referees for their thorough reports. The author is grateful to Nick Andersen for his insightful comments. [99]{} <span style="font-variant:small-caps;">S. Ahlgren and N. Andersen</span>, Kloosterman sums and Maass cusp forms of half integral weight for the modular group, IMRN, rnw234, (2018). <span style="font-variant:small-caps;">S. Ahlgren and A. Dunn</span>, Maass forms and the mock theta function $f(q)$, https://arxiv.org/abs/1806.01187. <span style="font-variant:small-caps;">S. Ahlgren and N. Andersen</span>, Algebraic and transcendental formulas for the smallest parts function. Adv. Math. 289 (2016), 411–437. <span style="font-variant:small-caps;">N. Andersen and W. Duke</span>, Modular invariants for real quadratic fields and Kloosterman sums, https://arxiv.org/pdf/1801.08174.pdf. <span style="font-variant:small-caps;">V. Blomer</span>, Sums of Hecke eigenvalues over values of quadratic polynomials. IMRN, 2008, no. 16, rnn059, 29pp. <span style="font-variant:small-caps;">R. W. Bruggeman</span>, Kloosterman sums for the modular group. In Number theory in progress, Vol. 2 (Zakopane-Kóscielisko, 1997), pages 651–674. de Gruyter, Berlin, 1999. <span style="font-variant:small-caps;">B. Cipra</span>, On the Niwa–Shintani theta-kernel lifting of modular forms. Nagoya Math. J., 91 (1983), 49–117. <span style="font-variant:small-caps;">P. Deligne</span>, La conjecture de Weil, I. Publ. Math., 43, (1973), 273–308. <span style="font-variant:small-caps;">W. Duke</span>, Hyperbolic distribution problems and half-integral weight Maass forms, Invent. Math., 92, (1988),73–90. <span style="font-variant:small-caps;">W. Duke, J.B Friedlander and H. Iwaniec</span>, The subconvexity problem for Artin $L$-functions Invent. Math, 149, (2002), no. 3, 489-577. <span style="font-variant:small-caps;">W. Duke, J.B Friedlander and H. Iwaniec</span>, Weyl sums for quadratic roots, Int. Math. Res. Not. IMRN, (2012), no. 11, 2493–2549. NIST Digital Library of Mathematical Functions. http://dlmf.nist.gov/, Release 1.0.10 of 2015-08- 07. Online companion to [@OLBC]. <span style="font-variant:small-caps;">S. Ganguly and J. Sengupta</span>, Sums of Kloosterman sums over arithmetic progressions. Int. Math. Res. Not. IMRN, (1):137–165, 2012. <span style="font-variant:small-caps;">D. Goldfeld and P. Sarnak</span>, Sums of Kloosterman sums, Invent. Math. 71, (1983), no. 2, 243–250. <span style="font-variant:small-caps;">D. R. Heath Brown</span>, Arithmetic applications of Kloosterman sums, Nieuw Arch. Wiskd., (5), 1(4):380–384, 2000. <span style="font-variant:small-caps;">H.H Kim</span>, Functoriality for the exterior square of $\text{GL}_4$ and the symmetric fourth of $\text{GL}_2$, J. Amer. Math. Soc., 16, (2003), no. 1, 139–183. With appendix 1 by D. Ramakrishnan and appendix 2 by Kim and P. Sarnak. <span style="font-variant:small-caps;">E. M. Kiral</span>, Opposite-sign Kloosterman sum zeta function, Mathematika, 62, (2016), no. 2, 406–429. <span style="font-variant:small-caps;">N. Koblitz</span>, Introduction to elliptic curves and modular forms. Graduate texts in mathematics, 97. Springer–Verlag, New York, 1984. <span style="font-variant:small-caps;">M.I. Knopp</span>, Modular functions in analytic number theory, Markham Publishing Co., Chicago, Ill., 1970, MR 0265287 (42 \#198). <span style="font-variant:small-caps;">N. V. Kuznetsov</span>, The Petersson conjecture for cusp forms of weight zero and the Linnik conjecture. Sums of Kloosterman sums. Mat. Sb. (N.S.), 111, (1980), no. 153 :334-383. <span style="font-variant:small-caps;">J.V. Linnik</span>, Additive problems and eigenvalues of the modular operators. In Proc. Internat. Congr. Mathematicians (Stockholm, 1962), pages 270–284. Inst. Mittag-Leffler, Djursholm, 1963. <span style="font-variant:small-caps;">K. Ono</span>, The web of modularity: arithmetic of the coefficients of modular forms and $q$–series. CBMS Regional Conference Series in Mathematics, 102. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2004. viii+216 pp. ISBN: 0-8218-3368-5. <span style="font-variant:small-caps;">F. W. J. Olver, D. W. Lozier, R. F. Boisvert, and C. W. Clark, editors.</span> NIST Handbook of Mathematical Functions. Cambridge University Press, New York, NY, 2010. Print companion to [@DL]. <span style="font-variant:small-caps;">W. d. A. Pribitkin</span>, A generalization of the Goldfeld-Sarnak estimate on Selberg’s Kloosterman zeta-function. Forum Math., 12, (2000), no. 4, 449–459. <span style="font-variant:small-caps;">N.V. Proskurin</span>, On the general Kloosterman sums, Journal of the Mathematical Sciences 129, (2006), no. 3, 3874–3889. <span style="font-variant:small-caps;">H. Rademacher</span>, On the partition function $p(n)$. Proc. London Math. Soc., 43, (4):241–254, 1936. <span style="font-variant:small-caps;">H. Rademacher</span>, On the expansion of the partition function in a series. Ann. of Math. (2), 44:416–422, 1943. <span style="font-variant:small-caps;">H. Rademacher</span> Topics in analytic number theory. Springer-Verlag, New York-Heidelberg, 1973. Edited by E. Grosswald, J. Lehner and M. Newman, Die Grundlehren der mathematischen Wissenschaften, Band 169. <span style="font-variant:small-caps;">P. Sarnak and J. Tsimerman</span>, On Linnik and Selberg’s conjecture about sums of Kloosterman sums, Algebra, arithmetic, and geometry: in honor of Yu. I. Manin. Vol. II, 619–635, Progr. Math., 270, Birkhauser Boston, Inc., Boston, MA, 2009. <span style="font-variant:small-caps;">P. Sarnak</span>, Additive number theory and Maass forms. In Number theory (New York,1982), volume 1052 of Lecture notes in Math., pages 286–309. Springer, Berlin, 1984. <span style="font-variant:small-caps;">P. Sarnak</span> Some applications of modular forms, Cambridge Tracts in Math., vol 99., Cambridge University Press, 1990. <span style="font-variant:small-caps;">A. Selberg</span>, On the estimation of Fourier coefficients of modular forms. In Proc. Sympos. Pure Math., Vol. VIII, pages 1–15. Amer. Math. Soc., Providence, R.I., 1965. <span style="font-variant:small-caps;">G. Shimura</span>, On modular forms of half integral weight, Ann Math. 97, (1973), 440–481. <span style="font-variant:small-caps;">G. Watson</span>, A treatise on the Theory of Bessel Functions. Cambridge University Press, (1966). <span style="font-variant:small-caps;">A. Weil</span>, On some exponential sums, Proc. Nat. Acad. Sci. U. S. A., 34:204–207, 1948. <span style="font-variant:small-caps;">F. Waibel</span>, Fourier coefficients of half–integral weight cusp forms and Waring’s problem, preprint. <span style="font-variant:small-caps;">Y. Yang</span>, Modular forms of half integral weights on $\text{SL}_2(\mathbb{Z})$, Nagoya Math. J., 215 (2014),1–66.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this [Letter]{}, we report detections of SiO $v=3$ $J=1\rightarrow 0$ maser emission in very long baseline interferometric (VLBI) observations towards 4 out of 12 long-period variable stars: WX Psc, R Leo, W Hya, and T Cep. The detections towards WX Psc and T Cep are new ones. We also present successful astrometric observations of SiO $v=2$ and $v=$3 $J=1\rightarrow 0$ maser emissions associated with two stars: WX Psc and W Hya and their position-reference continuum sources: J010746.0$+$131205 and J135146.8$-$291218 with the VLBI Exploration of Radio Astrometry (VERA). The relative coordinates of the position-reference continuum source and SiO $v=3$ maser spots were measured with respect to those of an SiO $v=2$ maser spot adopted as fringe-phase reference. Thus the faint continuum sources were [inversely]{} phase-referenced to the bright maser sources. It implies possible registration of multiple SiO maser line maps onto a common coordinate system with 10 microarcsecond-level accuracy.' author: - \| Hiroshi <span style="font-variant:small-caps;">Imai</span>, Miyako <span style="font-variant:small-caps;">Oyadomari</span>, Sze Ning <span style="font-variant:small-caps;">Chong</span>, Akiharu <span style="font-variant:small-caps;">Nakagawa</span>,\ Tomoharu <span style="font-variant:small-caps;">Kurayama</span>, Jun-ichi <span style="font-variant:small-caps;">Nakashima</span>, Naoko <span style="font-variant:small-caps;">Matsumoto</span>, Takumi <span style="font-variant:small-caps;">Nagayama</span>,\ Tomoaki <span style="font-variant:small-caps;">Oyama</span>, Shota <span style="font-variant:small-caps;">Mizuno</span>, Shuji <span style="font-variant:small-caps;">Deguchi</span>, and Se-Hyung <span style="font-variant:small-caps;">Cho</span> title: 'Pilot VLBI Survey of SiO $v=$3 $J=1\rightarrow 0$ Maser Emission around Evolved Stars' --- Introduction ============ Silicon monoxide (SiO) maser emission has been used as an important probe of the dynamical structure and the physical condition of the inner circumstellar envelopes (CSEs) of asymptotic giant branch (AGB) and post-AGB stars. The pumping mechanism of the SiO masers is still an open question and understanding of it is essential to the diagnostics of the CSEs through observed behaviors of clump clusters of the masers such as temporal variations of the flux density, angular distribution, and three-dimensional velocity structure. SiO maser emissions of $v=1$ $J=1\rightarrow 0$, $v=2$ $J=1\rightarrow0$, and $v=1$ $J=2\rightarrow1$ have been main targets of very long baseline interferometric (VLBI) observations (e.g., [@sor04] and references therein). The $v=3$ $J=1\rightarrow 0$ maser line is also a unique target ([@ima10], hereafter Paper [I]{}; [@des12]). This transition is located at an energy level higher by $\sim$4 000 cm$^{-1}$ ($\sim$5 800 K) than the rotational transitions in the vibrational ground state and it needs considerably strong excitation in the gas at a temperature of 2 000–3 000 K of the surface of AGB and post-AGB stars. Observations of this maser line may be a good test for currently most plausible maser pumping model (line-overlapping, [@sor04] and references therein). However, its detection in VLBI observations is difficult due to its extreme weakness [@cho96; @nak07]. By the way, precise measurement of relative positions of maser spots in the different maser transitions is essential for correctly deducing the pumping mechanism of SiO masers. Therefore, the registration technique of multiple SiO maser line maps onto a common coordinate system is always an interesting issue and should be improved. In any technique, accurate determination of the absolute coordinates of maser spots and accurate evaluation of the errors associated with instrumental factors (e.g., accuracy of VLBI station coordinates on the terrestrial reference frame) are major factors to reduce the map registration uncertainty. Straightforward measurement of maser spot positions with respect to extragalactic continuum sources (e.g., [@zha12] and references therein) yields a 10 microarcsecond($\mu$as)-level accuracy of the map registration. Note that such a position-reference continuum source is often too faint to detect within a VLBI coherence time in the 40 GHz band ($\lesssim$2 min). Therefore, a special technique for phase-referencing coherent integration is required [@ima13]. In this [Letter]{}, we report new detections of SiO $v=3$ $J=1\rightarrow 0$ maser emission towards two long-period variable stars: WX Piscium (WX Psc) and T Cephei (T Cep) and two repetitive detections of the $v=3$ masers towards R Leonis (R Leo)(its first detection was reported by [@des12]) and W Hydrae (W Hya) in VLBI observations with the telescopes of Nobeyama Radio Observatory (NRO) and the VLBI Exploration of Radio Astrometry (VERA). The relatively high-sensitivity and short baselines including the NRO 45 m telescope enabled us to detect the faint $v=3$ masers. We also report successful phase-referencing VLBI observations of SiO $v=2$ and $v=3$ $J=1\rightarrow 0$ maser emissions associated with WX Psc and W Hya conducted with the VERA telescopes in the dual-beam mode. With VERA’s baselines (1000–2300 km), only compact unresolved maser spots are detectable, so it is difficult to analyze and discuss the difference in the maser distributions between the different transitions of SiO masers. Instead, the VERA astrometry provides a good anchor of position reference for SiO map registration as mentioned above. This [Letter]{} focuses on technical points of VERA astrometry for yielding a 10 $\mu$as-level accuracy of map registration and pays attention to providing a guideline for such astrometric observations in the 40 GHz band in which continuum source flux densities are usually lower than those in lower frequency bands and below a detection limit of $\sim$300 mJy with VERA within VLBI coherence time. ------ -------------------- ------ --------------------------------------------- ---------------------- --------------- ------------------------- ------------------- Pair Scan $V_{{\rm ref}, v=2}$ Synthesized 1-$\sigma$ noise level ID Source name (hr) R.A (J2000) Decl. (J2000) (km s$^{-1}$) beam pattern (mJy beam$^{-1}$) 1 WX Psc 2.6 01$^{\rm h}$06$^{\rm m}$25$^{\rm s}$.99 $+$123553.4 7.4 0.38$\times$0.69, $-$39 194, 201 J010746.0$+$131205 01$^{\rm h}$07$^{\rm m}$45$^{\rm s}$.961872 $+$131205.19067 ... 0.47$\times$0.64, $-$48 3.9 2 W Hya 2.8 13$^{\rm h}$49$^{\rm m}$01$^{\rm s}$.9311 $-$282204.560 40.4 0.44$\times$1.17, $-$19 184, 184 J135146.8$-$291218 13$^{\rm h}$51$^{\rm m}$46$^{\rm s}$.838765 $-$291217.65002 ... 0.43$\times$1.15, $-$18 10.6 ------ -------------------- ------ --------------------------------------------- ---------------------- --------------- ------------------------- ------------------- Total integration time.\ LSR velocity of the SiO $v=2$ $J=1\rightarrow 0$ maser spot adopted as phase-reference.\ Major and minor axis length in unit of mas and position angle.\ In maser image cubes, emission-free spectral channel maps of the $v=2$ and $v=3$ lines were chosen for the calculation.\ Observations and data reduction =============================== We conducted VLBI observations of SiO $v=2$ and $v=3$ $J=1\rightarrow$0 maser emissions (at rest frequencies of 42.820582 and 42.519340 GHz, respectively) on 2012 March 24–25 and 2012 May 20–21 for 25 hr each towards 12 stars in total: WX Psc, AP Lyn, U Ori, VY CMa, R Leo, RS Vir, W Hya, U Her, RU Her, V1111 Oph, V4120 Sgr, and T Cep. They were selected from the sources detected in the SiO $v=3$ $J=1\rightarrow 0$ maser emission by @cho96. Three or four VERA$^{1}$ 20 m telescopes and the 45 m NRO[^1] were simultaneously operated; the former for observing the target maser and position-reference continuum sources using the dual-beam receiving system of VERA and the latter for the maser sources only but with higher sensitivity. Both of the VERA and NRO telescopes also observed bright continuum calibrators every 40 min for calibration of instrumental group-delay and fringe-phase residuals and bandpass characteristics. The observed signals, received in left-hand circular polarization, were digitized in four levels and divided into 16 base band channels (BBCs) each with a band width of 16 MHz. In the VERA stations, they were recorded with both the SONY DIR1000 and DIR2000 recorders at rates of 128 and 1024 M bits s$^{-1}$, respectively. Two of the BBCs were assigned to the SiO $v=2$ and $v=3$ masers in one beam and others to the position-reference continuum source emission in other beam. The DIR1000 recording was also made in NRO and accepted only the two SiO maser BBCs. The data correlation was processed with the Mitaka FX correlator, in which each of the maser BBCs was split into 512 spectral channels, corresponding to a velocity spacing of 0.22. On the other hand, each of the continuum BBCs was split into 32 spectral channels. In March, each of maser–continuum source pairs was observed for only 2–3 hr. Because the DIR1000 and DIR2000 data were available from only three and four stations, respectively, the imaging and the astrometry of maser sources were difficult. After the normal calibration procedures as mentioned soon later, we obtained the SiO $v=2$ and $v=3$ maser spectra with high sensitivity by coherently integrating the whole visibility data. In May, the observations similar to those in March were conducted, but the maser sources with the $v=3$ detections in March had longer scans (for 4–5 hr) for more surely successful maser source astrometry with VERA. In this [Letter]{}, we focus on the spectra of SiO $v=2$ and $v=3$ $J=1\rightarrow 0$ maser emissions towards WX Psc, R Leo, W Hya, and T Cep with the successful $v=3$ detections in the NRO–VERA baselines in March and the successful astrometric results of WX Psc and W Hya in the VERA dual-beam observations in May. The results of the whole observations including all SiO $v=2$ and $v=3$ maser maps will be published in a separate paper. Data reduction and image synthesis were made using the NRAO AIPS package. In order to conduct the astrometry using faint continuum sources, the [inverse]{} phase-referencing technique was adopted, which is described in more detail in @ima13[^2]. First, calibration of visibility amplitudes for antenna gains and bandpass characteristics and that of visibility phases for instrumental group-delay and phase residuals were made in a standard manner by using scans on the continuum calibrators. Then, fringe fitting and self-calibration were performed using visibilities in the velocity channel (Column 6 of Table \[tab:sources\]), which includes a $v=2$ bright maser spot (velocity component). The Doppler velocity is given with respect to the local standard of rest (LSR). The solutions were applied to the data in other velocity channels of the $v=2$ maser and those of the $v=3$ maser as well as to the data of the position-reference source. In the May session, thus we obtained the maser image cubes and the reference source maps with naturally weighted visibilities, which yielded a typical size of the synthesized beam of 0.5 milliarcsecond (mas). Source name Offsets (mas) ------------- ------------------------------------- WX Psc ($-24.24\pm0.05$, $-453.14\pm0.03$) W Hya ($240.00\pm0.01$, $263.67\pm0.03$) : Results of the VERA astrometry[]{data-label="tab:astrometry"} Measured (R.A., decl.) offsets with respect to the phase-tracking center shown in Table \[tab:sources\].\ (160mm,50mm)[fig1.eps]{} (150mm,50mm)[fig2.eps]{} (175mm,80mm)[fig3.eps]{} Results and Discussion {#sec:results} ====================== Figure \[fig:spectra\] shows the SiO $v=2$ and $v=3$ $J=1\rightarrow 0$ maser spectra with the successful $v=3$ detections. All of these maser sources were observed in the nearly stellar light maxima (during the light curve phase of $\phi\approx$0.8–1.0). The $v=3$ maser was again detected toward W Hya (with a pulsation period of $P\approx$360 d) three years after the first VLBI detection (Paper I). They are consistent with the suggestion that the $v=3$ maser should have strong correlation with the stellar light curve [@cho07]. The $v=3$ masers seem to be associated with relatively bright $v=2$ maser components on the spectra, however as mentioned later, they are not necessarily spatially coincident with each other. Table \[tab:sources\] gives parameters of the individual pairs of maser and position-reference sources with the successful astrometric results. It includes the parameters of the observation, signal correlation, and data reduction. The counterpart continuum sources of these maser sources, J010746.0$+$131205 and J135146.8$-$291218, respectively, are separated by 1.03 and 0.69 from the masers. Figure \[fig:continuums\] shows the phase-referenced images of the two position-reference sources. The 1-$\sigma$ noise levels of the images are 3.4 and 10.1 mJy beam$^{-1}$, respectively. The measured position offset of the continuum source (in Figure \[fig:continuums\]) is inversely equal to that of the phase-reference $v=2$ maser spot from the phase-tracking center of the maser source data (Columns 4 and 5 in Table \[tab:sources\]). Table \[tab:astrometry\] gives the determined coordinate offsets of the phase-reference $v=2$ maser spots. Because the absolute coordinates of the reference sources are determined with an accuracy better than 0.5 mas (e.g. [@pet08])[^3], those of the reference maser spots are also determined in this level of accuracy. The position drift of the $v=3$ maser map with respect to the $v=2$ map appears proportionally to the position offset of the phase-reference maser spot and the frequency difference between the $v=2$ and $v=3$ masers ([@gwi92]; Paper I). Using the derived position offsets mentioned above, this position drift was corrected for the $v=3$ maps.[^4] In the present VLBI observations, the map registration of the $v=2$ and $v=3$ maser lines has an uncertainty less than 50 $\mu$as, small enough to discuss the maser pumping models. Figure \[fig:masers\] shows the composite maps of SiO $v=2$ and $v=3$ $J=1\rightarrow 0$ maser lines in WX Psc and W Hya. The maser emissions were significantly spatially resolved; the cross-power flux densities of the $v=2$ and $v=3$ masers were much lower than the total-power flux densities ($\sim$400 and $\sim$50 Jy in WX Psc and $\sim$1000 and $\sim$350 Jy in W Hya, respectively). Comparing with those of previous observations (e.g., [@sor04] and Paper [I]{}, respectively), only for unresolved, compact maser spots, we find that their relative positions have roughly persisted for a decade. Assuming positions of the central stars at ($\sim$13, $\sim$5) and ($\sim$25, $\sim$30) \[mas\] for WX Psc and W Hya, respectively, the distance to the $v=3$ spots from the star is roughly equal to that of the $v=2$ spots, consistent with the previous observations (Paper [I]{}; [@des12]). As already mentioned, the $v=3$ and $v=2$ maser spots were not spatially coincident, in contrast to the results of Paper I and @des12. These results favor a collisional pumping scheme for SiO masers (e.g., [@loc92; @hum02]). However, it is premature to draw ring-shaped distribution models in Figure \[fig:masers\] and to discuss the correlation of the $v=2$ and $v=3$ emission distributions in the present results in which other fainter and more extended maser spots are missing. Thus we cannot rule out a scheme of line-overlapping between SiO and 2o ro-vibrational transitions, not only for $v=2$ $J=1\rightarrow 0$ [@sor04] but also $v=3$ $J=1\rightarrow 0$. @cho07 consider the most plausible overlap pair between the SiO $v=2$ $J=0 \rightarrow$ $v=3$ $J=1$ and the 2o $\nu_{2}=2$ $5_{05} \rightarrow$ $\nu_2=1$ $6_{34}$ lines. This overlapping can excite the SiO $v=3$ $J=1\rightarrow$0 maser. Temporal variation in spot displacement of different maser lines as predicted by @hum02 can be examined by future monitoring observations with combination of high precision astrometry as those demonstrated in this paper and high dynamic imaging of the maser sources (e.g., [@yi05]). Here we emphasize that the present success in long-time coherent integration using the [inverse]{} phase-referencing technique is a milestone for VERA astrometry using maser–continuum source pairs in the 40 GHz band. In the maser source astrometry with the Very Long Baseline Array (VLBA)(e.g. [@zha12]), SiO maser sources have been often used as phase-reference because of their detections at signal-to-noise ratios higher than those of continuum sources. On the other hand, taking into account more complicated procedures of the inverse phase-referencing technique [@ima13], a relatively high signal recording rate of VERA has favored continuum sources as phase-reference. Therefore, successful astrometric observations had been limited to nearly ideal cases in which both of maser and continuum sources should be bright enough to detect within a VLBI coherence time (e.g., [@cho09]). Such a condition empirically demands flux densities higher than $\sim$20 Jy and 300 mJy for SiO maser and 40 GHz continuum sources for VERA astrometry, respectively. The present results prove that a 30 mJy-level continuum source is detectable with the [inverse]{} phase-referencing technique by using data of a bright SiO maser source. This has advantage over the continuum-reference case in which even a bright, well-studied SiO maser source was dropped out from the target list of VERA astrometry because of the absence of bright continuum sources within 0.5–2.2 of the maser source for the VERA dual-beam observations. As a result of the milestone achievement, a statistically moderate number ($>$10) of well-studied SiO maser sources harboring multiple transitions of SiO masers, including $v=3$ $J=1\rightarrow 0$, will become targets for the study on pumping mechanism of SiO masers in VLBI and for the trigonometry with VERA. Moreover, in such a situation, we have to concern only about the dispersion of maser spot distributions, which is much larger than the uncertainty of relative positions among the different lines of SiO masers. This is helpful for unambiguous interpretation of the pumping mechanism of the SiO masers. We acknowledge all staff members and students who have helped in array operation and in data correlation of VERA. HI has been supported for stay at ICRAR by the Strategic Young Researcher Overseas Visits Program for Accelerating Brain Circulation funded by Japan Society for the Promotion of Science (JSPS). JN has been supported by the Research Grants Council of Hong Kong (project code: HKU 703308P, HKU 704209P, HKU 704710P, and HKU 704411P), and the Small Project Funding of the University of Hong Kong (201007176004). Cho, S.-H., Kaifu, N., & Ukita, N. 1996, , 115, 117 Cho, S.-H., Lee, C. W., & Park, Y.-S. 2007, , 657, 482 Choi, Y. K., Hirota, T., Honma, M., & Kobayashi, H. 2009, in Approaching Micro-Arcsecond Resolution with VSOP-2: Astrophysics and Technologies, ASP Conf. Ser., 402, eds. Hagiwara, Y., Fomalont, E., Tsuboi, M., & Murata, Y., p363 Desmurs, J.-F., Bujarrabal, V., Lindqvist, M., Alcolea, J., Soria-Ruiz, R., & Bergman, P. 2012, in Proceedings of the IAU Symp. 287, eds. Booth, R.SS., Humphreys, E. M. L., & Vlemmings, W. H. T. (Cambridge University Press: Cambridge), p252 Gwinn, C. R., Moran, J. M., & Reid, M. J. 1992, , 393, 149 Humphreys, E. M. L., Gray, M. D., Yates, J. A., Field, D., Brown, G. H., & Diamond, P. J. 2002, , 386, 256 Imai, H., Sakai, N., Nakanishi, H., Sakanoue, H., Honma, M., Miyaji, T. 2013, , in press Imai, H., Nakashima, J., Deguchi, S., Yamauchi, A., Nakagawa, A., & Nagayama, T. 2010, , 62, 431 (Paper [I]{}) Lockett, P., & Elitzur, M. 1992, , 399, 704 Nakashima, J., & Deguchi, S. 2007, , 669, 446 Petrov, L., Kovalev, Y. Y., Fomalont, E., & Gordon, D. 2008, , 136, 580 Soria-Ruiz, R., Alcolea, J., Colomer, F., Bujarrabal, V., Desmurs, J.-F., Marvel, K. B., & Diamond, P. J. 2004, , 426, 131 Yi, J., Booth, R. S., Conway, J. E., & Diamond, P. J. 2005, , 432, 531 Zhang, B., Reid, M. J., Menten, K. M., & Zheng, X. W. 2012, , 744, 23 [^1]: The NRO and VERA/Mizusawa VLBI observatory are branches of the National Astronomical Observatory of Japan, an interuniversity research institute operated by the Ministry of Education, Culture, Sports, Science and Technology. [^2]: We used the ParselTongue/python pipeline scripts through the data reduction, which are now available in the wiki page: http://milkyway.sci.kagoshima-u.ac.jp/groups/vcon\_lib/wiki/9fbfd/Data\_Analysis.html . See also the ParselTongue wiki page: http://www.jive.nl/dokuwiki/doku.php?id=parseltongue: parseltongue . [^3]: See also URL: http://astrogeo.org/vlbi/solutions/rfc\_2012b . [^4]: Phase correction using the AIPS task CLCOR before fringe-fitting with calibrator scans can perform the equivalent position correction.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We complete the study of the topological period-index problem over $8$ dimensional finite CW complexes started in a preceding paper. More precisely, we determine the sharp upper bound of the index of a topological Brauer class $\alpha\in H^3(X;{\mathbb{Z}})$, where $X$ is of the homotopy type of an $8$ dimensional finite CW complex and the period of $\alpha$ is divisible by $4$.' address: 'School of Mathematics and Statistics, the University of Melbourne, Parkville VIC 3010, Australia' author: - Xing Gu bibliography: - 'tpip8IIref.bib' title: 'The Topological Period-Index Problem over $8$-complexes, ' --- introduction ============ This paper is a sequel to Gu [@Gu], where we investigated the topological period-index problem (TPIP) over finite CW complexes of dimension $8$, and determined the upper bound of indices of topological Brauer classes with period $n$ not divisible by $4$. In this paper we give a complete answer to the TPIP over finite $8$-complexes by studying the case $4\|n$. For a path-connected topological space $X$, let $\operatorname{Br}(X)$ be the topological Brauer group defined in [@An1], whose underlying set is the Azumaya algebras (i.e. bundles of complex matrix algebras over $X$) modulo the Brauer equivalence: $\mathscr{A}_0$ and $\mathscr{A}_1$ are called Brauer equivalent if there are vector bundles $\mathscr{E}_0$ and $\mathscr{E}_1$ such that $$\mathscr{A}_0\otimes\operatorname{End}(\mathscr{E}_0)\cong \mathscr{A}_1\otimes\operatorname{End}(\mathscr{E}_1).$$ The multiplication is given by tensor product. Azumaya algebras over $X$ of degree $r$ are classified up to isomorphism by the collection of isomorphism classes of $PU_r$-principal bundles over $X$, i.e., the cohomology set $H^{1}(X;PU_r)$, where $PU_r$ is the projective unitary group of degree $r$. Consider the short exact sequence of Lie groups $$\label{ses1} 1\rightarrow S^1\rightarrow U_r\rightarrow PU_r\rightarrow 1$$ where the arrow $S^1\rightarrow U_r$ is the inclusion of scalars. Then the Bockstein homomorphism $$\label{Bockstein} H^{1}(X;PU_r)\rightarrow H^{2}(X;S^1)\cong H^{3}(X;\mathbb{Z})$$ associates an Azumaya algebra $\mathscr{A}$ to a class $\alpha\in H^{3}(X;\mathbb{Z})$. The exactness of the sequence (\[ses1\]) implies that 1. $\alpha\in H^{3}(X;\mathbb{Z})_{\textrm{tor}}$, the subgroup of torsion elements of $ H^{3}(X;\mathbb{Z})$, and 2. the class $\alpha$ only depends on the Brauer equivalence class of $\mathscr{A}$. An Azumaya algebra associated to a Brauer class may alternatively be described as follows. Let $\mathscr{A}$ be an Azumaya algebra of degree $r$ over a finite CW complex $X$. As discussed above, one can associate a $PU_r$-bundle, i.e., a homotopy class of maps $X\rightarrow\mathbf{B}PU_r$, where $\mathbf{B}PU_r$ is the classifying space of $PU_r$. An elementary computation yields $H^3(\mathbf{B}PU_r;{\mathbb{Z}})\cong{\mathbb{Z}}/r$. The Azumaya algebra $\mathscr{A}$ is associated to a class $\alpha\in H^{3}(X;\mathbb{Z})$ if and only if the following homotopy commutative diagram $$\label{eq:lift} \begin{tikzcd} &\mathbf{B}PU_r\arrow{d}\\ X\arrow[ur,dashed]\arrow{r}{\alpha}&K(\mathbb{Z},3) \end{tikzcd}$$ exists such that the dashed arrow represents the (isomorphism class) $PU_r$-bundle classifying $\mathscr{A}$, the vertical arrow corresponds to a generator of $H^3(\mathbf{B}PU_r)\cong{\mathbb{Z}}/r$, and the bottom arrow corresponds to the cohomology class $\alpha$. Therefore, we established a homomorphism $\textrm{Br}(X)\rightarrow H^{3}(X;\mathbb{Z})_{\textrm{tor}}$, and we call $H^{3}(X;\mathbb{Z})_{\textrm{tor}}$ the cohomological Brauer group of $X$, and sometimes denote it by $\textrm{Br}'(X)$. Serre [@Gr] showed that when $X$ is a finite CW complex, this homomorphism is an isomorphism. Hence, for any $\alpha\in H^{3}(X;\mathbb{Z})_{\textrm{tor}}$, there is some $r$ such that a $PU_r$-torsor over $X$ is associated to $\alpha$ via the homomorphism (\[Bockstein\]). Let $\operatorname{per}(\alpha)$ denote the order of $\alpha$ as an element of the group $H^{3}(X;\mathbb{Z})$. Serre [@Gr] also showed $\operatorname{per}(\alpha)\|r$ for all $r$ such that there is a $PU_r$-torsor over $X$ associated to $\alpha$ in the way described above. Let $\operatorname{ind}(\alpha)$ denote the greatest common divisor of all such $r$, then in particular we have $$\label{per div ind} \operatorname{per}(\alpha)\|\operatorname{ind}(\alpha).$$ The preceding definitions are motivated by their algebraic analogs. We refer to the introduction of [@An] for the algebraic version of the definitions as well as more algebraic backgrounds. The period-index conjecture, stated as follows, plays a key role in the study of Brauer groups: \[algconj\] Let $k$ be either a $C_d$-field or the function field of a $d$-dimensional variety over an algebraically closed field. Let $\alpha\in \textrm{Br}(k)$, and suppose that $\operatorname{per}(\alpha)$ is prime to the characteristic of $k$. Then $$\operatorname{ind}(\alpha)\|\operatorname{per}(\alpha)^{d-1}.$$ This conjecture has the following topological analog first considered by Antieau and Williams in [@An]: For a given Bruaer class $\alpha$ of a finite CW complexes $X$, find the sharp lower bound of $e$ such that $$\operatorname{ind}(\alpha)\|\operatorname{per}(\alpha)^e$$ holds for all finite CW complex $X$ in $\textrm{C}$ and all elements $\alpha\in\textrm{Br}(X)$. Notice that such an $e$ exists if and only if ${\operatorname}{per}{\alpha}$ and ${\operatorname}{ind}(\alpha)$ have the same prime divisors, which follows from Corollary 3.2, [@An1]. For further explanations of the preceding definitions and backgrounds on the topological period-index problem, see [@An], [@An1], [@Cr] and [@Gu]. All definitions and notations in this paper are consistent with those in [@Gu]. More precisely, the expression $\epsilon_p(n)$ denotes the greatest common divisor of a prime $p$ and a positive integer $n$, and we let $\beta_n$ be the canonical generator of $H^3(K(\mathbb{Z}/n,2);\mathbb{Z})$, i.e., the image of the identity class in $H^2(K(\mathbb{Z}/n,2);\mathbb{Z}/n)$ under the Bockstein homomorphism. In [@Gu] we have shown the following \[last\] Let $X$ be a topological space of homotopy type of an $8$-dimensional connected finite CW-complex, and let $\alpha\in H^{3}(X;\mathbb{Z})_{\operatorname{tor}}$ be a topological Brauer class of period $n$. Then $$\label{bound} \operatorname{ind}(\alpha)\|\epsilon_{2}(n)\epsilon_{3}(n)n^3.$$ In addition, if $X$ is the $8$-th skeleton of $K(\mathbb{Z}/n,2)$, and $\alpha$ is the restriction of the fundamental class $\beta_n\in H^{3}(K(\mathbb{Z}/n,2),\mathbb{Z})$, then $$\begin{cases} \operatorname{ind}(\alpha)=\epsilon_{2}(n)\epsilon_{3}(n)n^3,\quad\textrm{$4\nmid n$,}\\ \epsilon_{3}(n)n^3\|\operatorname{ind}(\alpha), \quad\textrm{$4\|n$.} \end{cases}$$ In particular, the sharp lower bound of $e$ such that $\operatorname{ind}(\alpha)\|n^{e}$ for all $X$ and $\alpha$ is $4$. The goal of this paper is to improve Theorem \[last\] and show the following \[main\] Let $X$ be a topological space of homotopy type of an $8$-dimensional connected finite CW-complex, and let $\alpha\in H^{3}(X;\mathbb{Z})_{\operatorname{tor}}$ be a topological Brauer class of period $n$. Then $$\label{bound1} \begin{cases} \operatorname{ind}(\alpha)\|2\epsilon_{3}(n)n^3,\textrm{ if }n\equiv 2\pmod{4},\\ \operatorname{ind}(\alpha)\|\epsilon_{3}(n)n^3,\textrm{ otherwise}. \end{cases}$$ In addition, if $X$ is the $8$-th skeleton of $K(\mathbb{Z}/n,2)$ and $\alpha$ is the restriction of the fundamental class $\beta_n\in H^{3}(K(\mathbb{Z}/n,2),\mathbb{Z})$, then $$\begin{cases} \operatorname{ind}(\alpha)=2\epsilon_{3}(n)n^3,\textrm{ if }n\equiv 2\pmod{4},\\ \operatorname{ind}(\alpha)=\epsilon_{3}(n)n^3,\textrm{ otherwise}. \end{cases}$$ In particular, the sharp lower bound of $e$ such that $\operatorname{ind}(\alpha)\|n^{e}$ for all $X$ and $\alpha$ is $4$. One readily sees that the only thing remains to show is the following \[mainprop\] Let $X$ be the $8$-th skeleton of $K(\mathbb{Z}/n,2)$ and $\alpha\in H^3(X;{\mathbb{Z}})$ the restriction of $\beta_n$. When $4\|n$, we have $$\operatorname{ind}(\alpha)=\epsilon_3(n)n^3.$$ Theorem \[main\] may be interpreted in terms of the twisted complex $K$-theory and the twisted Atiyah-Hirzebruch spectral sequences (AHSS). For a CW complex $X$ and a class $\alpha\in H^3(X;{\mathbb{Z}})$, the twisted complex $K$-theory associates the pair $(X,\alpha)$ to a graded ring $K^(X)_{\alpha}$ in a contravariant manner. For $\alpha=0$, it is the usual complex $K$-theory $K^(X)$. The readers may refer to [@At] and [@At1] for more backgrounds. As for the usual complex $K$-theory, there is a twisted AHSS of the pair $(X,\alpha)$ converging to the twisted K-theory $K^(X)_{\alpha}$. We denote it by $({\widetilde}{E}_^{,}, {\widetilde}{d}_^{,})$. The spectral sequence satisfies $${\widetilde}{E}_2^{s,t}\cong H^s(X;K^t({\operatorname}{pt}))$$ where $K^$ denotes the complex topological $K$-theory in the usual sense. By the Bott Periodicity Theorem, we have $$\label{AH} {\widetilde}{E}_2^{s,t}\cong \begin{cases} H^s(X;\mathbb{Z}), t \textrm{ even},\\ 0, t \textrm{ odd}. \end{cases}$$ The twisted AHSS is considered in [@An], [@An1], [@An2] and [@Gu]. The following result can be derived immediately from [@An1]. \[AH diff\] Let $X$ be a connected finite CW-complex and let $\alpha\in\textrm{Br}(X)$. Consider ${\widetilde}{E}_{}^{,}$, the twisted Atiyah-Hirzebruch spectral sequence of the pair $(X,\alpha)$ with differentials ${\widetilde}{d}_{r}^{s,t}$ with bi-degree $(r, -r+1)$. In particular, ${\widetilde}{E}_{2}^{0,0}\cong\mathbb{Z}$, and any ${\widetilde}{E}_{r}^{0,0}$ with $r>2$ is a subgroup of $\mathbb{Z}$ and therefore generated by a positive integer. The subgroup ${\widetilde}{E}_{3}^{0,0}$ (resp. ${\widetilde}{E}_{\infty}^{0,0}$) is generated by $\operatorname{per}(\alpha)$ (resp. $\operatorname{ind}(\alpha))$. Following Theorem \[AH diff\], it can be easily deduced from Theorem B of [@An] and Theorem \[last\] that when $X=K(\mathbb{Z}/n,2)$ and $\alpha=\beta_n$, the differentials ${\widetilde}{d}_r^{0,0}$ are surjective onto entries on the $E_2$-page for $r<7$, and when $4\nmid n$, also for $r=7$. Theorem \[main\], however, provides the first known example of a ${\widetilde}{d}_r^{0,0}$ that is NOT surjective onto some ${\widetilde}{E}_2^{s,t}$. Suppose $4\|n$. In the twisted Atiyah-Hirzebruch spectral sequence $({\widetilde}{E}_^{,}, {\widetilde}{d}_^{,})$ of the pair $(K(\mathbb{Z}/n,2),\beta_n)$, the image of ${\widetilde}{d}_7^{0,0}$ is $2{\widetilde}{E}_2^{7,-6}$. This is easily deduced from the preceeding paragraph and Theorem \[main\], once we recall the cohomology of $K(\mathbb{Z}/n,2)$, as in Section 2. See Figure \[twisted AH\]. (0,0) grid (7,-6); (0,-6)–(0,0)–(8,0); at (3,0) [$3$]{}; at (5,0) [$5$]{}; at (7,0) [$7$]{}; at (0,-2) [$-2$]{}; at (0,-4) [$-4$]{}; at (0,-6) [$-6$]{}; at (0,0) [$\mathbb{Z}$]{}; at (3,0) [$\mathbb{Z}/n$]{}; at (5,0) [$\mathbb{Z}/\epsilon_{2}(n)n$]{}; at (7,0) [$\mathbb{Z}/\epsilon_{3}(n)n$]{}; (0,0) \[thick, ->\] to (3,-2); (0,0) \[thick, ->\] to (5,-4); (0,0) \[thick, ->\] to (7,-6); The technical core of this paper is to solve a homotopy lifting problem of the same nature as the one displayed in (\[eq:lift\]), with some modifications for technical convenience. We recall some more notations from [@Gu]. Let $m,n$ be integers. Then $\mathbb{Z}/n$ is a closed normal subgroup of $SU_{mn}$ in the sense of the following monomorphism of Lie groups: $$\mathbb{Z}/n\hookrightarrow SU_{mn}: t\mapsto e^{2\pi\sqrt{-1}t/n}\mathbf{I}_{mn},$$ where $\mathbf{I}_r$ is the identity matrix of degree $r$. We denote the quotient group by $P(n,mn)$, and consider its classifying space $\mathbf{B}P(n,mn)$. It follows immediately from Bott’s periodicity theorem that we have $$\label{Bott} \pi_{i}(\mathbf{B}P(n,mn))\cong \begin{cases} \mathbb{Z}/n,\quad\textrm{if $i=2$},\\ \mathbb{Z},\quad\textrm{if $2<i<2mn+1$, and $i$ is even,}\\ 0, \quad\textrm{if $0<i<2mn$, and $i$ is odd.}\\ \end{cases}$$ In the context of this paper, the space $\mathbf{B}P(n,mn)$ may be considered as a refinement of $\mathbf{B}PU_{mn}$. For a finite CW-complex $X$ and a topological Brauer class $\alpha$ of period $n$, there is an element $\alpha'$ in $H^2(X;\mathbb{Z}/n)$, which is sent to $\alpha$ by the Bockstein homomorphism. Consider the following lifting diagram $$\label{lift'} \begin{tikzcd} &\mathbf{B}P(n,mn)\arrow{d}\\ X\arrow[ur,dashed]\arrow{r}{\alpha'}&K(\mathbb{Z}/n,2) \end{tikzcd}$$ where the vertical arrow is the projection from $\mathbf{B}P(n,mn)$ to the bottom stage of its Postnikov tower. We have the following \[lift to BP\]\[Proposition 4.3, [@Gu]\] Let $X$, $\alpha$ be as above, and suppose that $H^2(X;\mathbb{Z})=0$. Then $\alpha$ is classified by an Azumaya algebra of degree $mn$ if and only if the lifting in diagram (\[lift’\]) exists. In Section 2 we recall some facts on Eilenberg-Mac Lane spaces. Section 3 is a collection of lemmas on the cohomology of $\mathbf{B}P(n,mn)$. In Section 4 we consider a k-invariant in the Postnikov decomposition of $\mathbf{B}P(n,mn)$, for $4\|n$ and suitable $m$, leading to the proof of Proposition \[mainprop\] and Theorem \[main\]. recollection of facts on eilenberg - mac lane spaces ==================================================== All the facts recalled here are either well-known or easily deduced from [@Ca]. The integral cohomology of $K(\mathbb{Z}/n,2)$ in degree $\leq 8$ is isomorphic to the following graded commutative ring: $$\label{K(Z/2,2)} \mathbb{Z}[\beta_n,Q_n,R_n,\rho_n]/(n\beta_n,\epsilon_{2}(n)\beta_n^2, \epsilon_{2}(n)nQ_n,\epsilon_{3}(n)nR_n,\epsilon_3(n)\rho_n),$$ where $\operatorname{deg}(\beta_n)=3, \operatorname{deg}(Q_n)=5, \operatorname{deg}(R_n)=7$, and $\operatorname{deg}(\rho_n)=8$. In other words, there is exactly one generator in each of the degrees $3,5,6,7$, which are, respectively, $\beta_n, Q_n, \beta_n^2, R_n$, of order $$n,\epsilon_{2}(n)n,\epsilon_{2}(n),\epsilon_{3}(n)n,$$ and $2$ generators in degree $8$, $\beta_nQ_n$ and $\rho_n$, of order $\epsilon_2(n)$ and $\epsilon_3(n)$, respectively. (See (2.5) of [@Gu].) For $n\geq3$, the ring $H^{}(K(\mathbb{Z},n);\mathbb{Z})$ in degree $\leq n+3$ is isomorphic to the following graded rings: $$\label{K(Z,n)} \begin{cases} \mathbb{Z}[\iota_n,\Gamma_n]/(2\Gamma_n),\ n>3,\textrm{ even},\\ \mathbb{Z}[\iota_n,\Gamma_n]/(2\iota_n,2\Gamma_n),\ n>3,\textrm{ odd},\\ \mathbb{Z}[\iota_3,\Gamma_3]/(2\Gamma_3, \Gamma_3-\iota_3^2),\ n=3, \end{cases}$$ where $\iota_n$, of degree $n$, is the so-called fundamental class, and $\Gamma_n$, of degree $n+3$, is a class of order $2$.(See (2.1) of [@Gu].) \[stable Gamma\] The classes $\Gamma_n\in H^{n+3}(K(\mathbb{Z},n);\mathbb{Z})$ as above for all $n\geq 3$, stabilize to the same stable cohomology operation $\operatorname{Sq}_{\mathbb{Z}}^3\in H^3(K(\mathbb{Z});\mathbb{Z})$ of order $2$, where $K(R)$ denotes the Eilenberg - Mac Lane spectrum associated to a unit ring $R$. Moreover, the mod $2$ reduction of $\operatorname{Sq}_{\mathbb{Z}}^3$ is the well-understood Steenrod square $\operatorname{Sq}^3$. In other words, the following diagram in the homotopy category of spectra commutes: $$\label{spectra} \begin{tikzcd} K(\mathbb{Z})\arrow{d}\arrow{r}{\operatorname{Sq}_{\mathbb{Z}}^3}&\Sigma^3K(\mathbb{Z})\arrow{d}\\ K(\mathbb{Z}/2)\arrow{r}{\operatorname{Sq}^3}&\Sigma^3K(\mathbb{Z}/2) \end{tikzcd}$$ where the vertical arrows are the obvious ones. For a proof see Lemma 2.1 of [@Gu]. the group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ ============================================= As in the introduction we denote by $P(n,mn)$ the quotient group of the following inclusion of Lie groups: $$\mathbb{Z}/n\hookrightarrow SU_{mn}: t\mapsto e^{2\pi\sqrt{-1}t/n}\mathbf{I}_{mn},$$ where $\mathbf{I}_r$ is the identity matrix of degree $r$, as well as its classifying space $\mathbf{B}P(n,mn)$. In this section we study the cohomology group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$. Let us remind ourselves of the following notation: for a simply connected space $X$, let $X[k]$ denote the $k$th level of the Postnikov tower of $X$. Consider the Postnikov tower of $\mathbf{B}P(n,mn)$ for $\epsilon_2(n)n\|m$, $n>1$: $$\label{Postnikov} \begin{tikzcd} &\mathbf{B}P(n,mn)[6]\arrow{d}{=}\\ K(\mathbb{Z},4)\arrow{r}&\mathbf{B}P(n,mn)[5]\simeq K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4)\arrow{d}\arrow{r}{\kappa_5}&K(\mathbb{Z},7)\\ &K(\mathbb{Z}/n,2)\arrow{r}{\kappa_3=0}&K(\mathbb{Z},5) \end{tikzcd}$$ where $\kappa_3$ and $\kappa_5$ are the k-invariant of the space $\mathbf{B}P(n,m)$. The equation $\kappa_3=0$ follows from Proposition 4.11 of [@Gu]. Consequently, $\mathbf{B}P(n,mn)[5]$ is a trivial fibration over $K({\mathbb{Z}}/n,2)$ with fiber $K({\mathbb{Z}},4)$. Consider the projection $$\mathbf{B}P(n,mn)\rightarrow\mathbf{B}P(n,mn)[6]\simeq K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4),$$ where $\epsilon_2(n)n\|m$. For future reference we take notes of the induced homomorphism between integral cohomology rings as follows: $$\label{bookkeeping} \beta_n\times 1\mapsto x_1',\quad R_n\times 1\mapsto R_n(x_1'),\quad 1\times\iota_4\mapsto e_2',\quad 1\times\Gamma_4\mapsto\operatorname{Sq}_{\mathbb{Z}}^3(e_2'),$$ where $x_1'$ and $e_2'$ are the additive generators of $H^3(\mathbf{B}P(n,mn);\mathbb{Z})\cong\mathbb{Z}/n$ and $H^4(\mathbf{B}P(n,mn);\mathbb{Z})\cong\mathbb{Z}$, respectively. Here $R_n$ is the generator of $H^7(K({\mathbb{Z}}/n,2);{\mathbb{Z}})$ as in (\[K(Z/2,2)\]), regarded as a cohomology operation in the obvious way. By the construction of Postnikov towers we have \[H7\] Suppose $\epsilon_2(n)n\|m$. Then we have $$H^7(\mathbf{B}P(n,mn);\mathbb{Z})\cong H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4))/(\kappa_5).$$ In particular, the group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is generated by $R_n(x_1')$, $x_1'e_2'$, and $\operatorname{Sq}_{\mathbb{Z}}^3(e_2')$. Consider the short exact sequence of Lie groups $$1\rightarrow\mathbb{Z}/n\rightarrow SU_{mn}\rightarrow P(n,mn)\rightarrow 1,$$ from which arises a fiber sequence $$\mathbf{B}SU_{mn}\rightarrow\mathbf{B}P(n,mn)\rightarrow K(\mathbb{Z}/n,2)$$ considered at the beginning of Section 6 of [@Gu]. We denote the associated cohomological Serre spectral sequence by $(^SE_{}^{,},{^Sd}_{}^{,})$. For $k>2$, it is well-known that $$H^(\mathbf{B}SU_k;{\mathbb{Z}})=\mathbb{Z}[c_2,\cdots,c_k],$$ where $c_i$ is the $i$th universal Chern class of degree $2i$. \[Sd\_5\] Suppose $\epsilon_{2}(n)n\|m$. The differential $^Sd_5^{0,4}=0$. In particular, $c_2\in{^SE}_{2}^{0,4}$ is a permanent cocycle. Diagram (\[Postnikov\]) implies $$H^5(\mathbf{B}P(n,mn);\mathbb{Z})\cong H^5(K(\mathbb{Z}/n,2);\mathbb{Z})\cong\mathbb{Z}/\epsilon_2(n)n.$$ Hence we have $^Sd_5^{0,4}=0$. There is no other non-trivial differential out of $^SE_{}^{0,4}$ for obvious degree reasons, so $c_2$ is a permanent cocycle. \[SE\] Suppose $\epsilon_{2}(n)n\|m$. Recall that $H^{3}(K(\mathbb{Z}/n,2);\mathbb{Z}))\cong\mathbb{Z}/n$ is generated by an element $\beta_n$, and that $H^{7}(K(\mathbb{Z}/n,2);\mathbb{Z}))\cong\mathbb{Z}/\epsilon_{3}(n)n$ is generated by $R_n$. In the spectral sequence ${^{S}E}_{}^{,}$, we have ${^{S}d}_{3}^{0,6}(c_3)=2c_{2}\beta_{n}$ with kernel generated by $$\frac{n}{\epsilon_{2}(n)}c_3,$$ and $${^{S}d}_{7}^{0,6}(\frac{n}{\epsilon_{2}(n)}c_3)= \frac{\epsilon_{3}(n)\epsilon_3(m/n)}{\epsilon_3(mn)}nR_n.$$ All the other differentials out of ${^{S}E}_{}^{0,6}$ are trivial. See Figure \[\^S E figure\] for the differentials mentioned in the lemmas above. ; (0,6)–(0,0)–(7,0); at (3,0) [$3$]{}; at (5,0) [$5$]{}; at (7,0) [$7$]{}; at (0,4) [$4$]{}; at (0,6) [$6$]{}; at (0,4) [$\mathbb{Z}$]{}; at (0,6) [$\mathbb{Z}$]{}; at (3,0) [$\mathbb{Z}/n$]{}; at (5,0) [$\mathbb{Z}/\epsilon_{2}(n)n$]{}; at (7,0) [$\mathbb{Z}/\epsilon_{3}(n)n$]{}; (0,4) \[dashed, ->\] to (5,0); (0,6) \[thick, ->\] to node \[above\] [$\times 2$]{}(3,4); (0,6) \[dashed, ->\] to (7,0); \[cardH\^7BP\] Suppose $\epsilon_{2}(n)n\|m$. 1. The element $\operatorname{Sq}^3_{\mathbb{Z}}(e_2')$ is a linear combination of $x_1'e_2'$ and $R_n(x_1')$. 2. The element $R_n(x_1')$ is of order $$\frac{\epsilon_{3}(n)\epsilon_3(m/n)n}{\epsilon_3(mn)}.$$ 3. The order of the group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is $$\frac{\epsilon_2(n)\epsilon_{3}(n)\epsilon_3(m/n)n}{\epsilon_3(mn)}.$$ As indicated in Figure \[\^S E figure\], we have the exact sequence $$0\rightarrow {^SE}_{\infty}^{7,0}\rightarrow H^7(\mathbf{B}P(n,mn);\mathbb{Z}) \rightarrow {^SE}_{\infty}^{3,4}\rightarrow 0,$$ where ${^SE}_{\infty}^{7,0}$ and ${^SE}_{\infty}^{3,4}$ are generated by $R_n(x_1')$ and $x_1'e_2'$, respectively. Hence (1) follows. The statement (2) is an immediate a consequence of Lemma \[SE\]. To verify the statement (3), it suffices to check the order of $^SE_{\infty}^{3,4}$ and $^SE_{\infty}^{7,0}$, which also follows from Lemma \[SE\]. the k-invariant $\kappa_5$ ========================== Consider the space $\mathbf{B}P(n,mn)[6]$, the $6$th level of the Postnikov tower of $\mathbf{B}P(n,mn)$. We assume throughout this section that $\epsilon_{2}(n)n\|m$, $n>1$, for when this holds, we have the homotopy equivalence (Proposition 4.11, [@Gu]) $$\label{eq:directprod} \mathbf{B}P(n,mn)[6]=\mathbf{B}P(n,mn)[5]\simeq K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4).$$ The goal of this section is to prove Proposition \[mainprop\], therefore Theorem \[main\], by determining the k-invariant $$\kappa_5\in H^7(\mathbf{B}P(n,mn)[5];{\mathbb{Z}})\simeq H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);{\mathbb{Z}})$$ for $\epsilon_{2}(n)n\|m$ and $4\|n$. The equations (\[K(Z/2,2)\]) and (\[K(Z,n)\]) together with the K[ü]{}nneth formula give us $$\label{H^7(KK)} \begin{split} &H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})\\=&(R_n\times 1)\oplus (\beta_n\times\iota_4)\oplus (1\times\Gamma_4)\\ \cong&\mathbb{Z}/\epsilon_3(n)n\oplus\mathbb{Z}/n\oplus\mathbb{Z}/2, \end{split}$$ where $R_n\times 1$, $\beta_n\times\iota_4$ and $1\times\Gamma_4$ generate the three summands, respectively. When $n$ is even, it follows from Theorem 6.8 of [@Gu] that up to an invertible coefficient, we have $$\label{k_51} \kappa_5\equiv\frac{\epsilon_3(n)m}{\epsilon_2(m)\epsilon_3(m)n}\lambda R_n\times 1+ \lambda_2\beta_n\times\iota_4+1\times\Gamma_4\mod 2-\operatorname{torsions},$$ an element in $H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})$, where $\lambda$ is invertible in $\mathbb{Z}/\epsilon_3(n)n$. To narrow down the choices of $\kappa_5$, we have the following \[Gamma\] Suppose $\epsilon_2(n)n\|m$, $n>1$. In $H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})$ we have $$\kappa_5\equiv 1\times\Gamma_4\mod (R_n\times1, \beta_n\times\iota_4).$$ Assume that the lemma is false. Since $1\times\Gamma_4$ is of order $2$, we have $$\kappa_5\in (R_n\times1, \beta_n\times\iota_4).$$ Therefore, we have $$\begin{split} &H^7(\mathbf{B}P(n,mn);\mathbb{Z})\\ \cong&H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})/(\kappa_5)\\ \cong& (R_n\times 1)\oplus (\beta_n\times\iota_4)\oplus (1\times\Gamma_4)/(\kappa_5)\\ \cong& [(R_n\times 1)\oplus (\beta_n\times\iota_4)]/(\kappa_5)\oplus (1\times\Gamma_4), \end{split}$$ but this violates (1) of Lemma \[cardH\^7BP\], which asserts that $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is generated by $R_n(x_1')$ and ${\operatorname}{Sq}^3_{{\mathbb{Z}}}(e_2')$. To prove Proposition \[mainprop\], we only consider the case $4\|n$. More precisely, it suffices to consider the case that $X$ is the $8$th skeleton of $K({\mathbb{Z}}/n,2)$ for $n$ divisible by $4$, and determine ${\operatorname}{ind}(\alpha)$ for $\alpha$ the restriction of $\beta_n\in H^3(K({\mathbb{Z}}/n,2);{\mathbb{Z}})$. As a consequence of Theorem \[last\], the above ${\operatorname}{ind}(\alpha)$ is either $2\epsilon_3(n)n^3$ or $\epsilon_3(n)n^3$. We write $\alpha=\sum_{p}\alpha_p$ where $p$ runs over all prime divisors of ${\operatorname}{per}(\alpha)$ and the period of $\alpha_p$ is a power of $p$. Then Theorem 3 of [@An2] asserts ${\operatorname}{ind}(\alpha)=\prod_p{\operatorname}{ind}(\alpha_p)$, with each ${\operatorname}{ind}(\alpha_p)$ a power of $p$. Hence it suffices to determine ${\operatorname}{ind}(\alpha_2)$. In other words, we have reduced the problem to the case $n=2^r$, $m=2^{2r}$ or $2^{2r+1}$, with $r>1$. We assume this for the rest of this section: \[convention\] $$n=2^r,\ m=2^{2r}\textrm{ or }2^{2r+1},\ r>1.$$ \[lambda\_2not2\] In $H^7(K({\mathbb{Z}}/n,2)\times K({\mathbb{Z}},4);{\mathbb{Z}})$, We have $$\kappa_5\equiv\lambda_2\beta_n\times\iota_4\mod (R_n\times1, 1\times\Gamma_4),$$ where the coefficient $\lambda_2$ is invertible in $\mathbb{Z}/n$. We argue by contradiction. Suppose that $\lambda_2$ is not invertible in $\mathbb{Z}/n$. Notice that, for our choice of $m$ and $n$, equation (\[k\_51\]) implies $$\kappa_5\equiv 2^{r-1}\lambda R_n\times 1+ \lambda_2\beta_n\times\iota_4+1\times\Gamma_4\mod 2-\operatorname{torsions}.$$ Since $\lambda_2$ is not invertible, $\kappa_5$ has order less than $2^r$. On the other hand, it follows from (\[H\^7(K\K)\]) that the group $H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})$ has order $2^{2r+1}$. Therefore the group $$H^7(\mathbf{B}P(n,mn);\mathbb{Z})\cong H^7(K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4);\mathbb{Z})/(\kappa_5)$$ has order greater than $2^{r+1}$, contradicting Lemma \[cardH\^7BP\]. Notice that $\kappa_5$ is determined by the Postnikov tower merely up to multiplication by an invertible coefficient. By the choice we have made of $n,m$, this means that we are free to multiply $\kappa_5$ by any odd integer. Hence, we are enabled by Lemma \[lambda\_2not2\] to normalize (\[k\_51\]) by fixing $\lambda_2=1$: $$\kappa_5=2^{r-1}\lambda R_n\times 1+\beta_n\times\iota_4+1\times\Gamma_4\mod\textrm{a $2$-torsion in $(R_n\times1, 1\times\Gamma_4)$},$$ where $\lambda$ is odd. However, since $R_n$ is of order $2^r$, and $2^{r-1}\equiv 2^{r-1}\lambda\mod 2^{r}$ for all odd integer $\lambda$, the preceding equation becomes $$\label{lambdainv} \kappa_5=2^{r-1}R_n\times1+\beta_n\times\iota_4+1\times\Gamma_4\mod\textrm{a $2$-torsion in $(R_n\times1, 1\times\Gamma_4)$}.$$ Combining (\[lambdainv\]) with Lemma \[Gamma\], we have $$\label{k_5} \kappa_5\equiv2^{r-1}R_n\times 1+\beta_n\times\iota_4+1\times\Gamma_4\mod \textrm{a $2$-torsion in $(R_n\times1)$}.$$ Moreover, we have the following \[statement\] The abelian group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is additively generated by $R_n(x_1')$, $x_1'e_2'$ and $\operatorname{Sq}_{\mathbb{Z}}^3(e_2')$, modulo the relation $$\label{k_5'} \mu R_n(x_1')+x_1'e_2'+\operatorname{Sq}_{\mathbb{Z}}^3(e_2')=0,$$ where $\mu$ is either $0$ or $2^{r-1}$. Moreover, only one of the two possible relations holds. Only the uniqueness requires a proof. Indeed, if both relations hold, then we have $$2^{r-1}R_n(x_1')=0\textrm{ and }x_1'e_2'+\operatorname{Sq}_{\mathbb{Z}}^3(e_2')=0.$$ Hence the abelian group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is generated by $R_n(x_1')$ and $\operatorname{Sq}_{\mathbb{Z}}^3(e_2')$, whose orders divide $2^{r-1}$ and $2$, respectively. Therefore the order of the group $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ divides $2^r$, violating Lemma \[cardH\^7BP\]. We turn to the hard work of determining $\mu$, for which we will need an auxiliary space $Y$ to be defined and studied as follows. Recall the homotopy equivalence (\[eq:directprod\]) $$\mathbf{B}P(n,mn)[6]=\mathbf{B}P(n,mn)[5]\simeq K(\mathbb{Z}/n,2)\times K(\mathbb{Z},4).$$ Consider the following map $$\mathbf{B}P(n,mn)\rightarrow\mathbf{B}P(n,mn)[6]\rightarrow K(\mathbb{Z},4)$$ such that both arrows above are the obvious projections. We may normalize this map so that it represents the generator $e_2'$ of $H^4(\mathbf{B}P(n,mn);{\mathbb{Z}})$. We denote by $Y$ its homotopy fiber. In other words, we have a fiber sequence $$\label{fiber} Y\rightarrow\mathbf{B}P(n,mn)\xrightarrow{e_2'} K(\mathbb{Z},4)$$ where the second arrow is an additive generator $e_2'$ of $H^4(\mathbf{B}P(n,mn);{\mathbb{Z}})$. The space $Y$ plays a key role in determining the coefficient $\mu$. By construction the second arrow in (\[fiber\]) induces an isomorphism of homotopy groups $$\pi_4(\mathbf{B}P(n,mn))\cong \pi_4(K(\mathbb{Z},4)).$$ This isomorphism lies in the long exact sequence of homotopy groups of the fiber sequence (\[fiber\]), from which, together with (\[Bott\]), we deduce $$\label{piY} \pi_i(Y)\cong \begin{cases} \mathbb{Z}/n, \textrm{ if }n=2,\\ \mathbb{Z},\quad\textrm{if $6\leq i<2mn+1$, and $i$ is even},\\ 0, \quad\textrm{if $0<i<2mn$, and $i$ is odd, or }i=4.\\ \end{cases}$$ By de-looping the last term of (\[fiber\]) we obtain another fiber sequence $$\label{mainfiber} K(\mathbb{Z},3)\xrightarrow{h}Y\rightarrow\mathbf{B}P(n,mn).$$ Consider the projection from $Y$ to the bottom level of its Postnikov tower $$g: Y\rightarrow K(\mathbb{Z}/n,2).$$ \[H\^6Y\] 1. The induced homomorphisms $$g^: H^k(K(\mathbb{Z}/n,2);\mathbb{Z})\rightarrow H^k(Y;\mathbb{Z})$$ are isomorphisms for $k\leq 5$. 2. The homomorphism $H^6(g;\mathbb{Z})$ is injective. Furthermore, we have $$H^6(Y;\mathbb{Z})=g^(H^6(K(\mathbb{Z}/n,2);\mathbb{Z}))\oplus (\omega)\cong H^6(K(\mathbb{Z}/n,2);\mathbb{Z})\oplus\mathbb{Z},$$ where $\omega$ generates the summand $\mathbb{Z}$. 3. The induced homomorphism $$g^:H^7(K(\mathbb{Z}/n,2);\mathbb{Z})\rightarrow H^7(Y;\mathbb{Z})$$ is surjective. 4. $$H^6(Y;\mathbb{Z}/2)=g^(H^6(K(\mathbb{Z}/n,2);\mathbb{Z}/2))+ (\bar{\omega}),$$ where an integral cohomology class with an overhead bar denotes its reduction in cohomology with coefficients in $\mathbb{Z}/2$. It follows from (\[piY\]) that we have the following partial picture of its Postnikov tower $$\begin{tikzcd} &Y[7]\arrow{d}{=}\arrow{r}&K(\mathbb{Z},9)\\ K(\mathbb{Z},6)\arrow{r}&Y[6]\arrow{d} &\\ &Y[2]=K(\mathbb{Z}/n,2)\arrow{r}&K(\mathbb{Z},7) \end{tikzcd}$$ The statements (1), (2) and (3) follows from a simple observation on the fiber sequence $$K(\mathbb{Z},6)\rightarrow Y[6]\rightarrow K(\mathbb{Z}/n,2)$$ and the Serre spectral sequence associated to it. Finally, (4) follows from (2), (3), and K[ü]{}nneth formula. Consider the fiber sequence (\[mainfiber\]). Let $(E_{}^{,}(\mathbb{Z}), d_{}^{,})$ and $(E_{}^{,}(\mathbb{Z}/2), \bar{d}_{}^{,})$ denote the associated cohomological Serre spectral sequences with coefficients in $\mathbb{Z}$ and $\mathbb{Z}/2$, respectively. We first consider the case with coefficients in $\mathbb{Z}/2$. This is easier since $\mu\equiv 0$ mod $2$ whatsoever. We study the homomorphism $$h^:H^6(Y;\mathbb{Z}/2)\rightarrow H^6(K(\mathbb{Z},3);\mathbb{Z}/2)$$ with the spectral sequence $E_{}^{,}(\mathbb{Z}/2)$, from which, with a little luck, we obtain enough information on the homomorphism $$h^:H^6(Y;\mathbb{Z})\rightarrow H^6(K(\mathbb{Z},3);\mathbb{Z})$$ to determine a particular differential of the spectral sequence $E_{}^{,}(\mathbb{Z})$, which in turn determines the coefficient $\mu$. As in Lemma \[H\^6Y\], we use overhead bars to denote the mod $2$ reductions of integral cohomology classes. \[Z/2\] The homomorphism $$h^:H^6(Y;\mathbb{Z}/2)\rightarrow H^6(K(\mathbb{Z},3);\mathbb{Z}/2)$$ is surjective. More precisely, we have $$\label{rhoZ/2} h^(\bar{\omega})=\bar{\iota}_3^2.$$ Consider the spectral sequence $(E_{}^{,}(\mathbb{Z}), d_{}^{,})$. Refer to Figure \[Z/2 figure\] for the relevant differentials. It follows from (1) of Lemma \[H\^6Y\] that $d_4^{0,3}$ is the first nontrivial differential out of the bidegree $(0,3)$ and it is an isomorphism: $$\label{d_4Z} d_4^{0,3}: E_{4}^{0,3}(\mathbb{Z})\cong\mathbb{Z}\xrightarrow{\cong}\mathbb{Z}\cong E_{4}^{4,0}(\mathbb{Z})$$ sending the generator $\iota_3$ of $H^3(K(\mathbb{Z},3);\mathbb{Z})$ to $\pm e_2'$, the generator of $H^4(\mathbf{B}P(n,mn);\mathbb{Z})$. Passing to $(E_{}^{,}(\mathbb{Z}/2), \bar{d}_{}^{,})$ as shown in Figure \[Z/2 figure\], it follows from (\[d\_4Z\]) that $$\bar{\iota}_3\in H^3(K(\mathbb{Z},3);\mathbb{Z}/2)\cong E_{2}^{0,3}(\mathbb{Z}/2)$$ is transgressive (cf. Section 6.2 of [@Mc]) such that $$\label{d_4Z/2} \bar{d}_{4}^{0,3}(\bar{\iota}_3)=\bar{e}_2'\in H^4(\mathbf{B}P(n,mn);\mathbb{Z}/2)\cong E_{2}^{4,0}(\mathbb{Z}/2)\cong E_{4}^{4,0}(\mathbb{Z}/2).$$ Therefore, $\bar{\iota}_3^2=\operatorname{Sq}^3(\bar{\iota}_3)$ is also transgressive and we have $$\label{d_7Z/2} \bar{d}_{7}^{0,6}(\bar{\iota}_3^2)=\bar{d}_{7}^{0,6}(\operatorname{Sq}^3(\bar{\iota}_3)) =\operatorname{Sq}^3(\bar{d}_{4}^{0,3}(\bar{\iota}_3))=\operatorname{Sq}^3(\bar{e}_2').$$ It follows from (\[d\_4Z/2\]) that we have $$\label{d_4^3,3Z/2} \bar{d}_{4}^{3,3}(\bar{x}_1'\bar{\iota}_3)=\bar{x}_1'\bar{e}_2'\in H^7(\mathbf{B}P(n,mn);\mathbb{Z}/2)\cong E_{2}^{7,0}(\mathbb{Z}/2)\cong E_{4}^{7,0}(\mathbb{Z}/2).$$ Taking the mod $2$ reduction of $\kappa_5$ as in (\[k\_5\]), we have $\mu\equiv 0$ mod $2$ since $r>1$. Therefore it follows from Lemma \[statement\] that we have $$\label{relationZ/2} \bar{x}_1'\bar{e}_2'+\operatorname{Sq}^3(\bar{e}_2')=0\in H^7(\mathbf{B}P(n,mn);\mathbb{Z}/2).$$ This relation, together with (\[d\_4\^3,3Z/2\]), shows $$\operatorname{Sq}^3(\bar{e}'_2)\equiv0\in E_{5}^{7,0}(\mathbb{Z}/2)\cong H^7(\mathbf{B}P(n,mn);\mathbb{Z}/2)/(\bar{x}_1'\bar{e}_2').$$ Then it follows from (\[d\_7Z/2\]) that $\bar{d}_{7}^{0,6}(\bar{\iota}_3^2)=0$. Hence, $\bar{\iota}_3^2$ is a permanent cocycle, which proves that $\bar{\iota}_3^2$ is in the image of $h^$, and in particular, that $h^$ is surjective in dimension $6$ and with coefficients in $\mathbb{Z}/2$. To verify (\[rhoZ/2\]), notice that $g$ factors as $$Y\rightarrow\mathbf{B}P(n,mn)\rightarrow K(\mathbb{Z}/n,2)$$ since the map $Y\rightarrow\mathbf{B}P(n,mn)$ induces an isomorphism $\pi_2(Y)\cong\pi_2(\mathbf{B}P(n,mn))$. In particular, it follows that $$g\circ h: K(\mathbb{Z},3)\rightarrow Y\rightarrow K(\mathbb{Z}/n,2)$$ is null homotopic. On the other hand, recall that (4) of Lemma \[H\^6Y\] asserts $$H^6(Y;\mathbb{Z}/2)=g^(H^6(K(\mathbb{Z}/n,2);\mathbb{Z}/2))+ (\bar{\omega}),$$ so $h^$ takes the first direct summand to $0$. It then follows from the surjectivity of $h^$ that we have $h^(\bar{\omega})=\bar{\iota}_3^2$, and we conclude. ; (0,6)–(0,0)–(7,0); at (4,0) [$4$]{}; at (7,0) [$7$]{}; at (3,3) [$(3,3)$]{}; at (0,3) [$3$]{}; at (0,6) [$6$]{}; at (4,0) [$\bar{e_2'}$]{}; at (0,3) [$\bar{\iota}_3$]{}; at (3,3) [$\bar{\iota}_3\bar{x_1'}$]{}; at (7,0) [$\bar{e_2'}\bar{x_1'}, \operatorname{Sq}^3(\bar{e_2'})$]{}; at (0,6) [$\bar{\iota}_3^2$]{}; (0,3) \[thick, ->\] to (4,0); (3,3) \[thick, ->\] to (7,0); (0,6) \[dashed, ->\] to (7,0); Passing to the integral cohomology, we have the following \[Z\] The homomorphism $$h^:H^6(Y;\mathbb{Z})\rightarrow H^6(K(\mathbb{Z},3);\mathbb{Z})$$ is surjective. Furthermore, $h^(\omega)=\iota_3^2$. Since $H^6(K(\mathbb{Z},3);\mathbb{Z})\cong\mathbb{Z}/2$ has only two elements, $0$ and $\iota_3^2$, we have either $h^(\omega)=\iota_3^2$ or $h^(\omega)=0.$ Lemma \[Z/2\] shows that the latter is impossible. We proceed to determine $\kappa_5$. \[determinek\_5\] Under Convention \[convention\], we have $$\kappa_5=\beta_n\times\iota_4+1\times\Gamma_4.$$ Consider the spectral sequence $(E_{}^{,}(\mathbb{Z}), d_{}^{,})$. (The picture of the differentials looks the same as Figure \[Z/2 figure\], but with all the overhead bars removed, and $\operatorname{Sq}^3$ replaced by $\operatorname{Sq}_{\mathbb{Z}}^3$.) For obvious degree reasons the only possibly nontrivial differentials landing in bidegree $(7,0)$ are $d_{7}^{0,6}$ and $d_{4}^{3,3}$. It follows from (\[d\_4Z\]) and the Leibniz rule that we have $$d_4^{3,3}(x_1'\iota_3)=x_1'e_2',$$ and $$\label{E_6Z} E_{5}^{7,0}(\mathbb{Z})=E_{7}^{7,0}(\mathbb{Z})\cong H^7(\mathbf{B}P(n,mn);\mathbb{Z})/(x_1'e_2')$$ Recall (\[d\_4Z\]): $$d_4^{0,3}: E_{4}^{0,3}(\mathbb{Z})\cong\mathbb{Z}\xrightarrow{\cong}\mathbb{Z}\cong E_{4}^{4,0}(\mathbb{Z}),\ \iota_3\mapsto\pm e_2',$$ which, in particular, asserts that $\iota_3$ is transgressive. Therefore, so is $\operatorname{Sq}_{\mathbb{Z}}^3(\iota_3)=\iota_3^2$. (See Proposition \[stable Gamma\].) Furthermore, since transgressions commute with stable cohomology operations, we have $$\label{iota_3^2} d_7^{0,6}(\iota_3^2)=d_7^{0,6}(\operatorname{Sq}_{\mathbb{Z}}^3(\iota_3))= \operatorname{Sq}_{\mathbb{Z}}^3(d_4^{0,3}(\iota_3))=\operatorname{Sq}_{\mathbb{Z}}^3(e_2'),$$ where the last step follows from (\[d\_4Z\]). On the other hand, it follows from Lemma \[Z\] that $\iota_3^2$ is a permanent cocycle. Therefore, (\[E\_6Z\]) and (\[iota\_3\^2\]) implies that $\operatorname{Sq}_{\mathbb{Z}}^3(e_2')$ is in the subgroup of $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ generated by $x_1'e_2'$, i.e., $$\label{nu} \nu x_1'e_2'+\operatorname{Sq}_{\mathbb{Z}}^3(e_2')=0,$$ for some integer $\nu$. It follows from Lemma \[statement\] that $2x_1'e_2'=0$. Therefore we only need to chose from $\nu=0$ and $\nu=1$. If $\nu=0$, then $\operatorname{Sq}_{\mathbb{Z}}^3(e_2')=0$. Applying Lemma \[statement\] again, we have $$x_1'e_2'=\mu R_n(x_1'),$$ which implies that $H^7(\mathbf{B}P(n,mn);\mathbb{Z})$ is generated by $R_n(x_1')$, contradicting (2) and (3) of Lemma \[cardH\^7BP\]. Therefore, we have $\nu=1$ and $$x_1'e_2'+\operatorname{Sq}_{\mathbb{Z}}^3(e_2')=0\in H^7(\mathbf{B}P(n,mn);\mathbb{Z}).$$ By Lemma \[statement\], we conclude. Let $X={\operatorname}{sk}_8(K({\mathbb{Z}}/n,2))$ be the $8$th skeleton of $K({\mathbb{Z}}/n,2)$, and $\alpha\in H^3(X;{\mathbb{Z}})$ the restriction of the canonical generator $\beta_n\in H^3(K({\mathbb{Z}}/n,2);{\mathbb{Z}})$. Recall that it suffices to work under Convention \[convention\]: $$n=2^r,\ m=2^{2r}\textrm{ or }2^{2r+1},\ r>1.$$ It follows from Theorem \[last\] that ${\operatorname}{ind}(\alpha)$ is either $2^{3r}$ or $2^{3r+1}$. It suffices to show ${\operatorname}{ind}(\alpha)=2^{2r}$, for which we proceed to study the homotopy lifting problem discussed in the Introduction (\[lift’\]): $$\label{lifting} \begin{tikzcd} &\mathbf{B}P(2^r,2^{3r})\arrow{d}\\ X\arrow[r,hook]\arrow[ur,dashrightarrow]&K(\mathbb{Z}/2^r,2) \end{tikzcd}$$ where the bottom arrow is the obvious inclusion. By construction, this diagram is equivalent to $$\label{lifting1} \begin{tikzcd} &\mathbf{B}P(2^r,2^{3r})[7]\arrow{d}\\ K(\mathbb{Z}/2^r,2)\arrow{r}{=}\arrow[ur,dashrightarrow]&K(\mathbb{Z}/2^r,2) \end{tikzcd}$$ By (\[eq:directprod\]), we have the following $$\label{lifting2} \begin{tikzcd} &\mathbf{B}P(2^r,2^{3r})[7]\arrow{d}{=}&\\ &\mathbf{B}P(2^r,2^{3r})[6]\arrow{d}&\\ &\mathbf{B}P(2^r,2^{3r})[5]\simeq K(\mathbb{Z}/2^r,2)\times K(\mathbb{Z},4)\arrow{d}\arrow{r}{\kappa_5}&K(\mathbb{Z},7)\\ K(\mathbb{Z}/n,2^r)\arrow{r}{f_3=\operatorname{Id}}\arrow{ur}{f_5}\arrow[uur,bend left,dashrightarrow]&K(\mathbb{Z}/2^r,2)& \end{tikzcd}$$ where the map $f_5$ is the obvious inclusion. Therefore $f_5^$ annihilates all cohomology classes of $K(\mathbb{Z},4)$ in positive degrees, in particular, $\iota_4$ and $\Gamma_4$. Therefore, by Proposition \[determinek\_5\], we have $$f_5^(\kappa_5)=f_5^*(\beta_{2^r}\times\iota_4+1\times\Gamma_4)=0,$$ and the dashed arrow in (\[lifting2\]) exists. Therefore, by Proposition \[lift to BP\] we have ${\operatorname}{ind}(\alpha)\neq 2^{3r+1}$, and we have proved Proposition \[mainprop\]. It remains to show the divisibility relations (\[bound1\]). For an arbitrary finite CW complex $X$ of dimension $8$ and $\alpha\in H^3(X;{\mathbb{Z}})_{\textrm{tor}}$ of period $n$, take $\alpha'\in H^2(X;{\mathbb{Z}}/n)$ such that the Bockstein homomorphism takes $\alpha'$ to $\alpha$. Then $\alpha'$, up to homotopy, gives rise to a map $X\rightarrow K({\mathbb{Z}}/n,2)$. Now apply Proposition \[mainprop\] and compare the twisted AHSSs of the pairs $(X,\alpha)$ and $(K({\mathbb{Z}}/n,2),\beta_n)$. Theorem \[main\] follows from the naturality of the twisted AHSSs together with Theorem \[AH diff\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $K$ be a compact subset in the complex plane and let $A(K)$ be the uniform closure of the functions continuous on $K$ and analytic on $K^{\circ}$. Let $\mu$ be a positive finite measure with its support contained in $K$. For $1 \leq q < \infty$, let $A^{q}(K,\mu)$ denote the closure of $A(K)$ in $L^{q}(\mu)$. The aim of this work is to study the structure of the space $A^{q}(K,\mu)$. We seek a necessary and sufficient condition on $K$ so that a Thomson-type structure theorem for $A^{q}(K,\mu)$ can be established. Our theorem deduces J. Thomson’s structure theorem for $P^{q}(\mu)$, the closure of polynomials in $L^{q}(\mu)$, as the special case when $K$ is a closed disk containing the support of $\mu$.[^1]' author: - \| [Zhijian Qiu ]{}\ \ [Department of mathematics, Southwestern Univ of Finance and Economics]{}\ [Chengdu 610072, China, E-mail:[email protected] ]{} title: 'The Structure of the Closure of the Rational Functions in $L^{q}$($\mu$)' --- 46E30 30H05, 30E10, 46E15 Introduction {#introduction .unnumbered} ============ Let $1 \leq q < \infty$ and let $\mu$ be a positive finite (regular Borel) measure with compact support in the complex plane [C]{}. Let $K$ be a compact subset that contains the support of $\mu$. The purpose of this paper is to investigate the following problem: > What is the closure of A(K) in $L^{q}(\mu)$? This is a very difficult question to get a complete answer. For a given measure $\mu$, the answer depends on $K$. Let $K$ be a closed disk that contains the support of $\mu$. Then every $f$ in $A(K)$ can be uniformly approximated by polynomials, and hence $A^{q}(K,\mu) = P^{q}(\mu)$, which is the closure of polynomials in $L^{q}(\mu)$. In this case, J. Thomson proved a structure theorem for $P^{q}(\mu)$ in [@jm 1991]. Roughly speaking, Thomson’s theorem says that there exists a Borel partition $\{\Delta_{n}\}_{n=0}$ of the support of $\mu$ such that $$P^{q}(\mu) = L^{q}(\mu\|\Delta_{0}) \oplus P^{q}(\mu\|\Delta_{1})\oplus ... \oplus P^{q}(\mu\|\Delta_{n}) \oplus ...,$$ where each $P^{q}(\mu\|\Delta_{n}) $ is identified with a space consisting of analytic functions on a simply connected domain $U_{n}$ via a so-called [evalaution map]{}. The union, $\cup U_{n}$, is known as the set of analytic bounded point evaluations ($abpe$s) for $P^{q}(\mu)$. In this special case, since $A(K)$ is the uniform closure of polynomials, Thomson’s theorem shows that $A(K)$ is dense in $ L^{q}(\mu)$ if and only if $P^{q}(\mu)$ ($= A^{q}(K,\mu)$ has no $abpe$s (this special result answered an open question raised by D. Sarason in 1972 [@sar]). For an arbitrary compact subset $K$, the author shows in [@qiu1] that it is still true: $A(K)$ is dense in $L^{q}(\mu)$ if and only if $ A^{q}(K,\mu)$ has no $abpe$s. However, the corresponding result for $A^{q}(K,\mu)$ is not always possible even the set of $abpe$s for $A^{q}(K,\mu)$ is not empty. J. Conway and N. Elias give an example in [@ce] that shows the set of $abpe$s for $A^{q}(K,\mu)$ is a simply connected domain $U$, but $A^{q}(K,\mu)\cap L^{\infty}(\mu)$ can not be identified with $H^{\infty}(U)$ via the evaluation map. For $R^{q}(K,\mu)$, the closure in $L^{q}(\mu)$ of $R(K)$ (which is the uniform closure of the rational functions with poles off $K$), the situation is worse since the set of $abpe$s may be empty even $R(K)$ is not dense in $L^{q}(\mu)$ (see [@bre1]). Assuming the existence of $abpe$s for $R^{q}(K,\mu)$, Conway and Elias proved a structure-type theorem for $R^{q}(K,\mu)$ under additional conditions. But their result does not imply Thomson’s theorem. Then, can we have a structure theorem for $A^{q}(K,\mu)$ or $R^{q}(K,\mu)$ that is beyond the polynomial case and that also covers Thomson’s theorem? Prior our work, it was unknown whether such a structure theorem for $A^{q}(K,\mu)$ is possible. Thomson was unable to offer any result for $R^{q}(K,\mu)$ that is beyond $P^{q}(\mu)$ (that is, the disk case in our setting) (see [@jm p. 505]). To tackle the problem, we first need to restrict our effort on those $K$ such that the components of $K^{\circ}$ are finitely connected. In fact, the author shows in [@comm Theorem 2] even when $K$ is a simplest kind of infinitely connected domains, such as a “road-runner”, our main theorem (Theorem \[t:main\]) for $A^{q}(K,\mu)$ could fail. In this paper, we seek a necessary and sufficient condition on $K$ so that a Thomson type of structure theorem holds for $A^{q}(K,\mu)$. A domain is called a circular domain if its boundary consists of finitely many disjoint circles. We call a domain $U$ multi-nicely connected if there is a circular domain $W$ and a conformal map $\alpha$ from $W$ onto $U$ such that $\alpha$ is almost 1-1 on $\partial W$ with respect to the arclength measure. Our main theorem, Theorem \[t:main\], extends Thomson’s theorem to $A^{q}(K,\mu)$ in the case when the components of $K^{\circ}$ are multi-nicely connected and the harmonic measures of the components of $K^{\circ}$ are mutually singular. We also show that the condition of Theorem \[t:main\] is necessary. If every $f$ in $H^{\infty}(K^{\circ})$ a pointwise limit of a bounded sequence in $A(K)$, then $K$ satisfies the condition of Theorem \[t:main\]. In particular, when $K$ is such that $R(K)$ is a hypo-Dirichlet algebra [@as; @ga], $K$ satisfies the condition of Theorem \[t:main\] (in this case, $R^{q}(K,\mu)=A^{q}(K,\mu)$). If the complement of $K$ has only finitely many components (note, $K^{\circ}$ may still has infinitely many components in this case), then $R(K)$ is a hypo-Dirichlet algebra and hence $K$ satisfies the hypothesis of Theorem \[t:main\]. Since a quite large of class of $R^{q}(K,\mu)$ satisfies our conditions, Theorem \[t:main\] is also a theorem for the rational $R^{q}(K,\mu)$. Since the polynomial case is just the disk case in our setting and since a general compact subset $K$ is much more complicated than a disk in nature, one can expect the extension needs much more work. To get a structure theorem for $A^{q}(K,\mu)$ (here $K$ is an arbitrary compact subset), we need more than Thomson’s technic and method. In fact, we need a new Thomson-type approximation scheme as developed in [@qiu1] that takes all of what is used in Thomson’s paper [@jm]. In addition, we need what was not involved in the case of $P^{q}(\mu)$: we make extensive use of results and technics related to uniform algebra or rational approximation: such as peak points, harmonic measures, hypo-Dirichlet algebra, multi-nicely connected domains, representing measures, pointwise bounded approximation, etc (these concepts are not needed for $P^{q}(\mu)$). So, besides Thomson’s technic and method, we need a significant part of the theories from the uniform algebras and the rational approximation to get the work done. Combination for Thomson’s technic and uniform rational approximation theory is the key to prove Theorem \[t:main\]. However, not every thing we need in uniform algebra theory is ready for us. We have to prove some results in that theory by ourself. In doing so, we first introduce the concept of multi-nicely connected domains, then we prove Proposition \[p:ABS2\] and an interesting result in uniform algebra, Lemma \[l:key\], which is crucial for us to prove Lemma \[l:key2\]. That is one of our key lemmas and it is needed to prove another key lemma, Lemma \[l:scheme\]. The rest of paper is to use these two lemmas and results in [@qiu1] and [@comm] to prove the main theorem and extend those lemmas and results that were proved for the polynomial case in [@jm]. So far, we are unable to offer any other proof that is less involved with the theory of uniform algebra. Actually, due to the nature of this problem, we believe the rational approximation theory is the right tool in study this type of problems. Now we would like to point out the relation between this paper and some other related papers. This paper is the sequel of the author’s work [@qiu1; @comm]. Thomson’s paper consist of two parts of important results. One is to give a sufficient and necessary condition on when $\nabla P^{q}(\mu)$ is not empty and another one is to have a structure theorem for $P^{q}(\mu)$. In [@qiu1], we only study the problem of when $\nabla A^{q}(K,\mu)$ ($\nabla R^{q}(K,\mu)$) is not empty and we show it is empty if and only if $A(K)$ is dense in $L^{q}(\mu)$. In [@comm], our effort was primarily to establish the result that is a part of 4) and 5) of Theorem \[t:main\] in this paper and to solve a problem in [@ce]. In this paper, our effort is to establish a full version of Thomson’s theorem for $A^{q}(K,\mu)$ ($R^{q}(K,\mu)$) and this paper is based on [@qiu1; @comm]. The readers may notice that our theorem (Theorem \[t:main\]) not only completely covers Thomson’s theorem, it also has more important consequences, such as 4) and 5) (which are not in [@jm] and are important facts. One needs to have them when applying it to operator theory. For example, see [@equi] and [@quas]). In [@ce], Conway and Elias studied the same problem as that in this paper. However, since their theorem is based on the assumption that $K$ is the closure of $\nabla R^{q}(K,\mu)$ and $R^{q}(K,\mu)$ is pure, so their result does not cover Thomson’s theorem. For a given measure $\mu$, one can not tell when their conditions are satisfied. In contrast, our paper deals with arbitrary measures just as [@jm] does. In 1972, D. Sarason [@sar] established a structure theorem for $P^{\infty}(\mu)$, the weak star closure of polynomials in $L^{\infty}(\mu)$, which has a similar form to that of our theorem. Preliminaries ============= For a compact subset $K$ in the complex plane [C]{}. Let $C(K)$ denote the algebra of continuous functions on $K$. For an open subset $G$ in the sphere [C]{}$_{\infty}$ whose boundary does not contains $\infty$, let $A(G)$ be the closed subalgebra of $C(\overline{G})$ that consists of functions continuous on $\overline{G}$ and analytic on $G$. Notice that $A(\overline {\Omega}) \neq A(\Omega)$ in general. A point $w$ in [C]{} is called an analytic bounded point evaluation ($abpe$) for $A^{q}(K,\mu)$ if there is a neighborhood $G$ of $w$ and $c > 0$ such that for all $\lambda\in G$ $$\|f(\lambda)\|\leq c \hspace{.05in} \\|f\\|_{L^{q}(\mu)}\hspace{.05in} \mbox{for all} \hspace{.05in} f\in A(K).$$ So the map, $f \rightarrow f(\lambda)$, extends to a functional in $A^{q}(K,\mu)^{}$. Thus, there is a (kernel) function $k_{\lambda}$ in $A^{q}(K,\mu)^{}$ such that $f(\lambda) = \int f k_{\lambda} \hspace{.05in}d\mu,\hspace{.05in} f \in A(K)$. Clearly, the set of $abpe$s is open. For each $f \in A^{q}(K,\mu)$, let $\hat{f}(\lambda) = \int f k_{\lambda} \hspace{.05in}d\mu$. Then $\hat{f}(\lambda)$ is analytic on the set of $abpe$s. The $abpe$s for $R^{q}(K,\mu)$ are define similarly. We shall use $\nabla A^{q}(K,\mu)$ to denote the set of $abpe$s for $A^{q}(K,\mu)$. The following is one of the main results in [@qiu1] which our main theorem relies on: \[t:approx\] Let $K$ be compact subset of [C]{} and let $\mu$ be a positive finite measure supported on $K$. Then $A(K)$ is dense in $L^{q}(\mu)$ if and only if $\nabla A^{q}(K,\mu) = \emptyset$. For the case when $K$ is the polynomially convex hull of the support of $\mu$, the above theorem is a consequence of Thomson’s theorem. However, since it was known long before Thomson’s paper [@jm] that there is a compact $K$ and a measure $\mu$ on $K$ such that $R(K)$ is not dense in $L^{q}(\mu)$ but $R^{q}(K,\mu)$ has no $abpe$, and since $$P^{q}(\mu) \subset R^{q}(K,\mu) \subset A^{q}(K,\mu)$$ always holds, Theorem 1.1 was unexpected before [@qiu1]. Somehow, it was a surprise that the theorem is true for the spaces on the both sides of the inequality above, but fails for the spaces between. [Nicely connected domains]{}. Following Glicksburg [@glic], we call a domain $\Omega$ nicely connected if it is multi-nicely connected and if it is simply connected. [Harmonic measures.]{} Let $\Omega$ be a domain in the extended plane [C]{}$_{\infty}$ such that it is solvable for the Dirichlet problem and $\infty$ is not in $\partial\Omega$. For $u \in C(\partial \Omega)$, let $\hat{u} =$ sup $\{f: f$ is subharmonic on $\Omega$ and $\limsup_{z\rightarrow a} f(z) \leq u(a), a \in \partial \Omega\}$. The function $\hat{u}$ turns out to be harmonic on $\Omega$ and continuous on $\overline{\Omega}$, and the map $u\rightarrow \hat{u}(z)$ defines a positive linear functional on $C(\partial \Omega)$ with norm one, so the [Riesz representing theorem]{} implies that there is a probability measure $\omega_{z}$ on $\partial \Omega$ such that $$\hat{u}(z) = \int _{\partial \Omega} u d \omega_{z}, \hspace{.2in} u \in C(\partial \Omega).$$ The measure $\omega_{z}$ is called the harmonic measure of $\Omega$ evaluated at $z$. The harmonic measures evaluated at two different points are boundedly equivalent. We shall use $\omega_{\Omega}$ to denote a harmonic measure of $\Omega$. [Hypodirichlet algebras]{}. A closed subalgebra $\cal B$ of $C(K)$ is said to be a hypo-Dirichlet algebra, if the uniform closure of $ Re(\cal B)= \{ Re(f): f \in \cal B\}$ has finite codimension in $C_{R}(K) = \{f: f\in C(K)$ and $f$ is real$\}$ and the linear span of log$\|\cal B^{-1}\|$ is uniformly dense in $C_{R}(K)$, where $\cal B^{-1}$ is the subset in $\cal B$ consisting of invertible elements. A function algebra $\cal B$ is called a Dirichlet algebra if $Re(B)$ is uniformly dense in $C_{R}(K)$. Clearly, a Dirichlet algebra is also a hypo-Dirichlet algebra. An good example is that $R(K)$ is hypo-Dirichlet if [C]{} $\setminus K$ has only finitely many components. This covers a large class of domains that have been studied. If $R(K)$ is a hypo-Dirichlet algebra, then $A(K)=R(K)$ [@ga p. 116]. [Peak points]{}. A point $a\in K$ is a peak point for a function algebra $\cal B \subset C(K)$ if there is a function in $\cal B$ such that $f(a) = 1$ at $z = a$ and $\|f(z)\| < 1$ for $z \neq a$. [Pure and irreducible spaces.]{} The space $A^{q}(K,\mu)$ is called [pure]{} if there is no Borel subset $\Delta$ of $supp(\mu)$ such that the restriction of $A(K)$ on $\Delta$ is dense in $L^{q}(\mu\|\Delta)$. An observation is that for any $A^{q}(K,\mu)$, there is a Berel partition $\{\Delta_{0}, \Delta_{1}\}$ of the support of $\mu$ such that $A^{q}(K,\mu\|\Delta_{1})$ is pure and $$A^{q}(K,\mu) = L^{q}(\mu\|\Delta_{0}) \oplus A^{q}(K,\mu\|\Delta_{1}).$$ The space $A^{q}(K,\mu)$ is said to be [irreducible]{} if it contains no nontrivial characteristic functions. So an irreducible space must be pure. [Nontangential limits]{}. Let $G$ be a bounded domain that is conformally equivalent to a circular domain $W$ in the plane and let $u$ be a conformal map from $W$ onto $\Omega$. Then $u$ has well-defined boundary values on $\partial W$, which are equal to the nontangential limits of $u$ for almost every point on $\partial W$ with respect to $\omega_{W}$ (the harmonic measure of $W$). We still use $u$ to denote the boundary value function. Now, if $E$ is a Borel subset of $\partial W$ such that $u$ is 1-1 on $E$ [a.e.]{} $ [\omega_{W}]$, then each $f\in H^{\infty}(G)$ has nontangential limits almost everywhere on $u(E) $ with respect to $\omega_{G}$. That is, $$f(a) = \lim_{z\rightarrow u^{-1}(a)} f\circ u(z) \hspace{.05in} \mbox{{\em a.e.} on} \hspace{.05in} u(E) \hspace{.05in}\mbox{ with respect to }\hspace{.05in} \omega_{G}.$$ So, if $\mu$ is such that $\mu \ll \omega_{G}$ on on $u(E)$, then each $f\in H^{\infty}(G)$ has nontangential limits on $u(E)$ almost everywhere with respect to $\mu$. The Main Result =============== In this section, we introduce our main result, Theorem \[t:main\]. Recall that the connectivity of a finitely connected domain is defined to be the number of the components of its complement. \[t:main\] Let $K$ be a compact subset and let $\mu$ be a finite positive measure supported on $K$. If each of the components of $K^{\circ}$ is multi-nicely connected and the harmonic measures of the components of $K^{\circ}$ are mutually singular, then there exists a $Borel$ partition $\{ \Delta_{n}\}_{n=0}^{\infty} $ of $supp(\mu)$ such that $$A^{q}(K,\mu) = L^{q}(\mu \| \Delta_{0}) \oplus A^{q}(K, \mu \| \Delta_{1}) \oplus ... \oplus A^{q}(K, \mu \| \Delta_{n}) \oplus ...$$ and for each $n \geq 1$, if $U_{n} $ denotes $ \nabla A^{q}(K, \mu\|\Delta_{n})$, then 1\) $\overline{U}_{n} \supset \Delta_{n} $ and $A^{q}(K, \mu \| \Delta_{n}) = A^{q}(\overline{U}_{n}, \mu \| \Delta_{n})$; 2\) each $U_{n}$ is a finitely connected domain that is conformally equivalent to a circular domain $W_{n}$; the connectivity of $U_{n}$ does not exceed the connectivity of the component of $K^{\circ}$ that contains $U_{n}$; 3\) the map $e$, defined by $e(f)=\hat{f}$, is an isometrical isomorphism and a weak-star homeomorphism from $A^{q}(K_{n}, \mu\|\Delta_{n}) \bigcap L^{\infty}(\mu\|\Delta_{n})$ onto $H^{\infty}(U_{n})$; 4\) $\mu\| \partial U_{n} \ll \omega_{U_{n}}$; and if $u_{n}$ is a conformal map from $W_{n}$ onto $U_{n}$, then for each $f\in H^{\infty}(U_{n})$ has nontangential limits on $\partial U_{n}$ [a.e.]{} $[\mu] $ and $$e^{-1}(f)(a) = \lim_{z\rightarrow u_{n}^{-1}(a)} f\circ u_{n}(z) \hspace{.05in} \mbox{\em{a.e.} on} \hspace{.05in}\partial U_{n} \hspace{.05in}\mbox{ with respect to }\hspace{.05in} \mu\|\partial U_{n};$$ 5\) for each $f\in H^{\infty}(U_{n})$, if let $f^{}$ be equal to its nontangential limit values on $\partial U_{n}$ and let $f^{} = \hat{f}$ on $U_{n}$, then the map $m$, defined by $m(f) = f^{}\|\Delta_{n}$, is the inverse of the map $e$. Thomson proved 1), 2) and 3) of Theorem \[t:main\] in the case when $A^{q}(K,\mu)=P^{q}(\mu)$. For the polynomial case, 4) is the main result in [@oy]. The author proved 4) for $A^{q}(K,\mu)$ with a different method in [@comm]. 3\) clearly implies that each $A^{q}(K,\mu_{n})$ is irreducible. The condition on $K$ is the best possible one. What we mean here is that in order to have Theorem \[t:main\] holds for any positive finite measure supported on $K$, it is necessary and sufficient that each component of $K^{\circ}$ is multi-nicely connected and the harmonic measures of the components of $K^{\circ}$ are mutually singular. Now we outline a proof for this fact. Let $\Omega$ be a component of $K^{\circ}$ and let $\mu$ be a harmonic measure for $\Omega$. By Theorem 3 of [@comm], the map $f\rightarrow \hat{f}$, from $A^{q}(\overline{\Omega}, \mu) \bigcap L^{\infty}(\mu)$ onto $H^{\infty}(\Omega)$ is surjective if and only if $\Omega$ is a multi-nicely connected domain. Hence we know that $\Omega$ must be multi-nicely connected. Now let $\Omega_{1}$ and $\Omega_{2}$ be two components of $K^{\circ}$. We want to show that $\omega_{\Omega_{1}}$ and $\omega_{\Omega_{2}}$ are mutually singular. Set $\mu=\omega_{\Omega_{1}} +\omega_{\Omega_{2}}$. Then it is easy to see that $A^{q}(K,\mu)$ is pure. Since the harmonic measure at a given point is a representing measure, it follows by the definition of $abpe$ and the Harnack’s inequality that $ \nabla A^{q}(K,\mu) \supset \Omega_{1} \cup \Omega_{2}. $ Since each $\Omega_{i}$ has no boundary slit, it follows clearly that $\nabla A^{q}(K,\mu) = \Omega_{1} \cup \Omega_{2}.$ If Theorem \[t:main\] holds for $A^{q}(K,\mu)$, then $$A^{q}(K,\mu) = A^{q}(\overline{\Omega}_{1},\mu_{1}) \oplus A^{q}(\overline{\Omega}_{2},\mu_{2}),$$ where $\mu_{i}$, $i=1, 2$, are as in Theorem \[t:main\]. Let $v$ be a conformal map of $\Omega_{1}$ onto a circular domain $W$. Theorem \[t:main\] implies that there exists $v_{e}\in A^{q}(\overline{\Omega}_{1}, \mu_{1})$ such that $\hat{v}_{e} = v$. Set $\eta = \mu_{1} \circ v_{e}^{-1}$. According to Lemma 2 in [@comm], $\eta$ is a measure on $\partial W$ such that $A^{q}(\overline {W},\eta)$ is irreducible and $\nabla A^{q}(\overline{W},\eta) = W$. Moreover, $\eta \ll \omega_{W}$ by Lemma 3 of [@comm]. On the other hand, since $W$ is circular and since $\nabla A^{q}(\overline{W},\eta) = W$, it is easy to see that $A^{\infty}(\overline{W}, \eta)$, the weak-star closure of $A(\overline{W})$ in $L^{\infty}(\eta)$, is equal to $\widetilde H^{\infty}(W)$, which is the image of the map $f\rightarrow \tilde f$ from $H^{\infty}(W)$ into $L^{q}(\omega_{W})$ (where $\tilde f$ is the boundary value function of $f$ on $\partial W$). Since the support of $\mu_{1}\subset \partial \Omega_{1}$, it follows by a classical result that \[$\omega_{W}] = [\eta]$. Now, applying Lemma 3 in [@comm], we conclude that $[\omega_{\Omega_{1}}]=[\mu_{1}]$. Similarly, we have $[\mu_{2}]=[\omega_{\Omega_{2}}]$. But $\mu_{1}$ and $\mu_{2}$ are mutually singular, therefore $ \omega_{\Omega_{1}}$ and $ \omega_{\Omega_{2}}$ must be mutually singular. The proof of the main result ============================ For each $f \in A^{q}(K,\mu),\hspace{.03in} \hat{f}=f$ on $\nabla A^{q}(K,\mu)$ a.e. $[\mu]$. Let $ a$ be an $abpe$ and choose a sequence $\{f_{n}\} $ in $A(K)$ such that $f_{n}\rightarrow f $ in $L^{q}(\mu)$. Since $f_{n}\rightarrow \hat{f}$ uniformly in a neighborhood of $a$, it follows (by passing a sequence if necessary) that $f_{n}\rightarrow f$ [a.e.]{} $[\mu]$ and consequently $$\hat{f}(a) = \lim_{n\rightarrow \infty} \hat{f}_{n} = \lim_{n\rightarrow \infty} f_{n} = f(a) \hspace{.07in}a.e. \hspace{.07in} [\mu].$$ The following lemma is elementary too. \[l:elem1\] If $f\in L^{\infty}(\mu) \cap A^{q}(K,\mu)$ and $g\in A^{q}(K,\mu)$, then $fg \in A^{q}(K,\mu)$ and $\widehat{fg} = \hat{f}\hat{g}$. The next lemma is proved in [@comm]. \[l:ISO\] Let $\Omega = \nabla A^{q}(K,\mu)$. If $\Omega$ is finitely connected, then every component of ([C]{} $\setminus \Omega$) has nonempty interior. [Representing measures]{}. Let $\cal B$ be a closed subalgebra of $C(K)$. A complex representing measure of $\cal B$ for $a\in K$ is a finite measure $\nu$ on $K$ such that $$f(a) = \int f d\nu, \hspace{.05in} f\in \cal B.$$ A [representing measure]{} for $a$ is a probability measure that satisfies the above condition. Note, if $a$ is a peak point, then the only representing measure for $a$ is the point mass $\delta_{a}$. [The sweep of a measure]{}. Let $G$ be a domain that is regular for the Dirichlet problem and let $\mu$ be a measure on $\overline{G}$. The sweep of $\mu$ is the unique positive measure $\tilde{\mu}$ on $\partial G$ that satisfies $\int _{\overline {G}}\tilde{u} \hspace{.05in}d\mu = \int_{\partial G} u d\tilde{\mu}, \hspace{.05in} u \in C(\partial G)$, where $\tilde{u}$ is the solution of the Dirichlet problem for $u$. A simple fact is that if $\mu$ is a measure on $\overline{G}$, then $\tilde{\mu} = \mu\|\partial G + \widetilde{\mu\|G}$. \[ABS1\] Let $K$ be a compact subset such that the components of $K^{\circ}$ are multi-nicely connected and the harmonic measures of the components of $K^{\circ}$ are mutually singular. Let $\Omega$ be a component of $K^{\circ}$. If $A^{q}(K,\mu)$ is pure and if $K^{\circ}$ is dense in $K$, then $\mu\|\partial \Omega \ll\omega_{\Omega}$. Let $\{\Omega_{j}\}_{j=0}^{\infty}$ be the collection of the components of $K^{\circ}$. Fix an integer $j\geq 0$ and let $E$ be a component of $\partial \Omega_{j}$. Let $G_{j}$ be the unique simply connected domain in the sphere [C]{}$_{\infty}$ that has $E$ as its boundary and contains $\Omega_{j}$. Since $\Omega_{j}$ is multi-nicely connected, $G_{j}$ must be nicely connected. For $i \neq j$, let $G_{i}$ be the bounded simply connected domain that contains $\Omega_{i}$ and whose boundary is a component of $\partial \Omega_{i}$. Clearly, $G_{i}$ is also nicely connected. Now let $\Omega$ be the union of those $G_{i}$’s for which $G_{i} \cap G_{j} = \emptyset$ (different $G_{i}$’s are either disjoint or one contains other). Set $ G= \Omega \bigcup G_{j}$. Then each component of $G$ is equal to some $G_{i}$. Let $\{G_{i_{k}}\}$ be the collection of all the components of $G$. Then our hypothesis on $K$ implies that the harmonic measures of the components of $G$ are mutually singular. It follows from [@davi] that $A(G)$ is a Dirichlet algebra on $\partial G$. Hence, every point in $\partial G$ is a peak point for $A(G)$ and every trivial Gleason part of $A(G)$ consists of a single point. Therefore, $\{G_{i_{k}}\}$ is the collection of all the nontrivial Gleason parts of $A(G)$. Let $\eta \perp A(G)$. By the [Abstract F. and M. Riesz Theorem]{} [@ga], $\eta = \sum_{m\geq 0} \eta_{m}$, where each $\eta_{m} \perp A(G)$, $\eta_{m} \ll v_{m}$ for a representing measure $ v_{m}$ at some point $a_{m}$ in $\overline{G}$, $v_{m}$’s are mutually singular. Let $a \in \partial G$. Then $a$ is a peak point. Let $f\in A(G)$ be a peak function for $a$. Then $f^{n}(z) \rightarrow \chi_{\{a\}}$ pointwise, and thus $0 = \int \lim_{n\rightarrow \infty} f^{n} d \eta_{m} = \eta_{m}(\{a\})$. Hence $ a_{m} \in G$ (otherwise, $\nu_{m}$ is the point mass at $a_{m}$ and hence $\nu_{m} (\overline{G} - \{a_{m}\}) =0$. So we conclude that $\eta_{m}(\overline{G}) = \eta_{m}(\overline{G} - \{a_{m}\}) + \eta_{m}(\{a_{m}\}) = 0+0=0$, a contradiction). Let $G_{i_{k_{m}}}$ be the component that contains $a_{m}$ and let $\tilde{v}_{m}$ be the sweep of $v_{m}$ on $\partial G_{i_{k_{m}}}$. Then for each $ g \in A(G)$ $$\begin{aligned} \int_{\partial G} g(z) d \tilde{v}_{m} = \int_{\partial G_{i_{k_{m}}}} g(z) d \tilde{v}_{m} = \int_{\overline{G}_{i_{k_{m}}}} g(z) d v_{m} = \int_{\partial G_{i_{k_{m}}}} g(z) d \omega_{a_{m}},\end{aligned}$$ where $\omega_{a_{m}}$ is the harmonic measure of $G_{i_{k_{m}}}$ evaluated at $a_{m}$. It follows by the uniqueness that $ \tilde{v}_{m}= \omega_{a_{m}}$. Hence, we have $$\begin{aligned} \eta\|\partial G \ll (\sum v_{a_{m}})\|\partial G \ll \sum \tilde{v}_{a_{m}}\|\partial G = \sum \omega_{a_{m}}\|\partial G\end{aligned}$$ In particular, we have that $\eta\|E \ll \omega_{\Omega_{j}} \|E $. Finally, suppose that $g\in L^{p}(\mu)$ such that $\int f g d \mu = 0$, for $f \in A(K)$, where $\frac{1}{q} + \frac{1}{p} = 1$. Then $g \perp A(G)$ as well. Hence, $g\mu \|E \ll \omega_{\Omega_{j}}\|E$. This implies that $(g\mu)_{s}\|E = 0$, where $(g\mu)_{s}$ is the singular part of the [Lebesgue decomposition]{} of $g\mu$ with respect to $\omega_{\Omega_{j}}\|E$. Consequently, we have $g\perp \chi_{\Delta\cap E}$, where $\Delta$ is the carrier of $\mu_{s}\|\partial \Omega_{j}$ and $\mu_{s}$ is the singular part of the [Lebesgue decomposition* ]{} of $\mu$ with respect to $\omega_{\Omega_{j}}$. Now an application of [Hahn-Banach theorem]{} yields $\chi_{\Delta\cap E} \in A^{q}(K,\mu)$. By purity, $\chi_{\Delta\cap E} = 0$ a.e. $[\mu]$ and therefore $\mu\|E \ll \omega_{\Omega_{j}} \|E $. Since $E$ is an arbitrary component of $\partial \Omega_{j}$, it follows that $\mu\|\partial \Omega_{j} \ll \omega_{\Omega_{j} }$. \[p:ABS2\] Let $K$ be a compact subset such that the components of $K^{\circ}$ are multi-nicely connected and the harmonic measures of $K^{\circ}$ are mutually singular. Let $\{\Omega_{j}\}_{j=0}^{\infty}$ denote the collection of the components of $K^{\circ}$. Set $\omega = \sum_{j=0}^{\infty} \frac{1}{2^{j}} \omega_{\Omega_{j}}$. If $A^{q}(K,\mu)$ is pure, then $\mu\|\partial K \ll\omega$. By Lemma 17.10 in [@conw p. 246], there exists a function $g\in A^{q}(K,\mu)^{\perp}$ such that $\|f\|\mu \ll \|g\|\mu$ for each $f\in A^{q}(K,\mu)^{\perp}$. Since $A^{q}(K,\mu)$ is pure, we see that $g \neq 0$ on $supp(\mu)$ [a.e.]{} $[\mu]$. Thus $[\|g\|\mu] = [\mu]$ [a.e.]{} $[\mu]$. Set $\nu = \|g\|\mu$. Let $a\in K -\overline{K^{\circ}}$ be such that $\int \frac{d \|\nu\|} {\|z-a\|} < \infty$ and let $f$ be a peak function for $a$. For each integer $n\geq 1$, $\frac{1-f^{n}(z)}{z-a} \in A(K)$, so we have $$\hat{\nu}(a) = \lim_{n\rightarrow \infty} \int \frac{1-f^{n}(z)}{z-a} d\nu = 0$$ Since the set $\{a: \int \frac{d \|\nu\|} {\|z-a\|} < \infty\}$ has full area measure in the plane, it follows by a well-known fact (see Theorem \[t:light\] or the comments after it) that $\nu$ is the zero measure off $\overline{K^{\circ}}$. Finally, since $\mu\|\partial \Omega_{j} \ll \omega_{\Omega_{j}}$ for each $j$ (by Lemma \[ABS1\]), the conclusion follows. The proof of the next lemma can be found in [@b_apr]. \[l:key\] Let $\Omega$ be an open subset whose boundary does not contain $\infty$. Suppose the components of $\Omega$ are multi-nicely connected and the harmonic measures of the components of $\Omega$ are mutually singular. Let $\{\Omega_{j}\}_{j=0}^{\infty}$ be the collection of the components of $\Omega$. If all but finitely many components of $\Omega$ are simply connected, then $A(\Omega)$ is boundedly pointwise dense in $H^{\infty}(\Omega)$. For a finite positive measure $\mu$, let $A^{\infty}(K,\mu)$ denote the weak-star closure of $A(K)$ in $L^{\infty}(\mu)$. Let $\Omega$ be a multi-nicely connected domain. Let $W$ be a circular domain that is conformally equivalent to $\Omega$ and let $\phi$ be a conformal map of $W$ onto $\Omega$. Then the boundary value function $\tilde{\phi}$ has a well-defined inverse, $\tilde{\phi}^{-1}$, on $\partial \Omega$. Moreover, $\tilde{\phi}^{-1}\in A^{\infty} (\overline{\Omega}, \omega_{\Omega})$. Since $\phi$ is almost 1-1 on $\partial W$ with respect to $\omega_{W}$, it is apparent that $\tilde{\phi}$ has a well-defined inverse function on a set of full $\omega_{\Omega}$ measure. By Lemma \[l:key\] $A(\Omega)$ is boundedly pointwise dense in $H^{\infty}(\Omega)$. Choose a bounded sequence $\{f_{n}\}$ in $A(\Omega)$ such that $f_{n}\rightarrow \phi^{-1}$ on $\Omega$. Then one can show that $f_{n} \rightarrow \tilde{\phi}^{-1}$ in the weak-star topology of $L^{\infty}(\omega_{\Omega})$. Hence $\tilde{\phi}^{-1} \in A^{\infty}(\overline{W}, \omega_{\Omega})$. The following is one of our key lemmas. \[l:key2\] Suppose that each component of $K^{\circ}$ is multi-nicely connected and the harmonic measures of the components of $K^{\circ}$ are mutually singular. Let $U$ be a component of $\nabla A^{q}(K,\mu)$ and let $\Omega$ be the component of $K^{\circ}$ that contains $U$. Set $\tau = \mu\|\overline{\Omega}$. Then $ A^{q}(\overline{\Omega},\tau)\subset A^{q}(K,\mu)$ and $ U \subset \nabla A^{q}(\overline{\Omega},\tau)$. We first assume that $A^{q}(K,\mu)$ is pure. Let $\{\Omega_{i}\}_{i=0}^{\infty}$ be the collection of all the components of $K^{\circ}$. Without loss of generality, let $\Omega_{0} = \Omega$. Suppose $h\in A^{q}(\overline{\Omega}, \tau)$ and choose a sequence $\{r_{n}\}$ in $A(\overline{\Omega})$ such that $r_{n}\rightarrow h $ in $L^{q}(\tau)$. Let $\omega = \sum_{i} \frac{1}{2^{i}} \omega_{\Omega_{i}}$. Fix a function $r_{n}$. Extend both $h$ and $r_{n}$ to be functions on the whole plane by defining their values to be zero off $\overline{\Omega}$. We claim that there exists a sequence $\{q_{n}\}$ in $A(K^{\circ})$ such that it weak-star converges to $r_{n}$ in $L^{\infty}(\omega)$. For each $i\geq 1$, let $G_{i}$ be the bounded simply connected domain that contains $\Omega_{i}$ and whose boundary is a component of $\partial \Omega_{i}$. Let $G$ be the union of $\Omega$ with those domains $G_{i}$ that do not intersect $\Omega$. Then all but finitely many components of $G$ are simply connected domains and each component of $G$ is multi-nicely connected. By Lemma \[l:key\], $A(G)$ is boundedly pointwise dense in $H^{\infty}(G)$. Thus, there exists a bounded sequence $\{q_{m}\}$ in $A(G)$ so that it pointwise converges to $r_{n}$ on $G$. Now for given $\epsilon > 0$, let $f\in L^{1}(\omega)$. Then $$\|\int_{\partial K} f(r_{n}-q_{m}) d\omega \| \leq \|\int_{\cup_{o}^{k} \partial \Omega_{i}} f(r_{n}-q_{m}) d\omega \| + \frac{\epsilon}{2} \hspace{.05in} \mbox{for all} \hspace{.05in} m$$ whenever $k$ is sufficiently large. Observe that $\{q_{m}\}$ weak-star converges to $r_{n}$ in $L^{\infty}(\omega_{\Omega_{i}})$ for each $i$. Thus, $$\|\int_{\partial K} f(r_{n}-q_{m}) d\omega \| \leq \frac{\epsilon}{2} + \frac{\epsilon}{2} \hspace{.05in}\mbox{when} \hspace{.05in}m \hspace{.05in}\mbox{is sufficiently large.}$$ Hence, for each $ f\in L^{1}(\omega) $ we have $$\lim_{n\rightarrow \infty} \int_{\partial K} f(r_{n}-q_{m}) d\omega = 0.$$ That is, $\{q_{m}\}$ weak-star converges to $r_{n}$ in $L^{\infty}(\omega)$. This proves the claim. Next we show that $r_{n}$ belongs to the weak-star closure of $A(\overline{K^{\circ}})$ in $L^{\infty}(\mu)$. By Proposition \[p:ABS2\], $\mu\|\partial K \ll \omega$. Thus $$\lim_{m\rightarrow} \int_{\partial K} f q_{m} \hspace{.05in}d\mu = \int_{\partial K} f r_{n} d \mu, \hspace{.05in}f \in L^{1}(\mu).$$ Since $\{q_{m}\}$ is bounded and pointwise converges to $r_{n}$ on $K^{\circ}$, it follows by the bounded convergence theorem that $$\lim_{n\rightarrow } \int fq_{m} \hspace{.05in}d\mu = \int f r_{n} \hspace{.05in}d\mu, \hspace{.05in} f \in L^{1}(\mu).$$ Therefore, $r_{n}$ belongs the weak-star closure of $A(\overline{K^{\circ}})$ in $L^{\infty}(\mu)$. But this closure is contained in the weak closure of $A(\overline{K^{\circ}})$ in $L^{q}(\mu)$. Because a convex set is norm-closed if and only if it is weakly closed in $L^{q}(\mu)$, we have $r_{n}$ is in $A^{q}(\overline{K^{\circ}},\mu)$. Now for each $n \geq 1$, choose $x_{n} $ in $A(\overline{K^{\circ}})$ such that $$\\|r_{n}- x_{n} \\|_{A^{q}(\overline{K^{\circ}},\mu)}\| \leq \frac{1}{n}. $$ Then $$\begin{aligned} \\|x_{n}-h\\|_{L^{q}(\mu)} & = \\|x_{n}-r_{n}\\|_{L^{q}(\mu)} + \\|r_{n}-h\\|_{L^{q}(\mu)} \\ & \leq \frac{1}{n} + \\|r_{n}-h\\|_{L^{q}(\tau)} \rightarrow 0, \hspace{.07in} \mbox{as}\hspace{.05in}\rightarrow \infty.\end{aligned}$$ Thus we have that $h\in A^{q}(\overline{K^{\circ}},\mu)$. Because $h$ is an arbitrary element in $A(\overline{\Omega},\tau)$, we conclude that $A^{q}(\overline{\Omega},\tau) \subset A^{q}(\overline{K^{\circ}},\mu).$ Since $A^{q}(K,\mu)$ is pure and since $\mu$ is supported on $\overline{K^{\circ}}$ (for $\mu\|\partial K \ll \omega$ and $\omega$ is supported on $ \partial K^{\circ}$), we have that $$A^{q}(K,\mu) = A^{q}(\overline{K^{\circ}},\mu).$$ Hence $A^{q}(\overline{\Omega},\tau) \subset A^{q}(K,\mu)$. Now let $b \in U$. Then $b$ is an $abpe$ for $A^{q}(K,\mu)$ and thus there exists $d>0$ and a small open disk $D_{b} \subset U$ such that for all $r\in A(K)$ $$\|r(a)\| \leq d\{\int \|r\|^{q} d \mu \}^{\frac{1}{q}}, \hspace{.05in} a \in D_{b}.$$ Let $y \in A(\overline{\Omega})$ and extend $y$ to be zero off $\overline{\Omega}$. Then $y \in A^{q}(K,\mu)$ and so there is a sequence $\{q_{n}\}$ in $A(K)$ so that it converges to $y$ in $L^{q}(\mu)$. Then $\{q_{n}\}$ converges to $y$ uniformly on $D_{b}$. Hence, it follows by the expression above that for all $y\in A(\overline{\Omega})$ $$\|y(a)\| \leq d \{ \int \|y\|^{q} d\tau \}^{\frac{1}{q}},\hspace{.05in} a\in D_{b}.$$ Thus $ a\in \nabla A^{q}(\overline{\Omega},\tau)$. By the definition of $abpe$, $U\subset \nabla A^{q}(\overline{\Omega},\tau)$. If $A^{q}(K,\mu)$ is not pure, let $\mu=\mu_{0}+\mu_{1}$ be the decomposition so that $A^{q}(K,\mu) = L^{q}(\mu_{0}) \oplus A^{q}(K,\mu_{1})$ and $A^{q}(K,\mu_{1})$ is pure. Then $$\nabla A^{q}(K,\mu) \supset \nabla A^{q}(K,\mu_{1}) \supset U.$$ So the conclusion of the lemma follows. A function $f$ analytic at $\infty$ can be written as a power series of the local coordinate $\frac{1}{z-z_{0}}$ at $\infty$: $$f(z)= f(\infty)+\frac{a_{1}}{z-z_{0}}+\frac{a_{2}}{(z-z_{0})^{2}} + ... .$$ The coefficient $a_{1}$ is called the derivative of $f$ at $\infty$ and is denoted by $f^{'}(\infty)$. It is easy to see that $f^{'}(\infty) = \lim_{z\rightarrow \infty} z(f(z)-f(\infty))$. Define $\beta(f,z_{0}) = a_{2}$. The next lemma is elementary. \[l:elem\] Let $\delta > 0$ and let $a\ in $ [C]{}. Let $B(a,\delta)= \{z: \|z-a\| \leq \delta \}$. If $f$ is a bounded analytic function on [C]{}$_{\infty}\setminus B(a,\delta)$, then $ \|f^{'}(\infty)\| \leq \delta \\|f\\|_{\infty}$ and $\|\beta(f,a)\| \leq \delta^{2}\\|f\\|_{\infty}$. [Thomson’s Scheme.]{} [^2] Now, we introduce an approximation scheme originally developed by J. Thomson in [@jm]. For an integer $k \geq 1$, let $\{S_{kp} \}_{p=1}^{\infty}$ be the collection of all open squares with sides $2^{-k}$, parallel to the coordinate axes and corners at the points whose coordinates are both integral multiples of $2^{-k}$. A finite sequence $\{S_{i} \}_{i=1}^{n}$ of squares is called a path of squares if the interior of $\cup \overline{S}_{i}$ is connected. In this case we say $S_{1}$ and $S_{n}$ are joined by a path of squares. The collection of $\{S_{kp} \}_{p=1}^{\infty}$ is called the $k$-th generation of squares. Let $\phi$ be a nonnegative function in $L^{1}($$)$. An open square $S$ is said to be [light]{} with respect to $\phi$ if $\int_{S}\phi \hspace{.05in} d \hspace{.02in}\mbox{Area} \leq \mbox{[Area}(S)]^{2}$. Now we begin with the scheme. Let $a \in $ [C]{} and let $S$ be a square in $\{S_{kp} \}_{p=1}^{\infty}$ such that $a \in \overline S$. Color $S$ yellow and let $\Gamma_{k} = \partial S$. We then move to the squares in the next generation. First, color green every light square in $\{S_{(k+1)p} \}_{p=1}^{\infty}$ that lies outside $\Gamma_{k}$ and has a side on $\Gamma_{k}$. Second, color green every light square that can be joined to a green square in the first step by a path of light squares in $\{S_{(k+1)p} \}_{p=1}^{\infty}$. Now if there is an unbounded green path (that is made up by infinitely many squares), then this coloring process ends. Otherwise, let $\gamma_{k+1}$ be the boundary of the polynomially convex hull [^3] of the union of $\Gamma_{k}$ and the closure of the green squares. We then color red every square $S$ in $\{S_{(k+1)p} \}_{p=1}^{\infty}$ if $S$ is outside $\gamma_{k+1}$ and $S$ has a side on $\gamma_{k+1}$. After that, color a square $T$ yellow if $T$ is outside $\gamma_{k+1}$ and $T$ has no side lying on $\gamma_{k+1}$ and the distance from $T$ to some red square in $\{S_{(k+1)p} \}_{p=1}^{\infty}$ is less or equal to $(k+1)^{2} 2^{-(k+1)}$. Now let $\Gamma_{k+1}$ be the boundary of the polynomially convex hull of the union and the closure of the colored squares in the $(k+1)$-th generation. To this step, the coloring process in $(k+1)$-th generation of squares is completed. Next we continue this process to the $(k+2)$-th generation of squares and keep this process to all higher generations unless there is an unbounded green path in the coloring scheme in some $(m+l)$-th generation $(l \geq 1)$. We use $(\phi, a, k)$ to denote this colored scheme. [Light and heavy points]{}. For a nonnegative function $\phi \in L^{1}($$)$, a point $\lambda $ in [C]{} is called light (with respect to $\phi$) if there exists $\delta > 0$ such that for each $\delta_{0} \leq \delta$ $$\{z: \|z-a\| = \delta_{0} \} \cap \{\mbox{all colored squares in}\hspace{.05in} (\phi,k,a)\} \neq \emptyset,$$ whenever $k$ is a sufficiently large integer. If a point is not light, then it is called a heavy point. The construction of our colored scheme is exactly the same as that in [@jm]. But the light and heavy points improved ’[light route to $\infty$’]{} and ’[heavy barrier]{}’ in Thomson’s original work. Let us explain the difference: For a given $\phi$, if there is an unbounded green path in the colored scheme $(\phi,k,a)$ for every $k$, it is said that there is a sequence of light routes from $a$ to $\infty$. This is essentially the definition of ’light’ points in Thomson’s paper. Because most of the light points for a given $\phi$ in our definition don’t have a sequence of light routes from $a$ to $\infty$, the set of the light points is much larger than the set of points that have a sequence of light routes from $\infty$. The Cauchy transform of a measure (with compact support) $\mu$ is defined as $\hat{\mu}(z) = \int \frac{ d\mu(w)}{w-z}$. Because $\frac{1}{z}$ is local integrable with respect to the $\mbox{area}$ measure, it follows that $\hat{\mu}(z)$ is defined everywhere except a subset of zero area. The following is a practically useful result coming out of our light point concept [@qiu1 Theorem 2.4]. \[t:light\] Let $\mu$ be a finite measure with compact support. Let $V$ be an open subset in [C]{}. If every point in $V$ is light with respect to $\|\hat{\mu}\|$, then $\|\mu\|(V) = 0$. The above theorem generalizes a well-known result in the theory of uniform approximation: [if $\hat{\mu} = 0$ [a.e.]{} on an open subset with respect to the area measure, then the restriction of $\mu$ on the open subset is zero.]{} The following is the key lemma in [@qiu1 Lemma 2.3]. \[l:cru\] Let $\nu_{1}, ..., \nu_{k}$ be finite measures. Let $\phi(z) = \max\{\|\hat{\nu}_{i}(z)\|: 1\leq j\leq k \}$. If $ a$ is a light point with respect to $\phi$, then there is an arbitrarily small positive number $\delta$ such that for any $\epsilon > 0$ and $\alpha,\beta \in \{z: \|z-1\|\leq 1\} $, there is a function in $C($[C]{}$_{\infty}$) that has the following properties: 1) $\\|f\\|_{\infty} \leq C$ (a universal constant), 2) $f$ is analytic on $\{z: \|z-a\| > \delta \}$, 3) $f(\infty) = 0$, 4) $f^{'}(\infty) = \alpha \delta$, 5) $\beta(f,a) = \beta \delta^{2}$, 6) $ \|\int f d\nu_{j}\|\leq \epsilon$ for all $1\leq j\leq k$. [Vitushkin covering]{}. For a natural number $k$, let $\{S_{kl} \}_{l=1}^{\infty}$ is the $k$-th generation of squares with sides of length $2^{-k}$. For each $S_{kl}$, let $F_{kl}$ be the square obtained by enlarging $S_{kl}$ $\frac{5}{4}$ times. The collection $\{F_{kl} \} $ is called a regular Vitushkin covering of the plane. We suppress $k$ and let $z_{l}$ be the center of $F_{l}$. Then there exists a $C^{1}$ partition of unity $\{\phi_{l}\}$ subordinate to $\{F_{l}\}$ with $\\|\mbox{grad} \phi_{l}\\| \leq 100\hspace{.05in} 2^{k}$ such that $$\sum_{l}^{\infty} \min(1, \frac{2^{-3k}}{\|z-z_{l}\|^{3}}) \leq C \min\{1, \frac{2^{-k}}{dist(z, \cup_{l} F_{l})} \}, \hspace{.3in}z\in \mbox{\bf C}.$$ One may consult [@conw] for a proof of the inequality. \[l:scheme\] Suppose that each component of $K^{\circ}$ is multi-nicely connected and the harmonic measures of the components are mutually singular. Let $U$ be a component of $\nabla A^{q}(K,\mu)$ and let $f \in H^{\infty}(U)$. Then there exists a function $h \in A^{q}(K,\mu)\cap L^{\infty}(\mu)$ such that $\hat{h}(z) = f(z)$ on $U$ and $h = 0$ off $\overline{U}$. First, we assume that $K$ is finitely connected, $K=\overline K^{\circ}$ and $K^{\circ}$ is connected. Let $\{x_{j}\}$ be a countable dense subset of $A^{q}(K,\mu)^{\perp}$. For an integer $k\geq 1$, let $\Phi(z)= \mbox{max}\{\|(\widehat{x_{j}\mu})(z)\|: j\leq k\}$. Let $\{F_{l}\}$ be the regular Vitushkin covering of squares with sides of length $\frac{5}{4} 2^{-k}$ and center $z_{l}$. Then there is a $C^{1}$ partition $\{\phi_{l}\}$ subordinate to the covering $\{F_{l}\}$. For each $l$, let $f_{l}= T_{\phi_{l}} f = \frac{1}{\pi} \int \int \frac{f(z) - f(w)}{z-w} \frac{\partial \phi_{l}}{\partial \overline{z}} \hspace{.05in}d\mbox{ Area}. $ Then $f_{l}$ is analytic off $F_{l}$, $f_{l}(\infty)=0$, and $\\|f_{l}\\|_{\infty} \leq 2\\|\mbox{grad} \phi_{l}\\|\mbox{diam} [supp(\phi_{l}] \sup\{\|f(z)-f(w)\|: z,w \in supp(\phi_{l}\} \leq C_{0}, $ where $C_{0}$ is a positive universal constant. Let $l$ be such that $F_{l} \cap \partial U \neq \emptyset$ and let $a \in \partial U\cap F_{l}$. We claim that $a$ is a light point with respect to $\Phi$. In fact, first let $a\in K^{\circ}$. Set $V=K^{\circ}\setminus \partial U$, it follows by Lemma 3.3 in [@qiu1] that $\|(\widehat{x_{j}\mu})(z)\|=0$ on $V$ for each $j$ and thus $\Phi=0$ on $V$. By Lemma 3.7 in [@qiu1] we see $a$ is light.[^4] Now suppose that $a \in \partial K$. Since $\nu_{j} \perp A(K) \supset R(K)$, we have that $\hat{\nu}_{j} = 0$ off $K$. Thus, again it follows from Lemma 3.7 in [@qiu1] that $a$ is also a light point. This proves the claim. Next let $d_{l} = \frac{1}{2} 2^{-k}$ and let $B(a, d_{l})$ be open disk having radius $d_{l}$ and the center at $a$. Applying Lemma \[l:elem\] to $B(a, d_{l})$, it follows that $\|f_{l}^{'}(\infty)\| \leq C_{0} d_{l} \hspace{.05in} \mbox{and} \hspace{.05in} \beta(f_{l},z_{l}) \leq C_{0} d_{l}^{2}. $ Let $\alpha=\frac{f_{l}^{'}(\infty)}{C_{0}d_{l}}$ and $\beta= \frac{\beta(f_{l},z_{l}) }{C_{0}d_{l}^{2}}$. Then $\|\alpha\| \leq 1$ and $\|\beta\| \leq 1$. Let $n$ be the number of those $F_{l}$’s for which $F_{l} \cap \partial U \neq \emptyset$. Then $n$ is a positive integer. Because $a$ is a light point, applying Lemma \[l:cru\] with $\alpha, \beta, \frac{1}{nk C_{0}}$, then there exists a function $g_{l}$ in $C($[C]{}$_{\infty})$ that is analytic off $\overline{B(a,d_{l})}$ and satisfies: 1) $\\|g_{l}\\|\leq C_{1}$ ($C_{1}$ is a positive universal constant), 2) $g_{l}(\infty)=0$, 3) $g_{l}^{'}(\infty)=\alpha d_{l}$, 4) $\beta(g_{l}, z_{l})= \beta d_{l}^{2}$, 5) $\|\int g_{l} x_{j} \hspace{.05in}d\mu\| \leq \frac{1}{nk C_{0}}$ for all $ 1 \leq j \leq k$. Set $h_{l} = C_{0} g_{l}$. Then $h_{l}$ has the following properties: 1) $ \\|h_{l}\\| \leq C_{0}C_{1}$, 2) $h_{l}$ is analytic off $\overline{B(a,d_{l})}$, 3) $\|\int h_{l}x_{j} d \mu\| \leq \frac{1}{nk}$ for $j\leq k$, 4) $h_{l}-f_{l}$ has a triple zero at $\infty$, that is, $(h_{l}-f_{l})(\infty)=0$, $(h_{l}-f_{l})^{'}(\infty)=0$ and $\beta(h_{l}-f_{l}, z_{l}) = 0$. Let $\delta_{l}$ be the length of a side of $F_{l}$. Since $a\in F_{l}$, it is evident that $B(a,d_{l})$ is contained in the square with center $z_{l}$ and sides of length $ 2 \delta_{l}$. $f_{l} - h_{l}$ is clearly analytic off $\{z: \|z-z_{l}\| \leq 2 \delta_{l} \}$. Since $f_{l}-h_{l}$ has a triple zeros at $\infty$, $(z-z_{l})^{3}(f_{l} - h_{l})$ is also analytic off $\{z: \|z-z_{l}\| \leq 2 \delta_{l} \}$. So the maximum principle implies that $$\|(z-z_{l})^{3}(f_{l} - h_{l})\| \leq 2^{3}\delta_{l}^{3} \\|f_{l}- h_{l}\\|_{\infty} \leq C_{0}(C_{1}+1)2^{3} \delta_{l}^{3} \hspace{.2in} \mbox{whenever}\hspace{.05in} \|z-z_{l}\| \geq 2 \delta_{l}.$$ Let $C_{2} = 8C_{0}(C_{1}+1)$. Then, for each $z$ $ \|f_{l}(z) - h_{l}(z)\| \leq \mbox{min} (C_{2}, \frac{C_{2}\delta^{3}} {\|z-z_{l}\|^{3}}). $ So it follows that for every $z$ (the first sum is taken over those $l$’s for which $F_{l} \cap \partial U \neq \emptyset$), $$\begin{aligned} \sum \|f_{l}- h_{l}\| & \leq \sum_{l}^{\infty} \mbox{min} (C_{2}, \frac{C_{2}\delta^{3}} {\|z-z_{l}\|^{3}}) = \sum_{l}^{\infty} \mbox{min} (C_{2}, \frac{C_{2}(\frac{5}{4})^{3}2^{-3k}} {\|z-z_{l}\|^{3}}) \\& \leq C_{2} (\frac{5}{4})^{3} \mbox{min}\{1, \frac{2^{-k}} {\mbox{dist}(z, \cup F_{l})}\}.\end{aligned}$$ Notice that $f$ is analytic on those $F_{l}$’s for which $F_{l} \cap \partial U = \emptyset$. It follows that $$f_{l} = T_{\phi_{l}} f = \frac{1}{\pi} \int \int \frac{f(z) - f(w)}{z-w} \frac{\partial \phi_{l}(z)}{\partial \overline{z}} \hspace{.05in}d\mbox{ Area} = - \frac{1}{\pi} \int \int \frac{\partial f(z)}{\partial \overline{z}}\frac{\phi_{l}(z)}{z-w} \hspace{.05in}d\mbox{ Area} =0$$ for those $l$’s. So there are only finitely many $f_{l}$’s that are not zero. Now define $h_{l} =0$ if $l$ is such that $f_{l} =0$ and set $y_{k} = f + \sum_{l}(h_{l} - f_{l})$. Then $y_{k} = \sum_{l}h_{l}$. For any $z$ off $\partial U$, it is clear that $ \mbox{dist} (z, \cup F_{l}) \rightarrow \mbox{dist} (z,\partial U) \hspace{.05in} \mbox{as} \hspace{.05in} k \rightarrow \infty, $ and hence it follows from the above inequalities that $y_{k} \rightarrow f(z)$ for each $z$ in\ [C]{} $\setminus \partial U$. According to 3), we have $ \| \int y_{k} x_{j} \hspace{.05in}d\mu\| \leq \frac{1}{k}, \hspace{.05in}\mbox{for} \hspace{.05in} 1\leq j\leq k. $ Notice that $$\|y_{k}\| \leq \|f\| + \|\sum_{l}(h_{l}-f_{l})\| \leq \\|f\\|_{\infty}+ C_{2}(\frac{5}{4})^{3},$$ so $\{y_{k}\}$ is a bounded sequence. Since the weak-star topology on the unit ball of the dual space of a separable Banach space is metrizable, it follows by Alaoglu’s theorem that there exists a subsequence $\{y_{k_{j}}\}$ that weak-star converges to some $h\in L^{\infty}(\mu)$. According to the last inequality above, we have that $ \int h x_{j} \hspace{.05in}d\mu = 0 \hspace{.05in}\mbox{ for all} \hspace{.05in}j\geq 1. $ Consequently, we have that $h\in A^{q}(K,\mu)$. Because $y_{k_{j}} \rightarrow f$ pointwise off $\partial U$ and $f = 0$ off $\overline U$, we have $h=0$ off $\overline U$. Finally, we show that $\hat{h} = f$ on $U$. If $supp(\mu)$ contains an open subset of $G$, then this is easy to see this is true (since $y_{k} \rightarrow f$ pointwise on $U$ and hence $f = h = \hat{h}$ [a.e.]{} $[\mu]$ on $G$. Because both $f$ and $\hat{h}$ are analytic on $U$, hence $f = \hat{h}$ on $U$). Otherwise, let $G$ be open so that $\overline{G}\subset U$ and let $\rho = \mu+ Area\|\overline{U}$. Then $\\| \cdot \\|_{\mu}$ and $\\|\cdot \\|_{\rho}$ are equivalent norms. By the definition of $abpe$, we see that $U\subset \nabla A^{q}(K,\mu) = \nabla A^{q}(K,\rho). $ Therefore, there exists $f_{1} \in A^{q}(K,\rho) \cap L^{\infty}(\rho)$ such that $\hat{f}_{1} = f$ and $f_{1} = 0$ off $\partial U$. Now set $h= f_{1}\|supp(\mu)$. We show $h$ is the desired function. Let $\{f_{n}\} \subset A(K)$ such that $f_{n} \rightarrow f_{1}$ in $L^{q}(\rho)$. Then $f_{n}\rightarrow h$ in $L^{q}(\mu)$. Thus, $f_{n}\rightarrow \hat{h}$ uniformly on $G$. Since $f_{n}\rightarrow f$ uniformly on $G$ as well, it follows that $\hat{h} = f$ on $G$. Because $U$ is the union of such open subsets $G$, we conclude that $\hat{h} = f$ on $U$. Now we consider a general $K$ that satisfies the hypothesis of this lemma. Let $\Omega$ be the component of $K^{\circ}$ that contains $U$. Then the multi-nicely connectivity of $\Omega$ insures that there is circular domain $W$ and a conformal map $v$ from $\Omega$ onto $W$ such that $v$ is almost 1-1 on $\partial \Omega$ with respect the harmonic measure of $\Omega$. Let $\mu= \mu_{0} +\tau$ be the decomposition such that $A^{q}(K,\tau)$ is pure and $A^{q}(K,\mu) = L^{q}(\mu_{0}) \oplus A^{q}(K,\tau). $ By Proposition \[p:ABS2\], $\tau\|\partial \Omega$ is absolutely continuous with respect to the harmonic measure. Extend $v$ to $\overline \Omega$ by defining its boundary values as its nontangential limits and set $\nu=\tau\circ v^{-1}$. It is easy to check that $v(U)$ is a component of $\nabla A^{q}(\overline W,\nu)$ and there is $h_{1}\in A^{q}(\overline W,\nu)$ such that $\hat{h}_{1} =f\circ v^{-1}$. Set $h=h_{1} \circ v^{-1}$. Then, $h\in A^{q}(\overline \Omega,\tau)$ and it is straightforward to verify that $\widehat h= f$. Extend $h$ to be a function on $K$ by defining $h=0$ off $\overline \Omega$. By Lemma \[l:key2\] $h \in A^{q}(\overline{\Omega},\mu\|\overline{\Omega}) \subset A^{q}(K,\mu). $ Clearly, $h$ does the job. \[l:EASY\] If$a\in \nabla A^{q}(K,\mu)$, then $\frac{f(z)-\hat{f}(a)}{z-a} \in A^{q}(K, \mu)$ for each $f\in A^{q}(K,\mu)$. Let $W=\nabla A^{q}(K,\mu)$. Then there exists $\{f_{n}\} \subset A(K)$ such that $f_{n} \rightarrow f$ in $L^{q}(\mu)$ and so $f_{n} \rightarrow \hat{f}$ uniformly on compact subset of $W$. Thus $\frac{f_{n}(z) - f_{n}(a)} {z-a} \rightarrow \frac{f(z) - \hat{f}(a)} {z-a} $ uniform on a small closed disk $ B(a,\delta) \subset W$. Note, $$\begin{aligned} \int_{K}\|\frac{f_{n}(z)-f_{n}(a)}{z-a}-\frac{f(z)-\hat{f}(a)}{z-a} \|^{q} d \mu & \leq M\int_{K} \|f_{n}(z) - f(z) - (f_{n}(a) -\hat{f}(a))\|^{q} d \mu \\& + \int_{B(a,\delta)} \|\frac{f_{n}(z) - f_{n}(a)} {z-a} - \frac{f(z) - \hat{f}(a)} {z-a} \|^{q} d \mu,\end{aligned}$$ where $M=\sup_{z\in K\setminus B(a,\delta)}\|\frac{1}{z-a}\|^{q} $. Thus, $\frac{f_{n}(z) - f_{n}(a)} {z-a} \rightarrow \frac{\hat{f}(z) - \hat{f}(a)} {z-a} $ in $L^{q}(\mu)$. Since $\frac{f_{n}(z) - f_{n}(a)} {z-a} \in A(K)$ for each $n$, the conclusion of the lemma follows. \[l:ABPE\] Suppose that $\nu \perp A(K)$ and $supp(\nu) \subset K$. Let $U$ be a component of $ K^{\circ} \setminus supp(\nu)$. If $\hat{\nu}(a) \neq 0 $ at some $a\in U$, then $U\subset \nabla A^{1}(K,\|\nu\|)$. Clearly $\hat{\nu}(z) = \int \frac{1}{z-w} d\nu(w)$ is analytic on $U$. Observe that for $f\in A(K)$, $\frac{f(z)-f(a)}{z-a} \in A(K)$ for every $a\in K^{\circ}$. Suppose that $\hat{\nu}(a) \neq 0$ for some $a\in U$. Then there exists a small closed disk $\overline{B(a,\delta)} \subset U$ so that $\hat{\nu}(z) \neq 0$ on $\overline{B(a,\delta)}$. For each $\lambda \in B(a,\delta)$, $\int \frac{f(z) - f(\lambda)}{z- \lambda} d \nu = 0$ and hence $$f(\lambda) = \frac{1}{\hat{\nu}(\lambda)} \int \frac{f(z)}{z-\lambda} d\nu, \hspace{.05in} \mbox{for every}\hspace{.05in} f \in A(K).$$ Since $\overline{B(a,\delta)}$ does not interest $supp(\nu)$, we see that $$\|f(\lambda)\| \leq c \\|f\\|_{L^{1}(\|\nu\|)}\hspace{.05in} \mbox{ for some} \hspace{.05in}c > 0 \hspace{.05in}\mbox{ on}\hspace{.05in} \hspace{.05in}B(a,\delta).$$ Hence, $a\in \nabla A^{1}(K,\|\nu\|)$. Since the zeros of $\hat{\nu}$ is isolated on $U$, it follows by Lemma \[l:ISO\] that $U \subset \nabla A^{1}(K,\|\nu\|)$. Let $h \in A^{q}(K,\mu)^{\perp}$ and set $\nu= h\mu$. Then $ \nabla A^{1}(K,\|v\|) \subset \nabla A^{q}(K,\mu). $ For $f\in A(K)$, by Hölder’s inequality $ \\|f\\|_{L^{1}(\|\nu\|)} \leq \\|h\\|_{L^{p}(\|\nu\|)} \\|f\\|_{L^{q}(\mu)} , $ where $\frac{1}{q}+\frac{1}{p}=1$. The the conclusion of the lemma clearly follows \[p:ABPE\] Let $\mu$ be a positive finite measure with $supp(\mu) \subset K$. Let $U$ be a component of $ K^{\circ}\setminus supp(\mu)$. If $U \cap \nabla A^{q}(K,\mu) \neq \emptyset$, then $U\subset \nabla A^{q}(K,\mu)$. Let $a\in U \cap \nabla A^{q}(K,\mu) $. Then there is $g\in A^{q}(K,\mu)^{\perp}$ such that $\int \frac{g d \mu}{z-a} \neq 0$. Clearly, $$g\mu \perp A(K)\hspace{.08in} \mbox{ and }\hspace{.08in} U \subset K^{\circ}\setminus supp(g\mu).$$ So it follows from the previous lemmas that $$U \subset \nabla A^{1}(K,\|g\mu\|) \subset \nabla A^{q}(K,\mu).$$ \[l:unique\] Let $\Omega = \nabla A^{q}(K,\mu)$ and let $U$ be a component of $\nabla A^{q}(K,\mu)$. Suppose that $A^{q}(K,\mu)$ is pure. Let $f\in A^{q}(K,\mu)$. If $\hat{f}=0$ and $f=0$ off $\partial U$ [a.e.]{} $[\mu]$, then $f=0$. Since $ A^{q}(K,\mu)$ is pure, as we argued in the proof of Proposition \[p:ABS2\], there exists $h \in A^{q}(K,\mu)^{\perp}$ such that $h \neq 0$ [a.e.]{} $[\mu]$ and $h = 0 $ off $\overline{K^{\circ}}$. So $h \perp A(\overline{K^{\circ}})$ as well. Let $\nu = hf \mu$. Then $\nu$ is a measure such that it is perpendicular to $A(K)$ and $supp(\nu) \subset \partial U$. We show that $\hat{\nu}(a) =0$ off $\partial U$. Let $W$ be a component of $K^{\circ} \setminus supp(\nu)$. We claim that $\hat{\nu}(z) = 0$ on $W$. Suppose $\hat{\nu}(a) \neq 0$ for some $a \in W$. According to Lemma \[l:ABPE\], $$W \subset \nabla A^{1}(K,\|\nu\|) \subset \nabla A^{q}(K,\mu).$$ By Lemma \[l:EASY\] and the hypothesis, we have that $\frac{f}{z-a} \in A^{q}(K,\mu)$. But $h \perp A(K)$. So we conclude that $\hat{\nu}(a) = 0$, which contradicting our assumption above. Therefore, $\hat{\nu} = 0$ on $W$. In particular, $\hat{\nu} = 0$ on $K^{\circ} \setminus \partial U$. It is easy to see that $ \overline{K^{\circ} \setminus \partial U} \supset \overline{K^{\circ}}.$ So, by the continuity we have that $\hat{\nu} = 0$ on $\overline{K^{\circ}} - \partial U$. Because $h \perp A(\overline{K^{\circ}}) \supset R(\overline{K^{\circ}})$, $\hat{\nu} = 0$ off $\overline{K^{\circ}}$. Hence, we conclude that $\hat{\nu} = 0$ off $\partial U$. Now, according to our definition, it is apparent that every point off $\partial U$ is a light point with respect to $\|\hat{\nu}\|$. So it follows from Lemma 3.7 in [@qiu1] that every point in $\partial U$ is light as well. Consequently, every point in the plane [C]{} is a light point. Applying Theorem \[t:light\], we conclude that $v = hf\mu = 0$. Since $h \neq 0$ [a.e.]{} on $K$, $f$ must be the zero function in $L^{q}(\mu)$. So we are done. \[l:finite\] Let $K$ be a compact subset in [C]{} such that each component of $K^{\circ}$ is finitely connected. Let $\mu$ be a positive finite measure supported on $K$. Then each component $U$ of $\nabla A^{q}(K,\mu)$ is a finitely connected domains conformally equivalent to a circular domain in the plane. Moreover, the connectivity of $U$ does not exceed the connectivity of the component of $K^{\circ}$ that contains $U$. Suppose $\nabla A^{q}(K,\mu) \neq \emptyset$. Let $U$ be a component of $\nabla A^{q}(K,\mu)$ and let $\Omega$ be the component of $K^{\circ}$ that contains $U$. Let $M$ be the connectivity of $\Omega$. Now suppose $F$ is a component of [C]{}$\setminus U$. We claim that $ F\cap (\mbox{ {\bf C}}\setminus \Omega ) \neq \emptyset. $ First, if $F$ is unbounded, this is obvious. So we assume $F$ is a bounded subset in the plane and assume that $ F\cap (\mbox{ {\bf C}} \setminus \Omega ) = \emptyset$. Then $F\subset \Omega$. Since $U$ is a connected domain, $F$ is polynomially convex (this means the complement of $F$ is connected). Since $U$ is finitely connected, there exists a Jordan curve $\gamma$ in $U$ such that $F $ is contained in $V$, the bounded Jordan domain enclosed by $\gamma$. Let $f\in A^{q}(K,\mu)$ and choose a sequence of functions $\{r_{n}\}$ in $A(K)$ such that $ r_{n} \rightarrow f$ in $L^{q}(\mu)$. Since $\gamma$ is contained in $ U \subset \nabla A^{q}(K,\mu)$, it follows that $r_{n} \rightarrow f$ uniformly on $\gamma$. Also, it is clear that we can choose $\gamma$ such that $dist(F, \gamma)$ small enough that the closure of $V$ is contained in $\Omega \subset K^{\circ}$. Then each $r_{n}$ is analytic on $\overline{V}$ and thus the maximum principle implies that $\{r_{n}\}$ uniformly converges to a function $h$ near $F$. By the definition of $abpe$s, we have that $F \subset \nabla A^{q}(K,\mu)$. But $F\cap \partial U \neq \emptyset$, and hence we conclude that $\partial U \cap \nabla A^{q}(K,\mu) \neq \emptyset$, which is a contradiction. Hence $ F\cap (\mbox{ {\bf C}} \setminus \Omega ) \neq \emptyset$. Now let $\{E_{i}\}$ be the collection of all the components of ([C]{} $ \setminus \Omega$) that intersect $F$. Then $F \cup (\cup E_{i})$ is connected compact subset and $$[ F \cup (\cup E_{i})] \cap U = \emptyset.$$ So $F \cup (\cup E_{i}))$ is contained a component of [C]{} $\setminus U$ that contains $F$. Hence $F \cup (\cup E_{i}) = F$ and therefore $E_{i} \subset F$ for each $i$. Consequently, each component of [C]{} $\setminus U$ contains at least a component of ([C]{} $\setminus \Omega$). Since the number of the components of [C]{} $\setminus \Omega$ is $M$, we see that the number of the components of [C]{} $\setminus U$ is less than Finally, since $U$ is finitely connected and since $\partial U$ contains no single-point component (by Lemma \[l:ISO\]), it follows by a classical result [@tsuji Tsuji, p. 424] that $U$ is conformally equivalent to a circular domain. The next two propositions are Theorem 1 and Theorem 3 in [@comm], respectively. We include them for readers convenience and self contained. \[p:abs\] Let $A^{q}(K,\mu)$ be irreducible. Let $U =\nabla A^{q}(K,\mu)$ be a finitely connected domain. If the map $e$, defined by $e(f) = \hat{f}$, from $A^{q}(K,\mu)\cap L^{\infty}(\mu)$ to $H^{\infty}(U)$ is surjective, then $\mu\|\partial U \ll \omega_{U}$, the harmonic measure of $U$. \[p:nont\] Let $A^{q}(K,\mu)$ be irreducible. Let $U =\nabla A^{q}(K,\mu)$ be a finitely connected domain and let $u$ be a conformal map from a circular domain $W$ onto $U$. If the map $e$, defined by $e(f) = \hat{f}$, from $A^{q}(K,\mu)\cap L^{\infty}(\mu)$ to $H^{\infty}(U)$ is surjective, then for each $f\in H^{\infty}(U)$ $$e^{-1}(f)(a) = \lim_{z\rightarrow u^{-1}(a)} f\circ u(z) \hspace{.05in} \mbox{a.e. on} \hspace{.05in}\partial U \hspace{.05in}\mbox{ with respect to }\hspace{.05in} \mu\|\partial U.$$ Moreover, $A(U) \subset A^{q}(K,\mu)$. . Let $\mu= \mu_{0} +\tau$ be the decomposition such that $A^{q}(K,\tau)$ is pure and $$A^{q}(K,\mu) = L^{q}(\mu_{0}) \oplus A^{q}(K,\tau).$$ Suppose $A(K)$ is not dense in $L^{q}(\mu)$. Then $\tau \neq 0$ in the decomposition. According to Theorem \[t:approx\], $\nabla A^{q}(K,\tau) \neq \emptyset$. Let $\{U_{n}\}_{n=1}^{\infty}$ be the components of $\nabla A^{q}(K,\tau)$. For each $n\geq 1$, by Lemma \[l:scheme\] there exists $f_{n}$ in $A^{q}(K,\tau)\cap L^{\infty}(\tau)$ such that $\hat{f_{n}} = \chi_{U_{n}}$ and $f_{n} = 0$ off $U_{n}$. Since $U_{n}$’s are pairwise disjoint, we have $\widehat{f_{n}f_{m}} =\hat{f}_{n} \hat{f}_{m} = 0. $ It follows by Lemma \[l:unique\] that $f_{n}f_{m} = 0$. Similarly, since $\hat{f}_{n}^{2} = \hat{f_{n}}$, we get that $ f^{2}_{n} = f_{n}.$ Therefore, we conclude that $f_{n}=\chi_{\Delta_{n}}$ for some Borel subset $\Delta_{n}$. But $\tau(\Delta_{n} \cap \Delta_{m}) = 0$ (because $f_{n}f_{m}=0$). Thus $\Delta_{n}$’s can be chosen to be pairwise disjoint. Moreover, since $f_{n} = 0$ off $\overline U_{n}$, we can also require that $\overline \Delta_{n} \subset \overline U_{n} $. For each $n\geq 1$, let $K_{n}=\overline{U}_{n}$ and let $\mu_{n} = \tau\|\Delta_{n}$. We claim that $U_{n} = \nabla A^{q}(K_{n},\mu_{n}). $ Let $\Omega$ be the component of $K^{\circ}$ that contains $U_{n}$. By Lemma \[l:key2\], $A^{q}(\overline{\Omega},\tau\|\overline{ \Omega})\subset A^{q}(K,\tau)$. Note, every function $f$ in $A^{q}(\overline{\Omega}, \tau\|\overline{ \Omega})$ has zero values off $ \overline \Omega$. Clearly, $f_{n} $ ($\chi_{\Delta_{n}})$ belongs to $ A^{q}(\overline{\Omega},\tau\|\overline{ \Omega})$ also. Set $F_{n}= U_{n} \cup (\Delta_{n}\setminus U_{n}). $ Because $\overline \Delta_{n} \subset \overline U_{n} $, we see that $\chi_{F_{n} } = f_{n} \in A^{q}(\overline{\Omega},\tau\|\overline{ \Omega})$. So by Lemma \[l:elem1\], $\chi_{F_{n}}f \subset A^{q} (\overline{\Omega}, \tau\|\overline{\Omega})$ for each $f\in A(\overline{\Omega}) $. This implies that $ A^{q}(\overline{\Omega},\mu_{n})= A^{q}(\overline{\Omega},\tau\|F_{n} ) \subset A^{q}(\overline{\Omega},\tau). $ Let $a \in U_{n}$. Then there exists $c>0$ and an open open disk $D_{a} \subset U_{n}$ such that for all $f\in A(\overline \Omega)$ $$\|f(a)\| \leq c\{\int \|f\|^{q} d \tau \}^{\frac{1}{q}}, \hspace{.05in} a \in D_{a}.$$ Let $g \in A(\overline \Omega)$. Then there is a sequence $\{q_{i}\}_{i=1}^{\infty}$ in $A(\overline \Omega)$ so that $q_{i}\rightarrow g\chi_{F_{n}}$ in $L^{q}(\tau)$. Then $q_{i}\rightarrow g$ uniformly on $D_{a}$. Hence, it follows that for $a\in D_{a}$ $$\|g(a)\| = \lim\|q_{i}(a)\| \leq c \lim \{\int\|q_{i}\|^{q} d\tau\}^{\frac{1}{q}}= c\{\int g\chi_{F_{n}} d\tau \}^{\frac{1}{q}} = \hspace{.05in} c\{ \int \|g\|^{q} \hspace{.05in}d\mu_{n} \}^{\frac{1}{q}}.$$ Thus $ a\in \nabla A^{q}(\overline{\Omega},\mu_{n})$. Therefore, $U_{n}\subset \nabla A^{q}(\overline{\Omega},\mu_{n})$. By Lemma \[l:key2\], $\nabla A^{q}(\overline \Omega,\mu_{n}) \subset \nabla A^{q}(K,\tau)$, so we see (notice that $U_{n}$ is a component of $\nabla A^{q}(K,\tau)$) that $ U_{n} = \nabla A^{q}(\overline \Omega,\mu_{n}). $ The hypothesis and Lemma \[l:finite\] together imply that $U_{n}$ is a finitely connected domain. It is also easy to see that $A^{q}(\overline \Omega,\mu_{n})$ is irreducible. Applying Proposition \[p:nont\], we have $A(U_{n}) \subset A^{q}(\overline \Omega, \mu_{n})$. Consequently, $A^{q}(\overline U_{n},\mu_{n}) \subset A^{q}(\overline \Omega, \mu_{n})$. Therefore, we conclude that $A^{q}(\overline \Omega, \mu_{n})= A^{q}(\overline U_{n},\mu_{n}) $. Hence, $U_{n} = \nabla A^{q}(K_{n},\mu_{n}).$ This proves the claim. Since each $A^{q}(K_{n},\mu_{n})$ is contained in $A^{q}(K,\mu) $ and since $\{A^{q}(K_{n},\mu_{n})\}$ are pairwise othrogonal, we have $$A^{q}(K,\mu) \supset L^{q}(\mu_{0}) \oplus A^{q}(K_{1}, \mu \| \Delta_{1}) \oplus ... \oplus A^{q}(K_{n}, \mu \| \Delta_{n}) \oplus ... .$$ For the other direction of the equality, let $f$ be the pointwise limit of $\{ \sum_{i=1}^{n} f_{i} \}_{n=1}^{\infty}$. Then the bounded convergence theorem implies that $f\in A^{q}(K,\tau)$. We show that $1 - f = 0$ a.e. $[\tau]$. Otherwise, there exists a Borel subset $E$ of the support of $\tau$ such that $0 \neq \chi_{E} = 1 - f$. Since both $1$ and $f$ are in $A^{q}(K,\tau)$, we have that $\chi_{E} \in A^{q}(K,\tau)$. By the purity, we have $L^{q}(\tau\|E) \neq A^{q}(K,\tau\|E)$. So it follows by Theorem \[t:approx\] that $\nabla A^{q}(K,\tau\|E) \neq \emptyset$. But $\chi_{E} f_{n} = 0$ for each $n \geq 1$, thus we have $ \nabla A^{q}(K,\tau\|E) \cap (\cup U_{n}) = \emptyset. $ But by the definition of $abpe$s, $\nabla A^{q}(K,\tau\|E) \subset \nabla A^{q}(K,\mu) = \cup U_{n}. $ This is a contradiction, and hence $f - 1 =0$. Therefore, $\{ \Delta_{n}\}$ is a Borel partition. Let $g\in A^{q}(K,\tau)$. By the Lebesgue dominated convergence theorem, we conclude that $g= \lim_{n\rightarrow \infty} \sum_{i=1}^{n} f_{i} g$ in $ L^{q}(\tau)$. Therefore, $$A^{q}(K,\mu) \subset L^{q}(\mu_{0}) \oplus A^{q}(K_{1}, \mu \| \Delta_{1}) \oplus ... \oplus A^{q}(K_{n}, \mu \| \Delta_{n}) \oplus ... .$$ Consequently, $$A^{q}(K,\mu) = L^{q}(\mu_{0}) \oplus A^{q}(K_{1}, \mu \| \Delta_{1}) \oplus ... \oplus A^{q}(K_{n}, \mu \| \Delta_{n}) \oplus ... .$$ Now we prove the rest of Theorem \[t:main\]: For 1), since we have already proved $\overline U_{n} \subset \Delta_{n}$ above, we only need to show that $A^{q}(K,\mu_{n}) =A^{q}(\overline U_{n},\mu_{n})$. Because $A(K)\subset A(\overline U_{n})$, it follows by the definition of $abpe$ that $$U_{n} \subset \nabla A^{q}(\overline U_{n},\mu_{n}) \subset \nabla A^{q}(K,\mu_{n}).$$ Notice that $A^{q}(K,\mu_{n})\subset A^{q}(\overline U_{n},\mu_{n})$ and the latter is irreducible. So we see that $A^{q}(K,\mu_{n})$ is irreducible also. This implies that $\nabla A^{q}(K,\mu_{n})$ have only one component. Also, it is clear that $\nabla A^{q}(K,\mu_{n}) \subset \nabla A^{q}(K,\tau)$. So we conclude that $\nabla A^{q}(K,\mu_{n}) =U_{n}$. Thus, there is $h_{n}\in A^{q}(K,\mu_{n})$ such that $\hat{h}_{n} = \chi_{U_{n}}$ and $h_{n} =0 $ off $\overline U_{n}$. By the uniqueness, we get that $h_{n} = f_{n} = \chi_{\Delta_{n}}$. As we proved above, $f = \sum_{n=1}^{\infty} fh_{n}$ for each $f\in A^{q}(K,\tau)$. Hence, we conclude that $$\begin{aligned} A^{q}(K,\tau) & \subset A^{q}(K, \mu_{1}) \oplus ... \oplus A^{q}(K, \mu_{n}) \oplus ... \\ & \subset A^{q}(\overline U_{1}, \mu_{1}) \oplus ... \oplus A^{q}(\overline U_{n}, \mu_{n}) \oplus ... \\ & = A^{q}(K,\tau).\end{aligned}$$ Consequently, $A^{q}(K,\mu_{n}) = A^{q}(\overline U_{n},\mu_{n})$ for each $n \geq 1$. 2\) follows from Lemma \[l:finite\]. For 3), let $e$ be the map, $f \rightarrow \hat{f}$, from $L^{\infty}(\mu_{n}) \cap A^{q}(K_{n},\mu_{n})$ into $H^{\infty}(U_{n})$. Then $e$ is surjective by Lemma \[l:scheme\] and is injective by Lemma \[l:unique\]. Since $e(fg) = e(f) e(g)$, $e$ is an algebraic isomorphism between two commutative Banach algebras and thus $e$ is an isometry. Next we need to show that $e$ is a weak-star homeomorphism. To do this, we will argue as in [@ce]. Using [Krein-Smulian theorem]{} it suffices to show that $e$ is weak-star sequentially continuous. Recall that a sequence of functions in $H^{\infty}(U_{n})$ is weak-star Cauchy sequence if and only if it is uniformly bounded on $U_{n}$ and it is a Cauchy sequence in the topology of pointwise convergence. Let $\{h_{i}\}$ be a sequence in $A^{q}(K_{n},\mu_{n})\cap L^{\infty}(\mu_{n})$ that converges to zero in the weak star topology. By the uniform boundedness, $\{h_{i}\}$ is bounded and hence $\{e(h_{i})\}$ is also bounded. Let $a\in U_{n}$ and let $k_{a}$ be kernel function. Then $$\lim_{i\rightarrow \infty} e(h_{i}) = \lim_{i\rightarrow \infty} \hat{h}_{i}(a) = \lim_{i\rightarrow \infty} \int h_{i} k_{a} \hspace{.02in} d\mu_{n} =0.$$ So $e(h_{i})$ weak-star converges to zero. Therefore, $e$ is a weak-star homeomorphism. 4\) follows from Proposition \[p:abs\] and Proposition \[p:nont\]. For 5), by Proposition \[p:nont\], each $f\in H^{\infty}(U_{n})$ has nontangential limits almost everywhere on $\partial U_{n}$ with respect to $\mu\|\partial U_{n}$ and the nontangential limits are equal to $e^{-1}(f)$ [a.e.]{} $[\mu\|\partial U_{n}]$. Since $\hat{f}_{e} = f = f^{} $ [a.e.]{} $[\mu]$ on $U_{n}$, we see that $f^{}= e^{-1}(f) \|\Delta_{n}$ [a.e.]{} $[\mu]$. Evidently $$\begin{aligned} m(e(f)) = m(\hat{f}) & = \left\lbrace \begin{array}{c l} \hat{f} \hspace{.05in}\mbox{on}\hspace{.05in}U_{n} &\\ e^{-1}(\hat{f}) \hspace{.05in}\mbox{on}\hspace{.05in}\partial U_{n} \end{array} \right. \\ & = \left\lbrace \begin{array}{c l} f \hspace{.05in}\mbox{on}\hspace{.05in}U_{n} &\\ f \hspace{.05in}\mbox{on}\hspace{.05in}\partial U_{n} \end{array} \right. \\ & = f, \hspace{.2in}\mbox{for each}\hspace{.05in} f\in A^{q}(K_{n},\mu_{n}).\end{aligned}$$ Therefore, $m$ is the inverse map of $e$. So the proof of Theorem \[t:main\] is complete. [Remark:]{} This paper contains the best and close-up result in an once quite active research area. However, this paper has not been cited by any other authors but me, while [@jm] has been cited more than 80 times. I re-published this paper on Arxiv to hope to bring more attentions from future generation of mathematicians to the results in this paper, which I believe is great and can stay in history of mathematics. [Acknowledgement.]{} During 1994, I gave several talks on the result of this work in turn in the SouthEast Analysis Meeting at Virginia Tech, UC-Berkeley, AMS Summer Research Conference at Mt. HolyOak, Brown University and Wabash Conferences at IUPUI. When $R(K)$ is a hypodirichlet algebra, A slightly simper version of Theorem \[t:main\] was stated as the last theorem in my work [@qiu1]. [0000]{} P. Ahern, D. Sarason, [The $H^{q}$ spaces of a class of function algebras,]{} Acta Math., 117 (1967), 123-163. P. Ahern, D. Sarason, [On some hypo-Dirichlet algebras of analytic functions,]{} Amer. J. Math., 89 (1967), 932-941. , J. Akeroyd, E. Saleeby, [A class of $P^{t} (d\mu)$ spaces whose point evaluations very with $t$]{}, Proc. Amer. Math. Soc., Vol 127, No. 2 (1999), 537-542. J. Brenann, [Invariant subspaces and rational approximation]{}, J. Functional Analysis 7 (1971), 285-301. J. Conway, [The theory of subnormal operators]{}, Math. Surveys and Monographs Vol. 36, Amer. Math. Soc. 1991. J. Conway, N. Elias, [Analytic bounded point evaluations for spaces of rational functions]{}, J. Functional Analysis 117 (1993), 1-24. A. Davie, [Dirichlet algebras of analytic functions]{}, J. Functional Analysis 6 (1967), 348-356. T. Gamelin, [Uniform Algebras]{}, Prentice Hall, Englewood Cliffs, N.J., 1969. T. Gamelin, J.B. Garnett, [Constructive techniques in rational approximation]{}, Trans. Amer. Math. Soc. 143 (1969), 187-200. I. Glicksburg, [The abstract F. and M. Riese theorem, ]{} J. Functional Analysis 1 (1967), 109-122. S. Mergeljan, [On the completeness of system of analytic functions]{}, Amer. Math. Soc. Translation 19 (1962), 109-166. J. McCarthy, [Analytic structure for subnormal operators,]{} Integral Equation and Operator Theory, Vol. 13 (1990), 251-270. J. McCarthy, L. Yang, [Bounded point evaluations on the boundaries of L regions]{}, Indiana Univ. Math. J., Vol 43, No. 3, Fall 1994. T. Miller, R. Smith, [Nontangential limits of functions in some $P^{2}(\mu)$ spaces]{}, Indiana Univ. Math. J., Vol. 39, No. 1 (1990), 19-26. T. Miller, W. Smith, L. Yang, [Bounded point evaluations for certain $P^{t}(\mu)$ spaces]{}, Illinois J. Math., Vol 43, No. 1, Spring 1999. R. Olin, L. Yang, [The commutant of multiplication by $z$ on the closure of polynomials in $L^{q}(\mu)$]{}, J. Functional analysis, 134 (1995), 297-320. Zhijian Qiu, [Equivalence classes of subnormal operators]{}, J. Operator Theory, Vol 32 (1994), 47-75. Zhijian Qiu, [Density of polynomials]{}, Houston J. Math., Vol. 21 (1995), No. 1, 109-118. Zhijian Qiu, [Boundary values of analytic functions in the Banach spaces $P^{t}(\tau)$ on crescents]{}, Illinois Journal of Math., Vol 39, Number 2, Summer (1995), 305-322. Zhijian Qiu, [Approximation in the mean by rational functions]{}, Integral Equation and Operator Theory, Vol. 25 (1996), 235-252. Zhijian Qiu, [The commutant of rationally cyclic subnormal operators and rational approximation]{}, Integral Equation and Operator Theory, 27 (1997), No. 3, 334-346. Zhijian Qiu, [Carleson measures on circular domains]{}, Houston J. Math., Vol. 31 (2005), No. 4, 1199-1206. Zhijian Qiu, [A class of operators similar to the shift on $H^{2}(G)$]{}, Integral Equation and Operator Theory, 56 (2006), No. 3, 415-429. Zhijian Qiu, [On quasisimilarity of subnormal operators]{}, Science in China, Series A Math. 2007 Vol. 50, No. 2, 145-154. Zhijian Qiu, [Carleson measure and polynomial approximation]{}, Chinese Ann. of Math., Series A, 2007 Vol 28, No. 159-167. Zhijian Qiu, [On pointwise bounded approximation]{}, to appear. D. Sarason, [Weak-star density of polynomials]{}, J. Reine Angew. Math., 252 (1972), 1-15. J. Thomson, [Approximation in the mean by polynomials]{}, Ann of Math. 133 (1991), 477-507. T. Trent, [$ H^{2}(\mu)$ spaces and bounded point evaluations]{}, Pacific J. Math. 80, 279-292 M. Tsuji, [Potential theory in modern function theory]{}, Chelsa, New York, 1975. A. Vitushkin, [Analytic capacity of sets in problems of rational approximation]{}, Uspehi Mat. Nauk 22 (1967), No. 6 (138), 141-199= Russian Math. Surveys 22 (1967), 139-200. J. Walsh, [The approximation of harmonic functions by harmonic polynomials and harmonic rational functions]{}, Bull. Amer. Math. Soc. 3, 276-277. J. Wermer, [Analytic disks in maximal ideal spaces]{}, American J. of Mathematics, Vol. 86 (1964), 161-170. [^1]: This paper basically gives ultimate solution to the most important research problem raised in [@conw; @jm] and it is also very satisfactory generation of Thomson’s theorem for polynomials. It is the close-up work of this kind (in an once quite active research area) even it received no attentions since its publication in 2007. [^2]: In this paper, we don’t directly use this scheme. But we need the concept of light and heavy points and results related to them (Lemma \[l:cru\] and Theorem \[t:light\]). [^3]: The polynomially convex hull of a compact subset $K$ in the plane is defined as the union of $K$ and all the bounded components of [C]{}$\setminus K$. [^4]: By combining the proof Lemma 3.3 in [@qiu1] and that of Theorem 4.8 in [@jm], we can extend Lemma 3.3 in [@qiu1] so that it has the conclusion that Theorem 4.8 in [@jm] has (so it can be applied to $\sum_{j\leq k} \|x_{j}\mu\|$)). From this, we also see any point in $K^{\circ}$ that is not an $abpe$ must be light.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present the results of a broadband simultaneous campaign on the nearby low-luminosity active galactic nucleus M81\. From February through August 2005, we observed M81\ five times using the Chandra X-ray Observatory with the High-Energy Transmission Grating Spectrometer, complemented by ground-based observations with the Giant Meterwave Radio Telescope, the Very Large Array and Very Large Baseline Array, the Plateau de Bure Interferometer at IRAM, the Submillimeter Array and Lick Observatory. We discuss how the resulting spectra vary over short and longer timescales compared to previous results, especially in the X-rays where this is the first ever longer-term campaign at spatial resolution high enough to nearly isolate the nucleus (17pc). We compare the spectrum to our Galactic center weakly active nucleus Sgr A\, which has undergone similar campaigns, as well as to weakly accreting X-ray binaries in the context of outflow-dominated models. In agreement with recent results suggesting that the physics of weakly-accreting black holes scales predictably with mass, we find that the exact same model which successfully describes hard state X-ray binaries applies to M81\, with very similar physical parameters.' author: - Sera Markoff - 'Michael Nowak, Andrew Young, Herman L. Marshall, Claude R. Canizares' - Alison Peck - Melanie Krips and Glen Petitpas - Rainer Schödel - 'Geoffrey C. Bower' - Poonam Chandra - Alak Ray - 'Michael Muno, Sarah Gallagher and Seth Hornstein' - 'Chi C. Cheung' title: 'Results from an extensive simultaneous broadband campaign on the underluminous active nucleus M81\: further evidence for mass-scaling accretion in black holes ' --- Introduction ============ Black holes are a common feature in galaxies, spanning a huge range in mass, from the stellar-sized remnants scattered throughout the galaxy volume to the supermassive black hole often thought to lurk in the galaxy center. Black holes are known to operate over at least ten orders of magnitude in luminosity, and thus experience accretion rates that range from super-Eddington (defined with respect to an associated Eddington luminosity via $\dot M_{\rm Edd} \equiv L_{\rm Edd}/c^2$, where $L_{\rm Edd}\equiv {4\pi c G M m}/{\sigma_{\tau}}$) to extreme sub-Eddington. Clearly the accretion rate is the overall dominating factor determining the energy output; however, the accretion flow behavior of at least stellar mass black holes changes rather drastically between low to high accretion rates. Variations include the cyclic appearance and apparent subsequent quenching of jet outflows, alterations in accretion disk characteristics, and changes in the overall radiative efficiency. For a description of the most recent observations and their implications for theoretical models of accretion flows, see the reviews by [@RemillardMcClintock2006; @DoneGierlinskiKubota2007], and references therein. Above a few percent of the Eddington luminosity, for instance, accretion around black holes seems to be well-characterized by a dominant, thermally emitting “standard thin disk” [@ShakuraSunyaev1973]. Somewhere below this threshold there appears to be a transition to a radiatively inefficient state with some form of advective accretion and/or outflow [see, e.g., @MeyerMeyerHofmeister1994; @NarayanYi1994; @BlandfordBegelman1999; @QuataertGruzinov1999]. Although this transition has been predicted theoretically, the physical details and configurations of weakly accreting flows are still under significant debate [see, e.g. @Rykoffetal2007]. In particular there are open questions regarding fundamental plasma characteristics such as the coupling– and therefore respective temperatures– of the ions and electrons, viscosity and the role of magnetic fields, the accretion flow geometry, and the relationship of the accretion flow to outflows and jet production. For stellar mass black holes, we are learning a great deal by observing the transitions in real time between accretion states. Unfortunately, it is not possible to track such changes directly in supermassive black holes (SMBHs) as the relevant dynamical times scale approximately with the mass. Instead, ensemble comparisons can be made with large samples of accreting SMBHs that range from near-Eddington down to the intrinsically weakest active galactic nuclei (AGN), called low-luminosity AGN (LLAGN; @Heckman1980 [@Ho1999]). These sources are difficult to observe unless nearby, however, because of their intrinsically weak emission. Given its special role as the weakest observable active nucleus, Sgr A\, our Galactic center SMBH, has become the poster-child for a multitude of theoretical and observational studies. Several extensive multiwavelength campaigns [e.g. @Baganoffetal2003; @Eckartetal2004; @Anetal2005; @Yusef-Zadehetal2006] have well established the simultaneous broadband spectrum of Sgr A\, which provides a tight constraint on physical models. Over the course of these campaigns, however, some very unusual flaring behavior has been discovered in Sgr A\’s X-ray emission. No other black hole has shown similar flaring; however, currently we are unable to detect any other Sgr A\-like objects for comparison. Furthermore, the presence of jets has not been definitively confirmed or ruled out for Sgr A\, complicating our ability to use it as a test source for accretion/outflow relationships at low accretion rates [but see @MarkoffBowerFalcke2007]. The distinct lack of any thin accretion disk component in its spectrum also makes Sgr A\* unique compared to other LLAGN. What seems to be required is a “bridge” source, to help span the gap between Sgr A\* and other LLAGN in terms of accretion rate, while sharing as many other qualities as possible (mass, spectral features, galaxy type, etc.). Comparing Sgr A\* to such a source would help us determine what processes may be different or absent at the lowest accretion rates. A comparison of this type, however, would require simultaneous broadband data for the bridge source, of the quality now only associated with the Galactic center campaigns. The nucleus of the nearby galaxy M81 is an ideal candidate for such an extensive multiwavelength campaign on a LLAGN, for the following reasons. Although it is one of the intrinsically weakest LLAGN known, it is among the brightest because it is the nearest galaxy besides Centaurus A with a central AGN. Furthermore it is the nearest point-like LLAGN, similarly inside a spiral galaxy, for which reliable measurements of the black hole mass are available. Spectroscopy with the Hubble Space Telescope (HST) suggests a central mass of $7\times10^7$ [@Devereuxetal2003], and M81’s distance is only 3.6 Mpc [@Freedmanetal1994]. The nucleus, M81\* (following the convention based on Sgr A\), is associated with a compact radio core and exhibits both low-ionization emission line region (LINER; @Heckman1980 [@Ho1999]) and Seyfert 1 characteristics. In terms of radiative power, probable accretion rate, and the length of the jet emanating from the core, M81\ lies in the intermediate range between radio loud AGN and Sgr A\. The X-ray properties of M81\ are very typical for the LLAGN class [e.g. @Ho1999]. Its nonthermal X-ray luminosity is around a few $10^{-5} L_{\rm Edd}$, though its proximity means it is still bright. Similarly its spectral energy distribution (SED) displays no “big blue bump” yet does show evidence for double-peaked optical line emission [@Boweretal1996], suggesting the presence of a weak accretion disk. [HST]{} observations with STIS indicate that the disk is close to face-on with an inclination of 14$^\circ$ [@Devereuxetal2003]. As will be described in more detail below, M81\* exhibits significant variability across its SED on both short (daily) and long (monthly/yearly) time scales. M81\* has a low absorption column [@Pageetal2003], which allows UV and soft X-ray detection. Perhaps most importantly for the goals of this campaign, M81\* shares several important characteristics with Sgr A\, specifically its radio slope and polarization, that make it the ideal comparison source. As studies of our own Galactic center have shown, detailed, multi-wavelength observations of single objects such as M81\ are indispensable tools for understanding black hole accretion. In this paper we will present the results of a simultaneous, broadband, multiwavelength campaign on M81\, and discuss how it compares with Sgr A\, as well as its weakly accreting black hole counterparts in X-ray binaries (XRBs). The [Chandra]{} observations specifically resulted in the first gratings-resolution X-ray spectrum of an isolated LLAGN nucleus, which is the focus of a companion paper [@Youngetal2007]. The millimeter observations carried out with the Plateau de Bure Interferometer(PdBI) are also presented in more detail elsewhere [@Schoedeletal2007]. In Section \[sec:oldobs\] we summarize the results of previous observations of M81\* and in Section \[sec:obs\] we describe our new observations and analysis. We present the resulting broadband spectra in Section \[sec:newobs\] and their interpretation in the specific context of a jet-dominated model in Sections \[sec:model\] & \[sec:results\]. Section \[sec:discuss\] contains our discussion and conclusions. Previous Observations of M81\* {#sec:oldobs} ============================== The M81\* nucleus has been observed extensively in many wavebands (partly due to its proximity to supernova SN 1993J, which undergoes regular monitoring). In this section we will briefly summarize the previously known broadband characteristics of M81\* and discuss how they have provided the motivation for more detailed and higher resolution observations. Radio/millimeter Observations {#subsec:radio} ----------------------------- M81\* exhibits the signature flat/inverted radio spectrum associated with the compact cores of AGN. Such a spectrum is well-explained by a collimated, steady jet which radiates via self-absorbed synchrotron emission along its length [see, e.g. @BlandfordKoenigl1979; @Falcke1996]. Based on observations from 1.4–22.5 GHz using the VLA, several groups have observed an inverted spectrum ($\alpha\sim 0.0$–0.3, $F_\nu\propto\nu^{\alpha}$) with a flux in the range of $\approx80-300$ mJy [@Hoetal1999; @BietenholzBartelRupen2000; @Brunthaleretal2001; @BrunthalerBowerFalcke2006]. Based on four years of observations with the VLA, @Hoetal1999 note that M81\* undergoes frequent radio outbursts, with the underlying flux around $100$mJy and higher fluxes during flares. The larger flare events occur on time scales of months, and seem to roughly correspond with predictions of simple adiabatic expansion models [e.g. @vanderLaan1966] in which the variability moves towards lower frequencies as the flare amplitude decreases. Ho et al. also claim they detect intraday variability at a level of 10-60% amplitude changes. If the longer term radio flares are indeed expanding ejecta moving out along otherwise steady jet structures, there should be even higher amplitude variations in the millimeter regime. Consistent with this view, [@Sakamotoetal2001] observed the 3 mm flux to double within a single day. A one-sided jet system has been resolved in M81\. [@BietenholzBartelRupen2000] identified a stationary radio core with a very small (700 AU at 22 GHz), precessing (over 20$^\circ$) jet, the structure of which varies on relatively short timescales. Interestingly, this group found no significant intraday variations over a similar time frame as the Ho et al. study. Investigating this question of variability was one of our campaign’s goals for the lower frequencies. M81\’s radio luminosity is about four orders of magnitude brighter than Sgr A\’s, but its shape and polarization are quite similar. A key, and somewhat unusual, trait that these two sources share is the dominance of circular polarization up to ${\mathrel{ \rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}}}22$GHz in M81\, and $\sim 112$GHz in Sgr A\* [@Brunthaleretal2001; @Boweretal2002; @BrunthalerBowerFalcke2006]. Sgr A\* becomes increasingly linearly polarized towards its peak flux in the submm [@Boweretal2003; @Boweretal2005]; however, the characteristics of M81\* in the submm have not yet been determined. The interpretation of Sgr A\’s circular polarization is Faraday depolarization by the surrounding accretion flow. If the same physics is active in M81\, we would expect to detect linear polarization in the submm range as well. Infrared through Ultraviolet Observations {#subsec:uv} ----------------------------------------- Typical of LLAGN, M81\* lacks a bright optically thick “standard thin disk” component in the optical range, although double-peaked optical lines do suggest the presence of a weak disk [@Boweretal1996]. The low column ($N_h\sim5\times10^{20}$ cm$^{-2}$; @Pageetal2003) allows us to detect UV and soft X-rays from the nucleus, which suggests a nearly face-on disk. [@Maozetal2005] used the [HST ACS]{} to discover variable (by tens of %) UV emission, consistent with the results of [@HoFilippenkoSargent1996], who detected a weak and very steep ($\alpha\approx-2$) UV continuum. Instruments with less spatial resolution such as [Spitzer]{} [@Willneretal2004; @Murphyetal2006], MIRLIN [@Grossanetal2001] and ISOPHOT-S [@Satyapaletal2005] are all consistent with a steep ($\alpha\le-1.7$) non-stellar spectrum, similar to what is observed in the IR/UV in the LLAGN NGC 4258 [@Charyetal2000]. This is of interest because it suggests that in LLAGN, the UV is likely nonthermal emission, while the optical, and perhaps IR, contain the only potential signatures of a radiatively efficient, thin accretion disk. X-ray Observations {#subsec:xray} ------------------ M81\* has a persistent nonthermal power-law flux in the X-rays, with variations of factors of 3 or more over yearly time scales. [ASCA]{} has detected both long term variations, as well as 20–30% intraday variations which suggest that the source size is less than a few hundred gravitational radii ($r_{\rm g}\equiv GM/c^2$) [@IyamotoMakishima2001]. [ROSAT]{} also confirmed long term X-ray variability, with a factor of $\approx 2.5$ amplitude [@ImmlerWang2001]. A summary of these and more recent variability trends can be found in [@LaParolaetal2004]. M81\’s X-ray luminosity appears to vary between ($2$–$6)\times10^{41}$ ergs s$^{-1}$, which corresponds to $\sim (2$–$6) \times 10^{-5} L_{\rm Edd}$. It was not until observations with [BeppoSAX]{} that more detailed statements could be made about the nuclear X-ray emission properties. [@Pellegrinietal2000] observed both the short intraday as well as the long term variability over the 0.1–100keV band. However, due to the poor angular resolution, they could only place a lower limit of ${\mathrel{ \rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}}}80\%$ of the continuum originating in the nucleus. The [BeppoSAX]{} observations yielded several important new results which contributed to our interest in M81\ as a potential [Chandra HETGS]{} target. First, the data were consistent with no reflection component or blue bump, which would be unusual for a Seyfert 1 – a class of objects with which M81\* otherwise shares some qualities. The [BeppoSAX]{} results make it less likely that M81\* is a simple extension of the Seyfert 1 class to low luminosity. Although [BeppoSAX]{} did not detect reflection, it did detect emission and absorption features of highly ionized iron; however, these features were seemingly not correlated with the continuum luminosity. Second, there were problems reconciling the ionization with the low inferred accretion rate. [@Pellegrinietal2000] suggested that instead of ionization, there may instead be transmission through a highly photoionized medium close to the nucleus, such as a warm absorber. Third, while the observed X-ray powerlaw with $\Gamma\approx1.8$–1.9 was typical of bright Seyferts, there was no direct evidence for a thin accretion disk. A lack of a strong disk component is consistent with accretion being dominated by a radiatively inefficient accretion flow (RIAF); however, the question then arises whether such flows can account for the strong Seyfert-like powerlaw over the entire 0.1-100keV range at the low inferred accretion rates of M81\. [XMM-Newton]{} has confirmed these findings, at least in the 0.3-8keV band, and also detected redshifted Fe K$ \alpha$ as well as He- and H-like ionized iron [@Pageetal2003]. Because [Chandra]{} is the only X-ray mission with the spatial resolution to almost isolate the nucleus of M81\ (to within $\sim 17$ pc of the black hole), determining the nature of the line emission was one of our primary goals. As described in [@Youngetal2007], we indeed detect not just iron but many other low-metallicity species, as well as velocity broadening of some of these lines. The broadened line components are consistent with arising from regions close to the black hole, i.e., ${\mathrel{ \rlap{\raise 0.511ex \hbox{$<$}}{\lower 0.511ex \hbox{$\sim$}}}}10^{5}~GM/c^2$. Detected line features include those associated with fluorescence from cold material (Ar K$\alpha$, Si K$\alpha$, and Fe K$\alpha$), emission lines from a hot plasma (Ne [x]{}, Mg [xii]{}, Si [xiii]{}), and absorption lines (18.44Å and 20.74Å) that could be consistent with an outflowing wind. The plasma emission lines, specifically the Si [xiii]{} triplet line strength ratios, are consistent with a collisionally ionized plasma (although other models can not be ruled out). The focus of the [@Youngetal2007] work is on the X-ray spectra of M81\, specifically the aforementioned line features in relation to the X-ray continuum. Young et al. show that the X-ray spectra are consistent with the expectations of a somewhat simplified RIAF model. We did not consider the simultaneous radio and submm data in the context of those models. In this work we now include a description of the lower frequency (radio through sub-millimeter) observations, and consider the entire broadband spectrum in the context of an outflow-dominated model. Observations and Data Reduction {#sec:obs} =============================== The difficulty in observing a source as faint as M81\ with the [Chandra HETGS]{} lies in the required long integration times (300ksec) to adequately resolve the narrow (${\mathrel{ \rlap{\raise 0.511ex \hbox{$<$}}{\lower 0.511ex \hbox{$\sim$}}}}2500~{\rm km~s^{-1}}$ Full Width Half Maximum) line features. Additionally, aside from the Fe features found with [BeppoSAX]{} and [XMM]{} observations (which, due to the relatively poorer spatial resolution of these instruments, could have arisen from well outside the nucleus), the existence of line features from the innermost regions of LLAGN was uncertain. As such, at the time of our M81\* observations, no gratings observation of an LLAGN had been accepted in the [Chandra]{} Guest Observer program. Instead, we obtained a series of [HETGS]{} observations via the Guaranteed Time Observation program (PI: C. Canizares). In order to further constrain models and better understand the variability trends of M81\, we supplemented the [Chandra]{} program by proposing for simultaneous coverage with five ground-based instruments that span the lower frequencies: the Giant Meterwave Radio Telescope (GMRT), the Very Large Array/Very Large Baseline Array (VLA/VLBA), the Plateau de Bure Interferometer (PdBI) at IRAM, the Sub-millimeter Array (SMA), and Lick Observatory. Figure \[campoview\] gives an overview of the total campaign. For the two periods of greatest multi-wavelength overlap, Figures \[febview\] & \[julview\] give a closeup view of the coverage, and Tables \[febtimes\]– \[augtimes\] give the exact times in UT. In the following subsections we provide details of the individual instrument observations and data reduction. Low-frequency Radio Waves: GMRT {#subsec:gmrt} ------------------------------- The Giant Meterwave Radio Telescope (GMRT) is an aperture synthesis radio telescope [@Swarupetal1991] situated 80 km north of Pune in Western India at latitude $19^o \,06'$ and longitude $74^o\,03'$ E. The telescope operates at 233, 327, 610 and 1420 MHz bands, and consists of 30 fixed position, fully steerable paraboloid dishes of diameter 45 meters. Fourteen out of these 30 dishes are located within about a square kilometer of each other, the remaining 16 antennae form a “Y”-shaped array with northwest, northeast and southern arms spread over an area of 25 kilometers in diameter. The baselines in the central one square kilometer area are useful to map the extended emission of the source, whereas the wider baselines in the “Y” provide high angular resolution. We observed M81\ with the GMRT in 1420, 610 and 235 MHz bands on several occasions during the campaign. The total time spent on M81\* in the 235/610 bands was 5–8 hours, and 3–5 hours in the 1420 MHz band. The bandwidth at 1420 and 610 MHz was 16 MHz, divided into a total of 128 frequency channels, i.e., the default for the correlator. For the 243 MHz wave band the bandwidth was 6 MHz. Calibrator sources were used to remove the effects of instrumental variations in the measurements. 3C48, 3C286 and 3C147 were used as flux calibrators. 1035+564 was used as a phase calibrator in the 1420 MHz observations, whereas 0834+555 was used in the 610 and 235 MHz observations. The flux and phase calibrators were used for bandpass calibration as well. Flux calibrators were observed once or twice for 20–30 minutes during each observing session. Phase calibrators were observed for 5–6 minutes after every 25 minutes of observations. We used the Astronomical Image Processing System (AIPS) developed by NRAO for the data analysis, including the standard GMRT data reduction [see @ChandraRayBhatnagar2004 for details]. Standard flagging routines of AIPS were used to remove the bad antennas and corrupted data. About 25–30 antennae could be used in the radio interferometric setup at different observing epochs. Data were then calibrated and images and fields were formed by Fourier inversion and CLEANing using AIPS task IMAGR. We took into account the bandwidth smearing effects and wide field imaging, even though in this case M81\* is not spatially resolved, as a byproduct of analysis to derive the correct flux density of the nearby SN 1993J. Bandwidth effects were negligible for 1420 and 610 MHz bands and we averaged 100 central channels. For the 235 MHz observations, we divided the central 55–60 good channels and divided them into 4 sub-bands and stacked them together while imaging. To take care of the wide field imaging, we divided the whole field in 3 subfields for 1420 MHz and 610 MHz observations, and into 18 subfields for 235 MHz observations. A few rounds of self-calibrations were also performed in all the datasets to remove the phase variations. AIPS task FLATN was used to combine all the sub fields into one single image. Table \[m81\_obs\] gives details of the observations. The typical resolution of $3"$ at 1390 MHz at a distance of 3.6 Mpc for M81 corresponds to about 54 pc. Table \[sn1993\] shows a comparison of calibration observations of SN 1993J for both the GMRT and the VLA, indicating consistent flux levels. Centimeter Radio:\ VLA and VLBA+Effelsberg {#subsec:vla} ----------------------- The Very Large Array observed M81\* on 2005 14 February, 13 July, 19 July, and 14 August. Observations were obtained at 1.4, 8.4, 22, and 43 GHz on each of the days in continuum mode. Data were obtained in fast-switching mode between M81\* and the compact calibrators J1048+717 and J1056+701. We performed calibration of amplitude and phase variations on short time-scales using J1048+717 and transferred solutions to M81\* and J1056+701. Results for J1056+701 are, therefore, a check on variability of M81\. The amplitude scale was set by observations of 3C 286. Weather on 14 February and 13 July was poor making results at 22 and 43 GHz inaccurate. We determined average flux densities for each day as well as measuring flux density on short time scales. SN 1993J was in the field of view at 1.4 and 8.4 GHz. Its flux density was constant between the epochs. The Very Long Baseline Array and the Effelsberg 100m observed M81\ on 2005 13 July. Observations were made at 8.4 GHz with a sampling rate of 128 Mb/s, while attempts at 22 GHz failed due to poor weather. Standard self-calibration reduction techniques were performed and M81\* was imaged with a resolution of 0.6 mas. M81\* is clearly resolved into a compact and extended component (Fig. \[vlbafig\]). The peak flux density in the compact component is about 60 mJy. In the extended component, the peak flux density is about 10 mJy, but the total flux density in the whole region is $\approx2$–3 times that value. We fit the image with two elliptical Gaussians: the best fit compact component has major axis of 0.65mas, a minor axis of 0.57mas and a position angle of 81$^\circ$, and the best fit extended component has major axis 1.5 mas, minor axis 0.66mas and position angle 55$^\circ$. The separation between the two components is approximately 1 mas, and the total extension is about $10^4$ AU across. These results are similar to what was observed by [@BietenholzBartelRupen2000]. Our results are summarized in Table \[vla\_obs\]. We obtained limits on the linear polarization on 2005 13 July of 0.1% at all frequencies, but did not obtain accurate limits on the circular polarization of the source. Millimeter Radio: PdBI {#subsec:pdbi} ---------------------- The continuum radiation from M81\* at wavelengths of $\sim$3mm and $\sim$1mm was observed with the IRAM Plateau de Bure Interferometer (PdBI) on 2005 24 February, 14-15 July, and 19-20 July. The observation frequencies differ slightly between the individual epochs because they were fine-tuned in order to optimize phase stability depending on the weather conditions. A detailed description of the data, their reduction and calibration can be found in [@Schoedeletal2007]. The systematic absolute uncertainty of the flux calibration is 10-15% for the 3mm data, and 15-30% for the 1mm data (see Table 2 in Schödel et al.). Flux measurements were extracted from individual scans of 20min duration and detailed light curves were obtained that show significant variability of M81\* during the observations. If we add the systematic uncertainties with the statistical uncertainties in quadrature,the resulting overall uncertainty is around 20% (3mm) and 30% (1mm). Because the detected flux variations are likely real,and fully consistent with what is observed in other wavebands, including the systematics in this manner would seem to overestimate the actual uncertainties. For this reason we believe that taking the standard deviation of the flux is more representative of the total uncertainty in the context of this broadband system. For this reason we use the average fluxes and their standard deviations as obtained from the light curves presented in Schödel et al. (see their Table 3), which we list in Table \[tab:pdbi\]. The PdBI observed significant variability between the individual observing epochs, as well as intraday variability. The 3mm and 1 mm light curves from 2005 24 February — actually the best of the data obtained with the PdBI during the campaign — show a flux decrease with a significance of $>5\sigma$ over 5 hours that occurred at both wavelengths (see @Schoedeletal2007 for a detailed discussion). The lightcurves are presented in combination with those from the Submillimeter Array below in Figs. \[febpdbsma\] & \[1mmpdbsma\]. Submillimeter: SMA {#subsec:sma} ------------------ M81\* was observed at the Submillimeter Array[^1] (SMA; @HoMoranLo2004) on Mauna Kea on 2005 Feb 24, 2005 Jul 18 and 2005 Aug 14. Observations were also made on 2005 Jul 15, but the daytime phase stability was too poor to permit reliable calibration of the data. In all cases, 7 of the 8 SMA antennas were available. The observations on 2005 Feb 24 were made in good nighttime winter conditions, with optical depth towards zenith at 225GHz $\tau_{225}\sim$0.04 and 10% humidity. The summer observations suffered somewhat from afternoon atmospheric turbulence, and were made with $\tau_{225}\sim$0.05 and 40% humidity on 2005 Jul 18, and $\tau_{225}\sim$0.12 and 20% humidity on 2005 Aug 14. The SIS receivers were tuned to a center frequency of 345.796 GHz in the upper sideband for 2005 Feb 24, and 230.538 GHz in the upper sideband for 2005 Jul 18 and 2005 Aug 14. For the initial observations on 2005 Feb 24, one IF was configured with a higher spectral resolution to search for CO(3-2) at the systemic velocity of M81, but none was detected. This is consistent with the absence of CO(1-0) emission at the position of the core reported by [@Sakamotoetal2001]. For all observations, the full 4 GHz (2 GHz in each sideband separated by 10 GHz) were averaged to construct one continuum channel centered on 340.67 GHz and 225.42 GHz, respectively. The SMA data were calibrated using the MIR software package developed at Caltech and modified for the SMA. Gain calibration was performed using the nearby quasar 0958+655. Absolute flux calibration was performed using Callisto, and at least 2 of the following quasars– 0721+713, 0841+708, 0927+390, 1153+495– were observed hourly to ensure that any detected changes in the flux of M81\* were real, and not an artifact of the calibration. This is particularly important for the summer observations, which were made under poorer daytime conditions. The data were imaged using [difmap]{} to confirm that M81\* is unresolved at the 1$\farcs$5-3$\farcs$0 resolution of the SMA, following which the fluxes of both M81\* and the nearby quasars were determined by fitting a point source model to the data in the [${u,v}$]{} plane. The flux densities obtained are accurate to within 20%, based on the derived values for the quasars. As shown in Figs. \[smafeb\]–\[smaaug\], M81\* exhibited little to no variation on short timescales during the observations, although there may have been a brief dip in flux in the February data shortly preceding the apparent corresponding dip in the PdBI measurements shown in Fig. \[febpdbsma\]. This dip should be viewed with caution, given that the level of variation is not greater than the random variations seen in the calibrators, however the shape and timing are suggestive. More significant variations in flux were seen between the epochs, as shown in Fig. \[1mmpdbsma\]. The average flux on 2005 Feb 24 was 378.7$\pm$70.0 mJy at 340 GHz. For 2005 Jul 18 and 2005 Aug 14, we obtained 182.8$\pm$36.0 mJy and 91.5$\pm$15.3 mJy at 225 GHz, but it should be noted that the August data were obtained in substantially worse weather, with higher atmospheric opacity in addition to the normal summer daytime turbulence, and so the phase transfer from 0958+655, although only 4 degrees away, may not have been entirely successful, resulting in a lower flux density within a point source model. IR: Lick Observatory {#subsec:lick} -------------------- We observed M81$^$ with the infrared camera for adaptive optics at Lick (IRCAL) behind the laser-guide star adaptive optics (LGSAO) system on the Shane 3 m telescope on the nights of 2005 February 24 and 25 (see Tab. \[febtimes\] and Figs. \[campoview\] and \[febview\]). In the near-infrared, M81$^$ is an unresolved point source on top of a bright, extended background from the stars and gas in the nuclear regions. The high galactic background made use of the laser guide star necessary; attempts to use the bright nucleus as a natural guide star for AO correction failed. We cycled through observations in the $J$, $H$, and $K_{\rm s}$ bands, with more time spent on the $K_{\rm s}$ band where the correction is best. For each cycle a 5-point dither pattern was repeated twice. Each frame in the dither pattern was 50–90 s long in the $J$ band, 60 s long in the $H$ band, and 60 s long in the $K_s$ band. We created flats by taking the median of a series of images of the telescope dome with the lights turned on, and divided each frame by the flat. We then subtracted the sky background from each frame using a nearby, relatively blank field. The sky background was small compared to the bright extended emission from the galaxy. Co-added images were created by shifting and adding the images from each 5-point observing pattern. The shifts between the dithered images were empirically determined between frames using the centroid of a 2-dimensional Gaussian fitted to the emission from M81$^$ before summing the frames. Frames with an average $\sigma>11.2$ pixels (085) were excluded; this cutoff value was determined empirically by plotting the peak pixel value versus the average $\sigma$ for each frame. (This average $\sigma$ value is from a single Gaussian fit to the extended galaxy plus M81$^{}$.) Co-added images were then divided by an exposure map with the total exposure time per pixel to create fluxed images in units of counts sec$^{-1}$. Each night we obtained two images of a nearby star (GSC 04383–00224) in order to test the stability of the point-spread function (PSF), and these images revealed the PSF to be variable. For instance, in the $K_s$ band, the best PSF image was nearly diffraction-limited with a half-width half-maximum (HWHM) of 017, but the poorest on the same night had a HWHM of 025. No other point source was detected in the 20 field of view of M81$^$ that would allow us to track the shape of the PSF concurrent with each exposure, and we found that without empirical knowledge of the PSF we were unable to construct a self-consistent model of the light from the galaxy and the active nucleus. This prevented us from obtaining accurate point-source photometry of the nucleus, and thus constraining any variability. Therefore, we computed an upper limit to the $K_s$ band nuclear flux averaged over the course of the observations. To determine the upper limit, we measured the radial profile of the PSF calibration star, GSC 04383–00224, and used its 2MASS $K_{\rm s}$ magnitude to determine the photometric zeropoint for converting from counts to magnitudes using a large aperture (30 pixels = 228). At radii larger than 30 pixels, the background-subtracted, enclosed counts change by $<2\%$ while the noise increases. Given that M81$^$ lies on top of strong extended emission from the galaxy, we preferred to use a smaller aperture of 10 pixels (076) with the aperture correction determined from the calibration star; a 10-pixel radius circle encloses 79% of the counts within a 30-pixel aperture. The local background was estimated from the median in an annulus with inner radius of 30 pixels and outer radius of 40 pixels (304) and subtracted; this likely underestimates the true local background as the extended galaxy emission is also peaked at the position of M81$^{}$. The calibration star photometry, measured at three different airmasses (1.20, 1.32, and 1.36) was also used to calculate the local $K_{\rm s}$-band atmospheric extinction correction. Following this procedure, we found an upper limit value of 66.7 mJy in $K_{\rm s}$. This limit is displayed as the base of the arrow in Figure \[alldata\_nouls\]. X-rays: Chandra* {#subsec:chandra} ----------------- M81\* was observed by the Chandra X-ray observatory, with the High-Energy Transmission Grating Spectrometer [[HETGS]{}; @Canizaresetal2005] in place, on five separate occasions (see Tables 1–5). The [HETGS]{} consists of two sets of transmission gratings, the [High Energy Gratings]{} ([HEG]{}) covering the 0.8–10 keV bandpass with a spectral resolution of $\Delta \lambda = 0.012$Å FWHM, and the [Medium Energy Gratings]{} ([MEG]{}) covering the 0.4–8.0 keV bandpass with a spectral resolution of $\Delta \lambda = 0.023$Å FWHM. The angular resolution of Chandra, even with the insertion of the gratings, isolates X-ray emission from the central $< 1\arcsec$ around M81\. We do not utilize information from the $0^{\rm th}$ order (undispersed) spectrum, however, as the central image of the nucleus suffers from photon pileup [@Youngetal2007]. The Chandra* data were filtered for times of high background and spectra were extracted using the standard CIAO tools[^2], using v3.3 of the software and CALDB 3.2.2. Analyses were performed using ISIS version 1.4.7 [@Houck2002]. Our data preparation was identical to that described in @Youngetal2007. Specifically, for all X-ray analyses we separately combined the $\pm1^{\rm st}$ [HEG]{} spectra and the $\pm 1^{\rm st}$ order [MEG]{} spectra, and we utilized background files from narrow regions on either side of the respective gratings arms. As discussed in @Youngetal2007, we are confident that ${\mathrel{ \rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}}}90\%$ of the dispersed X-ray emission originates in an unresolved source in the nucleus of M81. Observational Results {#sec:newobs} ===================== The Broadband Spectrum of M81\* {#subsec:broad} ------------------------------- Figure \[alldata\_nouls\] shows the broadband spectrum comprised of all of our observations from the 2005 campaign. It is immediately clear that while there was some variability, particularly in the radio to mm, we did not see large variations in the X-ray or overall basic shape of the SED. The total average 7 GHz radio luminosity is $L_{\rm R}=8.2\times10^{36}$ erg s$^{-1}$ and the total average 2–10 keV X-ray luminosity is $L_{\rm X}=1.52\times10^{40}$ erg s$^{-1}$. The radio exhibited $\approx20$% variation about this average while the X-rays exhibited $\approx14$% variation. Therefore over the course of this half-year campaign, M81\* appeared to be more stable in both wavebands than has been reported in the past. Interestingly, however, the average $L_{\rm R}$ seen in our campaign is $\approx25$% lower than the average deduced from all prior campaigns, while our average $L_{\rm X}$ is about a factor of 5 less (see @Markoff2005). Either we have caught M81\* in a rather low, stable state, or previous observations (with notably larger fields of view) have included a large flux contribution from the surrounding diffuse medium. Fig. \[nulnuxrays\] shows the flux density for all X-ray observations, where we see that the largest change was a drop between July and August. The 2005 July 19 observations with the [HEG]{} showed a 0.8-7keV flux of $1.31\pm0.02\times10^{-11}~{\rm erg~cm^{-2}~s^{-1}}$— the highest value during our campaign— while the 2005 August 14 observations with the [HEG]{} dropped down to a value of $1.07\pm0.03\times10^{-11}~{\rm erg~cm^{-2}~s^{-1}}$. (See @Youngetal2007 for further discussions of the X-ray variability.) The radio through submm bands revealed significantly more variability in comparison to the X-ray band. Fig. \[radio\] shows the good data for all observations. The most reliable detection of significant intraday variability occurred at 3mm and 1mm in the PdBI observations on 24 February. A flux decrease of $\sim30$% with a significance of over $5\sigma$ was observed in both bands from 08h to 12h UT (see Fig. \[febpdbsma\] and @Schoedeletal2007). Such a time scale would suggest that the size of the emitting region is less than 20$r_{\rm g}$, if beaming were not involved. The expected beaming from our spectral fits discussed below is mild for the weak jet in this LLAGN; therefore, the mm variability still implies a relatively compact source. The SMA observations were successful in the 345 GHz band only on the first observing run in February. The resulting spectrum shows a suggestive upturn towards the submm (Fig. \[radio\], \[febfit\]). This steepening is similar to the spectral component seen in Sgr A\* that is referred to as the “submm bump”. In Sgr A\, this bump rapidly declines with decreasing wavelength towards the infrared (IR), and furthermore it varies simultaneously with the X-rays [@Eckartetal2004; @Eckartetal2006a]. This simultaneous variation suggests that for Sgr A\ the IR and X-ray emission both originate in regions close to the SMBH. In contrast, the variability detected in M81\* is both less pronounced and not as clearly correlated between the submm and X-ray. That being said, however, the low-frequency data are suggestive of waves of variability, with decreasing amplitudes, that appear to be moving from shorter to longer wavelengths over the half year of monitoring. Specifically, note that the peak at just under 10 GHz in 2005 February is gone by 2005 August, and does not appear to be associated with any lower frequency features. (The “peak” in the 2005 July 12-16 at just above 1 GHz is a mismatch between the GMRT and the VLA, which may be due to real intraday variability as there was $\sim 29$ hours between the two observations). One question that arises is whether or not the doubling of flux density at 43GHz that occurs between 2005 July 13 and 19 is associated with the 2005 July 14 peak at 80.5GHz moving to lower frequency over the ensuing several days. Likewise, is the 43GHz peaked bump on 19 July then seen moved to lower frequency (10–20GHz) and amplitude in the 2005 August 14 data? Such a “wave of variability” is consistent with expectations of adiabatic expansion. Similarly, the February observation with contiguous PdBI and SMA observations (Fig. \[febpdbsma\]) shows a clear dependence of variability amplitude on frequency, which is another expectation of adiabatic expansion. Comparison to Sgr A\* --------------------- Our new simultaneous data reinforce the similarities previously reported between M81\* and Sgr A\. In Fig. \[m81sgra\] we show the total M81\ campaign spectrum, including the non-simultaneous IR/O/UV data upper limits discussed in §\[subsec:uv\], and overplot the simultaneous Sgr A\* spectrum from [@Anetal2005] along with the various [Chandra]{} X-ray spectra [@Baganoffetal2001; @Baganoff2003; @Baganoffetal2003]. The Sgr A\* data have been scaled downward by a factor of $\approx10$ in order to ease visual comparison. M81\* and Sgr A\* show remarkable similarities in the radio frequencies. Even though for M81\, within the GMRT error bars, we see no direct evidence for the free-free absorption turnover observed in Sgr A\, the slightly inverted spectra in both sources are still classic indicators of synchrotron self-absorption effects in the jet core emission. The M81\* radio spectrum lies below that of Sgr A\* at higher frequencies, and thus is less inverted than that of Sgr A\. This is expected for M81\ in the context of self-absorbed, accelerating jet models as its lower inclination angle compared to Sgr A\* would result in a less inverted spectrum. The jets in Sgr A\* are presumed to be at a high inclination angle with respect to our line of sight (see the supporting evidence from the models presented by @MarkoffBowerFalcke2007 [@Meyeretal2007]). The spectra for M81\* and Sgr A\* both seem to peak near the submm range and then subsequently drop off towards the IR. Both sources also are consistent with sharing the same radio-IR power-law slope, if the M81\* IR upper-limits are indicative of the underlying intrinsic spectrum. M81\* and Sgr A\, however, clearly diverge from one another in the X-ray regime. On a relative scale, the M81\ X-ray spectrum always lies above that for Sgr A\, with the latter source in its rare, bright flare states still falling short of the X-ray/radio flux ratio of M81\. If M81\* indeed has a “submm bump” as in Sgr A\, this would be only the second source where such a feature is observed. In Sgr A\ this component has been associated alternatively with the base of a compact jet [@FalckeMarkoff2000; @Markoffetal2001], the regions of the accretion flow closest to the black hole [@Narayanetal1998; @YuanQuataertNarayan2003], or with an inner Keplerian disk [@LiuMelia2001]. Regardless of the exact geometry, in Sgr A\* this component now has been definitively associated with X-ray flares [@Eckartetal2004; @Eckartetal2006b], and thus is of significant interest for understanding high-energy processes within tens of $r_g$ of the Sgr A\* black hole. For the LLAGN class as a whole, it is important to understand if the flaring and coupling of the submm bump/X-ray emission in Sgr A\* is typical. The X-rays in M81\* have yet to show any flares of significant amplitude, while we have some evidence of variability in the submm. Thus in contrast to Sgr A\, the X-ray emission in M81\ may not be due to the same physical component that yields the variable submm. As further described below, this has implications for the theoretical modeling of M81\. These results for the submm band in particular are still tentative; more monitoring of this band as well as the millimeter range will determine whether an intrinsic submm bump is indicated. Massive amounts of dust are present in the region observed in these bands and could be contributing some level of flux. On the other hand, subsequent brief observations at 345 GHz using the SMA in 2005 (A. Peck, priv. comm.) show significant variability, yielding a flux density ranging from 300 mJy to 900 mJy over a period of 3 months from April to June 2006, consistent with the measurements made at the SMA in Feb 2005. These values are accurate to within about 20%, and thus the variability is clearly significant. These further data also support a rise in flux above the millimeter band. It is possible that the submm bump is a completely transient feature associated with flaring/ejecta, as has been suggested for Sgr A\. Confirming this strong variability as well as a detection of linear polarization would place stringent limits on any dust contribution, as well as the spectrum. With the advent of polarimetry at 345 GHz this year on the SMA, we have (with D. Marrone, PI) successfully proposed to search for linear polarization in M81\, which will hopefully resolve this issue in the near future. Jet-dominated spectral models {#sec:model} ============================= Model description {#subsec:jet_model} ----------------- One of the basic tenets of General Relativity is that black holes are essentially self-similar with regards to mass. An obvious possible consequence of this is a predictable scaling with mass of the accretion physics around black holes. If such a scaling exists, it would imply that the same underlying physical model could explain the continuum emission for stellar accreting black holes in XRBs as well as SMBHs. Likely complicating this simple picture are differences that could be introduced by accretion off one star compared to accretion off the winds of entire clusters of stars. Furthermore XRBs undergo state changes, which have yet to be clearly associated with the various AGN classes as would be naively expected if some of the AGN classes correspond to much longer lived state transitions. So far the best case for such a mapping, referred to as the “fundamental plane of black hole accretion”, is the correspondence in characteristics between the sub-Eddington low/hard state of XRBs and LLAGN [@MerloniHeinzDiMatteo2003; @FalckeKoerdingMarkoff2004; @KoerdingFalckeCorbel2006]. The hard state of XRBs is characterized by weak accretion disk emission, and the presence of steady, compact jets [e.g. @Fender2006]. These jets seem to increasingly dominate the power output of the system even as the total accretion rate decreases [@FenderGalloJonker2003]. Lending theoretical support to the idea of a fundamental plane, models of LLAGN as being dominated by outflows [e.g. @FalckeMarkoff2000; @Markoffetal2001; @Yuanetal2002] also have been quite successful at explaining the broad continuum properties of hard state XRBs [e.g @MarkoffFalckeFender2001; @Markoffetal2003]. More recently, these outflow-dominated models have been refined further to account for strong geometrical constraints, for example, such as signatures of reflection off of cool material. The outflow models were developed to take advantage of data from recent simultaneous, multi-wavelength monitoring campaigns. Statistical fitting of the simultaneous broadband continua of several hard state XRBs, including finer features such as fluorescent Fe lines in their X-ray bands, have produced consistent trends among the physical parameters determined for this class of sources [@MarkoffNowakWilms2005; @Migliarietal2007; @Galloetal2007; Maitra et al., in prep.]. In order to further test the principle of black hole accretion scaling with mass, and to enable a stronger physical comparison among M81\, Sgr A\, and their possible stellar-mass equivalents, we apply these same hard state XRB outflow-dominated models to the M81\ spectra from our campaign. The only significant difference from the model application to XRBs is that the input mass for M81\* is $\sim10^6$ times larger. A more detailed description of this model can be found in the appendix of [@MarkoffNowakWilms2005], and references therein; here we provide a brief summary. The model was designed to test the premise that the magnetized, outflowing compact accretion disk coronae such as described in [e.g. @Beloborodov1999; @MalzacBeloborodovPoutanen2001; @MerloniFabian2002] can comprise the footpoints of collimated jets. Although magnetic fields are assumed to play a global role, the dynamics of the model do not include magnetic acceleration. There are two primary reasons for this choice. First, because the exact role of the fields in the dynamics is still under debate, including a magnetic pressure term explicitly would add more assumptions (and thus free parameters) to the model. Secondly, the observations suggest that the steady jets in the weakly accreting black hole state are less accelerated than the transient jets occurring near the Eddington limit [e.g. @Fender2006]. The observations so far can thus be well-explained by a gas pressure dominated model. In addition, for M81\* and Sgr A\* in particular the inverted radio spectrum is also suggestive of adiabatic cooling in a jet with a broader opening angle ($\ge 20^\circ$), rather than a narrow jet as would be expected for magnetic collimation. Aside from these points, there are four basic assumptions in the model: 1) the total power in the jets scales with the total accretion power at the innermost part of the accretion disk, $\dot{M}c^2$, 2) the jets are freely expanding and only weakly accelerated via their own internal pressure gradients, 3) the jets contain cold protons which carry most of the kinetic energy, while leptons dominate the radiation, and 4) some or all of the originally thermally distributed particles are accelerated into a power-law which is maintained along the rest of the jet via distributed acceleration. The base of the jet consists of a small nozzle of constant radius where no bulk acceleration occurs. The nozzle absorbs our uncertainties about the exact nature of the relationship between the accretion flow and the jet, and fixes the initial value of most parameters. Beyond the nozzle, the jet expands laterally with its initial proper sound speed for a relativistic electron/proton plasma, $\gamma_{\rm s}\beta_{\rm s}c\sim0.4c$. The plasma is weakly accelerated by the resulting longitudinal pressure gradient force, allowing an exact solution for the velocity profile via the Euler equation [see @Falcke1996]. This results in a roughly logarithmic dependence of velocity upon distance from the nozzle, $z$. The velocity eventually saturates at large distances at Lorentz factors of $\Gamma_{\rm j}\ga$2-3. The size of the base of the jet, $r_0$, is a free parameter and once fixed determines the radius as a function of distance along the jet, $r(z)$. There is no radial dependence in this model, and the opening angle is thus fixed by the velocity profile as a function of distance along the jets. The model is most sensitive to the fitted parameter $N_{\rm j}$, which acts as a normalization. It dictates the power initially divided between the particles and magnetic field at the base of the jet, and is expressed in terms of a fraction of $L_{\rm Edd}$. Once $N_{\rm j}$ and $r_0$ are specified and conservation is assumed, the macroscopic physical parameters along the jet are determined. We assume that the jet power is roughly shared between the internal and external pressures. The radiating particles enter the base of the jet where the bulk velocities are lowest, with a quasi-thermal distribution of temperature $T_e$ (a fitted parameter). A significant fraction (here fixed at 75% based on results from the XRB fits mentioned above) of the particles are accelerated into a power-law tail at a location $z_{acc}$ (also a fitted parameter). The maximum energy the electrons can achieve is calculated explicitly by setting the local cooling and escape rates to the acceleration rate. Here we assume that the particles are accelerated via the most conservative case of diffusive shock acceleration with the magnetic field parallel to the shock normal [see @Jokipii1987]. The acceleration rate depends on the plasma parameters $u_{\rm sh}$ and $\xi$, the relative velocity of the shock to the plasma in the shock frame, and the ratio between the mean free path for scattering and the gyroradius, respectively. These terms enter into the acceleration rate as $\epsilon_{\rm sc}\equiv\left(u_{\rm sh}/c\right)^2/\xi$. This is a free parameter in our fits, and since $\xi$ is thought to lie in the range $\sim 10$–$1000$, it provides a consistency check on the plasma velocities. The particles in the jet radiatively cool via adiabatic expansion, synchrotron processes, and inverse Compton upscattering; however, adiabatic expansion is assumed to dominate the observed effects of cooling. While thermal photons from the accretion disk (via a multicolor blackbody model) are included as seed photons in the inverse Compton calculations, beaming reduces their energy density compared to the rest frame synchrotron photons (synchrotron self-Compton; SSC), except at the very base of the jet where they can be of the same order. The temperature $T_{\rm d}$ and total luminosity $L_{\rm d}$ of a thin accretion disk are also included as fitted parameters, but as this component is not an integral part of the outflow model, and not well constrained by this data set, we include these parameters mainly for a consistency check. Aside from those mentioned above, the other main fitted parameters are the ratio of nozzle length to its radius, $h_0$, and the equipartition parameter between the magnetic field and the radiating (lepton) particle energy densities, $k$. Physical parameters such as the mass, inclination angle and distance are fixed at the observed values. Table \[tab:params\] summarizes all the fitted jet parameters. This model has provided a very good statistical description of the broadband radio through X-ray data from several XRBs. The data from this campaign observing M81\* allows us to test the premise that sub-Eddington accretion in LLAGN can be described by the same physics as sub-Eddington accretion in low/hard state XRBs, even though the latter sources are over six orders of magnitude less massive than the former. This is the first time that our jet model has been applied to a “canonical” LLAGN, which thus allows for an interesting comparison to Sgr A\. Fitting methodology {#sec:fitting} ------------------- We explore two possible jet-model scenarios with our M81\ data. The first scenario is an exact analog to the XRB fits, where the particles are presumed to enter the nozzle in a quasi-thermal distribution that is later accelerated in the jet. This scenario was also explored for Sgr A\, where we determined that the acceleration must be either extremely weak or lacking, or that it must occur extremely far out in the jets, in order to prevent the presence of a predicted power-law violating the observed IR flux values [@MarkoffBowerFalcke2007]. The second scenario assumes that the particles enter the jets in a full power-law distribution, as may occur if jet formation is associated with a shock in the inner regions of the accretion flow [@Koideetal2000]. This scenario was also explored for Sgr A\, where in order to be consistent with the quiescent spectrum the acceleration must be sufficiently weak to result in an electron particle distribution power-law no harder than $p\sim3.8$ [@MarkoffBowerFalcke2007]. In exploring these models, fits to the individual ObsIds for the [Chandra]{} data were simultaneously carried out over the 0.5–7keV band for the [MEG]{} data and the 0.7–8keV band for the [HEG]{} data. Each dataset was grouped to have a minimum signal-to-noise of 5 and a minimum of four wavelength channels per bin. Because there are hundreds of X-ray data points compared to the radio/submm data, we initially set the radio/submm data errors to $<1\%$ to “weight” the significance of the sparser lower-frequency data in the fitting procedure. But because the jet model produces a predominantly smooth radio spectrum (i.e., it does not account for the waves of variability likely moving through the radio spectrum), and because the radio error bars are for the most part small compared to the observation-to-observation variability, in our final joint radio/X-ray fits we add 20% error bars in quadrature to the statistical error bars for the radio data. In this manner we decrease the likelihood that the jet models fall into local minima, and ensure that they instead more fairly represent the average radio properties. We also present a series of fits wherein we simultaneously consider the radio and X-ray data from the entire campaign. We again add 20% error bars in quadrature to the statistical error bars of the radio data. For the [Chandra]{} data we use the [ISIS]{} [combine\_datasets]{} function[^3] to combine the [HEG]{} data into a single spectrum and to combine the [MEG]{} data into a single spectrum. For each of these spectra we further group the data such that there is a minimum of 100 counts and a minimum of 32 energy channels per bin. The addition of 20% error bars to the radio and optical data acts to subsume both instrinsic variability, and allows for any systematic calibration differences among the different detectors. However, one might worry that the radio and optical data would then apply little statistical leverage on the fits. In fact, owing to the 2.5 to 9 orders of magnitude energy differences between the X-ray and UV/radio data, their effect far outweighs the simple contribution calculated to the overall $\chi^2$. For example, if one pivots a power-law with $\Delta \Gamma = 0.01$ at 1keV, the change in slope would yield well over a 20% difference in the radio regime (i.e., greater than our added error bars) while giving only a 2% difference at 8keV (i.e., substantially less than the X-ray data bars). The jet models are thus predominantly constrained by the radio data via this effect, rather than their contribution to the $\chi^2$ statistic. In order to illustrate this explicitly, in Fig. \[radioresids\] we present the radio residuals for a fit performed solely in the X-ray regime, using jet model parameters somewhat different than the typical (based on fitting X-ray binaries) values discussed here. The model provides an extremely good fit to the X-ray data, with a change of $\Delta \chi^2 = 1.2$ compared to our best fit from the broadband spectrum. This model, however, fails completely in the optical and radio regime, even with our expanded radio error bars. Thus the simultaneous radio/submm data especially are crucial for excluding large regions of “reasonable” parameter space for the jet models. The X-ray data alone simply cannot constrain jet physics. The resulting parameter differences between our canonical fits and the two fits utilizing non-simultaneous upper limits in the infra-red and optical, further emphasizes that broadening the multi-wavelength coverage is clearly very desirable. The jet model fits do present a number of “local minima”, which are also evidenced in the error bars for some parameter fits. Occasionally we find large parameter error bars– again an indication of inherent degeneracies in this complex theoretical model that we expect would be reduced with broader wavelength coverage. We also occasionally find parameter error bars that are unusually small. This is sometimes attributable to the fact that the radio data points do not represent a smoothed, averaged behavior, whereas the theoretical model does represent such an idealization to some extent. If a fit fortuitously passes almost exactly through a few radio points, the fits occasionally become very localized to that fit local minima. Such a local minimum might be slightly removed from a separate local minimum associated with a different subset of the radio points. The added 20% radio error bars tends to reduce, but not entirely eliminate, this behavior of the fits. Ultimately a time-dependent model is required to completely account for these effects. For the X-ray band specifically, @Youngetal2007 fit a series of emission and absorption lines to the combined M81\* [Chandra]{} spectra. To account for the presence of these features in the spectra, for all of our fits we added the complete set of lines from @Youngetal2007, but with there wavelengths frozen to those found in that work, and their widths frozen to $10^{-4}$eV. Only the line amplitudes were allowed to vary in the fits. Results and interpretation {#sec:results} -------------------------- Figs. \[febfit\] & \[febfit\_xray\] present our best fits of each type of model to the most comprehensive single observation in February, while Figures \[mongo\] & \[mongo\_x\] show the fits to all campaign data combined. Non-simultaneous data points from HST [@Maozetal2005] have, however, been included with error bars increased to 20% in order to guide the fits in the observational gap that otherwise extends over six decades of photon frequency. Figs. \[mongo\_ul\] & \[mongo\_ul\_x\] present additional fits which also include the upper limits from the simultaneous Lick Observatory observations, as well as several non-simultaneous detections taken from the literature (but here used as upper limits since they likely include non-nuclear emission) for ISO, MIRLIN, HST and Spitzer [@Grossanetal2001; @Gordonetal2004; @Murphyetal2006]. All of these upper limit data were treated as measured points at their upper values, with additional 20% error bars applied. Given the similarity in the shape of the submm through optical/UV data from M81\* compared to what is seen in Sgr A\* and NGC 4258, it seemed worthwhile investigating scenarios that included these limits as estimates of the true, underlying spectra. The values for the fitted parameters with 90% confidence errors for all fits, including the individual data set fits not shown here, are presented in Tables \[tab:PL\_FINAL\_a\]– \[tab:MXSW\_FINAL\_b\]. It is clear that the model provides a very good description of the individual observations, and to a lesser extent the combined data set. The latter data set obviously includes variations, especially for the non-X-ray data, for which the steady-state model cannot account. What is most striking about these fits is that the overall values for the fitted parameters fall into remarkably similar ranges compared to the parameter values for hard-state XRBs and Sgr A\. We first consider the model with an initial power-law electron distribution. The power-law electron distribution has previously been applied to spectra from Sgr A\, but not to any XRB spectra. Compared to the the range of Sgr A\* fits explored in [@MarkoffBowerFalcke2007], the fits to the M81\* spectra a similarly find a very compact jet base and an electron temperature of $\approx 10^{11}$ K. There are, however, significant differences which are likely influenced by the lower X-ray/radio flux ratio in Sgr A\, as well as the fact that Sgr A\, as a fraction of Eddington luminosity, is over four orders of magnitude lower compared to M81\. Whereas M81\ shows indications of a weak accretion disk, no such component has been observed in Sgr A\. Whether assumed to occur near the base, as for the power-law (PL) fits, or further out in the jets, as in the Maxwellian (MXW) fits, the M81\ spectra prefer solutions with more efficient particle acceleration. That is, the electron power-law index $p=2.4$–2.8, whereas for Sgr A\* $p$ is always $>3$. The M81\* particle distribution also shows more cooling than in Sgr A\, as indicated by the $\sim 10$ times smaller $\gamma_{\rm max}$ in M81\. These differences go far to explain why the weak jets of M81\* have been easier to discern than those in Sgr A\. Beyond the fact that we are not viewing M81\ through the Galactic plane scattering screen, more particle acceleration leads to more optically thin jet emission, which in turn increases the size of the photosphere at a given frequency. Thus M81\’s jets, with radiating particles accelerated into a canonical power-law, would be predicted to have a much larger photosphere at the same frequency than Sgr A\’s. The equipartition parameter, $k$, in M81\* is very similar to the values found in XRBs, favoring mild magnetic domination of the internal energy. Sgr A\* on the other hand seems to be the only source we have studied so far which favors a stronger magnetic energy density ($k\sim15$), bucking the trend seen here in M81\* and in XRBs for a correlation between total jet power and equipartition parameter. Because we have no information about the non-thermal X-ray spectrum of Sgr A\* in quiescence, however, this parameter may not be well constrained, but it is something to keep in mind for future explorations. The fitted jet nozzle scale height in M81\* is not well-constrained by these fits. However, the fits that include upper limits as estimates of the IR through optical spectra clearly select out a more elongated base/corona. Thus to strongly constrain the value of the nozzle scale height, further simultaneous millimeter through IR/optical observations are necessary. Looking at the Maxwellian model fits, which have been applied to both Sgr A\* and hard state XRBs, we see the same trends when comparing to the Sgr A\* fits. For the MXW cases, we can also compare acceleration efficiency by the need for a power-law tail in the best-fit electron spectrum. Both M81\* and XRBs are consistent with a high rate, here frozen to 75%, as compared to Sgr A\* where only a small fraction of particles are accelerated. In Sgr A\, the contribution of the tail also is minimized by acceleration occurring at quite large distances from the jet base. In contrast, the acceleration in M81\ occurs at the same location as seen in XRBs ($\sim 10-100\, r_g$). In fact, even though we are probing a fairly low fractional Eddington luminosity compared to the previously fit XRBs, nearly all fitted parameters for the MXW fits to M81\ fall into the exact same ranges as those seen in hard state XRBs. The only major parameter difference is the electron temperature, which is a factor of $\sim 2-3$ lower in XRBs. These results thus provide very strong support for the mass-scaling of accretion physics, at least for weakly accreting black holes. By comparing the fits to individual data sets, we can look for meaningful correlations in the parameters on timescales impossible to probe in XRBs, as weeks to months in M81\ would correspond roughly to sub-second variations in a typical XRB. Overall there are more obvious correlations among the MXW fit parameters than compared to the PL fit parameters. This fact leads us to consider the MXW fits as more likely probing more real physical effects. However, simultaneous constraints in the IR through UV are necessary to conclusively break this degeneracy. A selection of the strongest correlations are shown in Figs. \[eddratpars\] & \[equippars\]. Given the lack of simultaneous constraints on the IR through UV range, it is somewhat surprising to see some trends linking intrinsic jet parameters to the accretion disk parameters. There is a clear anti-correlation between the disk flux and the energy input to the jet. This may in fact be similar to the observed anti-correlation between soft and hard X-ray fluxes seen in hard state XRBs, such as Cyg X-1 [see @Wilmsetal2006]. For M81\, the higher disk flux is driven by an increase in the fitted disk temperature, but given that we do not have simultaneous data directly in the spectral region where the disk spectrum is most prominent (see Fig. 14), we cannot in fact be sure that the correlation is not systematic. More interesting are the correlations detected between the equipartition parameter $k$ and other fitted parameters. Along with the total normalized power $N_{\rm jet}$, $k$ seems to be one of the most important parameters for the jet model. Because it dictates the distribution of energy between the radiating particles and the magnetic fields, it effects the synchrotron/inverse Compton ratio, and can in some sense compensate for losses in the particle energy density due to, for instance, temperature decreases. As the temperature goes down, more energy is needed is needed in the magnetic energy to maintain the same synchrotron emission, thus requiring an increase in $k$. Similarly, a wider jet base results in a lower particle number density and thus $k$ must again be increased to maintain the same radiative fluxes. The most interesting correlation, however, is that between the jet power and the equipartition. We seem to be seeing a trend towards stronger magnetic powers relative to the radiating particles with increasing total jet power. With the exception of Sgr A\ (but see above for a reason to possibly discount this source in this regard), this trend is also seen in individual fits to hard state XRBs [@MarkoffNowakWilms2005; @Migliarietal2007; @Galloetal2007; Maitra et al. in prep.]. One possible interpretation of this is that the magnetic fields are more efficiently generated at higher accretion rates. After increasingly building up, an explosive release of this energy at some critical accretion rate may be responsible for driving the transient, and much more relativistic, ejecta seen in transition to the XRB hard state. Conclusions {#sec:discuss} =========== Our simultaneous broadband campaign on the LLAGN M81\* has generated five individual spectra, spread over 6 months, as well as a combined spectrum, that can be readily compared to other LLAGN, such as Sgr A\, as well as hard-state, weakly accreting X-ray binaries. These data definitively confirm for the first time many species of line emission from the accretion flow of an LLAGN [@Youngetal2007], and we have confirmed previous detections of variability across the entire spectrum. The radio through submm in particular shows significant levels of both intraday and longer term variability. We also see indications for adiabatically decaying “flares” moving out along the jets. The simultaneous, broadband nature of these data has allowed us to fit the spectra with an outflow-dominated model developed for hard state XRBs, that also has been used to understand our extremely subluminous Galactic nucleus, Sgr A\. We find several interesting results based on these spectral fits. Compared to Sgr A\, M81\ is not only much more luminous, but also more of its accretion energy is funneled into accelerating and maintaining power-law distributions of the radiating particles. Otherwise, the geometry and particle thermal/minimum temperatures seem to be very consistent between these two sources. We are currently conducting a campaign of similar scope for the LLAGN NGC 4258 (Nowak et al., Reynolds et al., in prep.), which will provide a third object to this “sample” of extensive, simultaneous multi-wavelength datasets for LLAGN. The most remarkable result of our modeling is the discovery that M81\* seems to behave just like a hard state XRB, despite it being over six orders of magnitude more massive, and accreting at a fractional Eddington luminosity somewhat lower than for the jet model fits to XRBs. The best fit parameters all fall into the same range as those found for XRBs, with the exception of the particle initial temperature/minimum energy, and the nozzle length. We do not consider the latter significant, however, since it cannot be well constrained by this data set. The temperature difference, on the other hand, could be more notable since it is shared with fits from other LLAGN [@Markoffetal2001; @Yuanetal2002]. If particles in LLAGN enter the jets with a factor of 2–3 higher temperature than XRBs, this could be due to the lower cooling rates for the comparatively less compact LLAGN jets given the same power and size in mass-scaling units of $L_{\rm Edd}$ and $r_g$. Our results provide an independent confirmation of the mass-scaling accretion physics suggested by the fundamental plane of black hole accretion described in § \[sec:model\]. Finally, given the $>20\%$ variations seen in the radio within the campaign, as well as in comparison with prior observations, and the factor of $\sim5$ times smaller average X-ray flux, it is clear that the differences between average/non-simultaneous measurements and simultaneous multiwavelength observations may be quite important for LLAGN, and perhaps AGN as well. In particular these variations becomes relevant for the fundamental plane of black hole accretion. Sub-Eddington accreting black holes show a correlation with slope $\approx0.6$–0.7 between the logarithms of the radio and X-ray luminosities, with an effective mass-dependent normalization. The exact values of the coefficients are important for placing stringent limits on the processes responsible for the emission. To date, the coefficients primarily have been determined from samples of AGN/LLAGN with non-simultaneous or averaged values for the radio and X-ray luminosities. Given the now confirmed significant broadband variability of M81\, taking average values will likely lead to incorrect determinations for the correlation coefficients. To quantify this possibility, we use the quasi-simultaneously measured radio and X-ray luminosities from this campaign and redo the fundamental plane statistical analysis of [@Markoff2005]. We find a difference of 8% in the radio/X-ray correlation slope, and a difference of 50% in the mass-dependence coefficient! Our results therefore strongly argue for some level of care in conclusions based on non-(quasi)simultaneous observations. We acknowledge the Guaranteed Time program of the [Chandra Observatory]{}, without which we would not have had an opportunity to pursue this campaign. We also thank the staff of all participating observatories, who made these observations possible. The GMRT is run by the National Center for Radio Astrophysics of the Tata Institute of Fundamental Research. M.A.N. is supported by NASA Grant SV3-73016. PC gratefully acknowledges support from the National Radio Astronomy Observatory (NRAO) through a Jansky Fellowship. NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.. [77]{} natexlab\#1[\#1]{} , T., [Goss]{}, W. M., [Zhao]{}, J.-H., [Hong]{}, X. Y., [Roy]{}, S., [Rao]{}, A. P., & [Shen]{}, Z.-Q. 2005, , 634, L49 , F. K. 2003, AAS/High Energy Astrophysics Division, 7, Baganoff, F. K., Bautz, M. W., Brandt, W. N., Chartas, G., Feigelson, E. D., Garmire, G. P., Maeda, Y., Morris, M., Ricker, G. R., Townsley, L. K., & Walter, F. 2001, Nature, 413, 45 Baganoff, F. K., Maeda, Y., Morris, M., Bautz, M. W., Brandt, W. N., & Burrows, D. N. 2003, , 591, 891 , A. M. 1999, , 510, L123 , M. F., [Bartel]{}, N., & [Rupen]{}, M. P. 2000, , 532, 895 , R. D., & [Begelman]{}, M. C. 1999, , 303, L1 , R. D., & [Königl]{}, A. 1979, , 232, 34 , G. A., [Wilson]{}, A. S., [Heckman]{}, T. M., & [Richstone]{}, D. O. 1996, , 111, 1901 , G. C., [Falcke]{}, H., [Sault]{}, R. J., & [Backer]{}, D. C. 2002, , 571, 843 , G. C., [Falcke]{}, H., [Wright]{}, M. C., & [Backer]{}, D. C. 2005, , 618, L29 , G. C., [Wright]{}, M. C. H., [Falcke]{}, H., & [Backer]{}, D. C. 2003, , 588, 331 , A., [Bower]{}, G. C., & [Falcke]{}, H. 2006, , 451, 845 , A., [Bower]{}, G. C., [Falcke]{}, H., & [Mellon]{}, R. R. 2001, , 560, L123 , C. R., [Davis]{}, J. E., [Dewey]{}, D., [Flanagan]{}, K. A., [Galton]{}, E. B., [Huenemoerder]{}, D. P., [Ishibashi]{}, K., [Markert]{}, T. H., [Marshall]{}, H. L., [McGuirk]{}, M., [Schattenburg]{}, M. L., [Schulz]{}, N. S., [Smith]{}, H. I., & [Wise]{}, M. 2005, , 117, 1144 , P., [Ray]{}, A., & [Bhatnagar]{}, S. 2004, , 612, 974 , R., [Becklin]{}, E. E., [Evans]{}, A. S., [Neugebauer]{}, G., [Scoville]{}, N. Z., [Matthews]{}, K., & [Ressler]{}, M. E. 2000, , 531, 756 , N., [Ford]{}, H., [Tsvetanov]{}, Z., & [Jacoby]{}, G. 2003, , 125, 1226 , C., [Gierli[ń]{}ski]{}, M., & [Kubota]{}, A. 2007, , 15, 1 , A., [Baganoff]{}, F. K., [Morris]{}, M., [Bautz]{}, M. W., [Brandt]{}, W. N., [Garmire]{}, G. P., [Genzel]{}, R., [Ott]{}, T., [Ricker]{}, G. R., [Straubmeier]{}, C., [Viehmann]{}, T., [Sch[ö]{}del]{}, R., [Bower]{}, G. C., & [Goldston]{}, J. E. 2004, , 427, 1 , A., [Baganoff]{}, F. K., [Sch[ö]{}del]{}, R., [Morris]{}, M., [Genzel]{}, R., [Bower]{}, G. C., [et al.]{} 2006, , 450, 535 , A., [Sch[ö]{}del]{}, R., [Meyer]{}, L., [Trippe]{}, S., [Ott]{}, T., & [Genzel]{}, R. 2006, , 455, 1 , H. 1996, , 464, L67 Falcke, H., Körding, E., & Markoff, S. 2004, , 414, 895 , H., & [Markoff]{}, S. 2000, , 362, 113 , R. 2006, [Jets from X-ray binaries]{} (Compact stellar X-ray sources), 381–419 , R. P., [Gallo]{}, E., & [Jonker]{}, P. G. 2003, , 343, L99 , W. L., [Hughes]{}, S. M., [Madore]{}, B. F., [Mould]{}, J. R., [Lee]{}, M. G., [Stetson]{}, P., [Kennicutt]{}, R. C., [Turner]{}, A., [Ferrarese]{}, L., [Ford]{}, H., [Graham]{}, J. A., [Hill]{}, R., [Hoessel]{}, J. G., [Huchra]{}, J., & [Illingworth]{}, G. D. 1994, , 427, 628 Gallo, E., Migliari, S., Markoff, S., Tomsick, J., Bailyn, C., Berta, S., Fender, R., & Miller-Jones, J. 2007, , in press (astro-ph/0707.0028) , K. D., [P[é]{}rez-Gonz[á]{}lez]{}, P. G., [Misselt]{}, K. A., [Murphy]{}, E. J., [et al.]{} 2004, , 154, 215 , B., [Gorjian]{}, V., [Werner]{}, M., & [Ressler]{}, M. 2001, , 563, 687 , T. M. 1980, , 87, 152 , L. C. 1999, , 516, 672 , L. C., [Filippenko]{}, A. V., & [Sargent]{}, W. L. W. 1996, , 462, 183 , L. C., [van Dyk]{}, S. D., [Pooley]{}, G. G., [Sramek]{}, R. A., & [Weiler]{}, K. W. 1999, , 118, 843 , P. T. P., [Moran]{}, J. M., & [Lo]{}, K. Y. 2004, , 616, L1 , J. C. 2002, in High Resolution X-ray Spectroscopy with XMM-Newton and Chandra , S., & [Wang]{}, Q. D. 2001, , 554, 202 , N., & [Makishima]{}, K. 2001, , 321, 767 , J. R. 1987, , 313, 842 , S., [Meier]{}, D. L., [Shibata]{}, K., & [Kudoh]{}, T. 2000, , 536, 668 , E., [Falcke]{}, H., & [Corbel]{}, S. 2006, , 456, 439 , V., [Fabbiano]{}, G., [Elvis]{}, M., [Nicastro]{}, F., [Kim]{}, D. W., & [Peres]{}, G. 2004, , 601, 831 , S., & [Melia]{}, F. 2001, , 561, L77 , J., [Beloborodov]{}, A. M., & [Poutanen]{}, J. 2001, , 326, 417 , D., [Nagar]{}, N. M., [Falcke]{}, H., & [Wilson]{}, A. S. 2005, , 625, 699 Markoff, S. 2005, , 618, L103 Markoff, S., Bower, G. C., & Falcke, H. 2007, , in press (astro-ph/0702637) , S., [Falcke]{}, H., & [Fender]{}, R. 2001, , 372, L25 , S., [Falcke]{}, H., [Yuan]{}, F., & [Biermann]{}, P. L. 2001, , 379, L13 , S., [Nowak]{}, M., [Corbel]{}, S., Fender, R., & Falcke, H. 2003, , 397, 645 , S., [Nowak]{}, M. A., & [Wilms]{}, J. 2005, , 635, 1203 , A., & [Fabian]{}, A. C. 2002, , 332, 165 , A., [Heinz]{}, S., & [di Matteo]{}, T. 2003, , 345, 1057 , F., & [Meyer-Hofmeister]{}, E. 1994, , 288, 175 Meyer, L., Schödel, R., Eckart, A., Duschl, W. J., Karas, V., & Dovciak, M. 2007, , [submitted]{} Migliari, S., Tomsick, J. A., Markoff, S., Kalemci, E., Bailyn, C., Buxton, M., Corbel, S., Fender, R. P., & Kaaret, P. 2007, , in press , E. J., [Braun]{}, R., [Helou]{}, G., [Armus]{}, L., [Kenney]{}, J. D. P., [Gordon]{}, K. D., [Bendo]{}, G. J., [Dale]{}, D. A., [Walter]{}, F., [Oosterloo]{}, T. A., [Kennicutt]{}, R. C., [Calzetti]{}, D., [Cannon]{}, J. M., [Draine]{}, B. T., [Engelbracht]{}, C. W., [Hollenbach]{}, D. J., [Jarrett]{}, T. H., [Kewley]{}, L. J., [Leitherer]{}, C., [Li]{}, A., [Meyer]{}, M. J., [Regan]{}, M. W., [Rieke]{}, G. H., [Rieke]{}, M. J., [Roussel]{}, H., [Sheth]{}, K., [Smith]{}, J. D. T., & [Thornley]{}, M. D. 2006, , 638, 157 , R., [Mahadevan]{}, R., [Grindlay]{}, J. E., [Popham]{}, R. G., & [Gammie]{}, C. 1998, , 492, 554 , R., & [Yi]{}, I. 1994, , 428, L13 , M. J., [Breeveld]{}, A. A., [Soria]{}, R., [Wu]{}, K., [Branduardi-Raymont]{}, G., [Mason]{}, K. O., [Starling]{}, R. L. C., & [Zane]{}, S. 2003, , 400, 145 , S., [Cappi]{}, M., [Bassani]{}, L., [Malaguti]{}, G., [Palumbo]{}, G. G. C., & [Persic]{}, M. 2000, , 353, 447 , E., & [Gruzinov]{}, A. 1999, , 520, 248 , R. A., & [McClintock]{}, J. E. 2006, , 44, 49 Rykoff, E. S., Miller, J. M., Steeghs, D., & Torres, M. A. P. 2007, , submitted (astro-ph/0703497) , K., [Fukuda]{}, H., [Wada]{}, K., & [Habe]{}, A. 2001, , 122, 1319 , S., [Dudik]{}, R. P., [O’Halloran]{}, B., & [Gliozzi]{}, M. 2005, , 633, 86 , R., [Krips]{}, M., [Markoff]{}, S., [Neri]{}, R., & [Eckart]{}, A. 2007, , 463, 551 , N. I., & [Sunyaev]{}, R. A. 1973, , 24, 337 , G., [Ananthakrishnan]{}, S., [Kapahi]{}, V. K., [Rao]{}, A. P., [Subrahmanya]{}, C. R., & [Kulkarni]{}, V. K. 1991, CURRENT SCIENCE V.60, NO.2/JAN25, P. 95, 1991, 60, 95 , H. 1966, , 211, 1131 , S. P., [Ashby]{}, M. L. N., [Barmby]{}, P., [Fazio]{}, G. G., [Pahre]{}, M., [Smith]{}, H. A., [Kennicutt]{}, R. C., [Calzetti]{}, D., [Dale]{}, D. A., [Draine]{}, B. T., [Regan]{}, M. W., [Malhotra]{}, S., [Thornley]{}, M. D., [Appleton]{}, P. N., [Frayer]{}, D., [Helou]{}, G., [Stolovy]{}, S., & [Storrie-Lombardi]{}, L. 2004, , 154, 222 , J., [Nowak]{}, M. A., [Pottschmidt]{}, K., [Pooley]{}, G. G., & [Fritz]{}, S. 2006, , 447, 245 Young, A. J., Nowak, M. A., Markoff, S., Marshall, H. L., & et al., C. C. R. 2007, , [submitted]{} , F., [Markoff]{}, S., [Falcke]{}, H., & [Biermann]{}, P. L. 2002, , 391, 139 , F., [Quataert]{}, E., & [Narayan]{}, R. 2003, , 598, 301 , F., [Bushouse]{}, H., [Dowell]{}, C. D., [Wardle]{}, M., [Roberts]{}, D., [Heinke]{}, C., [Bower]{}, G. C., [Vila-Vilar[ó]{}]{}, B., [Shapiro]{}, S., [Goldwurm]{}, A., & [B[é]{}langer]{}, G. 2006, , 644, 198 [lccc]{} GMRT & 235/610 MHz & 23 Feb 14:15 & 23 Feb 22:57\ GMRT & 1390 MHz & 24 Feb 13:56 & 24 Feb 16:23\ VLA & 1.4/8.4/22/43 GHz & 24 Feb 05:00 & 24 Feb 12:30\ PdBI & 115/230 GHz & 24 Feb 01:11 & 24 Feb 19:45\ SMA & 345 GHz & 24 Feb 04:44 & 24 Feb 17:20\ Lick & K/J/H band & 24 Feb 08:13 & 24 Feb 11:40\ Lick & H band & 25 Feb 09:28 & 25 Feb 09:49\ Chandra HETGS & 0.5–8 keV & 24 Feb 06:56 & 24 Feb 20:50 [lccc]{} GMRT & 235/610 MHz & 12 Jul 06:31 & 12 Jul 13:16\ GMRT & 1390 MHz & 15 Jul 04:05 & 15 Jul 08:11\ VLA & 1.4/8.4/22/43 GHz & 13 Jul 04:56 & 13 Jul 23:00\ VLBA & 8.4 GHz & 13 Jul 17:30 & 14 Jul 05:30\ PdBI & 80.5/241.4 GHz & 14 Jul 06:50 & 15 Jul 13:50\ SMA & 230 GHz & 15 Jul 21:25 & 16 Jul 04:38\ [Chandra HETGS]{} & 0.5–8 keV & 14 Jul 01:44 & 14 Jul 17:28\ [Chandra HETGS]{} & 0.5–8 keV & 14 Jul 19:25 & 15 Jul 13:20 [lccc]{} VLA & 1.4/8.4/22/43 GHz & 19 Jul 16:44 & 19 Jul 23:07\ PdBI & 86.2/218.2 GHz & 19 Jul 23:17 & 20 Jul 16:07\ SMA & 230 GHz & 18 Jul 21:25 & 19 Jul 03:25\ [Chandra HETGS]{} & 0.5–8 keV & 19 Jul 14:25 & 20 Jul 15:25 [lccc]{} VLA & 1.4/8.4/22/43 GHz & 14 Aug 13:19 & 14 Aug 19:15\ SMA & 230 GHz & 14 Aug 19:37 & 15 Aug 01:30\ [Chandra HETGS*]{} & 0.5–8 keV & 14 Aug 09:50 & 14 Aug 20:57 [lcccccc]{} 2005 Feb 24 & 1390 &27&75 & 5x3 & $80.1\pm2.1$ & 0.24\ 2005 Jul 15 & 1390 &26 & 80 & 5x3 & $114.8\pm1.1$ & 0.32\ 2005 Feb 23 & 610 &27 &80 & 8x7 & $67.4\pm1.3$ & 0.45\ 2005 Jul 12 & 610 &29 & 90 & 8x6 & $76.3\pm1.2$ & 0.42\ 2005 Jul 26 & 610 &27& 75 & 9x5 & $72.4\pm3.1$ & 0.49\ 2005 Feb 23 & 235 &30 & 55 & 18x13 & $93.5\pm14.9$ & 4.95\ 2005 Jul 12 & 235 &28& 70 & 19x12 & $61.8\pm20.9$ & 3.80\ [lccc]{} 2005 Jul 14 & VLA & 1.4& $9.64\pm0.56$\ 2005 Jul 14 & VLA & 8.4 & $2.43\pm0.29$\ 2005 Jul 15 & GMRT & 1.4 &$9.00\pm0.52$\ 2005 Aug 14 & VLA & 1.4 & $10.23\pm0.80$\ 2005 Aug 14 & VLA & 8.4 & $1.87\pm0.20$\ [lccc]{} 2005 Feb 24 & 43 & – & –\ 2005 Jul 13 & 43 & $66.5\pm8.5$ &\ 2005 Jul 19 & 43 & $143.4\pm3.1$ &\ 2005 Aug 14 & 43 & $82.1\pm2.4$ &\ 2005 Feb 24 & 22 &$101\pm20$ &\ 2005 Jul 13 & 22 & $86.4\pm5.8$ &\ 2005 Jul 19 & 22 & $133.6\pm0.6$ &\ 2005 Aug 14 & 22 & $109.6\pm1.4$ &\ 2005 Feb 24 & 8.4 &$176\pm20$ &\ 2005 Jul 13 & 8.4 &$89.7\pm2.6$ &\ 2005 Jul 19 & 8.4 &$105.1\pm0.4$ &\ 2005 Aug 14 & 8.4 &$112.5\pm0.1$ &\ 2005 Feb 24 & 1.4 & $75\pm10$ &\ 2005 Jul 13 & 1.4 &$91.9\pm0.3$ &\ 2005 Jul 19 & 1.4 &$92.6\pm0.9$ &\ 2005 Aug 14 & 1.4 &$89.5\pm0.6$ &\ [lccccc]{} 2005 Feb 24 & 115.3 & 6 & 90% & $88.0\pm11.7$ & –\ 2005 Jul 14 & 80.5 &5 & 50% & $241.2\pm33.8$ &\ 2005 Jul 19 & 86.2 &5 &98% & $118.7\pm11.4$ &\ 2005 Feb 24 & 230.5 & 6 & 90%&$85.6\pm17.8$ &\ 2005 Jul 14 & 241.4 &5 & 30%& $181.2\pm39.1$ &\ 2005 Jul 19 & 218.2 &5 & 90% & $74.8\pm13.3$ &\ [lccc]{} 2005 Feb 24 & 340.7 & 7 & $378.7\pm70.0$\ 2005 Jul 18 & 225.4 & 7 & $182.8\pm36.0$\ 2005 Aug 14 & 225.4 & 7 & $91.5\pm15.3$\ [ll]{}\ $N_{\rm jet}$ & Jet normalized power\ $r_0$ & Size of jet base\ $T_e$ & Temperature of particles entering base of jet\ $p$ & Energy index of electron power-law tail ($N_e\propto E_e^{-p}$)\ $k$ & Equipartition parameter: the ratio of energy density in the magnetic field\ & to energy density in the radiating particles\ $h_{\rm ratio}$ & Ratio of jet nozzle length to nozzle radius\ $L_{\rm d}$ & Accretion disk bolometric luminosity\ $T_d$ & Innermost accretion disk temperature\ \ \ $\gamma_{\rm max}$ & Maximum Lorentz factor of electron distribution\ \ \ $z_{\rm acc}$ & Location at which particles are first accelerated to produce powerlaw tail\ $\epsilon_{\rm sc}$ & Particle acceleration efficiency parameter; $\epsilon_{\rm sc} \equiv (u_{\rm sh}/c)^2 / \xi$ where $u_{\rm sh}$ is the relative\ & velocity of the shock to the plasma in the shock frame, and $\xi$ is the\ & ratio between the mean free path to the gyroradius. [cccccc]{} 6174 & & & & &\ 6346 & & & & &\ 6347 & & & & &\ 5601 & & & & &\ 5600 & & & & &\ ALL & & & & &\ ALL & & & & &\ [ccccccc]{} 6174 & & & & & 504.3/472 & 1.07\ 6346 & & & & & 529.4/536 & 0.99\ 6347 & & & & & 627.2/605 & 1.04\ 5601 & & & & & 818.4/782 & 1.05\ 5600 & & & & & 380.4/321 & 1.19\ ALL & & & & & 574.0/292 & 1.97\ ALL & & & & & 422.7/282 & 1.50\ [ccccccc]{} 6174 & & & & & &\ 6346 & & & & & &\ 6347 & & & & & &\ 5601 & & & & & &\ 5600 & & & & & &\ ALL & & & & & &\ ALL & & & & & &\ [cccccccc]{} 6174 & & & & & 497.0/471 & 1.06\ 6346 & & & & & 533.8/535 & 1.00\ 6347 & & & & & 608.4/604 & 1.01\ 5601 & & & & & 826.0/781 & 1.06\ 5600 & & & & & 322.8/320 & 1.01\ ALL & & & & & 576.1/291 & 1.98\ ALL & & & & & 379.4/281 & 1.35\ ![image](f8.eps){width="\textwidth"} ![image](f9.eps){width="\textwidth"} ![image](f10.eps){width="\textwidth"} ![image](f11.eps){width="\textwidth"} [^1]: The Submillimeter Array is a joint project between the Smithsonian Astrophysical Observatory and the Academia Sinica Institute of Astronomy and Astrophysics, and is funded by the Smithsonian Institution and the Academia Sinica. [^2]: <http://cxc.harvard.edu/ciao/> [^3]: For our purposes here, this is equivalent to adding the data via [ftools]{} functions outside of the fitting program; however, utilizing this function within [ISIS]{} allows more flexibility in modeling and plotting.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The most efficient modern optical communication is known as coherent communication and its standard quantum limit (SQL) is almost reachable with current technology. Though it has been predicted for a long time that this SQL could be overcome via quantum mechanically optimized receivers, such a performance has not been experimentally realized so far. Here we demonstrate the first unconditional evidence surpassing the SQL of coherent optical communication. We implement a quantum receiver with a simple linear optics configuration and achieve more than 90% of the total detection efficiency of the system. Such an efficient quantum receiver will provide a new way of extending the distance of amplification-free channels, as well as of realizing quantum information protocols based on coherent states and the loophole-free test of quantum mechanics.' author: - Kenji Tsujino - Daiji Fukuda - Go Fujii - Shuichiro Inoue - Mikio Fujiwara - Masahiro Takeoka - Masahide Sasaki title: Quantum receiver beyond the standard quantum limit of coherent optical communication --- [ ]{}[ ]{} [ ]{} [ ]{}[ ]{} [ ]{} [ ]{} Coherent communication systems achieve the best signal-to-noise ratio (SNR) in conventional optical communications based on laser light. It consists of coherent-state carriers with quadrature amplitude modulation (QAM) and a coherent receiver based on homodyne detections [@PMOCS]. QAM provides the largest signal distances in phase-space under the power constraint while a coherent receiver detects them at the standard quantum limited (SQL) sensitivity. The SQL in optical communication is naturally defined as the lowest average error probability obtainable by directly measuring the modulated physical observable of coherent states. For example, an ideal photon counter reaches the SQL for intensity modulation (IM) signals (often called the shot noise limit: SNL) while an ideal coherent receiver reaches the SQL for QAMs. However this is not the fundamental quantum limit of optical communication. One approach to overcome such a limit is to use eigenstates of the observables, such as photon number state for IM or (infinitely) squeezed state for QAM, as carriers [@Caves1994]. For noiseless channels, these non-classical carriers could completely circumvent the errors due to quantum noise and in principle realize an error-free communication. In practice, however, these states are extremely fragile to losses and easily turned out to be noisy mixed states, and thus will not work in realistic channels. Another approach is to keep using coherent states as carriers but optimize the measurement process. Coherent state is robust against the linear loss (which is inevitable in optical channels) and does not lose its coherence. Quantum detection theory has predicted that the minimum error bound is exponentially smaller than the SQL (Fig. \[fig:schematic\](a)) [@QDET] and its implementation schemes have also been proposed [@Kennedy1973; @Dolinar1973; @Sasaki1996; @Takeoka2006; @Takeoka2008]. However, though some proof-of-principle demonstrations have been reported [@Cook2007; @Wittmann2008; @Tsujino2010], no unconditional experimental evidence surpassing the SQL of coherent communication has been reported yet. In this paper, we demonstrate for the first time, a quantum receiver outperforming the SQL of coherent optical communication, i.e. the limit of current optical communications, in a lossy channel. We consider the simplest QAM signal set, binary phase-shift keyed (BPSK) coherent states $\{\|\alpha\rangle, \|$$-$$\alpha\rangle\}$ with equal prior probabilities. BPSK provides the largest signal distance within any binary signals under the average-power constraint. Our receiver discriminates these signals with the BER lower than that of the SQL. The physical process of a near-optimal quantum receiver was first suggested by Kennedy [@Kennedy1973] consisting of linear optics and photon counting. Soon after it was extended to the exactly optimal one by Dolinar [@Dolinar1973] (see also [@QDET]) via applying an ultrafast feedback process. A proof-of-principle of the Dolinar receiver was recently demonstrated [@Cook2007] where instead of the BPSK, intensity modulated (on-off keyed:OOK) coherent signals $\{\|0\rangle, \|\alpha\rangle\}$ were discriminated under the SNL. However, even if one could achieve the minimum error bound for an OOK signal discrimination, its error rate is still larger than the SQL for BPSK signals since the OOK is not an optimal modulation under the same power constraint (Fig. \[fig:schematic\](a)). Moreover, although in principle the feedback (or feedforward) based measurement can realize an arbitrary binary projection measurement [@Takeoka2006], it is still challenging to outperform the SQL for BPSK signals via the feedback approach since it requires a detector simultaneously fulfilling a very high detection efficiency and the operation speed faster enough than the optical pulse width. Instead of the feedback, the optimal detection with an optical nonlinear process was proposed [@Sasaki1996], but the required nonlinearity was unfortunately far from the current technology. Instead, we use a simpler receiver scheme without feedback such as Kennedy’s near-optimal receiver [@Kennedy1973]. The main obstacle to beat the SQL with the Kennedy receiver in practice is the fact that for weak signals, its attainable BER is comparable or even higher than that of the SQL [@Takeoka2008]. The receiver scheme employed here is based on the recent proposal [@Takeoka2008], which we call an optimal displacement receiver (ODR), solving the above problem and its operation principle was demonstrated experimentally [@Wittmann2008; @Tsujino2010] though their performances could not reach the ideal homodyne performance due to technical problems, mainly the low detection efficiency ($\lesssim 70$%) [@Wittmann2008; @Tsujino2010] and the low phase stability [@Tsujino2010]. The experiment demonstrated here employs the ODR scheme and achieves more than 90% of the total detection efficiency as well as enough phase stability, mode matching and low dark counts. To our knowledge, this is the highest total efficiency demonstrated in any photon-detection based quantum information (and single-photon level optical communication) protocols. Such a performance is realized by installing an efficient photon number resolving detector known as superconducting transition-edge sensor (TES) [@Lita2008; @Fukuda2011]), almost error-free linear optics, and the TES-based phase stabilization with quantum level signals. ![ (color online). (a) Bit error rates for the BPSK coherent signals at the quantum limit (blue solid), the standard quantum limit (SQL, red solid), detected by the optimal displacement receiver (ODR, black solid). As a reference, the blue dashed line is the quantum limit for an on-off keyed (OOK) signal. The vertical axis corresponds to the signal mean photon number reached to the receiver. (b) Schematic of a lossy optical channel and the quantum receiver based on the optimal displacement and photon counting. (c) Experimental setup. FBS: fiber beam splitter, AO: auxiliary oscillator, PRNG: pseudo random-number generator, TES: transition-edge sensor. PZT: piezo electric transducer, []{data-label="fig:schematic"}](scheme.eps){width="1\linewidth"} The SQL of the BPSK signal with the average power $\|\alpha\|^2$ is expressed as the BER of $P_{\rm SQL}=(1- {\rm erf}[\sqrt{2}\alpha])/2$ which is simply attained via a perfect homodyne detector. The optimal measurement achieving the quantum limit, on the other hand, is described by the projection onto the quantum superposition of signal states, $\|\pi_-\rangle = b_0 \|\alpha\rangle + b_1 \|$$-$$\alpha\rangle$ and its complement in the signal space $\|\pi_+\rangle$ [@QDET]. Here, $b_0 = -\sqrt{P_{\rm QL}/(1-e^{-4\alpha^2})}$ and $b_1 = \sqrt{(1-P_{\rm QL})/(1-e^{-4\alpha^2})}$, and $P_{\rm QL} = (1-\sqrt{1-e^{-4\alpha^2}})/2$ is the fundamental quantum limit of the BER. The superposition of coherent states is quite non-classical and is often regarded as optical “Schrödinger’s cat state” [@Yurke1986], which implies the implementation of such a projector is a non-trivial task. The BERs $P_{\rm SQL}$ and $P_{\rm QL}$ are compared in Fig. \[fig:schematic\](a) showing the high potential of the optimal quantum measurement. We implement such a measurement approximately by a simple quantum receiver [@Takeoka2008; @Wittmann2008]. Suppose that coherent signals from the sender are attenuated in a lossy channel to be $\|{\pm}\alpha\rangle$ and detected via the receiver. Our receiver consists of a linear displacement operation $\hat{D}(\beta) = \exp[\beta^* \hat{a} - \beta \hat{a}^\dagger]$ and a photon counting device announcing two outcomes, zero or non-zero photons. Depending on these outcomes, the signal state is projected onto $\|\omega_-\rangle = \hat{D}^\dagger (\beta) \|0\rangle = \|$$-$$\beta\rangle$ or its orthogonal space. Here $\beta$ is optimized such that $\|$$-$$\beta\rangle = e^{-\|\beta\|^2/2} (\|0\rangle - \beta\|1\rangle + \cdots)$ approximates the non-trivial superposition $\|\pi_-\rangle = e^{-\alpha^2/2} \{ (b_0-b_1) \|0\rangle + (b_0+b_1) \alpha \|1\rangle + \cdots \}$. In the ideal case, the optimal $\beta$ is given as the solution of the equality $\alpha = \beta \tanh (2\alpha\beta)$. This ODR outperforms the SQL for any $\|\alpha\|^2$ and nearly reaches the quantum limit as illustrated in Fig. \[fig:schematic\](a) [@Takeoka2008; @Wittmann2008]. Figure \[fig:schematic\](b) shows an experimental setup for the optical communication in a lossy channel with the ODR. Continuous wave light from an external-cavity laser diode (wavelength: 853 nm, linewidth: 300 kHz) is highly attenuated and modulated by an electro-optic modulator (EOM) to generate 20ns pulses with a 40 kHz repetition rate. The pulses are split into two paths for the signal and the auxiliary oscillator (AO) for the displacement operation. At the sender’s station, the signal amplitude control and the binary phase modulation are performed by two EOMs. The signal pulses are then propagated through a channel with the loss of ${-}7$ dB where the loss is introduced by coupling the signal beam into a single-mode fiber with a low efficiency. The displacement operation $\hat{D}(\beta)$ at the detection part is realized by interfering the signal pulse with the AO pulse through a highly transmissive fiber beam splitter (FBS). The AO for the displacement is prepared to be a coherent state $\|\beta/\sqrt{1-T}\rangle$ where $T$ is the transmittance of the FBS and chosen as $T>0.99$. The visibility at the FBS is measured to be $98.6 \pm 0.1\%$ which corresponds to the mode match factor of $\xi=0.993 \pm 0.001$. The output from the FBS is guided into a TES to detect photons. The superconducting TES enables us to resolve photon numbers with a very high detection efficiency. Its performance is characterized by the quantum efficiency and the energy resolution. Our TES is based on a titanium (Ti) superconductor and its quantum efficiency is $0.95\pm0.01$ at 853 nm. This is achieved by carefully designing the surface (anti-reflection) and backside (high-reflection) coatings and using a larger Ti device ($10 \times 10$ $\mu$m$^2$) which is coupled to the fiber with an almost unit efficiency [@Fukuda2011]. The energy resolution characterizes the photon number resolution and is degraded by the electrical noise on the readout voltage pulse derived from Johnson noise and phonon noise in TES. To increase the resolution, the noise in the output voltage waveform is reduced via a digital Wiener filter, which gives an energy resolution of 0.55 eV at 853 nm. The threshold discriminating the peaks of zero and non-zero photons is chosen in advance to minimize the incorrect guessing of photons due to the finite overlap of the energy distributions. The additional loss due to the overlap of the distributions was estimated to be 0.7%. We also directly observed the dark counts with the same threshold as $\nu = 0.003$/pulse. The relative path length of the signal and the AO modes is actively stabilized by the signal laser and the TES before measuring BERs. To circumvent additional losses stemming from the additional probe beam and optics, the phase locking loop is constructed by using the signal laser itself ($\|\alpha\|^2\approx2$) and photon counting via the TES with a digital feedback. At the phase locking step, The TES output is sent to a discriminator and the number of clicks are counted for each 10 ms for the feedback loop. The stability is estimated to be within $\pm 0.057$ rad. As a consequence, after including the optical losses before the TES (losses at the FBS and fiber splicings), the total detection efficiency in our setup amounts to $0.91\pm0.01$. ![ (color online). (a) Random signal modulation (bar) and raw measurement data of the ODR (plot is the TES output pulse height). (b) The output pulse height distribution for the signals $\|{-}\alpha\rangle$ (red) and $\|\alpha\rangle$ (blue). $\|\alpha\|^2=0.21$ and $\|\beta\|^2=0.59$. []{data-label="fig:RawData"}](result1.eps){width="1\linewidth"} ![ (color online). (a) The measured BERs for various displacements. $\|\alpha\|^2$ is fixed to be 0.21. (b) The measured BERs obtained by the optimal displacement receiver for various $\|\alpha\|^2$. Theoretical curves are given for the quantum limit (blue solid), the SQL (red solid), the SQL (homodyne) with $\eta=0.91$ (pink dashed), and the imperfect displacement receiver (gray dashed). The imperfections in the gray dashed lines are chosen as $\eta=0.91$, $\nu=0.003$, and $\xi=0.993$ (see the text for details). []{data-label="fig:BER"}](result2.eps){width="1\linewidth"} Figures \[fig:RawData\] and \[fig:BER\] show the main result of the BPSK signal discrimination via our quantum receiver. The phase of the signal is randomly prepared by a pseudo random number generator with equal probabilities while the amplitude of the displacement $\beta$ is kept in-phase with $\|\alpha\rangle$. Figure \[fig:RawData\](a) shows typical sequences of the signal modulation and the TES output. The dashed line indicates a threshold for the signal decision and the errors are highlighted by arrows. The TES output distribution with 10,000 trials is plotted in Fig. \[fig:RawData\](b) for $\|\alpha\|^2=0.21$ and $\|\beta\|^2=0.59$. The red and blue plots are the outputs for $\|{-}\alpha\rangle$ and $\|\alpha\rangle$, respectively. The asymmetric distribution clearly indicates the effect of the displacement since the original signals $\|\pm\alpha\rangle$ have the same photon number statistics. In Fig. \[fig:BER\](a), the $\beta$-dependence of BERs is shown for the signals with a mean photon number of $\|\alpha\|^2=0.21$. Each point is obtained by 10,000 measurements with error bars reflecting statistically estimated standard deviations of binomial distributions. For $\|\beta\|^2 > \|\alpha\|^2$, the probability of wrongly guessing $\|\alpha\rangle$ ($\|{-}\alpha\rangle$) is decreased (increased) by increasing $\|\beta\|^2$ and thus there is an optimal point minimizing the BER (the average errors) [@Tsujino2010]. The experimental BER shows the dependence on $\beta$ and clearly outperforms the SQL at the optimal $\beta$ (around 0.6 photons). Note that the data points do not employ any compensations of noises or detection losses. The experimental BERs make a good fit to the theoretical curve including imperfections of $\eta=0.91$, $\nu=0.003$, and $\xi=0.993$ (dashed line) [@Takeoka2008]. The experimental points are also compared to the SQL (homodyne detection) with the same detection efficiency ($\eta=0.91$), showing the superiority of our receiver scheme. In Fig. \[fig:BER\](b), experimentally observed BERs with optimal displacements are plotted for different signal photon numbers. Again, the performance of our receiver clearly outperforms the SQL ($\eta=1$) at $\|\alpha\|^2=0.21$, and is comparable or slightly better than the SQL for the signals with $\|\alpha\|^2 \le 0.4$. For higher signal photon numbers, the BERs deviate from the model calculation due to technical reasons, mainly the visibility degradation by the drift of polarizations in fibers. We should also note that the data shown here outperforms homodyne detection with the same efficiency ($\eta=0.91$) in a wide range of signal photon numbers. In summary, our result shows the first unconditional observation of the error rate surpassing the SQL in coherent communication, that is, exceeding the theoretical limit of current optical communication technology. This will open up a new way of extending a link distance in amplification-free channels such as deep-space optical links, as well as reducing the number of amplifiers in long-haul optical fiber communication. Our scheme could be extended to multiple modulation signals [@Bondurant1993] or another modulation format [@Guha2011]. While the receiver demonstrated here measures each signal separately, an important future work is to extend it to the quantum collective decoding [@Sasaki1997; @Buck2002; @Fujiwara2003] which collectively detects multiple pulses and is in principle reachable to the ultimate capacity bound in lossy optical channels [@Giovannetti2004]. From a technical point of view, the photon-level phase locking implemented here is an important step toward real communication and could also be useful for interferometer-based sensing applications. Finally, it should be noted that our receiver has an ability to resolve photon numbers. The displacement receiver with a very high efficiency and number resolving ability realized here is directly applicable to quantum information science and technologies such as a loophole-free test of quantum mechanics with continuous variable states [@Banaszek1999], an efficient receiver for coherent state-based quantum cryptography [@Wittmann2010], and quantum repeaters and computation based on entangled coherent states [@Ralph2003; @Spiller2006; @Azuma2010]. [Acknowledgements.]{} The authors thank H. Takahashi, E. Sasaki, and T. Itatani for their technical supports and J. Neergaard-Nielsen for reading of the manuscript. This work has been supported by MEXT Grant-in-Aid for Young Scientists (B) 22740720. [25]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop []{} (, ) @noop [, ()]{} @noop []{} (, ) @noop [ ()]{} @noop [ ()]{} @noop [**, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{}, @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [**, ()]{} @noop (),	{ "pile_set_name": "ArXiv" }
--- abstract: 'In hard interactions between external particles incident on a heat-bath, we show that large logarithms are generated when a radiated or absorbed gauge boson is collinear with the initial fermion momentum. These logarithms can be absorbed into process independent splitting/absorption probabilities. Unlike the zero-temperature case, however, they depend explicitly on the temperature and the scale of the interaction.' --- [Collinear Divergences at One-loop Order\ for External Particles in a Heat-bath.]{} [Saumen Datta[Theory Group, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India, Email— [email protected]]{}, Sourendu Gupta[Theory Group, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India, Email— [email protected]]{}, V. Ravindran[Theory Group, Tata Institute of Fundamental Research, Navrangpura, Ahmedabad 380009, India, Email— [email protected]]{}.\ ]{} In recent years the problem of calculating scattering cross sections or rates for external particles travelling in a thermal medium has become phenomenologically interesting [@pheno]. In a previous paper we have shown that to one-loop order there are no infrared divergences when two external particles collide in a heatbath kept at a temperature $T$ [@old]. We were able to resum the one-loop result and calculate the distribution of the mismatch in initial and final momenta due to soft radiation. The result was finite and contained some large logarithms of the type which usually arise from collinear emissions. In this paper we study this collinear behaviour in detail. We find that divergences exist, and are signalled by the familiar “large terms” of the order $\log(Q^2/m^2)$ ($Q^2$ is the scale of the process and $m$ is a regulating mass). We show that the collinear part gives rise to a certain universal splitting/absorption probability which explicitly depends on $Q$ and $T$. There has been recent interest in collinear singularities associated with soft particles [@many], because they might spoil the hard thermal loop resummation [@htl]. Since we study a hard particle in a heat-bath, we have nothing further to say about this. Consider the simplest scattering process in which the collinear singularities make their appearance. This is the inclusive cross section for the scattering of a charged electron with a space-like photon, $\gamma^$, in a QED heat-bath, which we consider in a perturbation theory of the gauge coupling. The initial electron momentum is denoted by $p$, the $\gamma^$ momentum by $q$ and the final electron momentum by $p'$. The 4-velocity of the heat-bath in any frame, $u$ ($u^2=1$), is a new vector in the problem [@wel82]. Compared to the $T=0$ case, there are extra scalars $p\cdot u$ and $q\cdot u$ which have to be taken into account. For the rest, we use the standard notation $$q^2=-Q^2\qquad{\rm and}\qquad p\cdot q={Q^2\over2x}, \label{int:kine}$$ and work in the limit $Q^2\gg m^2$, where $p^2=p'^2=m^2$. We also take $Q^2\gg T^2$ where $T$ is the temperature of the heat-bath. The inclusive cross section can be written as a perturbation expansion in the form $$\sigma=\alpha\left(\sigma_0+\alpha\sigma_1+\alpha^2\sigma_2+\cdots\right), \label{int:cross}$$ where $\alpha$ is the gauge coupling. At the lowest order in perturbation theory, one has only the interaction between $e$ and $\gamma^$, and hence $\sigma_0$ is independent of $T$. At higher orders, interactions with thermal photons have to be taken into account, and $\sigma_i$ ($i>0$) may change from its $T=0$ value. It is useful to write the cross section in the form $$\sigma\;=\;{\rho^{\mu\nu} W_{\mu\nu}\over4\sqrt{m^2Q^2+(p\cdot q)^2}}, \qquad{\rm where}\qquad \rho_{\mu\nu}=\sum_\lambda \epsilon^{\lambda}_\mu(q) \epsilon^\lambda_\nu(q). \label{int:sfdef}$$ $\epsilon^\lambda_\mu(q)$ is the polarisation vector of the off-shell photon in the polarisation state $\lambda$. The tensor $\rho$ is a density matrix for the polarisation states of $\gamma^$ and is symmetric in its indices. The quantity to be computed in perturbation theory is the symmetric part of the rank-2 tensor $W_{\mu\nu}$, which is the vacuum expectation value of the product of the electromagnetic current coupling to $\gamma^$. The computation is simplified by decomposing the tensor $W_{\mu\nu}$ into scalar functions multiplying all symmetric tensors which can be built out of the vectors in the problem. Furthermore, the gauge invariance of the current implies that only those tensors orthogonal to $q$ are relevant. There are four such tensors— $$\begin{array}{rl} T^1_{\mu\nu}\;=\;g_{\mu\nu} +{\displaystyle 1\over\displaystyle Q^2}q_\mu q_\nu\quad&\quad T^2_{\mu\nu}\;=\;P_\mu P_\nu\\ T^3_{\mu\nu}\;=\;U_\mu U_\nu\qquad\quad\;\quad&\quad T^4_{\mu\nu}\;=\;U_\mu P_\nu+U_\nu P_\mu.\\ \end{array}\label{apten:basis}$$ We have used a shorthand notation for the components of $p$ and $u$ orthogonal to $q$— $$P_\mu\;=\;p_\mu+{\displaystyle p\cdot q\over \displaystyle Q^2}q_\mu, \quad{\rm and}\quad U_\mu\;=\;u_\mu+{\displaystyle u\cdot q\over \displaystyle Q^2}q_\mu. \label{apten:ortho}$$ As a result, $$W_{\mu\nu}=\sum_{i=1}^4 W_i(x,Q^2,p\cdot u,q\cdot u) T^i_{\mu\nu}, \label{apten:sfdef}$$ and hence there are four “structure functions” in this problem. At $T=0$ only the two structure functions $W_1$ and $W_2$ appear. Even for $T>0$, at the leading order of perturbation theory $W_3=W_4=0$, since $\sigma_0$ does not contain any terms involving $u$. Hence the Callan-Gross relation [@cg] is also valid to this order, with corrections generated at higher orders, through the usual vacuum ($T=0$) processes, as well as by additional interactions with real thermal gauge bosons. 9truecm Reviews of the real-time thermal field theory techniques we use can be found in [@realtime]. Using these we can generate the processes contributing to $\sigma_1$ and the rules for their evaluation. The relevant diagrams with real thermal photon emission and absorption are shown in Figure (\[fg:feynr\]). For one photon emission (absorption) the two-body phase space measure can be taken to be $$d\Gamma_\pm\;=\; {1\over(2\pi)^4}d^4k 2\pi\delta^\pm(k^2)2\pi\delta^+((p+q-k)^2) B(k\cdot u), \label{apsud:measure}$$ where $k$ is the four-momentum of the thermal photon, $\delta^\pm(x^2)= \delta(x^2)\theta(\pm x_0)$ and $B(x)$ is the Bose distribution $1/[\exp(\|x\|/T)-1]$. The positive sign in eq. (\[apsud:measure\]) corresponds to the emission process and negative to absorption. Note that for the diagrams in Figure (\[fg:feynr\]), the vector $u$ appears only in the measure. This leads to a singularity as $k\cdot u\to0$. However, this is not a collinear singularity but the previously analysed soft singularity [@old]. Since $u^2=1$, a boost to the rest frame of $u$ can always be done. This gives us the correct interpretation of the divergence. The only collinear singularities then arise from the matrix elements. These can be identified as divergences in the limit $k\cdot p\to0$ when $m\to0$, and have the same origin as those occurring at $T=0$. We choose to work in a planar gauge [@ddt], which is a ghost-free gauge specified by the gauge fixing part of the Lagrangian— $${\cal L}_{gf}\;=\; -{1\over2 v^2}(v_\mu A^\mu)\partial^2(v_\nu A^\nu). \label{pert:gauge}$$ The vector $$v_\mu\;=\;C_1 p'+ C_2 p \label{pert:gaugedef}$$ defines the gauge choice. The coefficients $C_1$ and $C_2$ are chosen such that $v^2\ne0$ and the sum over polarisations of the real photons becomes $$d_{\alpha\beta}\;\equiv\; \sum_\lambda\epsilon_\alpha^\lambda(k)\epsilon_\beta^{\lambda }(k) \;=\;-g_{\alpha\beta}+{\displaystyle k_\alpha v_\beta +k_\beta v_\alpha\over k\cdot v}, \label{pert:completeness}$$ where $\epsilon_\mu^\lambda(k)$ is the polarisation vector of a photon with polarisation $\lambda$ and momentum $k$. It can be verified that in this gauge all collinear singularities come from the squares of the diagrams with the real photon attached to the initial fermion leg. For the first emission diagram, we obtain $$\begin{array}{rl} \|M\|^2\;=&\;-2e^4{\displaystyle1\over\displaystyle(p-k)^2}\\&{\rm Tr} \left[\gamma_\nu p'\!\!\!/ \gamma_\mu \left\{ p\!\!/\left({\displaystyle2p\cdot v\over\displaystyle k\cdot v}-1\right) +k\!\!/\left(1-{\displaystyle p\cdot v\over\displaystyle k\cdot v}\right) +v\!\!/{\displaystyle k\cdot p\over\displaystyle k\cdot v}\right\}\right]. \end{array}\label{coll:eq2}$$ We are interested in extracting the leading terms in the collinear limit, $p\cdot k\to0$. The most transparent way of doing this is to use the Sudakov parametrisation [@ddt]— $$k\;=\;(1-\rho)p+\beta(q+xp)+k_{\scriptscriptstyle T}, \qquad{\rm where}\qquad p\cdot k_{\scriptscriptstyle T}=q\cdot k_{\scriptscriptstyle T}=0. \label{coll:sud}$$ It is clear that this is a Lorentz invariant decomposition. The integration variables are changed to $\rho$, $\beta$ and the two independent components of $k_{\scriptscriptstyle T}$. The Jacobian is simply $${d^4k\over d\rho d\beta d^2k_{\scriptscriptstyle T}}\;=\;{Q^2\over2x}. \label{apsud:jacob}$$ The variables $\rho$ and $\beta$ are fixed by the $\delta$-function constraints in the measure $d\Gamma_+$. In the collinear limit, the solution which leads to $p\cdot k\to0$ as $k_{\scriptscriptstyle T}^2\to0$ is $$\rho\;=\;x+{\cal O}\left(k_{\scriptscriptstyle T}^2\right),\qquad\beta\;=\; {\displaystyle xk_{\scriptscriptstyle T}^2\over\displaystyle Q^2(1-x)} +{\cal O}\left(k_{\scriptscriptstyle T}^4\right). \label{apsud:coll}$$ Requiring $\rho$ and $\beta$ to be real, in the limit $m\to0$ we find $$0\;\le\;k_{\scriptscriptstyle T}^2\;\le\;{Q^2\over4x}(1-x). \label{apsud:limits}$$ The $\theta$-functions place no further restrictions, and may be dropped to give the phase space measure $$d\Gamma_+\;=\; {\displaystyle xd\rho d\beta d^2k_{\scriptscriptstyle T} \over\displaystyle 2Q^2(1-x)(2\pi)^2} \delta\left(\beta-{xz\over1-\rho}\right)\delta(\rho-x). \label{apsud:final}$$ Absorption is handled by making the change of variables $k_\mu\to-k_\mu$, and then writing a Sudakov parametrisation as before. $d\Gamma_-$ differs from $d\Gamma_+$ only in the change $$\delta(\rho-x)\;\longrightarrow\;\delta(\rho-2+x). \label{apsud:finalm}$$ In the collinear limit, the denominator on the right of eq. (\[coll:eq2\]) becomes zero, and hence some care is required in taking this limit. We retain the fermion mass, $m$, as a regulator in the denominator and write $${1\over(p-k)^2}\;=\;{1\over m^2(1-2\rho)-Q^2\beta/x}. \label{coll:den}$$ The most singular terms in the collinear limit can then be found by evaluating the trace in eq. (\[coll:eq2\]) for $m=0$ and neglecting terms in $k_{\scriptscriptstyle T}^2$ (and hence in $\beta$). After some straightforward manipulations, we find that the most singular contribution to the tensor $W_{\mu\nu}$ is $$\begin{array}{rl} -8e^4&\left[2xT^2_{\mu\nu} -{\displaystyle Q^2\over\displaystyle2x}T^1_{\mu\nu}\right]\\ &\;\;\int d\Gamma_+\,B[(1-\rho)p\cdot u] \left({\displaystyle1+\rho^2\over\displaystyle1-\rho}\right) {\displaystyle1\over\displaystyle m^2(1-2\rho)-Q^2\beta/x} \end{array}\label{coll:eq3}$$ The contribution from the corresponding absorption diagram, obtained by changing $k\to-k$ in the matrix element and using the phase space measure $d\Gamma_-$, is simply obtained by setting $\rho$ to $2-\rho$ is the above expression. In the chosen gauge, the squares of the other two diagrams do not have singular denominators and hence can be neglected in the collinear limit. The cross terms do have a singular denominator. However, the trace contains $$\stackrel{\displaystyle{\rm Lt}}{\scriptscriptstyle k\to(1-\rho)p} d_{\alpha\beta}(k)p^\alpha\;=\;0.$$ Hence this term can also be neglected. Two facts about eq. (\[coll:eq3\]) and its analogue for photon absorption are worth pointing out. First, both terms contain only the tensors $T^1$ and $T^2$. Hence there are no collinear contributions to the thermal structure functions $W_3$ and $W_4$ at this order. Second, the integrand has a probability interpretation. The part $(1+\rho^2)/(1-\rho)$ can be interpreted as the probability that an incoming electron radiates a photon carrying a fraction $1-\rho$ of the initial momentum. The Bose distribution factor is the probability that the radiated photon is indistinguishable from a thermal photon. The integrand has support on $0\le\rho\le1$, which is consistent with this interpretation. Similarly, the absorption process can also be interpreted as the product of two probabilities. The Bose distribution is the probability of finding a photon in the heat-bath carrying a fraction $\rho-1$ of the incoming electron’s momentum and the factor $(1+(2-\rho)^2)/(1-(2-\rho))$ is the probability of absorption. This integral has support on $1\le\rho\le2$. The integrals can be performed completely. The thermal leading log part is $$W_{\mu\nu}\;=\;4\pi\alpha^2 P(x,Q/T) \log\left({\displaystyle Q^2\over \displaystyle m^2}\right) {\displaystyle x\over\displaystyle Q^2} \left[2xT^2_{\mu\nu} -{\displaystyle Q^2\over\displaystyle2x}T^1_{\mu\nu}\right]. \label{coll:result}$$ We have introduced the finite temperature part of the “splitting function” $$P(x,Q/T)\;=\;{2\over\exp[(1-x)p\cdot u/T]-1}\, \left({\displaystyle 1+x^2\over\displaystyle1-x}\right). \label{coll:splitting}$$ This can be given an interpretation as the probability of an external electron splitting off a thermal photon. Unlike the case at $T=0$, this factor has an explicit dependence on $Q/T$. Note also that $P(x,Q/T)$ has a singularity as $x\to1$. This is the region of phase space where the thermal photon is soft. We have shown in [@old] that the cross section is finite in this limit provided virtual corrections are taken into account. Using this result, we can simply write down a regulated version of eq. (\[coll:splitting\]) as the splitting probability for the external electron. At $T=0$ the leading soft divergence in the real diagrams is logarithmic, and shows up in the splitting functions as a divergence of the form $1/(1-x)$, in the limit $x\to1$. It is cured by taking into account the virtual diagrams. The regulated form of the splitting functions is then given by the familiar prescription $$\int dx P_+(x) f(x)\;=\;\int dx P(x)\left[f(x)-f(1)\right], \qquad\qquad(T=0), \label{appl:zero}$$ where $f(x)$ is a test function. Note that the first moment of any distribution vanishes when convoluted with the splitting function so regularised. For $T>0$ the leading soft divergence is quadratic. This is signalled by a divergence of the form $1/(1-x)^2$ ($x\to1$) in the splitting functions. Its cancellation against virtual contributions has been shown in [@old]. The sub-leading logarithmic divergence has also been shown to cancel against virtual corrections [@indu]. Consequently, we just write down an appropriately regularised version of the $T>0$ part of the splitting function— $$\int dx P_+(x) f(x)\;=\;\int dx P(x) \left[f(x)-f(1)-(x-1) f'(1) \right]. \label{appl:def}$$ Note that the first two moments of $P_+$ vanish with this regularisation. The vanishing second moment implies that finite temperature effects do not change the expectation value of the parton’s momentum. This is expected [@old] and is a consequence of detailed balance. The universality of this additional finite temperature term in the splitting functions derived here can be easily checked. The calculation for Fermion pair-production is very similiar to the computation presented in this paper, and yields precisely the same collinear term derived here. In QED, since the electron mass is non-zero, our results are complete. However, in a real experiment we shall have to deal with a QCD heatbath. Our main results, eq. (\[coll:result\]), along with eqs. (\[coll:splitting\]) and (\[appl:def\]), can be carried over to this case with the simple replacement $\alpha^2\to C_{\scriptscriptstyle F}\alpha\alpha_{\scriptscriptstyle S}$. The crucial change is that $m=0$ for quarks and hence the results are singular. At $T=0$, these collinear singularities are handled by factoring them into universal quark distributions inside hadrons. A similiar procedure will have to be developed for external hadrons or jets impinging on a plasma. In order to complete this program, a suitable definition of the QCD running coupling at finite temperature [@baier] must be provided. At $T=0$, this is sufficient information to sum the one-loop iterated ladder diagrams into the DGLAP equations [@dglap]. For $T>0$ the situation is more complicated. This is clear from the fact that the analogue of the splitting function contains the dimensionless variable $Q/T$ in addition to $x$. At finite temperature and arbitrary scale $Q^2$, we are forced to consider two scales in the renormalisation group [@twoscale]. In various domains these simplify. For example, when $Q\gg T$, one expects to be able to use a single scale. In this limit, the regularised version of eq (\[coll:splitting\]) vanishes, and the evolution in $Q^2$ is the same as at $T=0$. However, the parton distributions at each $T$ must then be seperately measured. Only by keeping two scales can information at $T=0$ be evolved to $T>0$. This work is left to the future. [99]{} M. Gyulassy and M. Plümer, [Nucl. Phys.]{} B 346 (1990) 1;\ M. Gyulassy and X.-N. Wang, [Nucl. Phys.]{} B 420 (1994) 583;\ J.-C. Pan and C. Gale, [Phys. Rev.]{}, D 50 (1994) 3235;\ S. Gupta, [Phys. Lett.]{}, B 347 (1995) 381. S. Gupta, D. Indumathi, P. Mathews and V. Ravindran, [Nucl. Phys.]{} B 458 (1996) 189;\ H. A. Weldon, [Phys. Rev.]{}, D 49 (1994) 1579. R. Baier, S. Peigne and D. Schiff, [Z. Phys.]{}, C 62 (1994) 337;\ F. Flechsig and A. K. Rebhan, [Nucl. Phys.]{}, B 464 (1996) 279;\ P. Aurenche, F. Gelis, R. Kobes and E. Petitgirard, hep-ph/9604398 (to appear in [Phys. Rev.]{}, D), and hep-ph/9609256. R. D. Pisarski, [Phys. Rev. Lett.]{}, 63 (1989) 1129;\ E. Braaten and R. D. Pisarski, [Nucl. Phys.]{}, B 337 (1990) 569;\ J. Frenkel and J. C. Taylor, [Nucl. Phys.]{}, B 334 (1990) 199. H. A. Weldon, [Phys. Rev.]{}, D 26 (1982) 1394. C. G. Callan D. Gross, [Phys. Rev. Lett.]{}, 22 (1969) 156. N. P. Landsman and Ch. G. van Weert, [Phys. Rep.]{}, 145 (1987) 141;\ A. Niemi and G. Semenoff, [Nucl. Phys.]{}, B 230 (1984) 181. Dokshitzer, Yu. L. Dyakonov, D. I. and Troyan, S. I., [Phys. Rep.]{} 58 (1980) 269. D. Indumathi, Dortmund preprint DO-TH-96-09, hep-ph/9607206. R. Baier, B. Pire and D. Schiff, [Phys. Lett.]{}, B 238 (1990) 367. G. Altarelli and G. Parisi, [Nucl. Phys.]{}, B 126 (1977) 298;\ V. N. Gribov and L. N. Lipatov, [Sov. J. Nucl. Phys.]{}, 46 (1972) 438. J. Kapusta, [Finite Temperature Field Theory*]{}, Cambridge University Press, Cambridge, UK, 1989;\ C. Ford and C. Wiesendanger, preprint DIAS-STP 96-10, hep-ph/9604392.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this study, two different simple and intuitive semiempirical schemes for computing approximate non-adiabatic coupling vectors (NACVs) between the ground and excited electronic states are presented. The first approximation makes use of Mulliken transition charges, while the second is based on derivative coupling vectors between localized molecular orbitals. Both approximations lend themselves easily to implementation within a whole spectrum of semiempirical quantum-chemical semiempirical methods. Here we present the implementation within the tight-binding DFT and benchmark its performance against analytical TD-DFT NAC vectors for a range of planar fluorescent chromophores at the Franck-Condon point. The pattern of the atomic NAC vectors is often reproduced, but the relative magnitude and total length of the NAC vector are often in serious error. Although quantitative predictions are not possible, these simple and intuitive approximations allow to explain, in a qualitative way, trends in the electronic coupling in extended molecular systems and complex materials. In this context, we investigate how the non-adiabatic coupling depends on the delocalization length of an excitation in chromophoric oligomers based on a simple model. Finally, we make general qualitative predictions on the size dependence of the fluorescence quantum yields in extended molecular systems, and illustrate those on the example of triply fused porphyrin tapes with increasing length.' author: - Alexander Humeniuk - Roland Mitrić bibliography: - 'references.bib' title: 'Assessment of simple and intuitive semiempirical approximations for non-adiabatic coupling vectors in the frame of (LC)-TDDFTB' --- Introduction ============ Non-adiabatic coupling vectors (NACVs) play an important role in photochemistry. They describe the coupling between Born-Oppenheimer surfaces due to the nuclear kinetic energy and allow transitions between electronic states in the absence of radiation [@baer_book]. They are a vital ingredient in - non-adiabatic molecular dynamics simulations (surface hopping on “on the fly”) [@tapavicza2013ab], - searching for minimal energy conical intersections [@ragazos1992optimization] - and predicting non-radiative transition rates and fluorescence quantum yields [@valiev2018first]. The brute force method for computing non-adiabatic coupling vectors is numerical differentiation of the wavefunction with respect to the atomic positions, which requires at least $3 N_{\text{atoms}}$ electronic structure calculations. In the context of TD-DFT exact coupling vectors can be obtained analytically [@send2010first] in an efficient way, but the implementation of the method is complicated. Having a simple and intuitive approximation for NACVs that may be combined with any semiempirical electronic structure method is therefore highly desirable. In this article we compare two different semiempirical approximation for calculating non-adiabatic coupling vectors between the ground state and an excited state (usually $S_1$), which have been implemented in the frame of tight-binding DFT [@humeniuk2017dftbaby]. The first approximation is based on transition charges: In analogy with the transition dipole moment, the NACV is obtained simply from the transition charges and the molecular geometry. The second approximation, which has been propounded by Abad et.al. [@nacs_approx_MOs], is based on molecular orbitals: Non-adiabatic couplings between Kohn-Sham orbitals are constructed from gradients of the overlap and Hamiltonian matrix between localized atomic orbitals, which are readily available in tight-binding DFT, since the same quantities are needed for evaluation of the energy gradient. Abad et.al. tested their approximation in the vicinity of conical intersections, where the magnitude of the NACVs is largely determined by the small energy gap. At these photochemical funnels the non-adiabatic coupling diverges and the transfer of population between electronic states is usually extremely fast. It remains to be investigate how well the approximation performs when the $S_0-S_1$ energy gap is large such as at the $S_1$ minimum or the Franck-Condon point where the length of the NACVs and their orientation relative to the normal modes determines the non-radiative transition rate. The article is structured as follows: After a brief description of the approximations (in sections \[sec:nacs\_charges\] and \[sec:nacs\_mos\]), we investigate how the non-adiabatic coupling depends on the delocalization length of an excitation in chromophoric oligomers (in section \[sec:quantum\_yield\]). We then graphically compare the direction and magnitude of the approximate NACVs with their exact counterparts for a range of organic molecules with bright $\pi\pi^$ excitations (in section \[sec:comparison\]). Finally we make some qualitative predictions of fluorescence quantum yields in porphyrin tapes (section \[sec:tapes\]). Theory ====== Semiempirical approximations ---------------------------- The first-order non-adiabatic coupling vector between two electronic Born-Oppenheimer states $m$ and $n$ is $${\boldsymbol{\tau}}_{mn} = {\langle\! \Psi_m\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle}. \label{eqn:nacv_grad}$$ The coupling vector may be expressed as $${\langle\! \Psi_m\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle} = \frac{{\langle\! \Psi_m\! \mid} \frac{\partial {\hat{H}}}{\partial {\boldsymbol{R}}} {\mid\! \Psi_n\! \rangle}}{E_n - E_m} \label{eqn:nacv_ediff}$$ by differentiating the electronic Schrödinger equation on both sides with respect to the nuclear coordinates ${\boldsymbol{R}}$ and multiplying by ${\langle\! \Psi_m\! \mid}$ for $m \neq n$ and rearranging. The derivation of this expression requires that ${\hat{H}} {\mid\! \Psi_n\! \rangle} = E_n {\mid\! \Psi_n\! \rangle}$ is satisfied exactly, which is a much stronger statement than just requiring that Schrödinger’s equation is satisfied after projecting onto a finite basis set ${ \{ {\mid\! \Phi_i\! \rangle} \}_{i=1,\ldots,N_{\text{basis}}} }$: $${\langle\! \Phi_i\! \mid} {\hat{H}} {\mid\! \Psi_n\! \rangle} = \tilde{E}_n {\langle\! \Phi_i\! \mid\! \Psi_n\! \rangle} \quad \quad \forall i=1,\ldots,N_{\text{basis}}$$ Therefore eqn. (\[eqn:nacv\_ediff\]) is strictly correct only if a complete basis set is used. In finite basis sets additional Pulay terms [@pulay1969ab] have to be considered which arise from the dependence of the basis set on the nuclear coordinates. Nevertheless it is a good starting point for semiempirical approximations. ### \[sec:nacs\_charges\] Approximation based on transition charges Since the electronic Hamiltonian depends on the nuclear geometry only through the Coulomb attraction between nuclei and electrons, $${\hat{V}}_{ne} = \sum_A^{\text{atoms}} \sum_i^{\text{electrons}} \frac{-Z_A}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}}_i \vert},$$ the coupling vector (\[eqn:nacv\_ediff\]) on atom $A$ simplifies to $${\boldsymbol{\tau}}^A_{mn} = \frac{Z_A}{E_n - E_m} \int d{\boldsymbol{r}} \frac{{\boldsymbol{R}}_A - {\boldsymbol{r}}}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert^3} \rho_{mn}({\boldsymbol{r}}), \label{eqn:nacv_ediff_enuc}$$ where we have also introduced the transition density matrix $$\rho_{mn}({\boldsymbol{r}}) = N \int \ldots \int d{\boldsymbol{r}}_2 \ldots d{\boldsymbol{r}}_N \Psi_m^({\boldsymbol{r}},{\boldsymbol{r}}_2,\ldots,{\boldsymbol{r}}_N) \Psi_n({\boldsymbol{r}},{\boldsymbol{r}}_2,\ldots,{\boldsymbol{r}}_N).$$ By partial integration of eqn. (\[eqn:nacv\_ediff\_enuc\]) (see appendix \[sec:partial\_integration\]) the NACV turns into $${\boldsymbol{\tau}}^A_{mn} = \frac{-Z_A}{E_n - E_m} \int d{\boldsymbol{r}} \frac{\nabla \rho_{mn}({\boldsymbol{r}})}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \label{eqn:nacv_rho_deriv}.$$ This expression is very instructive since it shows that the coupling vector density is proportional to the gradient of the transition density. The largest contribution comes from points where ${\boldsymbol{r}} \approx {\boldsymbol{R}}_A$ due to the singularity of the Coulomb potential. Therefore we can say qualitatively that the non-adiabatic coupling vector on atom $A$ is approximately proportional to the gradient of the transition density around that atom. To derive a semiempirical approximation for $\tau^A_{mn}$ let us return to eqn. (\[eqn:nacv\_ediff\_enuc\]) and assume that the transition density may be approximated by atomic transition charges (monopoles) $$\rho_{mn}({\boldsymbol{r}}) \approx \sum_B q_B \delta({\boldsymbol{r}} - {\boldsymbol{R}}_B). \label{eqn:rho_monopole_approx}$$ where $\delta(\cdot)$ is Dirac’s $\delta$-function. This approximation is frequently employed in semiempirical methods such as tight-binding DFT [@koskinen2009density]. The transition charges $q_A$ may be fitted to reproduce the electrostatic potential generated by the transition density (using the CHELPG algorithm) [@madjet2006intermolecular] or they may be calculated as Mulliken transition charges from the transition density matrix. Substituting the monopole approximation (\[eqn:rho\_monopole\_approx\]) into eqn. (\[eqn:nacv\_ediff\_enuc\]) and using the property of the $\delta$-function, $\int \delta(x-x_0) f(x) dx = f(x_0)$, we get $${\boldsymbol{\tau}}^A_{mn} \approx \frac{Z_A}{E_n - E_m} \sum_{B \neq A} q_B \frac{{\boldsymbol{R}}_A - {\boldsymbol{R}}_B}{\vert {\boldsymbol{R}}_A - {\boldsymbol{R}}_B \vert^3}. \label{eqn:nacv_trchg}$$ The term where $A = B$ was excluded to avoid dividing by zero. Only valence electrons are usually included in semiempirical calculations. Then the bare nuclear charge $Z_A$ should be replaced by the charge of the atomic core $Z_A^{\text{core}}$ (nucleus and core electrons), for instance in the case of carbon $Z_A^{\text{core}}=4$ instead of $Z_A=6$. This approximation is completely analogous to how the transition dipole moment is calculated from the transition charges in the frame of TD-DFTB [@niehaus2001tight], $${\boldsymbol{\mu}}_{mn} \approx \sum_A q_A {\boldsymbol{R}}_A. \label{eqn:dipole_trchg}$$ The simplicity of the derived approximate expressions enables us to make some general statements about the properties of the NACVs. The direction and length of NACVs can be deduced qualitatively by inspecting the transition density or the distribution of the transition charges: - Coupling vectors are non-zero only on atoms which take part in an excitation. - The coupling vectors point roughly along the direction where the transition density changes most strongly. Thus, if there is a node in the transition density between two atoms, the NACV on the atom is perpendicular to the nodal surface. As a simple example consider the $\pi\pi^$ excitation in ethene (Fig. \[fig:ethene\_nacs\_charges\_trdensity\]). The transition charge is positive on one carbon, negative on the other and almost zero on the hydrogen atoms. Therefore the coupling vectors on the hydrogen atoms are zero. The transition charges change strongly from $+q$ to $-q$ when moving from one carbon to the other along the C$=$C bond. Therefore the NACVs on the carbons point along this bond. ![Ethene, $\pi\pi^$ transition** .[]{data-label="fig:ethene_nacs_charges_trdensity"}](ethene_nacs_charges_trdensity.png){width="90.00000%"} The approximation fails completely when the transition density cannot be adequately described by monopoles. For instance in water, the HOMO-LUMO transition, $4a_1 \leftarrow 1b_1$, has lobes of opposite sign below and above the molecular plane. The gradient of the transition density points perpendicularly to the molecular plane and is orthogonal to all vectors ${\boldsymbol{R}}_A - {\boldsymbol{R}}_B$. This implies that the coupling vector cannot be represented in the basis of bond vectors. The Mulliken transition charges are all zero, as is the approximate non-adiabatic coupling vector. In this case the approximation for the electric transition dipole given in Eq. \[eqn:dipole\_trchg\] is also incorrect. ### \[sec:nacs\_mos\] Approximation based on molecular orbitals Here we briefly recapitulate how NACVs are calculated in the local-orbital scheme proposed in Ref. [@nacs_approx_MOs] using the language of tight-binding DFT (DFTB). In DFTB a minimal basis set of valence atomic orbital is used. The molecular orbitals (MO) are linear combinations of these localized basis functions ${\mid\! \mu\! \rangle}$: $${\mid\! \psi_i\! \rangle} = \sum_{\mu} c_{\mu i} {\mid\! \mu\! \rangle}$$ The coefficients $c_{\mu i}$ for the molecular orbital $i$ are the eigenvector of the Kohn-Sham equation belonging to eigenenergy $\epsilon_i$: $$\sum_{\nu} \left( H^0_{\mu\nu} - \epsilon_i S_{\mu\nu} \right) c_{\nu i} = 0 \label{eqn:kohn_sham_nonscc}$$ Matrix elements of the Kohn-Sham Hamiltonian at the reference density, $$H^0_{\mu\nu} = {\langle\! \mu\! \mid} H^{\text{KS}}[\rho_0] {\mid\! \nu\! \rangle}$$ the overlap matrix elements $$S_{\mu\nu} = {\langle\! \mu\! \mid\! \nu\! \rangle}$$ and their gradients are obtained from Slater-Koster rules[@slater_koster]. With the help of eqn. (\[eqn:kohn\_sham\_nonscc\]) the authors of Ref. [@nacs_approx_MOs] derived an approximate expression for non-adiabatic coupling vectors between molecular orbitals: $$\begin{split} {\boldsymbol{d}}_{ij}^A & = {\langle\! \psi_i\! \mid\! \frac{\partial \psi_j}{\partial {\boldsymbol{R}}_A}\! \rangle} \\ & \approx \frac{1}{\epsilon_i - \epsilon_j} \sum_{\mu,\nu} c^_{\mu i} c_{\nu j} \left[ - \frac{\partial H^0_{\mu\nu}}{\partial {\boldsymbol{R}}_A} + \frac{\epsilon_i + \epsilon_j}{2} \frac{\partial S_{\mu\nu}}{\partial {\boldsymbol{R}}_A} \right] \label{eqn:coupling_mos} \end{split}$$ In time-dependent density functional theory and its tight-binding version, excited states are represented as linear combinations of singly excited Slater determinants: $${\mid\! \Psi_n\! \rangle} = \sum_{o \in \text{occ}} \sum_{v \in \text{virt}} C_{ov}^{(n)} {\mid\! \Psi_{o}^v\! \rangle}$$ Non-adiabatic coupling vectors between the many-electron ground and excited states are obtained by contraction of the single-particle coupling vectors ${\boldsymbol{d}}_{ij}^A$ with the coefficients $C^{(n)}$: $${\boldsymbol{\tau}}^A_{0n} = {\langle\! \Psi_0\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}_A}\! \rangle} = \sum_{o \in \text{occ}} \sum_{v \in \text{virt}} C_{ov}^{(n)} {\boldsymbol{d}}_{ov}^A \label{eqn:nacv_mos}$$ It is worthwhile to highlight some of approximations made in the above derivation: (a) Eqn. (\[eqn:kohn\_sham\_nonscc\]) neglects the dependence of the Hamiltonian on the density ($H[\rho] \approx H[\rho_0]$). This allowed to derive the relatively simple expression (\[eqn:coupling\_mos\]) for the coupling vectors. However, in our calculations we use expression (\[eqn:coupling\_mos\]) with MO coefficients obtained from solving the Kohn-Sham equations self-consistently. (b) In principle, the NACV contains a contribution from changes of the coefficients $\frac{\partial C^{(n)}_{ov}}{\partial {\boldsymbol{R}}}$, which is neglected. (c) The exact NACV diverges when the ground and excited state cross ($E_1 \approx E_0$), whereas for the approximate NACV this happens when the HOMO-LUMO gap closes ($\epsilon_{\text{HOMO}} \approx \epsilon_{\text{LUMO}}$). It is well-known that HOMO-LUMO gaps are often significantly larger than $S_0-S_1$ excitation energies obtained from TD-DFT due to the mixing of single excitations. This suggests that the approximation will work best when such many-body effects are small so that the $S_0-S_1$ transition predominantly has HOMO $\to$ LUMO character. Expression (\[eqn:coupling\_mos\]) is ideally suited for tight-binding DFT, since the gradients of the matrix elements can be constructed very efficiently at runtime from precalculated Slater-Koster tables. Since the same quantities are needed for assembling the gradient of the energy, which is needed in any molecular dynamics (MD) simulation, the computation of the NACVs comes at little additional cost. This should be contrasted with the computational cost of an alternative method for computing the non-adiabatic couplings in MD simulations: In the surface hopping method [@tully1990molecular] the electronic populations depend only on the scalar product between the NACV and the nuclear velocity vector. This scalar can be obtained directly from the overlap of the electronic wavefunctions at neighbouring timesteps [@mitric2009nonadiabatic] obviating the need for computing the NACVs: $${\langle\! \Psi_m\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle} \cdot \frac{d{\boldsymbol{R}}}{dt} \approx \frac{1}{\Delta t} {\langle\! \Psi_m(t)\! \mid\! \Psi_n(t+\Delta t)\! \rangle}$$ However, since each excited state is a linear combination of Slater determinants, the evaluation of the overlap entails a large number of determinants, rendering this scheme very expensive for large molecules, unless cutoff thresholds are used for culling determinants which contribute little to the overlap integral. \[sec:quantum\_yield\] Qualitative fluorescence quantum yield ------------------------------------------------------------- Approximations (\[eqn:nacv\_trchg\]) and (\[eqn:dipole\_trchg\]) provide qualitative guidelines on how to tune the electronic wavefunctions for increasing the fluorescence quantum yields. At the moment we will only focus on electronic effects, although vibrational effects can also be very important, as will become clear later on. If the vibrational wavefunction is neglected, according to Fermi’s Golden rule the rates for radiative (spontaneous emission) and non-radiative (internal conversion) decays are proportional to the lengths squared of the transition dipole and non-adiabatic coupling vectors, respectively, between the ground state $S_0$ and the first excited state $S_1$: $$\begin{aligned} k_{\text{rad}} &\propto \vert {\boldsymbol{\mu}}_{01} \vert^2 \\ k_{\text{IC}} &\propto \vert {\boldsymbol{\tau}}_{01} \vert^2 \end{aligned}$$ To increase the fluorescence quantum yield $$\text{QY} = \frac{1}{1 + \frac{k_{\text{IC}}}{k_{\text{rad}}}}$$ $k_{\text{rad}}$ needs to be maximized while $k_{\text{IC}}$ needs to be minimized. This can be achieved by - increasing the length of the transition dipole and/or - avoiding conical intersections, where $E_1 = E_0$ - and reducing the gradient of the transition density. To avoid the crossing of the energy levels of $S_1$ and $S_0$, the geometry should be rigid, so that we can assume there is a stable minimum on $S_1$ and the reorganization energy is small. Then it remains to maximize the length of the transition dipole moment and to minimize the gradient of the transition density. Since it is easier to analyze only one factor, we build a simple model, where the transition dipole is constant and only the length of the NAC vector changes. ### \[sec:1d\_model\] 1D model fluorophore Consider a linear molecule (e. g. a polyene) with $2 M$ atoms on an equidistant grid with spacing $h$ (see Fig. \[fig:1d\_model\_fluorophore\]a). For simplicity, we assume that each atom has a single $p_z$ orbital and contributes one electron. The $S_1$ state is a HOMO-LUMO transition. The $S_0-S_1$ transition density has nodes between all atoms, so that the transition charges alternate between positive and negative values. The atomic positions and transition charges for atom $i$ are given by $$\begin{aligned} x_i &=& i h \\ q_i &=& (-1)^i \frac{q}{M}\end{aligned}$$ For simplicity we set the nuclear charge to $Z=1$ and the excitation energy to $E_1 - E_0=1$. The transition dipole moment is independent of the number of atoms (see appendix \[sec:dipole\_constant\]): $$\vert {\boldsymbol{\mu}} \vert = \sum_{i=0}^{2M-1} q_i x_i = q h$$ The non-adiabatic coupling vector on the $i$-th atom is given by $$\tau^{(i)} = \frac{q}{M h^2} \sum_{j \neq i}^{2M-1} (-1)^j \frac{i-j}{\vert i-j \vert^3}.$$ The length of the total non-adiabatic coupling vector is given by $$\vert {\boldsymbol{\tau}} \vert = \sqrt{\sum_{i=0}^{2M-1} (\tau^{(i)})^2}.$$ and needs to be evaluated numerically. If we plot the length of ${\boldsymbol{\tau}}$ against the number of atoms $2 M$ that participate in the excitation (Fig. \[fig:1d\_model\_fluorophore\]b), we see that the rate for internal conversion can be minimized by spreading the transition charge over as many atoms as possible while maintaining the same electric transition dipole moment. Since the transition charges change sign every second atom, the gradient of the transition density can be reduced only if the charges themselves are small. In order to keep the same transition dipole moment, the number of atoms over which the excitation is delocalized needs to be increased. ![1D model fluorophore. a)* Linear molecule with 4 atoms. The transition density has nodes between neighbouring atoms. b) The ratio of the non-adiabatic coupling to the transition dipole moment decreases with the number of atoms $2 M$ taking part in the excitation.[]{data-label="fig:1d_model_fluorophore"}](linear_model_fluorophore.png){width="100.00000%"} Results ======= \[sec:comparison\] Comparison between approximate and exact NACVs ----------------------------------------------------------------- The two approximations for NACVs are tested for a series organic molecules with bright $\pi\pi^$ transitions. Many of the selected molecules are fluorescent dyes which have a stable lowest excited singlet state (with the exception of the polyenes). After optimizing the geometries at the AM1 level of theory, the lowest bright excited state was computed with TD-$\omega$B97XD/def2-SVP using Gaussian 16 [@g16]. Analytical NACVs were obtained in the frame of TD-DFT [@send2010first] via the keyword TD=NAC. These vectors serve as “exact” reference values against which the quality of the approximate vectors is measured. Approximate NACVs based on either Mulliken transition charges (according to eqn. (\[eqn:nacv\_trchg\])) or localized orbitals (according to eqns. (\[eqn:coupling\_mos\]) and (\[eqn:nacv\_mos\])) were computed in the frame of long-range corrected tight-binding DFT [@humeniuk2017dftbaby]. The comparison between the three types of NACVs is presented in a graphical way in Figs. \[fig:polyenes\_nacs\] to \[fig:porphyrin\_tapes\_nacs\] below. The components of the NAC vectors on each atom are shown as little red arrows. Since eigenvalue solvers produce eigenvectors with arbitrary global signs, only the relative orientation of the vectors to each other is important. A sign change in either the bra or the ket wavefunction is equivalent to flipping all vectors simultaneously. The NACVs were scaled by a factor (which is indicated in the upper left corner) so that the largest vectors in each figure has approximately the same length. We proceed by analyzing the quality of the semiempirical approximations as compared to the exact NACVs for each class of molecules: Trans-polyenes (Fig. \[fig:polyenes\_nacs\]) are the simplest conjugated systems which behave like the linear 1D model discussed above. The lowest bright state ($B_u$) is polarized along the molecular axis. The transition density has nodes between neighbouring carbon atoms, which coincide with the positions where the tip of one arrow touches the tail of the next. The individual arrows become shorter as the number of carbon atoms increases from ethene to hexatriene, reflecting the smearing out of the transition charges over a larger area. The transition charge (TC) approximation overestimates the NACVs by a factor of 5 but gets the orientation of the vectors right. In turn, the localized orbital (LO) approximation underestimates the NACVs by at least a factor of 10 and fails to predict the orientation. The cyanines Cy$N$ (Fig. \[fig:cyanine\_dyes\_nacs\]) are fluorescent cationic dyes that consist of a polymethine chain connecting two nitrogens which are part of an indole moiety. Cy3, Cy5 and Cy7 differ by the number of carbon atoms in the bridge, in Cy3B [@cooper2004cy3b] the polymethine chain is stabilized against deformation by additional aliphatic rings. In all cyanines the lowest bright excitation is localized on the polymethine chain, and consequently the NACVs are also limited to this region of the molecule. In the polymethine chain the orientation of the arrows alternates as is expected based on the location of the nodes in the transition density. The LO approximation predicts the position and orientation of the NAC vectors correctly but underestimates their magnitude by a factor of 3. The TC approximation yields NACVs that are spread out too much over non-chromophoric parts of the molecule, such as methyl groups in Cy3-Cy7 or the aliphatic rings in Cy3B. In the polymethine chain the NAC vectors all point in the same direction, but the total magnitude of the NACV is approximately correct. Dicyanovinyl-substituted squaraines (Fig. \[fig:squaraine\_dyes\_nacs\]) [@mayerhoffer2013synthesis] are another class of fluorescent dyes. In squaraine-O position 3 of the indole moiety is replaced by oxygen, whereas in squaraine-CMe a methyl group is added. The excitation is localized on the central four-membered ring and the adjacent methine groups. The LO approximation reproduces the orientation of the vectors accurately, except for those on the C$\!=\!$O group, which are far too short. The TC approximation places the largest NAC vectors on two opposite carbon atoms in the four-membered ring, although the coupling vectors at these positions should be zero. As with the cyanines a tendency of TC is observed to place large NAC vectors on atoms that are not part of the chromophore. Finally a selection of polycyclic aromatic hydrocarbons is considered in Figs. \[fig:aromatic\_hydrocarbons\_1\_nacs\] and \[fig:aromatic\_hydrocarbons\_2\_nacs\]. The couplings were calculated for the lowest excited state. Since the ordering of states can be method-dependent, the symmetry label is given in brackets. The ring systems give rise to complex patterns in the distribution of NAC vectors. In fluorene ($B_2$) and phenanthrene ($B_2$) the arrows are arranged in cycles around the outer six-membered rings. This pattern is reproduced by the LO approximation, whereas the TC pattern has no similarity with the exact results. In pyrene ($B_{1u}$), perylene ($B_{1u}$) and rubrene ($B_1$) and relative orientation of the vectors is reproduced correctly both by the LO and the TC approximations, however the relative magnitudes of the vectors differ considerably. The total magnitude of the coupling is severely underestimated by the LO approximation (by a factor of 3-10) and overestimated by the TC approximation (by as much as a factor of 10). In rubrene the excitation is strictly confined to the tetracene core. Inspite of this, the semiempirical approximations yield large vectors on the adjacent phenyl groups, which are perpendicular to the central tetracene. ![Polyenes.* Non-adiabatic coupling vectors computed using Furche’s analytic method (left) and the approximations based on transition charges (middle) or couplings between Kohn-Sham orbitals (right). The factor by which the vectors where scaled is shown in the upper left corner.[]{data-label="fig:polyenes_nacs"}](polyenes_nacs.png){width="100.00000%"} ![Cyanine dyes.[]{data-label="fig:cyanine_dyes_nacs"}](cyanine_dyes_nacs.png){width="100.00000%"} ![Squaraine dyes.[]{data-label="fig:squaraine_dyes_nacs"}](squaraine_dyes_nacs_renamed.png){width="100.00000%"} ![Aromatic hydrocarbons 1.[]{data-label="fig:aromatic_hydrocarbons_1_nacs"}](aromatic_hydrocarbons_1_nacs.png){width="100.00000%"} ![Aromatic hydrocarbons 2.[]{data-label="fig:aromatic_hydrocarbons_2_nacs"}](aromatic_hydrocarbons_2_nacs.png){width="100.00000%"} \[sec:tapes\] Porphyrin tapes ----------------------------- We will now test the predictions of the 1D model from section \[sec:1d\_model\] for the porphyrin tapes that were synthesized by the Tsuda group [@tsuda2001fully]. These tapes consist of triply-fused zinc-porphyrins (the structure is shown as an inset in Fig. \[fig:porphyrin\_tapes\_nacv\_tdip\_size\_dependence\]). The monomer units are linked through conjugation allowing the electrons to delocalize freely over the entire tape like particles in a box. The delocalization is reflected in the lowering of the excitation energy far into the infrared with increasing length. At the same time, delocalization of the transition density should also impact the magnitude of the electronic non-adiabatic coupling. ![Triply-fused porphyrin oligomers. NACVs between the lowest B$_{1u}$ state and the ground state computed exactly (left) and approximately (center and right).[]{data-label="fig:porphyrin_tapes_nacs"}](porphyrin_tapes_1-3_combined.png){width="100.00000%"} The transition dipole moments and NACVs were computed with long-range corrected TD-DFTB for the lowest $B_{1u}$ state, which is polarized along the long axis of the tape. For the monomer, dimer and trimer the NACVs are depicted in Fig. \[fig:porphyrin\_tapes\_nacs\]. As the conjugation extends over all porphyrin units, the transition dipole moment ${\boldsymbol{\mu}}$ grows approximately linearly with the size of the tape. However, since the transition charges are spread out over a larger area, the non-adiabatic coupling ${\boldsymbol{\tau}}$ grows sublinearly and saturates. The ratio between the lengths of the two vectors is shown in Fig. \[fig:porphyrin\_tapes\_nacv\_tdip\_size\_dependence\]. Since the tapes are also very rigid, one can expect that ultrafast internal conversion through conical intersections, which usually requires some local deformation of the geometry, is not the dominant non-radiative decay channel. Based on this analysis one would expect the long tapes to have extremely high fluorescence quantum yields. ![Porphyrin tapes. The ratio of the non-adiabatic coupling vector to the transition dipole moment (in a.u.) is shown as a function of the length $n$ of the porphyrin tape. Observe the similarity with Fig. \[fig:1d\_model\_fluorophore\]b. Since NACVs computed with localized orbitals are systematically too low, the semiempirical curve had to be scale by a factor of $8$ to agree with the DFT curve.[]{data-label="fig:porphyrin_tapes_nacv_tdip_size_dependence"}](porphyrin_tapes_nacv_tdip_size_dependence.png){width="70.00000%"} However, this is not the case. Article [@cho2002photophysical] actually shows that the non-radiative rate increases rapidly with the length of the tapes so that the fluorescence is quenched as compared to the monomer. At first this appears to contradict the fact that the ratio of the electronic non-adiabatic coupling to the transition dipole moment, $\tau / \mu$, decreases. However, the reason for the fluorescence quenching in the oligomers is the lower energy gap[@tittelbach1995measurements] and the vastly higher density of states. The reduction in the electronic non-adiabatic coupling per porphyrin unit is more than compensated by the increase of accessible final vibrational states. To verify this explanation we computed the radiative and non-radiative rates for the smallest porphyrin tapes using Fermi’s Golden Rule in the harmonic approximation following the steps of Ref. [@valiev2018first]. [^1] For the smallest tapes optimizations and frequency calculations on the ground state and the first excited state with $1B_{1u}$ symmetry at the TD-DFT level of theory are still feasible. The resulting rates and quantum yields are listed in table \[tbl:porphyrin\_tapes\_rates\]: The non-radiative rate jumps by orders of magnitude from the monomer to the dimer and increases further in the trimer. With the non-radiative rate increasing much faster than the radiative rate the quantum yield drops to zero, as observed in experiment. Since the sum over final vibrational states necessarily has to be truncated, the reported non-radiative rates are only a lower limit. Even then it is clear that the non-adiabatic coupling between vibrational states and the sheer density of states is responsible for the fluorescence quenching. T$_n$ $E_{\text{vert}}$ / eV $k_{\text{rad}}$ (s$^{-1}$) $k_{\text{nr}}$ (s$^{-1}$) $QY$ ------- ------------------------ ----------------------------- ---------------------------- --------- 1 2.35 9.6e+05 3.0e-07 1.0e+00 2 1.51 1.5e+07 8.2e+07 1.6e-01 3 1.19 3.4e+07 3.2e+09 1.0e-02 : Dependence of the vertical excitation energy $E_{\text{vert}}$, radiative rate $k_{\text{rad}}$, non-radiative rate $k_{\text{nr}}$ and fluorescence quantum yield $QY = \frac{k_{\text{rad}}}{k_{\text{rad}} + k_{\text{nr}}}$ on the number of porphyrin units $n$ in the triply-fused porphyrin tapes T$_n$.[]{data-label="tbl:porphyrin_tapes_rates"} Discussion ========== Judging the quality of the NAC vectors by visual inspection can be misleading since it suggests there is more agreement than there actually is. The symmetry of NAC vectors is related to the symmetry of the excited state. The relative orientation of the vectors in molecules with high symmetry, is therefore largely determined by the irreducible representation. In trans-butadiene ($C_{2h}$), for instance, only the relative orientation of two out of four vectors is not already fixed by symmetry. According to TD-DFT these two vectors not related by symmetry should be parallel, but the localized orbital method yields an antiparallel orientation (see Fig. \[fig:polyenes\_nacs\]). The orientation is thus entirely wrong and the magnitude is also wrong by a factor of 10. The localized orbital method tends to underestimate the magnitude of the vectors: In ethene the vectors are too short by a factor of 60, in the cyanine dyes by a factor of 3 and in the porphyrin tapes by a factor of 8. The large error for a system as simple as ethene is surprising. Eqn. (\[eqn:nacv\_rho\_deriv\]) and Fig. \[fig:ethene\_nacs\_charges\_trdensity\] showed that the non-adiabatic coupling in the $\pi\pi^*$ state is due to the gradient of the transition density which points along the C-C bond. The transition charge approximation fares a little bit better in predicting the magnitude of the coupling, but it fails in predicting the distribution of the vectors: In the cyanines the excitation is strictly localized on the polyene bridge, but large vectors can be found on two adjacent methyl groups. An extreme example of this are the porphyrin tapes, where the largest vector is placed on the zinc atom, which does not take part in the excitation at all. Comparison between the two approximations is hindered by the fact that one is derived from eqn. (\[eqn:nacv\_rho\_deriv\]) (gradient of transition density) and the other from eqn. (\[eqn:nacv\_grad\]) (gradient of excited state wavefunction), but the two expressions are only equivalent in the basis set limit, and the minimal valence basis of DFTB is a long way off a complete basis set. The LO approximation considers terms that arise because basis functions are attached to the nuclei (Pulay terms) but neglects changes of the excitation coefficients ($\frac{\partial C^{(n)}_{ov}}{\partial \mathbf{R}}$). The TC approximation is independent of the basis set and thus cannot account for Pulay terms. Some of the errors relative to TD-DFT might also be due to the tight-binding approximations: Semiempirical transition charges, excitation energies and molecular orbitals, which enter the expressions for the TC and LO approximations, differ from their ab initio counterparts. However, those sources of error are of minor importance. In fact, if we feed our TC approximation with transition charges that were fitted to reproduce the electrostatic potential of the TD-DFT transition density (using the PSPFFT library[@budiardja2011parallel] for solving the Poisson equation and the CHELPG algorithm [@breneman1990determining]), the resulting vectors are very similar to the tight-binding results. The valence basis set employed in DFTB is also not to blame. With a minimal STO-3G basis set the resulting TD-DFT NAC vectors are indistinguishable from the def2-SVP results. Conclusion ========== Two simple semiempirical approximations for non-adiabatic coupling vectors between excited singlet states and the ground state were implemented in the frame of (LC)-TDDFTB and compared with TD-DFT coupling vectors as benchmarks for a set of planar chromophores with bright $S_1$ states. The TC approximation is based on excitation energies, atom-centered transition charges and geometric information. In the LO approximation the coupling between many-body states is calculated from the coupling vectors between molecular orbitals. While easy to implement and highly efficient, both approximations are not accurate enough to predict the absolute magnitude of the non-adiabatic coupling vector. In particular the LO approximation underestimates couplings by one order of magnitude. Nevertheless, the region in the molecule where the coupling is large can often be identified. For a series of fused porphyrin tapes the reduction in the electronic coupling per porphyrin unit can be explained by the increasing delocalization of the excitation. As a general rule, spreading transition charges over a larger area reduces the electronic non-adiabatic coupling. This however, does not imply that the fluorescence quantum yield may be increased simply by enlarging the delocalization length, since larger $\pi$-system also have larger nuclear non-adiabatic couplings due to the increased density of states. The upshot is that quantitative NAC vectors cannot be obtained with these simple approximations. The implementation of analytical coupling vectors in the spirit of Ref. [@send2010first] can in principle be adapted to tight-binding DFT in analogy to the analytic gradients [@heringer2007analytical] but will require a major effort. The LO approximation is a first step in that direction. Without going to these lengths, the TC approximation might be improved upon by including higher multipoles to represent the transition density more faithfully away from the molecular plane. Acknowledgements {#acknowledgements .unnumbered} ================ A.H. and R.M. acknowledge funding by the European Research Council (ERC) Consolidator Grant DYNAMO (Grant No. 646737). \[sec:appendix\_nacv\_ediff\] Detailed derivation of eqn. (\[eqn:nacv\_ediff\]) ================================================================================ Starting from the electronic Schrödinger equation, $${\hat{H}} {\mid\! \Psi\! \rangle} = E_n {\mid\! \Psi_n\! \rangle},$$ we differentiate on both sides with respect to the nuclear coordinates $\boldsymbol{R}$, which leads to $$\frac{\partial {\hat{H}}}{\partial {\boldsymbol{R}}} {\mid\! \Psi_n\! \rangle} + {\hat{H}} {\mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle} = \frac{\partial E_n}{\partial {\boldsymbol{R}}} {\mid\! \Psi_n\! \rangle} + E_n {\mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle}.$$ Multiplying on both sides by ${\langle\! \Psi_m\! \mid} \times$ and using the orthogonality of Born-Oppenheimer states belonging to the same geometry, ${\langle\! \Psi_m\! \mid\! \Psi_n\! \rangle} = \delta_{mn}$, and the fact the ${\hat{H}}$ is Hermitian, ${\langle\! \Psi_m\! \mid} {\hat{H}} = E_m {\langle\! \Psi_m\! \mid}$, gives $${\langle\! \Psi_m\! \mid} \frac{\partial {\hat{H}}}{\partial {\boldsymbol{R}}} {\mid\! \Psi_n\! \rangle} + E_m {\langle\! \Psi_m\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle} = \frac{\partial E_n}{\partial {\boldsymbol{R}}} \delta_{mn} + E_n {\langle\! \Psi_n\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle}.$$ For $m \neq n$ this can be rearranged into $${\langle\! \Psi_m\! \mid\! \frac{\partial \Psi_n}{\partial {\boldsymbol{R}}}\! \rangle} = \frac{{\langle\! \Psi_m\! \mid} \frac{\partial {\hat{H}}}{\partial {\boldsymbol{R}}} {\mid\! \Psi_n\! \rangle}}{E_n - E_m} \label{eqn:nacv_ediff_derivation}$$ \[sec:partial\_integration\] Detailed derivation of eqn. (\[eqn:nacv\_rho\_deriv\]) ==================================================================================== The single-electron part of the electron-nuclear attraction is $$V_{ne}({\boldsymbol{r}}) = \sum_A \frac{-Z_A}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert}.$$ $$\begin{split} {\boldsymbol{\tau}}^A_{mn} &= \frac{1}{E_n - E_m} \int d{\boldsymbol{r}} \left(\nabla_A V_{en}({\boldsymbol{r}}) \right) \rho_{mn}({\boldsymbol{r}}) \\ &= \frac{-Z_A}{E_n - E_m} \int d{\boldsymbol{r}} \left(\nabla_A \frac{1}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \right) \rho_{mn}({\boldsymbol{r}}) \\ &= \frac{-Z_A}{E_n - E_m} \int d{\boldsymbol{r}} \left((-\nabla) \frac{1}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \right) \rho_{mn}({\boldsymbol{r}}) \end{split}$$ By the partial integration rule $\int_V (\nabla f) g = [f g]_{\partial V} - \int_V f (\nabla g)$ this becomes $$\begin{split} {\boldsymbol{\tau}}^A_{mn} &= \frac{Z_A}{E_n - E_m} \left\{ \cancel{\left[ \frac{\rho_{mn}({\boldsymbol{r}})}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \right]_{\infty}} - \int d{\boldsymbol{r}} \frac{\nabla \rho_{mn}({\boldsymbol{r}})}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \right\} \\ &= - \frac{Z_A}{E_n - E_m} \int d{\boldsymbol{r}} \frac{\nabla \rho_{mn}({\boldsymbol{r}})}{\vert {\boldsymbol{R}}_A - {\boldsymbol{r}} \vert} \end{split}$$ \[sec:dipole\_constant\] Transition dipole moment in the linear molecule ======================================================================== The following little calculation shows that the transition dipole moment is a constant, independently of the number of atoms $2 M$: $$\begin{split} \vert {\boldsymbol{\mu}} \vert &= \sum_{i=0}^{2M-1} q_i x_i = \sum_{i} (-1)^i \frac{q}{M} i h = \ q h \frac{1}{M} \sum_{i=0}^{2M-1} (-1)^i i \\ &= q h \frac{1}{M} \left\{ \sum_{j=0}^M (-1)^{2j} 2j + \sum_{j=0}^{M-1} (-1)^{2j+1} (2j+1) \right\} \\ &= qh \frac{1}{M} \left\{ \sum_{j=0}^{M-1} 2j - \sum_{j=0}^{M-1} 2 j + 2 M - \sum_{j=0}^{M-1} 1 \right\} = q h \end{split}$$ [^1]: Calculation of rates is based on the following approximations: The $S_0$ and $S_1$ potential energy surfaces are harmonic, share the same normal modes and frequencies but have different equilibrium geometries (shifted harmonic oscillators). Frequencies and normal modes are determined from a frequency calculation using $\omega$B97XD/dev2-SVG at the $S_0$ minimum. The Huang-Rhys factors are deduced from the gradient on $S_1$ at the Franck-Condon point. Total rates are obtained by summing over all transitions that start in the initial vibrational ground state on $S_1$ and lead to a final vibrational state on $S_0$. Radiative transitions may lead to any vibrational state lower in energy, while in non-radiative transitions the final vibrational states on $S_0$ have to be approximately isoenergetic with the initial vibrational state on $S_1$. Following Ref. [@santoro2007effective] final states are grouped into classes C$_n$ depending on the number $n$ of simultaneously excited modes. The summation is limited to classes $C_0-C_8$, which captures most of the radiative rate and a large fraction of the non-radiative rate. Modes are sorted in decreasing order by Franck-Condon factors and the number of modes from which excitations are allowed, is reduced, until there are no more than $10^8$ elements per class left.	{ "pile_set_name": "ArXiv" }
--- abstract: \| An eight-parametric family of complex connections on a class complex manifolds with Norden metric is introduced. The form of the curvature tensor with respect to each of these connections is obtained. The conformal group of the considered connections is studied and some conformal invariants are obtained. 2010 Mathematics Subject Classification: 53C15, 53C50. Keywords: complex connection, complex manifold, Norden metric. author: - Marta Teofilova title: Complex connections on conformal Kähler manifolds with Norden metric --- Introduction ============ Almost complex manifolds with Norden metric were first studied by A. P. Norden [@N] and are introduced in [@Gri-Mek-Dje] as generalized $B$-manifolds. A classification of these manifolds with respect to the covariant derivative of the almost complex structure is obtained in [@Gan] and two equivalent classifications are given in [@Gan-Mih; @Gan-Gri-Mih2]. An important problem in the geometry of almost complex manifolds with Norden metric is the study of linear connections preserving the almost complex structure or preserving both, the structure and the metric. The first ones are called almost complex connections, and the second ones are known as natural connections. A special type of natural connection is the canonical one. In [@Gan-Mih] it is proved that on an almost complex manifold with Norden metric there exists a unique canonical connection. The canonical connection (called also the $B$-connection) and its conformal group on a conformal Kähler manifold with Norden metric are studied in [@Gan-Gri-Mih2]. In [@Teo3] we have obtained a two-parametric family of complex connections on a conformal Kähler manifold with Norden metric and have proved that the curvature tensors corresponding to these connections coincide with the curvature tensor of the canonical connection. In the present work we continue our research on complex connections on complex manifolds with Norden metric by focusing our attention on the class of the conformal Kähler manifolds, i.e. manifolds which are conformally equivalent to Käher manifolds with Norden metric. We introduce an eight-parametric family of complex connections on such manifolds and consider their curvature properties. We also study the conformal group of these connections and obtain some conformal invariants. In the last section we give an example of a four-dimensional conformal Kähler manifold with Norden metric, on which the considered complex connections are flat. Preliminaries ============= Let $(M,J,g)$ be a $2n$-dimensional almost complex manifold with Norden metric, i.e. $J$ is an almost complex structure and $g$ is a pseudo-Riemannian metric on $M$ such that $$\label{11} J^{2}x=-x,\qquad g(Jx,Jy)=-g(x,y)$$ for all differentiable vector fields $x$, $y$ on $M$, i.e. $x,y\in \mathfrak{X}(M)$. The associated metric $\widetilde{g}$ of $g$ is given by $\widetilde{g}(x,y)=g(x,Jy)$ and is a Norden metric, too. Both metrics are necessarily neutral, i.e. of signature $(n,n)$. If $\nabla $ is the Levi-Civita connection of the metric $g$, the fundamental tensor field $F$ of type $(0,3)$ on $M$ is defined by $$\label{F} F(x,y,z)=g\left((\nabla _{x}J)y,z\right)$$ and has the following symmetries $$\label{Fp} F(x,y,z)=F(x,z,y)=F(x,Jy,Jz).$$ Let $\left\{ e_{i}\right\} $ ($i=1,2,\ldots ,2n$) be an arbitrary basis of $ T_{p}M$ at a point $p$ of $M$. The components of the inverse matrix of $g$ are denoted by $g^{ij}$ with respect to the basis $\left\{ e_{i}\right\} $. The Lie 1-forms $\theta $ and $\theta^{\ast}$ associated with $F$, and the Lie vector $\Omega$, corresponding to $\theta$, are defined by, respectively $$\label{1-3} \theta (x)=g^{ij}F(e_{i},e_{j},x), \qquad \theta^{\ast}=\theta \circ J, \qquad \theta (x)=g(x,\Omega ).$$ The Nijenhuis tensor field $N$ for $J$ is given by [@Ko-No] $$N(x,y)=[Jx,Jy]-[x,y]-J[Jx,y]-J[x,Jy].$$ It is known [@N-N] that the almost complex structure is complex if and only if it is integrable, i.e. iff $N=0$. A classification of the almost complex manifolds with Norden metric is introduced in [@Gan], where eight classes of these manifolds are characterized according to the properties of $F$. The three basic classes $\mathcal{W}_{i}$ ($i=1,2,3$) are given by $\bullet$ the class $\mathcal{W}_{1}$: $$\label{w1} \begin{array}{l} F(x,y,z)=\frac{1}{2n}\left[ g(x,y)\theta (z)+g(x,Jy)\theta (Jz)\right. \medskip \\ \quad \qquad \qquad \quad \left. +g(x,z)\theta (y)+g(x,Jz)\theta (Jy)\right]; \end{array}$$ $\bullet$ the class $\mathcal{W}_{2}$ of the special complex manifolds with Norden metric: $$\label{w2} F(x,y,Jz)+F(y,z,Jx)+F(z,x,Jy)=0,\quad \theta =0 \quad\Leftrightarrow\quad N=0,\quad \theta=0;$$ $\bullet$ the class $\mathcal{W}_{3}$ of the quasi-Kähler manifolds with Norden metric: $$\label{w3} F(x,y,z)+F(y,z,x)+F(z,x,y)=0.$$ The special class $\mathcal{W}_{0}$ of the Kähler manifolds with Norden metric is characterized by the condition $F=0$ ($\nabla J=0$) and is contained in each one of the other classes. Let $R$ be the curvature tensor of $\nabla $, i.e. $R(x,y)z=\nabla _{x}\nabla _{y}z-\nabla _{y}\nabla _{x}z-\nabla _{\left[ x,y\right] }z$. The corresponding (0,4)-type tensor is defined by $R(x,y,z,u)=g\left( R(x,y)z,u\right)$. A tensor $L$ of type (0,4) is called a curvature-like tensor if it has the properties of $R$, i.e. $$\label{L} \begin{array}{l} L(x,y,z,u)=-L(y,x,z,u)=-L(x,y,u,z),\medskip\\ L(x,y,z,u)+L(y,z,x,u)+L(z,x,y,u)=0. \end{array}$$ The Ricci tensor $\rho(L)$ and the scalar curvatures $\tau(L)$ and $ \tau^{\ast}(L)$ of $L$ are defined by: $$\begin{array}{c} \rho(L)(y,z)=g^{ij}L(e_{i},y,z,e_{j}),\medskip\\ \tau(L)=g^{ij}\rho(L)(e_{i},e_{j}),\quad \tau^{\ast}(L)=g^{ij}\rho(L) (e_{i},Je_{j}). \end{array} \label{Ricci, tao}$$ A curvature-like tensor $L$ is called a Kähler tensor if $$\label{Ka} L(x,y,Jz,Ju) = - L(x,y,z,u).$$ Let $S$ be a tensor of type (0,2). We consider the following tensors [@Gan-Gri-Mih2]: $$\label{psi} \begin{array}{l} \psi_{1}(S)(x,y,z,u) =g(y,z)S(x,u)-g(x,z)S(y,u) \smallskip\\ \phantom{\psi_{1}(S)(x,y,z,u)}+ g(x,u)S(y,z) - g(y,u)S(x,z), \medskip\\ \psi_{2}(S)(x,y,z,u) =g(y,Jz)S(x,Ju) - g(x,Jz)S(y,Ju) \smallskip\\ \phantom{\psi_{1}(S)(x,y,z,u)} + g(x,Ju)S(y,Jz) - g(y,Ju)S(x,Jz), \medskip\\ \pi_{1}=\frac{1}{2}\psi_{1}(g), \qquad \pi_{2}=\frac{1}{2}\psi_{2}(g),\qquad \pi_{3}=-\psi_{1}(\widetilde{g})=\psi_{2}(\widetilde{g}). \end{array}$$ The tensor $\psi_{1}(S)$ is curvature-like if $S$ is symmetric, and the tensor $\psi_{2}(S)$ is curvature-like if $S$ is symmetric and hybrid with respect to $J$, i.e. $S(x,Jy)=S(y,Jx)$. In the last case the tensor $\{\psi_1 - \psi_2\}(S)$ is Kählerian. The tensors $\pi_{1} - \pi_{2}$ and $\pi_{3}$ are also Kählerian. The usual conformal transformation of the Norden metric $g$ (conformal transformation of type I [@Gan-Gri-Mih2]) is defined by $$\label{conf} \overline{g}=e^{2u}g,$$ where $u$ is a pluriharmonic function, i.e. the 1-form $du\circ J$ is closed. A $\mathcal{W}_1$-manifold with closed 1-forms $\theta$ and $\theta^{\ast}$ (i.e. $\mathrm{d}\theta=\mathrm{d}\theta^\ast=0$) is called a conformal Kähler manifold with Norden metric. Necessary and sufficient conditions for a $\mathcal{W}_1$-manifold to be conformal Kählerian are: $$\label{cK} (\nabla_x\theta)y=(\nabla_y\theta)x,\qquad (\nabla_x\theta)Jy=(\nabla_y\theta)Jx.$$ The subclass of these manifolds is denoted by $\mathcal{W}_{1}^{\hspace{0.01in}0}$. It is proved [@Gan-Gri-Mih2] that a $\mathcal{W}_{1}^{\hspace{0.01in}0}$-manifold is conformally equivalent to a Kähler manifold with Norden metric by the transformation (\[conf\]). It is known that on a pseudo-Riemannian manifold $M$ ($\dim M=2n \geq 4$) the conformal invariant Weyl tensor has the form $$\label{Weyl} W(R)=R-\frac{1}{2(n-1)}\big \{\psi_{1}(\rho)-\frac{\tau}{2n-1}\pi_{1}\big \}.$$ Let $L$ be a Kähler curvature-like tensor on an almost complex manifold with Norden metric $(M,J,g)$, $\dim M=2n\geq 6$. Then the Bochner tensor $B(L)$ for $L$ is defined by [@Gan-Gri-Mih2]: $$\label{Bochner} \begin{array}{l} B(L)= L - \frac{1}{2(n-2)}\big\{\psi_{1}-\psi_{2}\big\}\big(\rho(L)\big)\medskip\\ \phantom{B(L)}+ \frac{1}{4(n-1)(n-2)}\big\{\tau(L)\big(\pi_{1}-\pi_{2}\big) + \tau^{\ast}(L)\pi_{3}\big\}. \end{array}$$ Complex Connections on $\mathcal{W}_1$-manifolds ================================================ \[def-complex\] A linear connection $\nabla^{\prime}$ on an almost complex manifold $(M,J)$ is said to be almost complex if $\nabla^{\prime}J=0$. We introduce an eight-parametric family of complex connections in the following On a $\mathcal{W}_1$-manifold with Norden metric there exists an eight-parametric family of complex connections $\nabla^{\prime}$ defined by $$\label{2-1} \begin{array}{l} \nabla_x^{\prime}y = \nabla_x y + Q(x,y), \end{array}$$ where the deformation tensor $Q(x,y)$ is given by $$\label{2-2} \begin{array}{l} Q(x,y)= \frac{1}{2n}\left[\theta(Jy)x-g(x,y)J\Omega\right]\medskip\\ \hspace{0.065in}\phantom{Q(x,y)}+ \frac{1}{n}\left\{\lambda_1\theta(x)y+\lambda_2\theta(x)Jy +\lambda_3\theta(Jx)y +\lambda_4\theta(Jx)Jy\right.\medskip\\ \phantom{Q(x,y)}\hspace{0.06in}+\lambda_5\left[\theta(y)x-\theta(Jy)Jx\right] + \lambda_6\left[\theta(y)Jx+\theta(Jy)x\right]\medskip\\ \left.\hspace{0.06in}\phantom{Q(x,y)}+\lambda_7\left[g(x,y)\Omega-g(x,Jy)J\Omega\right]+\lambda_8\left[g(x,Jy)\Omega+g(x,y)J\Omega\right]\right\}, \end{array}$$ $\lambda_i\in \mathbb{R}$, $i=1,2,...,8$. By (\[w1\]), (\[2-1\]) and (\[2-2\]) we verify that $(\nabla^{\prime}_xJ)y = \nabla^{\prime}_xJy - J\nabla^{\prime}_xy=0$, and hence the connections $\nabla^{\prime}$ are complex for any $\lambda_i\in \mathbb{R}$, $i=1,2,...,8$. Let us remark that the two-parametric family of complex connections obtained for $\lambda_1=\lambda_4$, $\lambda_3=-\lambda_2$, $\lambda_5=\lambda_7=0$, $\lambda_8=-\lambda_6=\frac{1}{4}$, is studied in [@Teo3]. Let $T^{\prime}$ be the torsion tensor of $\nabla^{\prime}$, i.e. $T^{\prime}(x,y)=\nabla^{\prime}_xy - \nabla^{\prime}_xy - [x,y]$. Taking into account that the Levi-Civita connection $\nabla$ is symmetric, we have $T^{\prime}(x,y)=Q(x,y)-Q(y,x)$. Then by (\[2-2\]) we obtain $$\label{T} \begin{array}{l} T^{\prime}(x,y)=\frac{1}{n}\left\{\left(\lambda_1-\lambda_5\right)\left[\theta(x)y-\theta(y)x\right] +\left(\lambda_2-\lambda_6\right)\left[\theta(x)Jy-\theta(y)Jx\right]\right.\medskip\\ \qquad\quad\hspace{0.02in}\left.+\left(\lambda_3-\lambda_6-\frac{1}{2}\right)\left[\theta(Jx)y-\theta(Jy)x\right] +\left(\lambda_4+\lambda_5\right)\left[\theta(Jx)Jy-\theta(Jy)Jx\right] \right\}. \end{array}$$ It is easy to verify the following $$\underset{x,y,z}{\mathfrak{S}}T^{\prime}(x,y,z)=\underset{x,y,z}{\mathfrak{S}}T^{\prime}(Jx,Jy,z)=\underset{x,y,z}{\mathfrak{S}}T^{\prime}(x,y,Jz)=0,$$ where $\mathfrak{S}$ is the cyclic sum by the arguments $x,y,z$. Next, we obtain necessary and sufficient conditions for the complex connections $\nabla^{\prime}$ to be symmetric (i.e. $T^{\prime}=0$). The complex connections $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]) are symmetric on a $\mathcal{W}_1$-manifold with Norden metric if and only if $\lambda_1=-\lambda_4=\lambda_5$, $\lambda_2=\lambda_3-\frac{1}{2}=\lambda_6$. Then, by putting $\lambda_1=-\lambda_4=\lambda_5=\mu_1$, $\lambda_2=\lambda_6=\lambda_3-\frac{1}{2}=\mu_2$, $\lambda_7=\mu_3$, $\lambda_8=\mu_4$ in (\[2-2\]) we obtain a four-parametric family of complex symmetric connections $\nabla^{\prime\prime}$ on a $\mathcal{W}_1$-manifold which are defined by $$\label{sym} \begin{array}{l} \nabla^{\prime\prime}_x y= \nabla_x y + \frac{1}{2n}\left[\theta(Jx)y+\theta(Jy)x-g(x,y)J\Omega\right]\medskip\\ \phantom{\nabla^{\prime}_x y}+\frac{1}{n}\left\{\mu_1\left[\theta(x)y+\theta(y)x-\theta(Jx)Jy-\theta(Jy)Jx\right]\right.\medskip\\ \phantom{\nabla^{\prime}_x y}+\mu_2\left[\theta(Jx)y+\theta(Jy)x+\theta(x)Jy+\theta(y)Jx\right]\medskip\\ \phantom{\nabla^{\prime}_x y}+\left.\mu_3\left[g(x,y)\Omega-g(x,Jy)J\Omega\right]+\mu_4\left[g(x,Jy)\Omega+g(x,y)J\Omega\right]\right\}. \end{array}$$ The well-known Yano connection [@Ya1; @Ya2] on a $\mathcal{W}_1$-manifold is obtained from (\[sym\]) for $\mu_1=\mu_3=0$, $\mu_4=-\mu_2=\frac{1}{4}$. \[def-nat\] A linear connection $\nabla^{\prime}$ on an almost complex manifold with Norden metric $(M,J,g)$ is said to be natural if $\nabla^{\prime}J=\nabla^{\prime}g=0$ (or equivalently, $\nabla^{\prime}g=\nabla^{\prime}\widetilde{g}=0$). From (\[2-1\]) and (\[2-2\]) we compute the covariant derivatives of $g$ and $\widetilde{g}$ with respect to the complex connections $\nabla^{\prime}$ as follows $$\label{2-5} \begin{array}{l} \left(\nabla^{\prime}_x g\right)(y,z)=-Q(x,y,z)-Q(x,z,y)\medskip\\ =-\frac{1}{n}\left\{ 2\left[\lambda_1\theta(x)g(y,z)+\lambda_2\theta(x)g(y,Jz)+\lambda_3\theta(Jx)g(y,z)\right.\right.\medskip \\ \left.+\lambda_4\theta(Jx)g(y,Jz)\right]+(\lambda_5+\lambda_7)\left[\theta(y)g(x,z)+\theta(z)g(x,y)\right.\medskip\\ \left.-\theta(Jy)g(x,Jz)-\theta(Jz)g(x,Jy)\right]+(\lambda_6+\lambda_8)\left[\theta(y)g(x,Jz)\right.\medskip\\ \left.+\theta(z)g(x,Jy)+\theta(Jy)g(x,z)+\theta(Jz)g(x,y)\right]\left.\right\},\medskip\\ \left(\nabla^{\prime}_x \widetilde{g}\right)(y,z)=-Q(x,y,Jz)-Q(x,Jz,y). \end{array}$$ Then, by (\[2-5\]) we get the following The complex connections $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]) are natural on a $\mathcal{W}_1$-manifold if and only if $\lambda_1=\lambda_2=\lambda_3=\lambda_4=0$, $\lambda_7=-\lambda_5$, $\lambda_8=-\lambda_6$. If we put $\lambda_8=-\lambda_6=s$, $\lambda_7=-\lambda_5=t$, $\lambda_i=0$, $i=1,2,3,4$, in (\[2-2\]), we obtain a two-parametric family of natural connections $\nabla^{\prime\prime\prime}$ defined by $$\label{nabla-n} \begin{array}{l} \nabla^{\prime\prime\prime}_x y = \nabla_x y +\frac{1-2s}{2n}\left[\theta(Jy)x-g(x,y)J\Omega\right] +\frac{1}{n}\left\{s\left[g(x,Jy)\Omega - \theta(y)Jx\right]\right.\medskip\\ \phantom{\nabla^{\prime}_x y}\left.+t\left[g(x,y)\Omega - g(x,Jy)J\Omega - \theta(y)x+\theta(Jy)Jx\right]\right\}. \end{array}$$ The well-known canonical connection [@Gan-Mih] (or $B$-connection [@Gan-Gri-Mih2]) on a $\mathcal{W}_1$-manifold with Norden metric is obtained from (\[nabla-n\]) for $s=\frac{1}{4}$, $t=0$. We give a summery of the obtained results in the following table -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Connection type Symbol Parameters -------------------------------------------------------------------------- ------------------------------- --------------------------------------------------------------------------------- $\begin{array}{l}\text{Complex}\end{array}$ $\nabla^{\prime}$ $\lambda_i\in\mathbb{R}$, $i=1,2,...,8$. $\begin{array}{l}\text{Complex}\smallskip\\ \text{symmetric}\end{array}$ $\nabla^{\prime\prime}$ $\begin{array}{c}\mu_i,\hspace{0.03in} i=1,2,3,4, \smallskip\\ \mu_1=\lambda_1=-\lambda_4=\lambda_5,\hspace{0.03in} \mu_2=\lambda_2=\lambda_6=\lambda_3-\frac{1}{2},\smallskip\\ \mu_3=\lambda_7,\hspace{0.03in} \mu_4=\lambda_8\end{array}$ $\begin{array}{l}\text{Natural}\end{array}$ $\nabla^{\prime\prime\prime}$ $\begin{array}{c}s,t,\smallskip\\ s=\lambda_8=-\lambda_6, \hspace{0.03in} t = \lambda_7=-\lambda_5,\smallskip\\ \lambda_i=0,\hspace{0.03in} i =1,2,3,4.\end{array}$ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Our next aim is to study the curvature properties of the complex connections $\nabla^{\prime}$. Let us first consider the natural connection $\nabla^0$ obtained from (\[nabla-n\]) for $s=t=0$, i.e. $$\label{nabla-0} \nabla^{0}_x y = \nabla_x y + \frac{1}{2n}\left[\theta(Jy)x - g(x,y)J\Omega\right].$$ This connection is a semi-symmetric metric connection, i.e. a connection of the form $\nabla_x y + \omega(y)x - g(x,y)U$, where $\omega$ is a 1-form and $U$ is the corresponding vector of $\omega$, i.e. $\omega(x)=g(x,U)$. Semi-symmetric metric connections are introduced in [@Ha] and studied in [@Im; @Ya3]. The form of the curvature tensor of an arbitrary connection of this type is obtained in [@Ya3]. The geometry of such connections on almost complex manifolds with Norden metric is considered in [@Si]. Let us denote by $R^0$ the curvature tensor of $\nabla^0$, i.e. $R^0(x,y)z=\nabla^0_x\nabla^0_y z - \nabla^0_y\nabla^0_x z - \nabla^0_{[x,y]}z$. The corresponding tensor of type (0,4) is defined by $R^0(x,y,z,u)=g(R^0(x,y,)z,u)$. According to [@Ya3] it is valid \[2\] On a $\mathcal{W}_1$-manifold with closed 1-form $\theta^{\ast}$ the Kähler curvature tensor $R^0$ of $\nabla^0$ defined by (\[nabla-0\]) has the form $$\label{R0} R^0=R - \frac{1}{2n}\psi_1(P),$$ where $$\label{R01} P(x,y)=\left(\nabla_x \theta\right)Jy + \frac{1}{2n}\theta(x)\theta(y)+\frac{\theta(\Omega)}{4n}g(x,y)+\frac{\theta(J\Omega)}{2n}g(x,Jy).$$ Since the Weyl tensor $W(\psi_1(S))=0$ (where $S$ is a symmetric (0,2)-tensor), from (\[R0\]) and (\[R01\]) we conclude that $$\label{WW} W(R^0)=W(R).$$ Thus, the last equality implies \[thW\] Let $(M,J,g)$ be a $\mathcal{W}_1$-manifold with closed 1-form $\theta^{\ast}$, and $\nabla^0$ be the natural connection defined by (\[nabla-0\]). Then, the Weyl tensor is invariant by the transformation $\nabla \rightarrow \nabla^0$. Further in this section we study the curvature properties of the complex connections $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]). Let us denote by $R^{\prime}$ the curvature tensors corresponding to these connections. If a linear connection $\nabla^\prime$ and the Levi-Civita connection $\nabla$ are related by an equation of the form (\[2-1\]), then, because of $\nabla g=0$, their curvature tensors of type (0,4) satisfy $$\label{33} \begin{array}{l} g(R^\prime(x,y)z,u) = R(x,y,z,u) + (\nabla_x Q)(y,z,u) - (\nabla_y Q)(x,z,u)\medskip\\ \phantom{g(R^\prime(x,y)z,u)} + Q(x,Q(y,z),u) - Q(y,Q(x,z),u), \end{array}$$ where $Q(x,y,z) = g(Q(x,y),z)$. Then, by (\[2-1\]), (\[2-2\]), (\[R0\]), (\[R01\]), (\[33\]) we obtain the relation between $R^{\prime}$ and $R^0$ as follows $$\label{h1} \begin{array}{l} R^{\prime}(x,y,z,u) = R^{0}(x,y,z,u) +g(y,z)A_1(x,u) - g(x,z)A_1(y,u)\medskip\\ + g(x,u)A_2(y,z) - g(y,u)A_2(x,z) -g(y,Jz)A_1(x,Ju)\medskip\\ + g(x,Jz)A_1(y,Ju)-g(x,Ju)A_2(y,Jz)+g(y,Ju)A_2(x,Jz)\medskip\\ +\left[\frac{\lambda_5\lambda_7 - \lambda_6\lambda_8}{n^2}\theta(\Omega) + \frac{\lambda_7 - \lambda_5 + 2(\lambda_5\lambda_8+\lambda_6\lambda_7)}{2n^2}\theta(J\Omega) \right]\{\pi_1-\pi_2\}(x,y,z,u)\medskip\\ -\left[\frac{\lambda_5\lambda_8 + \lambda_6\lambda_7}{n^2}\theta(\Omega) - \frac{\lambda_6 - \lambda_8 + 2(\lambda_5\lambda_7-\lambda_6\lambda_8)}{2n^2}\theta(J\Omega)\right]\pi_3(x,y,z,u), \end{array}$$ where $$\label{h2} \begin{array}{l} A_1(x,y) = \frac{\lambda_7}{n}\left\{\left(\nabla_x\theta\right)y+\frac{\lambda_7}{n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)]\right\}\medskip\\ \phantom{A_1(x,y)}+\frac{\lambda_8}{n}\left\{\left(\nabla_x\theta\right)Jy + \frac{1-2\lambda_8}{2n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)]\right\}\medskip\\ \phantom{A_1(x,y)}+\frac{\lambda_7(4\lambda_8-1)}{2n^2}[\theta(x)\theta(Jy)+\theta(Jx)\theta(y)],\bigskip\\ A_2(x,y)=-\frac{\lambda_5}{n}\left\{\left(\nabla_x\theta\right)y - \frac{\lambda_5}{n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)]\right\}\medskip\\ \phantom{A_2(x,y)}-\frac{\lambda_6}{n}\left\{\left(\nabla_x\theta\right)Jy+\frac{1+2\lambda_6}{2n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)]\right\}\medskip\\ \phantom{A_2(x,y)}+\frac{\lambda_5(4\lambda_6+1)}{2n^2}[\theta(x)\theta(Jy)+\theta(Jx)\theta(y)]. \end{array}$$ We are interested in necessary and sufficient conditions for $R^{\prime}$ to be a Kähler curvature-like tensor, i.e. to satisfy (\[L\]) and (\[Ka\]). From (\[psi\]), (\[h1\]) and (\[h2\]) it follows that $R^{\prime}$ is Kählerian if and only if $A_1(x,y)=A_2(x,y)$. Hence we obtain \[1\] Let $(M,J,g)$ be a conformal Kähler manifold with Norden metric, and $\nabla^{\prime}$ be the complex connection defined by (\[2-1\]) and (\[2-2\]). Then, $R^{\prime}$ is a Kähler curvature-like tensor if and only if $\lambda_7=-\lambda_5$ and $\lambda_8=-\lambda_6$. In this case, from (\[2-1\]) and (\[2-2\]) we obtain a six-parametric family of complex connections $\nabla^{\prime}$ whose curvature tensors $R^{\prime}$ have the form $$\begin{array}{l}\label{Rpr} R^{\prime} = R^0 + \frac{\lambda_7}{n}\left\{\psi_1 - \psi_2\right\}(S_1) + \frac{\lambda_8}{n}\left\{\psi_1-\psi_2\right\}(S_2) \medskip\\ \phantom{R^{\prime}}+ \frac{\lambda_7(4\lambda_8-1)}{2n^2}\left\{\psi_1-\psi_2\right\}(S_3) + \frac{\lambda_7(1-2\lambda_8)\theta(J\Omega)}{n^2}\left\{\pi_1-\pi_2\right\}\medskip\\ \phantom{R^{\prime}}+\frac{2\lambda_7\lambda_8\theta(\Omega)}{n^2}\pi_3, \end{array}$$ where $R^0$ is given by (\[R0\]), (\[R01\]), and $$\label{111} \begin{array}{l} S_1(x,y) = \left(\nabla_x\theta\right)y + \frac{\lambda_7}{n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)] - \frac{\lambda_7\theta(\Omega)}{2n}g(x,y) \medskip\\ \phantom{S_1(x,y)} + \frac{\lambda_7\theta(J\Omega)}{2n}g(x,Jy),\medskip\\ S_2(x,y) = \left(\nabla_x\theta\right)Jy + \frac{1-2\lambda_8}{2n}[\theta(x)\theta(y)-\theta(Jx)\theta(Jy)]\medskip\\ \phantom{S_2(x,y)}+\frac{\lambda_8\theta(\Omega)}{2n}g(x,y) + \frac{(1-\lambda_8)\theta(J\Omega)}{2n}g(x,Jy),\medskip\\ S_3(x,y) = \theta(x)\theta(Jy) + \theta(Jx)\theta(y). \end{array}$$ From (\[Rpr\]), (\[111\]) and Theorem \[1\] we get \[cor\] Let $(M,J,g)$ be a conformal Kähler manifold with Norden metric and $\nabla^{\prime}$ be the eight-parametric family of complex connections defined by (\[2-1\]) and (\[2-2\]). Then $R^{\prime}=R^0$ if and only if $\lambda_i=0$ for $i=5,6,7,8$. Let us remark that by putting $\lambda_i=0$ for $i=1,2,5,6,7,8$ in (\[2-2\]) we obtain a two-parametric family of complex connections whose Kähler curvature tensors coincides with $R^0$ on a $\mathcal{W}_1$-manifold with closed 1-form $\theta^{\ast}$. Theorem \[thW\] and Corollary \[cor\] imply On a conformal Kähler manifold with Norden metric the Weyl tensor is invariant by the transformation of the Levi-Civita connection in any of the complex connection $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]) for $\lambda_i=0$, $i=5,6,7,8$. Since for the Bochner tensor of $\{\psi_1-\psi_2\}(S)$ it is valid $B\left(\{\psi_1-\psi_2\}(S)\right)=0$, where $S$ is symmetric and hybrid with respect to $J$, from Theorem \[1\] and (\[psi\]) it follows $$\label{BR0} B(R^{\prime}) = B(R^0).$$ By this way we proved the truthfulness of the following \[thB\] Let $(M,J,g)$ be a conformal Kähler manifold with Norden metric, $R^{\prime}$ be the curvature tensor of $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]) for $\lambda_7=-\lambda_5$, $\lambda_8=-\lambda_6$, and $R^0$ be the curvature tensor of $\nabla^0$ given by (\[nabla-0\]). Then the Bochner tensor is invariant by the transformations $\nabla^0\rightarrow\nabla^{\prime}$. Conformal transformations of complex connections ================================================ In this section we study usual conformal transformations of the complex connections $\nabla^{\prime}$ defined in the previous section. Let $(M,J,g)$ and $(M,J,\bar{g})$ be conformally equivalent almost complex manifolds with Norden metric by the transformation (\[conf\]). It is known that the Levi-Civita connections $\nabla$ and $\overline{\nabla}$ of $g$ and $\overline{g}$, respectively, are related as follows $$\label{con-trans1} \overline{\nabla}_{x}y = \nabla_{x}y + \sigma(x)y + \sigma(y)x - g(x,y)\Theta,$$ where $\sigma(x)=du(x)$ and $\Theta=\textrm{grad}\hspace{0.02in} \sigma$, i.e. $\sigma(x)=g(x,\Theta)$. Let $\overline{\theta}$ be the Lie 1-form of $(M,J,\overline{g})$. Then by (\[1-3\]) and (\[con-trans1\]) we obtain $$\label{theta-bar} \overline{\theta} = \theta + 2n\big(\sigma\circ J\big),\qquad\quad \overline{\Omega}=e^{-2u}\big(\Omega + 2nJ\Theta\big).$$ It is valid the following \[lemma1\] Let $(M,J,g)$ be an almost complex manifold with Norden metric and $(M,J,\overline{g})$ be its conformally equivalent manifold by the transformation (\[conf\]). Then, the curvature tensors $R$ and $\overline{R}$ of $\nabla$ and $\overline{\nabla}$, respectively, are related as follows $$\label{Rbar} \overline{R}=e^{2u}\big\{R-\psi_{1}\big(V\big) - \pi_{1}\sigma\big(\Theta\big)\big\},$$ where $V(x,y)=\big(\nabla_{x}\sigma\big)y - \sigma(x)\sigma(y)$. Let us first study the conformal group of the natural connection $\nabla^0$ given by (\[nabla-0\]). Equalities (\[nabla-0\]) and (\[con-trans1\]) imply that its conformal group is defined analytically by $$\label{nabla0-bar} \overline{\nabla}^{\hspace{0.02in}0}_{x}y = \nabla^{0}_{x}y + \sigma(x)y.$$ It is known that if two linear connections are related by an equation of the form (\[nabla0-bar\]), where $\sigma$ is a 1-form, then the curvature tensors of these connections coincide iff $\sigma$ is closed. Hence, it is valid \[thR0\] Let $(M,J,g)$ be a $\mathcal{W}_1$-manifold with closed 1-form $\theta^{\ast}$. Then the curvature tensor $R^0$ of $\nabla^0$ is conformally invariant, i.e. $$\overline{R}^0 = e^{2u}R^0.$$ Further in this section let $(M,J,g)$ be a conformal Kähler manifold with Norden metric. Then $(M,J,\overline{g})$ is a Kähler manifold and thus $\overline{\theta}=0$. From (\[theta-bar\]) it follows $\sigma =\frac{1}{2n}(\theta \circ J)$. Then, from (\[2-1\]) and (\[2-2\]) we get $\overline{\nabla}^{\hspace{0.02in}\prime}=\overline{\nabla}$ and hence $\overline{R}^{\prime}=\overline{R}$ for all $\lambda_i\in\mathbb{R}$, $i=1,2,...,8$. In particular, $\overline{R}^\prime=\overline{R}^0$. Then, Theorems \[thB\] and (\[BR0\]) imply On a conformal Kähler manifold with Norden metric the Bochner curvature tensor of the complex connections $\nabla^{\prime}$ defined by (\[2-1\]) and (\[2-2\]) with the conditions $\lambda_7=-\lambda_5$ and $\lambda_8=-\lambda_6$ is conformally invariant by the transformation (\[conf\]), i.e. $$B(\overline{R}^{\prime})=e^{2u}B(R^{\prime}).$$ Let us remark that the conformal invariancy of the Bochner tensor of the canonical connection on a conformal Kähler manifold with Norden metric is proved in [@Gan-Gri-Mih2]. From Theorem \[thR0\] and Corollary \[cor\] we obtain Let $(M,J,g)$ be a conformal Kähler manifold with Norden metric and $\nabla^{\prime}$ be a complex connection defined by (\[2-1\]) and (\[2-2\]). If $\lambda_i=0$ for $i=5,6,7,8$, then the curvature tensor of $\nabla^{\prime}$ is conformally invariant by the transformation (\[conf\]). An Example ========== Let $G$ be a real connected four-dimensional Lie group, and $\mathfrak{g}$ be its corresponding Lie algebra. If $\{e_1,e_2,e_3,e_4\}$ is a basis of $\mathfrak{g}$, we equip $G$ with a left-invariant almost complex structure $J$ by $$\label{J1} Je_1 = e_3,\qquad Je_2 = e_4,\qquad Je_3=-e_1,\qquad Je_4=-e_2.$$ We also define a left-invariant pseudo-Riemannian metric $g$ on $G$ by $$\label{g1} \begin{array}{l} g(e_1,e_1)=g(e_2,e_2)=-g(e_3,e_3)=-g(e_4,e_4)=1,\medskip\\ g(e_i,e_j)=0,\quad i\neq j,\quad i,j=1,2,3,4. \end{array}$$ Then, because of (\[11\]), (\[J1\]) and (\[g1\]), $(G,J,g)$ is an almost complex manifold with Norden metric. Further, let the Lie algebra $\mathfrak{g}$ be defined by the following commutator relations $$\label{lie1} \begin{array}{l} [e_1,e_2]=[e_3,e_4]=0,\medskip\\ \lbrack e_1,e_4]=[e_2,e_3]=\lambda (e_1 + e_4) + \mu (e_2 - e_3),\medskip\\ \lbrack e_1,e_3]=-[e_2,e_4] = \mu (e_1 + e_4) - \lambda (e_2 - e_3), \end{array}$$ where $\lambda,\mu\in\mathbb{R}$. The well-known Koszul’s formula for the Levi-Civita connection of $g$ on $G$, i.e. the equality $$\label{K} 2g(\nabla_{e_i} e_j,e_k) = g([e_i,e_j],e_k) + g([e_k,e_i],e_j)+g([e_k,e_j],e_i),$$ and (\[g1\]) imply the following essential non-zero components of the Levi-Civita connection: $$\label{nabla1} \begin{array}{ll} \nabla_{e_1}e_1 = \nabla_{e_2}e_2 = \mu e_3 + \lambda e_4,\quad & \nabla_{e_3}e_3 = \nabla_{e_4}e_4 = -\lambda e_1 + \mu e_2,\medskip\\ \nabla_{e_1}e_3 = \mu (e_1 + e_4),\quad & \nabla_{e_1}e_4 = \lambda e_1 - \mu e_3,\medskip\\ \nabla_{e_2}e_3 = \mu e_1 + \lambda e_4,\quad & \nabla_{e_2}e_4 = \lambda (e_2 - e_3). \end{array}$$ Then, by (\[F\]), (\[Fp\]) and (\[nabla1\]) we compute the following essential non-zero components $F_{ijk}=F(e_i,e_j,e_k)$ of $F$: $$\label{Fijk} \begin{array}{l} F_{111} = F_{422} = 2\mu,\quad F_{222}=-F_{311} = 2\lambda,\medskip\\ F_{112} = -F_{214} = F_{314} = -F_{412} = \lambda,\quad F_{212}=-F_{114}=F_{312}=-F_{414}=\mu. \end{array}$$ Having in mind (\[1-3\]) and (\[Fijk\]), the components $\theta_i=\theta(e_i)$ and $\theta_i^\ast=\theta^\ast(e_i)$ of the 1-forms $\theta$ and $\theta^\ast$, respectively, are: $$\label{theta} \theta_2=\theta_3=\theta^\ast_1=-\theta^\ast_4 = 4\lambda,\qquad\theta_1=-\theta_4=-\theta^\ast_2=-\theta^\ast_3=4\mu.$$ By (\[1-3\]) and (\[theta\]) we compute $$\label{22} \begin{array}{c} \Omega = 4\mu(e_1+e_4) + 4\lambda(e_2 -e_3),\qquad J\Omega = 4\lambda(e_1 + e_4) - 4\mu(e_2 - e_3),\medskip\\ \theta (\Omega) = \theta(J\Omega) = 0. \end{array}$$ By the characteristic condition (\[w1\]) and equalities (\[Fijk\]), (\[theta\]) we prove that the manifold $(G,J,g)$ with Lie algebra $\mathfrak{g}$ defined by (\[lie1\]) belongs to the basic class $\mathcal{W}_1$. Moreover, by (\[nabla1\]) and (\[theta\]) it follows that the conditions (\[cK\]) hold and thus The manifold $(G,J,g)$ defined by (\[J1\]), (\[g1\]) and (\[lie1\]) is a conformal Kähler manifold with Norden metric. According to (\[nabla-0\]), (\[g1\]), (\[nabla1\]) and (\[theta\]) the components of the natural connection $\nabla^0$ are given by $$\label{nabla-01} \begin{array}{ll} \nabla_{e_1}^0 e_1 = - \nabla_{e_4}^0 e_1 = \mu e_2,\qquad & \nabla_{e_2}^0 e_1 = \nabla_{e_3}^0 e_1 = \lambda e_2,\medskip\\ \nabla_{e_1}^0 e_2 = -\nabla_{e_4}^0 e_2 = -\mu e_1,\qquad & \nabla_{e_2}^0 e_2 = \nabla_{e_3}^0 e_2 = - \lambda e_1,\medskip\\ \nabla_{e_1}^0 e_3 =-\nabla_{e_4}^0 e_3 = \mu e_4,\qquad & \nabla_{e_2}^0 e_3 = \nabla_{e_3}^0 e_3 = \lambda e_4,\medskip\\ \nabla_{e_1}^0 e_4 = -\nabla_{e_4}^0 e_4 = -\mu e_3,\qquad & \nabla_{e_2}^0 e_4 = \nabla_{e_4}^0 e_4 = - \lambda e_3. \end{array}$$ By (\[nabla-01\]) we obtain $R^0=0$. Then, by (\[R0\]) and (\[R01\]) the curvature tensor $R$ of $(G,J,g)$ has the form $$R = \frac{1}{4}\psi_1(A),\qquad A(x,y) = (\nabla_x\theta)Jy + \frac{1}{4}\theta(x)\theta(y).$$ Moreover, having in mind (\[111\]), (\[nabla1\]), (\[theta\]) and (\[22\]), we compute $S_1=S_2=S_3=0$. Hence, for the tensors $R^\prime$ of the complex connections $\nabla^\prime$ given by (\[Rpr\]), it is valid $R^\prime = 0$. The complex connections $\nabla^\prime$ defined by (\[2-1\]) and (\[2-2\]) are flat on $(G,J,g)$. [99]{} G. Ganchev, A. Borisov, Note on the almost complex manifolds with a Norden metric, Compt. Rend. Acad. Bulg. Sci. 39(5) (1986), 31–34. G. Ganchev, V. Mihova, Canonical connection and the canonical conformal group on an almost complex manifold with $B$-metric, Ann. Univ. Sofia Fac. Math. Inform., 81(1) (1987), 195–206. G. Ganchev, K. Gribachev, V. Mihova, $B$-connections and their conformal invariants on conformally Kähler manifolds with $B$-metric, Publ. Inst. Math. (Beograd) (N.S.) 42(56) (1987), 107–121. K. Gribachev, D. Mekerov, G. Djelepov, Generalized $\emph{B}$-manifolds, Compt. Rend. Acad. Bulg. Sci. 38(3) (1985), 299–302. A. Hayden, Subspaces of a space with torsion, Proc. London Math. Soc. 34 (1932), 27–50. T. Imai, Notes on semi-symmetric metric connections, Tensor N.S. 24 (1972), 293–296. S. Kobayshi, K. Nomizu, Foundations of differential geometry vol. 1, 2, Intersc. Publ., New York, 1963, 1969. A. Newlander, L. Nirenberg, Complex analytic coordinates in almost complex manifolds, Ann. Math. 65 (1957), 391–404. A. P. Norden, On a class of four-dimensional A-spaces, Russian Math. (Izv VUZ) 17(4) (1960), 145–157. S. D. Singh, A. K. Pandey, Semi-symmetric metric connections in an almost Norden metric manifold, Acta Cienc. Indica Math. 27(1) (2001), 43–54. M. Teofilova, Almost complex connections on almost complex manifolds with Norden metric, In: Trends in Differential Geometry, Complex Analysis and Mathematical Physics, eds. K. Sekigawa, V. Gerdjikov and S. Dimiev, World Sci. Publ., Singapore (2009), 231–240. K. Yano, Affine connections in an almost product space, Kodai Math. Semin. Rep. 11(1) (1959), 1–24. K. Yano, Differential geometry on complex and almost complex spaces, Pure and Applied Math. vol. 49, Pergamon Press Book, New York, 1965. K. Yano, On semi-symmetric metric connection, Rev. Roumanie Math. Pure Appl. 15 (1970), 1579–1586. Marta Teofilova\ Department of Geometry\ Faculty of Mathematics and Informatics\ University of Plovdiv\ 236 Bulgaria Blvd.\ 4003 Plovdiv, Bulgaria\ e-mail: `[email protected]`	{ "pile_set_name": "ArXiv" }
--- abstract: \| Regge action is represented analogously to how the Palatini action for general relativity (GR) as some functional of the metric and a general connection as independent variables represents the Einstein-Hilbert action. The piecewise flat (or simplicial) spacetime of Regge calculus is equipped with some world coordinates and some piecewise affine metric which is completely defined by the set of edge lengths and the world coordinates of the vertices. The conjugate variables are the general nondegenerate matrices on the 3-simplices which play a role of a general discrete connection. Our previous result on some representation of the Regge calculus action in terms of the local Euclidean (Minkowsky) frame vectors and orthogonal connection matrices as independent variables is somewhat modified for the considered case of the general linear group GL(4,R) of the connection matrices. As a result, we have some action invariant w. r. t. arbitrary change of coordinates of the vertices (and related GL(4,R) transformations in the 4-simplices). Excluding GL(4,R) connection from this action via the equations of motion we have exactly the Regge action for the considered spacetime. author: - \| V.M. Khatsymovsky\ [Budker Institute of Nuclear Physics]{}\ [of Siberian Branch Russian Academy of Sciences]{}\ [ Novosibirsk, 630090, Russia]{}\ [E-mail address: [email protected]]{} title: Affine connection form of Regge calculus --- PACS Nos.: 04.60.Kz; 04.60.Nc MSC classes: 83C27; 53C05 keywords: Einstein theory of gravity; minisuperspace model; piecewise flat spacetime; Regge calculus; affine connection; Palatini action; discrete connection Minisuperspace models allow one to work with a countable number of the degrees of freedom. This may be important in such an essentially nonlinear theory of gravity as GR. In quantum framework, minisuperspace approach works as a kind of lattice methods [@Ham'] allowing to solve the problem of the formal nonrenormalizability of the original continuum GR. The GR considered as some curved geometry theory can be most naturally formulated on a minisuperspace by restricting ourselves by the metric field distributions $g_{\lambda \mu} (x )$ describing the piecewise flat spacetimes composed of the flat 4D tetrahedra ([4-simplices]{}) or the [simplicial complex]{} [@piecewiseflat=simplicial']. These spacetimes can be chosen arbitrarily close in some sense to any given Riemannian spacetime, and GR on them is known as Regge calculus [@Regge']; see, e. g. review in Ref. [@RegWil']. The Regge action ($\int R \sqrt{g} {{\rm d}}^4 x$) is $$\label{S-Regge} 2 \sum_{\sigma^2}{\alpha (\sigma^2) A (\sigma^2)},$$ where $A (\sigma^2)$ is the area of the triangle (2-simplex) $\sigma^2$, $\alpha (\sigma^2)$ is the angle defect on this triangle, summation is over all the 2-simplices $\sigma^2$. The action and the geometry of the piecewise flat spacetime is completely defined by the full set of the edge lengths. Among contemporary approaches the Causal Dynamical Triangulations approach related to the Regge calculus has led to important results in quantum gravity [@cdt]. The Regge action and the corresponding equations of motion in terms of edge lengths seem to be simple from geometrical point of view, but rather complicated from the field theoretical point of view. In the continuum GR, certain simplification is achieved by adding some extra independent variables so that some new action in terms of these is classically (on the equations of motion) equivalent to the original Einstein-Hilbert action, but is of the first order contrary to the second order Einstein-Hilbert one. The discrete analogs of these first order formulations have been widely addressed. The discrete analogs of the connection and curvature, finite orthogonal rotations, were first considered in [@Fro]. An application to the discrete Hamiltonian analysis of gravity was discussed in [@Ban3]. In the paper [@our2], we have found the exact representation of the Regge action using the discrete analogs of the connection and curvature [@Fro]. An approximate version was considered in [@CasDadMag]. A first order form of Regge calculus was considered in [@Bar] where the extra independent variables are the interior dihedral angles of a simplex, with conjugate variables the areas of the triangles. Now we consider the possibility of using general linear rather than the orthogonal matrices as discrete connection variables. Thus, we are talking about some discrete version of the Palatini action [@Pal] depending on the metric and a general connection as independent variables. The Palatini action can be written in the form $$\begin{aligned} \label{Pal action} & & \hspace{0mm} S = \int R^\nu_{\lambda \nu \mu} g^{\lambda \mu} \sqrt {g} {{\rm d}}^4 x = \int [ - \Gamma^\nu_{\lambda \mu} \partial_\nu (\sqrt{g} g^{\lambda \mu}) + \Gamma^\nu_{\lambda \nu} \partial_\mu (\sqrt{g} g^{\lambda \mu}) \nonumber \\ & & + \sqrt{g} g^{\lambda \mu} ( \Gamma^\nu_{\lambda \mu} \Gamma^\rho_{\nu \rho} - \Gamma^\nu_{\lambda \rho} \Gamma^\rho_{\nu \mu} ) ] {{\rm d}}^4 x,\end{aligned}$$ where $g \equiv \det \\| g_{\lambda \mu} \\|$ and we assume the Euclidean metric signature for definiteness. Here, the metric $g_{\lambda \mu}$ and the Christoffel symbol $\Gamma^\lambda_{\mu \nu}$ are independent variables, and an additional requirement for that the equations of motion would give for $\Gamma^\lambda_{\mu \nu}$ the unique metric-compatible connection for $g_{\lambda \mu}$ is the requirement that the connection be torsion-free, $\Gamma^\lambda_{\mu \nu} = \Gamma^\lambda_{\nu \mu}$. Note, however, that $\Gamma^\lambda_{\mu \nu} {{\rm d}}x^\nu$ describes an infinitesimal transformation matrix of any vector being parallel transported from $x^\lambda$ to $x^\lambda + {{\rm d}}x^\lambda$. Some discrete analog of it would be a finite transformation matrix ${{\cal M}}^\lambda_{\sigma^3 \mu}$ for the transport across the 3-simplex $\sigma^3$, the 3-dimensional face between some two neighboring 4-simplices. Consequently, imposing the zero torsion condition require comparing the matrices ${{\cal M}}^\lambda_{\sigma^3 \mu}$ at the different $\sigma^3$s, that is, this condition is not local and is unnatural in the considered formalism. Fortunately, the requirement that the connection be torsion-free is not absolutely necessary for the GR action being reproduced via equations of motion in the Palatini formalism. Without this condition, we get for $\Gamma^\lambda_{\mu \nu}$ the unique metric-compatible part $\Gamma^\lambda_{\mu \nu} ( \{ g_{\lambda \mu} \} )$ plus some part $\Gamma_\nu \delta^\lambda_\mu $ with torsion which, however, does not contribute to the action. Thus, it is more natural to consider the completely unrestricted three-index variable $\Gamma^\lambda_{\mu \nu}$ as the continuum counterpart; the discrete counterpart will be the general nondegenerate transformation matrices $4 \times 4$ on the 3-simplices $\sigma^3$, ${{\cal M}}_{\sigma^3} \in $ GL(4, R). If the connection has support on the 3-simplices, this means that the metric is piecewise constant or constant in the 4-simplices. That is, the world coordinate frame is piecewise affine. The metric field $g_{\lambda \mu} (x )$ is completely defined by the edge lengths and the coordinates of the simplicial vertices $\sigma^0$. Earlier we have found a representation of the Regge action in terms of the edge vectors and orthogonal connection matrices between the Euclidean local frames in the 4-simplices [@our2]. Now we should, first, transform this expression from the local metric $diag (1, 1, 1, 1)$ to the general symmetric nondegenerate constant matrix $g_{\lambda \mu}$ in each 4-simplex. Second, we should generalize it from the orthogonal to general nondegenerate connection matrices. The considered expression for the Regge action connection representation in the local Euclidean frame formalism takes the form $$\begin{aligned} \label{Cartan-SO(4)} \hspace{0mm} S^{\rm discr}_{\rm SO(4)} = 2 \sum_{\sigma^2} A(\sigma^2 ) \arcsin \left [ \frac{ R^{ab}_{\sigma^2} ( \Omega ) l^c_{\sigma^1_1} l^d_{\sigma^1_2} }{4 A(\sigma^2 )} \epsilon_{abcd} \right ].\end{aligned}$$ Here, the edge vectors $l^a_{\sigma^1_1}$, $l^a_{\sigma^1_2}$ form the triangle $\sigma^2$, the area of $\sigma^2$ is $A(\sigma^2 )$ $=$\ $\sqrt{ l^2_{\sigma^1_1} l^2_{\sigma^1_2} - (l_{\sigma^1_1} l_{\sigma^1_2})^2} / 2 $. The [curvature]{} SO(4) matrix $R^{ab}_{\sigma^2} ( \Omega )$ on the triangles $\sigma^2$ is the product of the connection SO(4) matrices $\Omega_{\sigma^3}$s for the set of $\sigma^3$s meeting at $\sigma^2$ ordered along a closed path encircling $\sigma^2$ and passing through each of these (and only these) $\sigma^3$s, $$R_{\sigma^2} (\Omega ) = \prod_{ \{ \sigma^3 : ~ \sigma^3\supset\sigma^2 \} }{\Omega^{\epsilon (\sigma^2, \sigma^3)}_{\sigma^3}},$$ where $\epsilon (\sigma^2, \sigma^3) = \pm 1$ is some sign function. This path begins and ends in a 4-simplex $\sigma^4$. That is, $R^{ab}_{\sigma^2}$ is defined in (the frame of) this simplex. The edge vectors $l^a_{\sigma^1_1}$, $l^a_{\sigma^1_2}$ are also defined in this simplex. When we pass from the local Euclidean frames to the piecewise affine world one, the edge vectors $l^a$ are substituted by the corresponding world coordinate differences $\Delta x^\lambda$. The RHS of (\[Cartan-SO(4)\]) takes the form $$\begin{aligned} \label{Cartan-SO(4)-version} \hspace{0mm} 2 \sum_{\sigma^2} A(\sigma^2 ) \arcsin \left [ \frac{ {{\cal R}}^\lambda_{\sigma^2 \tau} ( {{\cal M}}) g^{\tau \mu} \Delta x^\nu_{\sigma^1_1} \Delta x^\rho_{\sigma^1_2} }{4 A(\sigma^2 )} \epsilon_{\lambda \mu \nu \rho} \sqrt{g} \right ]\end{aligned}$$ where ${{\cal R}}_{\sigma^2} ( {{\cal M}})$ is built of ${{\cal M}}_{\sigma^3}$s just as $R_{\sigma^2} ( \Omega )$ is built of $\Omega_{\sigma^3}$s. The generalization to arbitrary ${{\cal M}}_{\sigma^3}$s should provide the correct equations of motion. The dependence of ${{\cal R}}_{\sigma^2}$ on ${{\cal M}}_{\sigma^3}$ takes the form $(\Gamma_1 (\sigma^2 , \sigma^3 ) {{\cal M}}_{\sigma^3} \Gamma_2 (\sigma^2 , \sigma^3 ))^{\epsilon (\sigma^2, \sigma^3)}$, $\sigma^2 \subset \sigma^3$, $\Gamma_1 (\sigma^2 , \sigma^3 )$, $\Gamma_2 (\sigma^2 , \sigma^3 )$ are GL(4, R) matrices. Let us take for example $\epsilon (\sigma^2, \sigma^3) = + 1$ (that is, ${{\cal R}}_{\sigma^2}$ is linear in ${{\cal M}}_{\sigma^3}$). Applying ${{\cal M}}^\nu_{\sigma^3 \lambda} \partial / \partial {{\cal M}}^\nu_{\sigma^3 \mu}$ to (\[Cartan-SO(4)-version\]) gives $$\begin{aligned} \label{eq-motion-version} - \hspace{-10mm} \sum_{ \hspace{10mm} \{\sigma^2 : ~ \sigma^2 \subset \sigma^3 \} } \hspace{-10mm} [ \Gamma_2 (\sigma^2 , \sigma^3 ) \frac{v_{\sigma^2 } {{\cal R}}_{\sigma^2 }}{\cos \alpha (\sigma^2 )} \Gamma^{-1}_2 (\sigma^2 , \sigma^3 ) ]^\mu {}_{ \lambda}, ~~~ v_{\sigma^2 \lambda \mu} = \frac{1}{2} \sqrt{g} \epsilon_{\lambda \mu \nu \rho} \Delta x^\nu_{\sigma^1_1} \Delta x^\rho_{\sigma^1_2},\end{aligned}$$ for the contribution to the equations of motion. Here $\alpha (\sigma^2 )$ means just $\arcsin$ function in (\[Cartan-SO(4)-version\]). For the particular ansatz of the metric-compatible connection with ${{\cal R}}_{\sigma^2}$ rotating around $\sigma^2$ (by the angle $\alpha (\sigma^2 )$), $$\label{ansatz-identity} v_{\sigma^2 } {{\cal R}}_{\sigma^2 } + {{\cal R}}^{-1}_{\sigma^2 } v_{\sigma^2 } = 2 v_{\sigma^2 } \cos \alpha (\sigma^2 ), ~~~ {{\cal R}}_{\sigma^2 } v_{\sigma^2 } + v_{\sigma^2 } {{\cal R}}^{-1}_{\sigma^2 } = 2 v_{\sigma^2 } \cos \alpha (\sigma^2 ).$$ The expression (\[eq-motion-version\]) looks similar to the sum of the bivectors $v_{\sigma^2}$ of the faces $\sigma^2$ of $\sigma^3$, but not quite that. The situation can be corrected by the following replacement of the matrix describing the curvature in (\[Cartan-SO(4)-version\]), $${{\cal R}}_{\sigma^2 } \Longrightarrow \frac{1}{2} ( {{\cal R}}_{\sigma^2 } - {{\cal R}}^{-1}_{\sigma^2 } ).$$ Geometrically, the curvature is evaluated by parallel transport of the vector along a closed loop in the two possible opposite directions and comparing the results. This gives a discrete action, $$\begin{aligned} \label{Palatini} \hspace{0mm} S^{\rm discr}_{\rm GL(4,R)} = 2 \sum_{\sigma^2} A(\sigma^2 ) \arcsin \left \{ \frac{ [{{\cal R}}_{\sigma^2} - {{\cal R}}^{-1}_{\sigma^2} ]^\lambda {}_{ \tau} ( {{\cal M}}) g^{\tau \mu} \Delta x^\nu_{\sigma^1_1} \Delta x^\rho_{\sigma^1_2} }{8 A(\sigma^2 )} \epsilon_{\lambda \mu \nu \rho} \sqrt{g} \right \}.\end{aligned}$$ Here, to resume, $${{\cal R}}_{\sigma^2} ({{\cal M}}) = \prod_{ \{ \sigma^3 : ~ \sigma^3\supset\sigma^2 \} }{{{\cal M}}^{\epsilon (\sigma^2 \sigma^3)}_{\sigma^3}}$$ ($\epsilon (\sigma^2, \sigma^3) = \pm 1$ is some sign function), the product of the connection GL(4,R) matrices ${{\cal M}}_{\sigma^3}$s for the set of $\sigma^3$s meeting at $\sigma^2$ ordered along a closed path encircling $\sigma^2$ and passing through each of these (and only these) $\sigma^3$s. This path begins and ends in a 4-simplex $\sigma^4$, and the metric appearing in (\[Palatini\]) and the dual bivector $v_{\sigma^2 \lambda \mu}$ are taken just in this 4-simplex. Again, denote ${{\cal R}}_{\sigma^2} = (\Gamma_1 (\sigma^2 , \sigma^3 ) {{\cal M}}_{\sigma^3} \Gamma_2 (\sigma^2 , \sigma^3 ))^{\epsilon (\sigma^2, \sigma^3)}$, apply ${{\cal M}}^\nu_{\sigma^3 \lambda} \partial / \partial {{\cal M}}^\nu_{\sigma^3 \mu}$ to the action and take into account the following equality for differentiating the inverse matrix, $${{\cal M}}^\nu {}_\lambda \frac{\partial }{\partial {{\cal M}}^\nu {}_\mu} ({{\cal M}}^{-1})^\rho {}_\tau = - ({{\cal M}}^{-1})^\mu {}_\tau \delta^\rho_\lambda.$$ The equations of motion take the form $$\begin{aligned} \label{eq-motion} \hspace{-10mm} \sum_{ \hspace{10mm} \{\sigma^2 : ~ \sigma^2 \subset \sigma^3 \} } \hspace{-10mm} \epsilon (\sigma^2, \sigma^3) \left [ \Gamma_2 (\sigma^2 , \sigma^3 ) \frac{v_{\sigma^2 } {{\cal R}}^{\epsilon (\sigma^2, \sigma^3)}_{\sigma^2 } + {{\cal R}}^{- \epsilon (\sigma^2, \sigma^3)}_{\sigma^2 } v_{\sigma^2 } }{\cos \alpha (\sigma^2 )} \Gamma^{-1}_2 (\sigma^2 , \sigma^3 ) \right ] = 0.\end{aligned}$$ Now, with taking into account (\[ansatz-identity\]) for the particular ansatz of the metric-compatible connection, these equations express the closure of the (dual) bivectors\ $\Gamma_2 (\sigma^2 , \sigma^3 ) v_{\sigma^2} \Gamma^{-1}_2 (\sigma^2 , \sigma^3 )$, $\sigma^2 \subset \sigma^3$ fulfilled identically; the matrices $\Gamma_2$ serve to transport them all to the same 4-simplex, one of the two containing the given $\sigma^3$ (from which ${{\cal M}}_{\sigma^3}$ transports). Thus, the $S^{\rm discr}_{\rm GL(4,R)}$ (\[Palatini\]) can serve as an exact affine connection representation of the Regge action. The configuration of the considered system is completely described by the set of connection matrices on the 3-simplices ${{\cal M}}_{\sigma^3} \in {\rm GL(4,R)}$, edge lengths $l_{\sigma^1}$ and coordinates of the vertices $x^\lambda_{\sigma^0}$. The coordinates of the vertices $x^\lambda_{\sigma^0}$ define the contravariant edge vector $$\Delta x^\lambda_{\sigma^1} = x^\lambda_{\sigma^0_2 } - x^\lambda_{\sigma^0_1 }$$ for the edge $\sigma^1$, the difference between the coordinates of its ending vertices $\sigma^0_1$, $\sigma^0_2$. Knowing the contravariant edge vectors and edge lengths allows to find the metric tensor in each 4-simplex as a function of its ten edge lengths from simple system of ten linear equations, $$\Delta x^\lambda_{\sigma^1} \Delta x^\mu_{\sigma^1} g^{\lambda \mu } = l^2_{\sigma^1 }.$$ Any gauge transformation is an arbitrary change of coordinates of the vertices. This is analogous to diffeomorphisms in the continuum theory. Elementary gauge transformation is an arbitrary change of coordinates of any [given]{} vertex $\sigma^0$. When doing so, the vectors $\Delta x^\lambda_{\sigma^1 }$ are changed for the edges containing this vertex, $\sigma^1 \supset \sigma^0$. This can be described as general coordinate transformations acting in each 4-simplex containing the vertex, $\sigma^4 \supset \sigma^0$, $$\Delta x^\lambda_{\sigma^1} \Longrightarrow \Delta \tilde{x}^\lambda_{\sigma^1} = {{\cal N}}^\lambda_{\sigma^4 \mu} \Delta x^\mu_{\sigma^1}, ~~~ \sigma^4 \supset \sigma^1$$ and $$A_\lambda \Longrightarrow \tilde{A}_\lambda = A_\mu ({{\cal N}}^{-1}_{\sigma^4})^\mu {}_\lambda$$ for the covariant vectors in $\sigma^4$. This ${{\cal N}}_{\sigma^4}$ is restricted by the condition that it does not change the vectors of those edges of $\sigma^4$ which do not contain $\sigma^0$. Besides that, the matrices in the neighboring 4-simplices ${{\cal N}}_{\sigma^4_1}$ and ${{\cal N}}_{\sigma^4_2}$ should give the same result when acting on the vectors of their common edges $\sigma^1 \subset \sigma^4_1 \cap \sigma^4_2$. $S^{\rm discr}_{\rm GL(4,R)}$ is invariant w. r. t. such transformations: $${{\cal M}}^\lambda_{\sigma^3 \mu} \Longrightarrow \tilde{{{\cal M}}}^\lambda_{\sigma^3 \mu} = {{\cal N}}^\lambda_{\sigma^4_2 \nu} {{\cal M}}^\nu_{\sigma^3 \rho} ({{\cal N}}^{-1}_{\sigma^4_1})^\rho {}_\mu, ~~~ \sigma^3 = \sigma^4_1 \cap \sigma^4_2$$ (${{\cal M}}_{\sigma^3}$ acts from $\sigma^4_1$ to $\sigma^4_2$), the curvature matrix $${{\cal R}}^\lambda_{\sigma^2 \mu} \Longrightarrow \tilde{{{\cal R}}}^\lambda_{\sigma^2 \mu} = {{\cal N}}^\lambda_{\sigma^4 \nu} {{\cal R}}^\nu_{\sigma^2 \rho} ({{\cal N}}^{-1}_{\sigma^4})^\rho {}_\mu,$$ $\sigma^4$ is the considered above 4-simplex where ${{\cal R}}_{\sigma^2}$ is defined, and (\[Palatini\]) is formally invariant w. r. t. any ${{\cal N}}_{\sigma^4}$ acting in this $\sigma^4$. The invariance under the general transformation of the vertex coordinates says that these coordinates might be chosen by hand in some convenient way. In any given 4-simplex this can be done but procedure of a convenient extension to the other 4-simplices is unknown in general case. In some particular cases of simple structure this choice can be made globally. For example, there is the case of the periodic structure such that spacetime is decomposed into 4-cubes and each cube is decomposed with the help of its edges and diagonals into the 4-simplices [@RocWil]. Here we have, in fact, the 4-cubic lattice, the vertices of which can be numbered by integer valued Cartesian coordinates. Thereby, the action can be explicitly written in terms of the edge lengths as metric variables (and arbitrary nondegenerate connection matrices, of course). If the considered affine connection action is compared with its orthogonal connection form, the latter differs by fewer number of the connection degrees of freedom (6 parameters of the general orthogonal matrix instead of 16 elements of general 4 $\times$ 4 matrix) although at the expense of imposing additional (orthogonality) constraints, and by considerably larger tetrad/metric sector with a number of additional constraints. For example, consider the simplicial periodic structure [@RocWil]. There are 24 4-simplices in each 4-cube and we have some 16-component tetrad of edge vectors in the local frame of each 4-simplex $l^a_{\sigma^1_i \| \sigma^4}, i = 1, 2, 3, 4$. Simultaneously, a number of some constraints should be imposed of the type of $(l^a_{\sigma^1 \| \sigma^4_1})^2 = (l^a_{\sigma^1 \| \sigma^4_2})^2$ for the vectors of the same edge $\sigma^1$ defined in the different local frames of some two 4-simplices sharing this edge, $\sigma^1 \subset \sigma^4_1 \cap \sigma^4_2$. In contrast, in the affine connection formalism, we have the lengths of the 15 edges and 4 coordinates of the vertex per 4-cube. The coordinates of the vertex can be conveniently chosen in the overall lattice as, e. g. integer-valued coordinates, and we are left just with 15 edge lengths per cube/vertex. This may also have an advantage in a numerical simulation. Certain inconvenience associated with using the affine connection is that the freely chosen elements ${{\cal M}}_{\sigma^3} \in {\rm GL(4,R)}$ do not automatically enter the domain of definition of $S^{\rm discr}_{\rm GL(4,R)}$. This is because the argument of $\arcsin$ in (\[Palatini\]) can become in absolute value greater than unity if we rescale ${{\cal M}}_{\sigma^3}$s strongly enough. Therefore, we need to specially check whether we came out of the domain of definition of the action. An attractive feature of the affine connection may be a uniform description with the same group GL(4,R) of both the Riemannian and pseudo-Riemannian spacetimes instead of the two groups SO(4) and SO(3,1) in the local frame formalism. If the functional integral is considered, some typical value of the coefficient at $\arcsin$ in the functional integral exponent is $A / l^2_g$ where $A$ is a typical simplicial area and $l_g$ is the Plank scale $10^{-33} cm$. Therefore, at $A \gg l^2_g$ the greatest contribution, not suppressed by the oscillating exponential, comes from the region of small argument of $\arcsin$, and $\arcsin$ can be substituted by its argument. We have certain matrix analog of some Bessel function, which is transformed into an absolutely convergent integral by the transition to the contour integration in the complex plane, $$\label{func-int-model} \int^{\infty}_0 \exp \left [ i\frac{A}{l^2_g} \left ( x - \frac{1}{x} \right ) \right ] \frac{{{\rm d}}x}{x} \Longrightarrow \int^{\infty}_0 \exp \left [ -\frac{A}{l^2_g} \left ( x + \frac{1}{x} \right ) \right ] \frac{{{\rm d}}x}{x},$$ where A stands for a typical scale of the dual bivector $v_{\sigma^2 \lambda}{}^\tau / 2$ (\[eq-motion-version\]) in (\[Palatini\]) (now with the standard replacement $\sqrt{g} \to \sqrt{ - g}$ for the pseudo-Riemannian spacetime), that is, just the triangle area. Here we suggest that the functional integral measure includes the product over the 3-simplices $\prod_{\sigma^3} {{\cal D}}{{\cal M}}_{\sigma^3}$ where ${{\cal D}}{{\cal M}}= (\det {{\cal M}})^{-4} {{\rm d}}^{16} {{\cal M}}$ is the Haar measure on GL(4,R). Using multiplicative dependence of ${{\cal R}}$s on ${{\cal M}}$s and invariance of the Haar measure w. r. t. the multiplication we can make certain change of variables and replace some ${{\cal D}}{{\cal M}}_{\sigma^3}$s by ${{\cal D}}{{\cal R}}_{\sigma^2}$s for those ${{\cal R}}_{\sigma^2}$s that are not connected with each other (by the Bianchi identities [@Regge']). In (\[func-int-model\]), ${{\rm d}}x / x$ just models ${{\cal D}}{{\cal R}}$. (Of course, the genuine expression is not factorizable into the analogs of (\[func-int-model\]) so easily.) The above suggests that the contribution of areas exceeding the Plank scale is probably exponentially suppressed. Recasting to an absolutely convergent integral may also be important for a possible numerical simulation. In conclusion, using the affine connection form of Regge calculus may have interesting implications in both analytical and numerical analysis of classical and especially quantum GR. Acknowledgments {#acknowledgments .unnumbered} =============== The present work was supported by the Ministry of Education and Science of the Russian Federation [99]{} H. W. Hamber, Quantum Gravity on the Lattice, [Gen. Rel. Grav.]{} [41]{}, 817 (2009); ([Preprint]{} arXiv:0901.0964\[gr-qc\]). J. Cheeger, W. Müller, and R. Shrader, On the curvature of the piecewise flat spaces, [Commun. Math. Phys.]{} [92]{}, 405 (1984). T. Regge, General relativity theory without coordinates, [Nuovo Cimento]{} [19]{}, 558 (1961). T. Regge and R. M. Williams, Discrete structures in gravity, [Journ. Math. Phys.]{} [41]{}, 3964 (2000); ([Preprint]{} arXiv:0012035\[gr-qc\]). J. Ambjorn, A. Goerlich, J. Jurkiewicz, and R. Loll, Nonperturbative Quantum Gravity, [Physics Reports]{} [519]{}, 127 (2012); ([Preprint]{} arXiv:1203.3591\[hep-th\]). J. Fröhlich, Regge Calculus and Discretized Gravitational Functional Integrals, IHES preprint, 1981 (unpublished); in [Non-Perturbative Quantum Field Theory: Mathematical Aspects and Applications, Selected Papers]{}, 523 (Singapore: World Scientific, 1992). M. Bander, Hamiltonian lattice gravity. II. Discrete moving-frame formulation [Phys. Rev.]{} D [38]{}, 1056 (1988). V. M. Khatsymovsky, Tetrad and self-dual formulations of Regge calculus, [Class. Quantum Grav.]{} [6]{}, L249 (1989). M. Caselle, A. D’Adda, and L. Magnea, Regge calculus as a local theory of the Poincaré group, [Phys. Lett.]{} B [232]{}, 457 (1989). J. W. Barrett, First order Regge calculus, [Class. Quantum Grav.]{} [11]{} 2723 (1994); ([Preprint]{} arXiv:9404124\[hep-th\]). A. Palatini, Deduzione invariantiva delle equazioni gravitazionali dal principio di Hamilton, [R.C. Circ. Mat. Palermo]{} [43]{}, 203 (1919). M. Rocek, R.M. Williams, Quantum Regge calculus, [Phys. Lett.]{} B [104]{}, 31 (1981).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Compelling evidence–though yet no formal proof–has been adduced that the probability that a generic two-qubit state ($\rho$) is separable is $\frac{8}{33}$ (arXiv:1301.6617, arXiv:1109.2560, arXiv:0704.3723). Proceeding in related analytical frameworks, using a further determinantal moment formula of C. Dunkl (Appendix), we reach the conclusion that one-half of this probability arises when the determinantal inequality $\|\rho^{PT}\|>\|\rho\|$, where $PT$ denotes the partial transpose, is satisfied, and, the other half, when $\|\rho\|>\|\rho^{PT}\|$. These probabilities are taken with respect to the flat, Hilbert-Schmidt measure on the fifteen-dimensional convex set of $4 \times 4$ density matrices. We find fully parallel bisection/equipartition results for the previously adduced, as well, two-“re\[al\]bit” and two-“quater\[nionic\]bit”separability probabilities of $\frac{29}{64}$ and $\frac{26}{323}$, respectively. The computational results reported lend strong support to those obtained earlier–including the “concise formula” that yields them–most conspicuously amongst those findings being the $\frac{29}{64}, \frac{8}{33}$ and $\frac{26}{323}$ probabilities noted.' author: - 'Paul B. Slater' bibliography: - 'Dissection.bib' title: 'Symmetric Halves of the $\frac{8}{33}$-Probability that the Joint State of Two Quantum Bits is Disentangled' --- The problem of determining the probability that generic sets of bipartite/multipartite quantum states exhibit entanglement features of one form or another is clearly of intrinsic interest [@ZHSL; @simon; @BHN; @sbz; @ingemarkarol]. We have reported [@MomentBased; @slaterJModPhys] major advances, in this regard, with respect to the “separability/disentanglement probability” of two-qubit states, endowed with the flat, Hilbert-Schmidt (HS) measure [@szHS; @ingemarkarol]. (The alternative use of the theoretically-important Bures \[minimal monotone\] measure [@szBures; @ingemarkarol] has subsequently been investigated [@BuresHilbert; @Hybrid], but much less progress has so far been achieved in that area.) In particular, a concise formula [@slaterJModPhys eqs. (1)-(3)] $$\label{Hou1} P(\alpha) =\Sigma_{i=0}^\infty f(\alpha+i),$$ where $$\label{Hou2} f(\alpha) = P(\alpha)-P(\alpha +1) = \frac{ q(\alpha) 2^{-4 \alpha -6} \Gamma{(3 \alpha +\frac{5}{2})} \Gamma{(5 \alpha +2})}{3 \Gamma{(\alpha +1)} \Gamma{(2 \alpha +3)} \Gamma{(5 \alpha +\frac{13}{2})}},$$ and $$\label{Hou3} q(\alpha) = 185000 \alpha ^5+779750 \alpha ^4+1289125 \alpha ^3+1042015 \alpha ^2+410694 \alpha +63000 =$$ $$\alpha \bigg(5 \alpha \Big(25 \alpha \big(2 \alpha (740 \alpha +3119)+10313\big)+208403\Big)+410694\bigg)+63000$$ has been developed that yields for a given $\alpha$, where $\alpha$ is a random-matrix-Dyson-like-index [@dumitriu], the corresponding separability probability $P(\alpha)$. The setting $\alpha=1$ pertains to the fifteen-dimensional convex set of (standard, complex-entries) two-qubit density matrices, and the formula yields (to arbitrarily high numerical precision) $P(1) =\frac{8}{33}$ (cf. [@slater833] [@joynt eq. B7]). It is interesting to note that in this standard case [@steve], the probability seems of a somewhat simpler nature (smaller numerators and denominators) than the value $P(\frac{1}{2}) =\frac{29}{64}$ obtained for the nine-dimensional convex set of $4 \times 4$ (two-“rebit”) density matrices with real entries [@carl], or, the value $P(2) = \frac{26}{323}$ derived for the 27-dimensional convex set of $4 \times 4$ (two-“quaterbit”) density matrices with quaternionic entries [@asher2; @adler]. These simple rational-valued separability probabilities and the formula above that yields them were obtained through a number of distinct steps of analysis. First, based on extensive computations, C. Dunkl was able to obtain the (yet formally unproven) determinantal-moment formula [@MomentBased p. 30] (cf. [@zozor eq. (28)]) $$\begin{gathered} \label{nequalzero} \left\langle \left\vert \rho^{PT}\right\vert ^{n}\right\rangle =\frac {n!\left( \alpha+1\right) _{n}\left( 2\alpha+1\right) _{n}}{2^{6n}\left( 3\alpha+\frac{3}{2}\right) _{n}\left( 6\alpha+\frac{5}{2}\right) _{2n}}\\ +\frac{\left( -2n-1-5\alpha\right) _{n}\left( \alpha\right) _{n}\left( \alpha+\frac{1}{2}\right) _{n}}{2^{4n}\left( 3\alpha+\frac{3}{2}\right) _{n}\left( 6\alpha+\frac{5}{2}\right) _{2n}}~_{5}F_{4}\left( \genfrac{}{}{0pt}{}{-\frac{n-2}{2},-\frac{n-1}{2},-n,\alpha+1,2\alpha +1}{1-n,n+2+5\alpha,1-n-\alpha,\frac{1}{2}-n-\alpha};1\right) .\end{gathered}$$ (The brackets denote expectation with respect to Hilbert-Schmidt \[Euclidean\] measure, while $5F4$ indicates a generalized hypergeometric function. The partial transpose of $\rho$, obtained by transposing in place its four $2 \times 2$ blocks, is denoted by $\rho^{PT}$.) 7,501 of these moments ($n =0,1,\ldots 7500$) were employed as input to a Mathematica program of Provost [@Provost pp. 19-20], implementing a Legendre-polynomial-based-moment-inversion routine. From the high-precision, exact-arithmetic results obtained, we were able to formulate highly convincing, well-fitting conjectures (including the above-mentioned $\frac{8}{33}$ for $\alpha =1$) as to underlying simple rational-valued separability probabilities. Then, with the use of the Mathematica FindSequenceFunction command applied to the sequence ($\alpha =1, 2,\ldots,32$) of these conjectures, and simplifying manipulations applied to the lengthy Mathematica result generated, we derived a multi-term $7F6$ hypergeometric-based formula [@slaterJModPhys Fig. 3] (cf. [@karolkarol eq. (11)]), with argument $\frac{27}{64}= (\frac{3}{4})^3$, for the conjectured values. Then, Qing-Hu Hou applied [@slaterJModPhys Figs. 5, 6] a highly celebrated (“creative telescoping”) algorithm of Zeilberger [@doron] to this expression to obtain the concise separability probability formula ((\[Hou1\])-(\[Hou3\])) for $P(\alpha)$ itself. In the course of his work in obtaining the $5F4$-hypergeometric-based HS moment formula above–and a more general one still for $\left\langle \left\vert \rho^{PT}\right\vert ^{n}\left\vert \rho\right\vert ^{k}\right\rangle /\left\langle \left\vert \rho\right\vert ^{k}\right\rangle$–Dunkl employed certain “utility functions”, in particular [@MomentBased p. 26], $$\begin{aligned} F_{2}\left( n,k\right) & =\left\langle \left\vert \rho\right\vert ^{k}\left( \left\vert \rho^{PT}\right\vert -\left\vert \rho\right\vert \right) ^{n}\right\rangle /\left\langle \left\vert \rho\right\vert ^{k}\right\rangle. \end{aligned}$$ Recently, upon request, he was able to obtain the explicit formula (Appendix) $$\begin{aligned} F_{2}\left( n,k\right) & = \frac{\left(-\frac{1}{16}\right)^n (\alpha )_n \left(\alpha +\frac{1}{2}\right)_n (2 k+n+5 \alpha +2)_n}{\left(k+3 \alpha +\frac{3}{2}\right)_n \left(2 k+6 \alpha +\frac{5}{2}\right)_{2 n}}\\ & \times~_{4}F_{3}\left( \genfrac{}{}{0pt}{}{-\frac{n}{2},\frac{1-n}{2},k+1+\alpha,k+1+2\alpha }{1-n-\alpha,\frac{1}{2}-n-\alpha,n+2k+2+5\alpha};1\right) .\end{aligned}$$ We set $k=0$ in this formula, and once again applied the Legendre-polynomial-based-moment-inversion procedure of Provost [@Provost], in the same manner as in our previous studies. It was first necessary to note, however, that rather than the variable range $-\frac{1}{16} \leq \|\rho^{PT}\| \leq \frac{1}{256}$ employed in these earlier studies, the appropriate range would now be $-\frac{1}{16} \leq (\|\rho^{PT}\|-\|\rho\|) \leq \frac{1}{432}$. (Note that $432 = 2^4 \cdot 3^3$, as well as, of course, $16=2^4$ and $256 =2^8$.). The lower bound of $-\frac{1}{16}$ is achieved by Bell states–one example being the density matrix with $\frac{1}{2}$ in its four corners, and zeros elsewhere–and the upper bound of $\frac{1}{432}$ by a $4 \times 4$ density matrix with diagonal entries $\{\frac{1}{6}, \frac{1}{3},\frac{1}{3},\frac{1}{6}\}$ and (1,4) and (4,1)-entries equal to $-\frac{1}{6}$, and zeros otherwise. (Note that if we interchange the roles of $\|\rho^{PT}\|$ and $\|\rho\|$ in this last example, a value of $-\frac{1}{432}$, the lower bound on the domain of separability, is obtained for the variable $(\|\rho^{PT}\|-\|\rho\|)$ of interest.) We crucially rely throughout these series of analyses upon the proposition that $\|\rho^{PT}\|>0$ is both a necessary and sufficient condition for a two-qubit state to be separable [@augusiak; @Demianowicz]. (We note that the partial transpose of a $4 \times 4$ density matrix $\rho$ can possess at most one negative eigenvalue, so the non-negativity of $\|\rho^{PT}\|$–the product of the four eigenvalues of $\rho^{PT}$–is tantamount to separability.) For the subrange $[0, \frac{1}{432}]$ of $(\|\rho^{PT}\|-\|\rho\|)$, containing only separable states, employing $\alpha =1$ in the new hypergeometric-based formula immediately above, we obtained, based on 9,451 ($n =0, 1,\ldots 9,450$) moments, an estimate that was 0.50000004358 as large as $\frac{8}{33}$. The parallel calculations in the two-rebit ($\alpha = \frac{1}{2}$) and two-“quaterbit” ($\alpha = 2$) cases yielded counterpart estimates of 0.5000025687 and 0.5000000000177, respectively. (Differences in rates of convergence–much the same as observed in [@MomentBased]–can be attributed to the initial \[zero[th]{}-order\] assumption of the Legendre-polynomial-moment-inversion procedure that the probability distributions to be fitted are uniform in nature, rendering more sharply-peaked distributions more difficult to rapidly approximate well. [A fortiori]{}, for the $\alpha =4$ (conjecturally octonionic) value [@slaterJModPhys p. 9], $P(4)= \frac{4482}{4091349}$, the computed value here was, $0.500000000000000015 \times P(4).$) These outcomes, certainly, help to strongly bolster the validity of the (yet formally unproven) concise formula of Hou ((\[Hou1\])-(\[Hou3\])), yielding the full generic Hilbert-Schmidt two-qubit separability probabilities $P(\alpha)$. For the two-rebit, two-qubit and two-quaterbit probabilities over the extended interval $[-\frac{1}{432}, \frac{1}{432}]$, containing all separable and now some entangled states (and thus providing upper bounds on the total separability probabilities), the estimates, again based on 9,451 moments were 0.78082617689, 0.69244685258 and 0.601390039979. However, we did not discern any particular underlying common structure in these values. As examples of entangled states dense in $[-\frac{1}{432}, 0]$, Dunkl advanced a one-parameter ($s \in [-\frac{1}{108}, 0]$)) family of density matrices with diagonal entries $\frac{1}{2} -\frac{s}{2}, \frac{s}{2}, \frac{s}{2},\frac{1}{2} -\frac{s}{2}$, and (1,4)- and (4,1)-entries equal to $s$, and zeros elsewhere. In Fig. \[fig:BisectionPlot\] we display an estimate based on the first 51 $(n =0,\ldots,50)$ moments of the probability distributions in question as a function of $\alpha$ over the subrange $[-\frac{1}{108},\frac{1}{432}]$ of the full range $[-\frac{1}{16},\frac{1}{432}]$ of $(\|\rho^{PT}\|-\|\rho\|)$. The distributions are more sharply peaked for smaller $\alpha$ (nearer to $\alpha = \frac{1}{2}$ in the plot), as the larger values of $P(\alpha)$ for smaller $\alpha$ would indicate. ![\[fig:BisectionPlot\]Estimate based on first 51 moments of the probability distributions, as a function of the Dyson-index-like parameter $\alpha$, of the variable ($\|\rho^{PT}\|-\|\rho\|$)](BisectionPlot2.pdf) Let us note that these “half-separability-probabilities” of $\frac{29}{128}, \frac{4}{33}, \frac{13}{323}$, indicated above, appear, by Hilbert-Schmidt-based analyses of Szarek, Bengtsson and [Ż]{}yczkowski [@sbz], to be exactly equal to the “full-separability-probabilities” for the corresponding minimally-degenerate (boundary) generic two-rebit, two-qubit and two-quaterbit states (for which the determinant $\|\rho\|=0$). It would certainly be of interest to attempt to reproduce these probabilities–again employing $\alpha$ as a Dyson-index-like parameter–in a moment-based analysis (cf. [@MomentBased App. D.7]) analogous to that conducted above and previously in [@MomentBased; @slaterJModPhys]. (The range of $\|\rho^{PT}\|$ under the constraint $\|\rho\|=0$ would, first, have to be determined.) These last three authors had established that the set of positive-partial-transpose states for an arbitrary bipartite systems is “pyramid-decomposable” and hence, a body of constant height“. They stated that ”since our reasoning hinges directly on the Euclidean geometry, it does not allow one to predict any values of analogous ratios computed with respect to the Bures measure, nor other measures“ [@sbz p. L125]. Let us, then, pose the question of whether the ”symmetric halves" finding elucidated above is itself particular to the Hilbert-Schmidt (flat, Euclidean) metric or, contrastingly, does extend to the use of alternative metrics, such as the Bures (minimal monotone) metric [@szBures; @ingemarkarol]? Also, in need of clarification is the issue of whether or not the Dyson-index [ansatz]{} of random matrix theory [@dumitriu]–apparently applicable in the Hilbert-Schmidt case, as our various results so far would indicate–extends to other measures, as well (cf. [@BuresHilbert; @Hybrid]). Appendix (C. Dunkl) =================== Let$$g\left( k,n\right) :=\frac{\left( k+1\right) _{n}\left( k+1+\alpha \right) _{n}\left( k+1+2\alpha\right) _{n}}{2^{6n}\left( k+3\alpha +\frac{3}{2}\right) _{n}\left( 2k+6\alpha+\frac{5}{2}\right) _{2n}},$$ there is a multiplication relation:$$g\left( 0,k\right) g\left( k,n\right) =g\left( 0,k+n\right) .$$ Let$$h\left( k,n\right) :=~_{5}F_{4}\left( \genfrac{}{}{0pt}{}{-n,-k,\alpha,\alpha+\frac{1}{2},-2k-2n-1-5\alpha }{-k-n-\alpha,-k-n-2\alpha,-\frac{k+n}{2},-\frac{k+n-1}{2}};1\right) .$$ Then$$\begin{aligned} \left\langle \left\vert \rho\right\vert ^{k}\right\rangle & =g\left( 0,k\right) \\ \left\langle \left\vert \rho^{PT}\right\vert ^{n}\left\vert \rho\right\vert ^{k}\right\rangle /\left\langle \left\vert \rho\right\vert ^{k}\right\rangle & =g\left( k,n\right) h\left( k,n\right) \\ \left\langle \left\vert \rho^{PT}\right\vert ^{n}\left\vert \rho\right\vert ^{k}\right\rangle & =g\left( 0,k+n\right) h\left( k,n\right) .\end{aligned}$$ Define$$F_{2}\left( n,k\right) =\left\langle \left\vert \rho\right\vert ^{k}\left( \left\vert \rho^{PT}\right\vert -\left\vert \rho\right\vert \right) ^{n}\right\rangle /\left\langle \left\vert \rho\right\vert ^{k}\right\rangle ,$$ then$$\begin{aligned} F_{2}\left( n,k\right) & =\frac{1}{g\left( 0,k\right) }\sum_{j=0}^{n}\binom{n}{j}\left( -1\right) ^{n-j}\left\langle \left\vert \rho\right\vert ^{k+n-j}\left\vert \rho^{PT}\right\vert ^{j}\right\rangle \\ & =\frac{1}{g\left( 0,k\right) }\sum_{j=0}^{n}\binom{n}{j}\left( -1\right) ^{n-j}g\left( 0,k+n\right) h\left( k+n-j,j\right) \\ & =g\left( k,n\right) \sum_{j=0}^{n}\binom{n}{j}\left( -1\right) ^{n-j}h\left( k+n-j,j\right) .\end{aligned}$$ We will produce $F_{2}^{\prime}\left( n,k\right) :=\sum_{j=0}^{n}\binom {n}{j}\left( -1\right) ^{n-j}h\left( k+n-j,j\right) $ as a single sum (so that $F_{2}\left( n,k\right) =g\left( k,n\right) F_{2}^{\prime}\left( n,k\right) $). Let $n,m=0,1,2,\ldots$ and let $x$ be a variable, if $0\leq m\leq n$ then$$\sum_{j=0}^{n}\frac{\left( -n\right) _{j}}{j!}\left( -j\right) _{m}\left( x+j\right) _{m}=\left( -1\right) ^{m}\frac{\left( x\right) _{2m}}{\left( x\right) _{n}}\left( -n\right) _{m}\left( -m\right) _{n-m},$$ otherwise the sum is zero. If $m>n$ then $\left( -j\right) _{m}=0$ for $0\leq j\leq n$. Suppose $0\leq m\leq n$ then $\left( -j\right) _{m}=0$ for $0\leq j<m$ and the sum is over $m\leq j\leq n$. Thus $$\begin{aligned} \sum_{j=m}^{n}\frac{\left( -n\right) _{j}}{j!}\left( -j\right) _{m}\left( x+j\right) _{m} & =\left( -1\right) ^{m}\sum_{j=m}^{n}\frac{\left( -n\right) _{j}~j!}{j!\left( j-m\right) !}\frac{\left( x\right) _{j}\left( x+j\right) _{m}}{\left( x\right) _{j}}\\ & =\left( -1\right) ^{m}\left( x\right) _{m}\sum_{j=m}^{n}\frac{\left( -n\right) _{j}\left( x+m\right) _{j}}{\left( j-m\right) !\left( x\right) _{j}}.\end{aligned}$$ Change the index of summation $j=m+i$ then the sum equals$$\begin{aligned} & \left( -1\right) ^{m}\frac{\left( -n\right) _{m}\left( x\right) _{m}\left( x+m\right) _{m}}{\left( x\right) _{m}}\sum_{i=0}^{n-m}\frac{\left( m-n\right) _{i}\left( x+2m\right) _{i}}{i!\left( x+m\right) _{i}}\\ & =\left( -1\right) ^{m}\frac{\left( -n\right) _{m}\left( x\right) _{2m}}{\left( x\right) _{m}}\frac{\left( -m\right) _{n-m}}{\left( x+m\right) _{n-m}}\\ & =\left( -1\right) ^{m}\left( -n\right) _{m}\left( -m\right) _{n-m}\frac{\left( x\right) _{2m}}{\left( x\right) _{n}},\end{aligned}$$ by the Chu-Vandermonde sum. Observe that $\left( -m\right) _{n-m}=0$ for $2m<n$. Then$$\begin{aligned} F_{2}^{\prime}\left( n,k\right) & =\left( -1\right) ^{n}\sum_{j=0}^{n}\frac{\left( -n\right) _{j}}{j!}\\ & \times\sum_{i=0}^{n}\frac{\left( -j\right) _{i}\left( j-k-n\right) _{i}\left( \alpha\right) _{i}\left( \alpha+\frac{1}{2}\right) _{i}\left( -2k-2n-1-5\alpha\right) _{i}}{i!\left( -k-n-\alpha\right) _{i}\left( -k-n-2\alpha\right) _{i}\left( -\frac{k+n}{2}\right) _{i}\left( -\frac{k+n-1}{2}\right) _{i}}\\ & =\left( -1\right) ^{n}\sum_{i=0}^{n}\frac{\left( \alpha\right) _{i}\left( \alpha+\frac{1}{2}\right) _{i}\left( -2k-2n-1-5\alpha\right) _{i}}{i!\left( -k-n-\alpha\right) _{i}\left( -k-n-2\alpha\right) _{i}\left( -\frac{k+n}{2}\right) _{i}\left( -\frac{k+n-1}{2}\right) _{i}}\\ & \times\sum_{j=0}^{n}\frac{\left( -n\right) _{j}}{j!}\left( -j\right) _{i}\left( j-k-n\right) _{i}.\end{aligned}$$ Apply the lemma to the $j$-sum with $x=-k-n$ and $m=i$ to obtain$$\left( -1\right) ^{i}\left( -n\right) _{i}\left( -i\right) _{n-i}\frac{\left( -n-k\right) _{2i}}{\left( -n-k\right) _{n}}=\left( -1\right) ^{i}\frac{\left( -n\right) _{i}\left( -i\right) _{n-i}}{\left( -n-k\right) _{n}}2^{2i}\left( -\frac{k+n}{2}\right) _{i}\left( -\frac{k+n-1}{2}\right) _{i}$$ and thus $$F_{2}^{\prime}\left( n,k\right) =\frac{\left( -1\right) ^{n}}{\left( -n-k\right) _{n}}\sum_{i=0}^{n}\frac{\left( -n\right) _{i}\left( -i\right) _{n-i}\left( \alpha\right) _{i}\left( \alpha+\frac{1}{2}\right) _{i}\left( -2k-2n-1-5\alpha\right) _{i}}{i!\left( -k-n-\alpha\right) _{i}\left( -k-n-2\alpha\right) _{i}}\left( -1\right) ^{i}2^{2i}.$$ This is not in hypergeometric form because of the term $\left( -i\right) _{n-i}$; also the summation extends over $\frac{n}{2}\leq i\leq n$. Change the index $j=n-i$ then $$\begin{aligned} \frac{\left( -n\right) _{i}}{i!}\left( -i\right) _{n-i} & =\left( -1\right) ^{i}\frac{n!}{\left( n-i\right) !i!}\left( -1\right) ^{n-i}\frac{i!}{\left( 2i-n\right) !}=\left( -1\right) ^{n}\frac {n!}{j!\left( n-2j\right) !}\\ & =\left( -1\right) ^{n}\frac{2^{2j}}{j!}\left( -\frac{n}{2}\right) _{j}\left( \frac{1-n}{2}\right) _{j}$$ and the reversal formula is$$\begin{aligned} \left( x\right) _{i} & =\left( x\right) _{n-j}=\frac{\left( x\right) _{n-j}\left( x+n-j\right) _{j}}{\left( x+n-j\right) _{j}}\\ & =\left( -1\right) ^{j}\frac{\left( x\right) _{n}}{\left( 1-n-x\right) _{j}}.\end{aligned}$$ Thus$$\begin{aligned} F_{2}^{\prime}\left( n,k\right) & =\frac{\left( -1\right) ^{n}\left( \alpha\right) _{n}\left( \alpha+\frac{1}{2}\right) _{n}\left( -2k-2n-2k-1-5\alpha\right) _{n}}{\left( -n-k\right) _{n}\left( -k-n-\alpha\right) _{n}\left( -k-n-2\alpha\right) _{n}}\\ & \times\sum_{j=0}\frac{\left( -\frac{n}{2}\right) _{j}\left( \frac{1-n}{2}\right) _{j}\left( k+1+\alpha\right) _{j}\left( k+1+2\alpha\right) _{j}}{j!\left( 1-n-\alpha\right) _{j}\left( \frac{1}{2}-n-\alpha\right) _{j}\left( n+2k+2+5\alpha\right) _{j}}2^{2j+2n-2j}\\ & =\left( -1\right) ^{n}2^{2n}\frac{\left( \alpha\right) _{n}\left( \alpha+\frac{1}{2}\right) _{n}\left( n+2k+2+5\alpha\right) _{n}}{\left( k+1\right) _{n}\left( k+1+\alpha\right) _{n}\left( k+1+2\alpha\right) _{n}}\\ & \times~_{4}F_{3}\left( \genfrac{}{}{0pt}{}{-\frac{n}{2},\frac{1-n}{2},k+1+\alpha,k+1+2\alpha }{1-n-\alpha,\frac{1}{2}-n-\alpha,n+2k+2+5\alpha};1\right) ;\end{aligned}$$ a balanced sum. The formula was tested for $F_{2}\left( 2,k\right) $, also directly verified for $n=3$, arbitrary $\alpha$. Combining the front factors in $F_{2}\left( n,k\right) $ (from $g\left( k,n\right) $) we obtain$$\left( -1\right) ^{n}\frac{\left( \alpha\right) _{n}\left( \alpha +\frac{1}{2}\right) _{n}\left( n+2k+2+5\alpha\right) _{n}}{2^{4n}\left( k+3\alpha+\frac{3}{2}\right) _{n}\left( 2k+6\alpha+\frac{5}{2}\right) _{2n}}.$$ I would like to express appreciation to the Kavli Institute for Theoretical Physics (KITP) for computational support in this research. C. Dunkl supplied the crucial $4F3$-hypergeometric-based formula employed, as well as much useful advice.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Plutonium metal exhibits an anomalously large softening of its bulk modulus at elevated temperatures that is made all the more extraordinary by the finding that it occurs irrespective of whether the thermal expansion coefficient is positive, negative or zero — representing an extreme departure from conventional Grüneisen scaling. We show here that the cause of this softening is the compressibility of plutonium’s thermally excited electronic configurations, which has thus far not been considered in thermodynamic models. We show that when compressible electronic configurations are thermally activated, they invariably give rise to a softening of bulk modulus regardless of the sign their contribution to the thermal expansion. The electronically driven softening of the bulk modulus is shown to be in good agreement with elastic moduli measurements performed on the gallium-stabilized $\delta$ phase of plutonium over a range of temperatures and compositions, and is shown to grow rapidly at small concentrations of gallium and at high temperatures, where it becomes extremely sensitive to hydrostatic pressure.' author: - 'N. Harrison' title: 'Electronically driven collapse of the bulk modulus in $\delta$-plutonium ' --- Introduction ============ Plutonium (Pu) has the richest phase diagram amongst the metallic elements, and, as a consequence, has proven to be the most challenging to grasp.[@moore2009; @hecker2000; @smith1983; @hecker2004] In recent years, considerable advances have been made towards understanding some of Pu’s unusual thermodynamic properties, such as its anonamously enhanced electronic heat capacity for an element[@lashley2003; @shim2007; @zhu2007; @lanata2015] and its invar-like negative thermal expansion coefficient.[@lawson2002; @lawson2006; @harrison2019] The softening of the bulk modulus,[@soderlind2010; @suzuki2011; @freibert2012; @migliori2016; @nadal2010] by contrast, which occurs $\sim$ 50% more rapidly with increasing temperature (relative to the melting temperature) than regular solids,[@digilov2019; @ida1969; @born1939; @varshni1970; @rose1984; @anderson1989; @ida1970] has continued to remain a mystery. The increased likelihood that the entirety of the bulk modulus softening cannot be explained by phonons alone has led to the suggestion of an unconventional contribution originating from electronic degrees-of-freedom.[@migliori2016; @lawson2019] A natural candidate for electronic degrees-of-freedom in Pu is provided by its unstable $5f$-electron atomic shell, which has been shown to allow Pu to exist in a greater number of near degenerate electronic configurations[@eriksson1999; @wills2004; @svane2007] and oxidation states[@windorff2017] than other actinides and rare earths. Direct experimental evidence for the presence of multiple near-degenerate electronic configurations in $\delta$ phase Pu ($\delta$-Pu) has been provided by way of x-ray spectroscopy[@booth2012] and neutron scattering experiments.[@janoschek2015] While it has been argued on the basis of sophisticated electronic structure models that virtual valence fluctuations cause extensive mixing between these configurations,[@savrasov2001; @shim2007; @zhu2007; @lanata2015] a strong thermally-activated component has been indicated by way of thermal expansion experiments[@lawson2002; @lawson2006] and, more recently, by way of temperature-dependent magnetostriction experiments.[@harrison2019] We show here that a crucial factor in causing multiple electronic configurations to contribute significantly to the bulk modulus is their compressibility, which we show to give rise to a previously unknown yet significant electronic contribution to the bulk modulus when excited electronic configurations are thermally activated. Because this contribution is both negative and quadratic in the size of the difference in equilibrium volume (manifesting itself as a partial pressure) between the excited configurations and the lattice, it invariably leads to a softening of the bulk modulus with increasing temperature. Using a form for the free energy recently adapted from measurements of different thermodynamic quantities,[@lawson2006; @harrison2019] we show that the uncovered electronically driven softening of the bulk modulus is in agreement with temperature-dependent and Ga concentration-dependent resonant ultrasound spectroscopy results, in which Ga is used to stabilize the $\delta$ phase.[@suzuki2011; @freibert2012; @migliori2016; @soderlind2010] The softening is shown to become especially large in $\delta$-Pu stabilized with small concentrations of Ga at temperatures well above room temperature, where it is further predicted to undergo a collapse under hydrostatic pressure. Results ======= Origin of the bulk modulus softening ------------------------------------ Studies of the thermodynamic properties of Pu have shown that their temperature dependences can be modeled by considering a partition function of the form[@migliori2016; @harrison2019; @lawson2006; @lawson2002] $Z_{\rm el}=\sum_{i}{\rm e}^{-\frac{E_i}{k_{\rm B}T}}$, where $i$ refers to different electronic configurations with fixed energies $E_i$ and atomic volumes $V_i$. Use of such a partition function for modeling thermodynamic quantities is warranted under circumstances where the mixing between different electronic configurations attributable to valence fluctuations is sufficiently small for the higher energy configurations to be thermally activated.[@harrison2019; @lawrence1981; @wohlleben1984] We find, however, that whereas the consideration of $E_i$ and $V_i$ as fixed and independent quantities is a reasonable approximation for modeling the thermal expansion, heat capacity and the magnetostriction, [@harrison2019; @lawson2006; @lawson2002] this is not the case when considering the bulk modulus (see Fig. \[experimentalbulkmod\]). Since the relationship between $E_i$ and $V$ forms the basis of the definition of the bulk modulus of a material, neglect of this relationship has the potential to cause entire terms to be missing from the equation of state. Electronic structure calculations have shown that $E_i$ and $V$ are inextricably linked for each electronic configuration of Pu and other actinides,[@svane2007; @eriksson1999] making $E_i(V)$ a function of $V$, or, equivalently, $E_i(\nu)$ a function of the volume strain $\nu=\frac{V}{V_0}-1$ (see schematic in Fig. \[schematic\]). A more generalized form $$\label{electronictruncpartitionfunction} Z_{\rm el}(\nu)=\sum_{i}{\rm e}^{-\frac{E_i(\nu)}{k_{\rm B}T}}$$ for the partition function that preserves information relating to the compressibility is therefore required. ![[a]{} A comparison of experimental adiabatic bulk modulus data for different compositions $x$ of $\delta$-Pu$_{1-x}$Ga$_x$ as indicated[@soderlind2010; @migliori2016; @suzuki2011; @freibert2012] with the model calculations of $K_S=\gamma K_T$ (where $K_T=K_{\rm el}+K_{\rm ph}$ is the isothermal bulk modulus and $\gamma$ is plotted in the Appendix) for different compositions, as indicated. As a point of reference, the black dotted line is the fitted functional form of $K_{\rm ph}$ (for $b=$ 18); we have added this to $K_0=$ 37.7 GPa in order to bring it into alignment with the other curves at $T\approx$ 10 K. [b]{} A comparison of the experimentally measured change in temperature-dependent electronic contribution to the bulk moduli $\Delta K_{\rm el}=K_{\rm el}(T)-K_{\rm el}(280~{\rm K})$ for $x=$ 2.36%, 3.30% and 4.64% (colored squares), having subtracted the measured values at $T=$ 280 K and the calculated change $\Delta K_{\rm ph}=K_{\rm phl}(T)-K_{\rm ph}(280~{\rm K})$ in $K_{\rm ph}$ (again for $b=$ 18) relative to that at $T=$ 280 K (assuming $K_{\rm ph}$ to be independent of $x$), with the equivalent change $\Delta K_{\rm el}$ in the electronic contribution (colored lines) relative to that at $T=$ 280 K calculated using Equation (\[bulkmodulusequation\]). For non-integer values of $x$, calculated values of $K_{\rm el}$ are interpolated in $x$.[]{data-label="experimentalbulkmod"}](experimentalbulkmod092719.pdf){width=".35\textwidth"} ![Schematic $E(V)$ curves (lower axis) or $E(\nu)$ curves (upper axis) as described in the text according to Equation (\[cohesionvolume\]), showing an approximately parabolic form. The energy minima $E_{i,0}$ and the energies $E^\ast_i$ at $V=V_0$ or $\nu=0$, in relation to the volumes $V_i$ or volume strains $\nu_i$ are also indicated. []{data-label="schematic"}](schematic092719.pdf){width=".4\textwidth"} Electronic structure calculations have show that the $E_i(\nu)$ curves of the different configurations (illustrated in Fig. \[schematic\]) are approximately parabolic,[@svane2007; @eriksson1999] enabling them to be represented in the reduced form $$\label{cohesionsmall} \lim_{\nu-\nu_i\rightarrow0}E_i(\nu)=E_{i,0}+\frac{K_{i}~[\nu-\nu_i]^2}{2N}$$\ for small $\nu$, where $N$ is the atomic density and each of their contributions to the bulk modulus is given by $K_i(\nu)=N\frac{\partial^2E_i(\nu)}{\partial\nu^2}$ evaluated at $\nu=0$. Here, $E_{i,0}$ is the energy minimum and $\nu_i=\frac{V_i}{V_0}-1$ is the volume strain at the energy minimum for each configuration (see Fig. \[schematic\]). For the lowest energy (ground state) configuration, by definition $i=0$, $V_i=V_0$, $\nu_{i}=0$ and $E_{i,0}=0$. The excited electronic configurations, by contrast, are in nonequilibrium states, causing them each to exert a statistical partial pressure $P_i(\nu)=-N\frac{\partial E_i(\nu)}{\partial\nu}$ on the surrounding lattice. In the limit of small strain, this pressure is simply $$\label{effectivepressure} \lim_{\nu-\nu_i\rightarrow0}P_i(\nu)=-{K_{i}~[\nu-\nu_i]}.$$ The ensemble average of the volume changes associated with these partial pressures is what ultimately drives the positive and negative electronic contributions to the thermal expansion (see Appendix).[@harrison2019] We proceed to calculate the electronic contribution to the bulk modulus by taking the second derivative $K_{\rm el}=\frac{\partial^2F_{\rm el}}{\partial\nu^2}\big\|_T$ of the electronic component $F_{\rm el}=-k_{\rm B}T\ln Z_{\rm el}(\nu)$ of the free energy with respect to the volume strain. Using the volume-dependent partition function given by Equation (\[electronictruncpartitionfunction\]) and $E_i(\nu)$ curves, we obtain $$\label{bulkmodulusequation} K_{\rm el}={{\sum_{i}p_i(\nu)K_i(\nu)}}-\frac{\sum_{i}p_i(\nu)P^2_i(\nu)}{Nk_{\rm B}T} +\frac{\big[\sum_{i}p_i(\nu)P_i(\nu)\big]^2}{Nk_{\rm B}T},$$ where $$\label{probability} p_i(\nu)=Z_{\rm el}^{-1}{\rm e}^{-\frac{E_i(\nu)}{k_{\rm B}T}}$$ is the probability of occupancy for each configuration. The first term on the right-hand-side of Equation (\[bulkmodulusequation\]) is the probability-weighted sum of bulk moduli that has been assumed in prior models of the multiple electronic configurations in $\delta$-Pu.[@lawson2002; @lawson2006; @migliori2016] While large changes in $K_i(\nu)$ with $i$, (we discuss the extreme case where $K_i=0$ for an excited configuration in the Appendix)[@lawson2006; @lawson2019] have the potential to yield significant changes in $K_{\rm el}$ with temperature, the bulk moduli contributions of all of the electronic configurations obtained by density functional theory are found to all be very similar at $\nu=0$.[@eriksson1999; @svane2007] In Table \[table2\], we find these to have a mean value of $\bar{K}_i(\nu=0)=$ 28.2 GPa and a standard deviation of only $\sigma K_i(\nu=0)=$ 5.0 GPa. The first term in Equation (\[bulkmodulusequation\]) is therefore not expected to lead to significant changes of the bulk modulus with increasing temperature. The second term on the right-hand-side of Equation (\[bulkmodulusequation\]) has the potential to lead to much larger changes in the bulk modulus of $\delta$-Pu with increasing temperature, making it the primary motivation of the present study. The origin of this term is the statistical partial pressure $P_i(\nu)$ between thermally excited configurations and the ground state that occurs as a result of their equilibrium volume strains $\nu_i$ being nonzero. For sufficiently small total strains $\nu-\nu_i$, this partial pressure is linear as shown in Equation (\[effectivepressure\]). Because the partial pressure produces a negative quadratic contribution to the electronic bulk modulus in Equation (\[bulkmodulusequation\]), it implies that thermally fluctuating electronic configurations invariably lead to a softening of the bulk modulus irrespective of whether $\nu_i>0$, as for a positive contribution to the thermal expansion, or $\nu_i<0$, as for a negative contribution to the thermal expansion (see Appendix). The third term on the right-hand-side of Equation (\[bulkmodulusequation\]) is also determined by $P_i(\nu)$. However, because the probability factors $p_i(\nu)$ in this term are multiplied together, its overall contribution to the bulk modulus is weaker than that of the second term. For the effect of hydrostatic pressure on the bulk modulus, this we estimate by taking the third derivative of the free energy with respect to $\nu$ and considering $\frac{\partial}{\partial P}=-K_T^{-1}\frac{\partial}{\partial\nu}$, where $K_T$ is the isothermal bulk modulus, whereupon we obtain $$\label{pressurederivative} K^\prime=\frac{\partial K_{\rm el}}{\partial P}\bigg\|_T\approx\frac{1}{K_T[Nk_{\rm B}T]^2}\bigg[{{{\sum_{i}p_i(\nu)P_i^3(\nu)}}}-3{\big[\sum_{i}p_i(\nu)P^2_i(\nu)\big]\big[\sum_{i}p_i(\nu)P_i(\nu)\big]} +2{\big[\sum_{i}p_i(\nu)P_i(\nu)\big]^3}+\delta\bigg]$$ — once again assuming the parabolic approximation (see Appendix). The first term on the right-hand-side in Equation (\[pressurederivative\]), which originates from the derivative of the anomalous softening (i.e. the second) term in Equation (\[bulkmodulusequation\]), is found to dominate over the other terms. Its dominance implies that the sign and magnitude of the change in bulk modulus under pressure is determined almost entirely by the partial pressures of the electronic configurations, which in turn depend on the signs of $\nu_i$. The cubic dependence on $P_i(\nu)$ implies $K^\prime$ has a more extreme sensitivity to composition than $K_T$. The last term on the right-hand-side of Equation (\[pressurederivative\]), $\delta$, is a correction term (see Appendix) that vanishes in the limit where the bulk moduli of the electronic configurations are the same. Electronically driven softening estimates ----------------------------------------- We proceed to estimate the electronic contribution to the bulk modulus and its pressure derivative in Fig. \[electronicbulkmod\] from the multiple electronic configurations, by defining $E_i^\ast\approx E_{i,0}+K_i\nu_i^2/2N$ according to Equation (\[cohesionsmall\]) and using the approximation $P_i\approx K_i\nu_i$ according to Equation (\[effectivepressure\]).[@harrison2019] Low temperature specific heat measurements have shown that the Debye temperature $\Theta_{\rm D}\approx$ 100 K remains largely unchanged as a function of the Ga concentration $x$ used to stabilize the $\delta$ phase,[@harrison2019] which is consistent with the parabolic approximation given by Equation (\[cohesionsmall\]). We therefore assume that the bulk modulus of the ground state electronic configuration also remains unchanged, and adopt the value $K_0=$ 37.7 GPa found in $\delta$-Pu$_{1-x}$Ga$_x$ for $x=$ 2.36%[@suzuki2011] by way of resonant ultrasound measurements. Since the bulk moduli of the excited electronic configurations in $\delta$-Pu are unknown, yet are predicted to fall within a narrow range of possible values in Table \[table2\], we further assume the excited configurations to have bulk moduli $K_i$ that are similar to that of the ground state configuration (see Table \[table1\]). ![[a]{}, $K_{\rm el}$. of Pu$_{1-x}$Ga$_x$ as a function of temperature $T$ and Ga composition $x$ (colored lines) calculated by way of Equations (\[cohesionsmall\]) though (\[probability\]), using the parameters listed in Table \[table1\] for $x=$ 2% and $x=$ 7%. Parameters are extrapolated for $x=$ 0% and interpolated for $x=$ 4% and 6%. The inset shows $\nu_i^2$ for $i=1$ and 2 versus $V_0$, with the percentage of Ga ($x$) indicated for each point. It is important to note that because $P_i=0$ for the ground state configuration, the denominators of the second and third terms in Equation (\[bulkmodulusequation\]) do not lead to a divergence at $T=0$. [b]{}, Calculated pressure derivative of the bulk modulus $K^\prime$ according to Equation (\[pressurederivative\]).[]{data-label="electronicbulkmod"}](electronicbulkmod112019.pdf){width=".35\textwidth"} [c c c c c]{}\ Quantity & $x=$ 2% Ga & all $x$ & $x=$ 7% Ga & Units\ \[0.5ex\] $K_0=K_1=K_2$&&37.7&&GPa\ $\nu_0$&&0&&-\ $\nu_1$&3.7 $\times$ 10$^{-3}$&&2.1 $\times$ 10$^{-2}$&-\ $\nu_2$&-0.13(-0.17)&&-6.7 $\times$ 10$^{-3}$&-\ $E_0^\ast$&&0&& meV\ $E_1^\ast$&22.8&&41.0& meV\ $E_2^\ast$&125(121)&&70& meV\ $b$&&18&&-\ $T_0$&&1.39 $\times$ 10$^5$&&K\ $V_0$&24.57&&23.87&Å$^3$\ \[table1\] In calculating the electronic contribution to the bulk modulus, we use excitation energies $E_i^\ast$ and equilibrium volume strains $\nu_i$ determined[@harrison2019] (values listed in Table \[table1\]) from fitting thermal expansion[@lawson2002; @lawson2006] and temperature-dependent magnetostriction[@harrison2019] measurements, the results of which were further validated using heat capacity measurements.[@harrison2019; @lashley2003] The resulting calculations of the electronic contribution to the bulk modulus as a function of $T$ and $x$ using Equations (\[cohesionsmall\]) though (\[probability\]) are shown in Fig. \[electronicbulkmod\]a. Both of the previously determined[@harrison2019] excited electronic configurations ($E_1^\ast$ and $E_2^\ast$) are found to lead to discernible reductions in the bulk modulus with increasing temperature. Since $E_1^\ast$ leads to a positive thermal expansion[@harrison2019] whereas $E_2^\ast$ leads to a negative thermal expansion,[@lawson2002; @lawson2006] yet both lead to a softening of the bulk modulus, it is the combination of both of these terms that is responsible for the departures from simple Grüneisen scaling in $\delta$-Pu. The degree of bulk modulus softening with temperature is significantly more pronounced for the higher excitation energy $E_2^\ast=$ 125 meV (for $x=$ 2%) at small Ga concentrations due to the large ($\approx$ 13%) difference between its equilibrium volume $V_2$ and the ground state volume $V_0$. Very similar results at high temperatures would therefore be obtained on neglecting $E_1^\ast$ and $\nu_1$ and considering only the excitation energy $E_2^\ast$ and volume strain $\nu_2$ determined from the invar model fits.[@lawson2002; @lawson2006] (see Table \[table1\]). According to our calculation, this dominant excitation is predicted to yield a reduction in the bulk modulus that is as large as $\approx$ 8 GPa in pure $\delta$-Pu at 700 K; beyond this temperature the $\delta$ phase becomes unstable.[@hecker2000] Turning to the effect of hydrostatic pressure on the bulk modulus, since the leading term on the right-hand-side of Equation (\[pressurederivative\]) varies as the cube of the partial pressure, $K^\prime$ is found to be strongly dependent on $x$. The invar contribution, which is characterized by a negative partial pressure, clearly dominates, leading to the prediction of a dramatic collapse of the bulk modulus under pressure and at high temperatures for small concentrations of Ga in Fig. \[electronicbulkmod\]b. Comparison with experiment -------------------------- In order to compare the calculations against experimental data, we must also include the phonon contribution to the bulk modulus, which is known to reduce the bulk modulus of most materials by $\approx$ 20% upon reaching $T=T_{\rm m}/2$,[@digilov2019; @ida1969; @born1939; @varshni1970; @rose1984; @anderson1989; @ida1970] where $T_{\rm m}$ is the melting temperature ($T_{\rm m}\approx$ 912 K in Pu). While a universal model able to accurately describe the reduction in bulk modulus $K_{\rm ph}$ attributable to phonons in all materials has yet to be developed,[@digilov2019; @ida1969; @born1939; @varshni1970; @rose1984; @anderson1989; @ida1970] the model of Ida[@ida1969; @ida1970] has been shown to provide a good description of the heat capacity of $\delta$-Pu at temperatures above room temperature[@lawson2019] – most notably an observed upturn in the heat capacity above $\sim$ 600 K. Since the electronic and phonon contributions to the free energy are additive, this should, to a first approximation, be similarly true for derivatives, in which case $K_T=K_{\rm el}+K_{\rm ph}$ (see Appendix) for the isothermal bulk modulus. To compare with the adiabatic bulk modulus $K_S$ obtained by ultrasound measurements, we use the fact that $K_S=\gamma K_T$, where $\gamma\approx1$ for $\delta$-Pu (see Appendix). In comparing the calculation with the experimental bulk modulus data, only the phonon scaling coefficient $b$ is adjusted. The remainder of the parameters are taken from published results (tabulated in Table \[table1\]). Figure \[experimentalbulkmod\]a shows that on combining the electronic and lattice vibration contributions, a phonon coefficient $b=$ 18 (see Table \[table1\] and Appendix) yields a $K_S$ that closely follows the temperature-dependence of bulk modulus measured in $\delta$-Pu$_{1-x}$Ga$_x$ (with $x=$ 2.36%)[@suzuki2011] over a broad range of temperatures. Establishing further confidence in the model is the finding that the value of the phonon coefficient $b=$ 18 that best fits the phonon part of the bulk modulus is very similar to that $b=$ 16 that best fits the anharmonic phonon contribution to the high temperature heat capacity.[@lawson2019] For this value of $b$, $K_{\rm ph}$ also accounts for an $\approx$ 20% reduction in $K_S$ with temperature at $T_{\rm m}/2\approx$ 456 K (the remainder coming from the reduction in $K_{\rm el}$), therefore making the phonon contribution to the softening comparable to that in other materials.[@digilov2019; @ida1969; @born1939; @varshni1970; @rose1984; @anderson1989; @ida1970] Having established the approximate form of the (assumed) $x$-independent phonon contribution $K_{\rm ph}$ to the bulk modulus softening, we can proceed to subtract this contribution from measurements of $K_S$ in other samples so as to isolate the electronic contribution $K_{\rm el}$ to the bulk modulus softening, and to investigate its changes with temperature and composition $x$. Figure \[experimentalbulkmod\]b shows the $T$-dependence of the residual electronic contribution $K_{\rm el}$ for samples of three different compositions,[@soderlind2010] $x=$ 2.36%, 3.30% and 4.64%, after having subtracted $K_{\rm ph}$ as well as an offset (see Appendix) to bring the measured curves into alignment at $T=$ 280 K. Despite the limited range in temperature of these measurements, significant differences in the temperature dependences are clearly discernible. On comparing the model predictions of $K_{\rm el}$ calculated for the same $x$ compositions using Equation (\[bulkmodulusequation\]) with the experimental curves in Fig. \[experimentalbulkmod\]b, we find them to be in excellent agreement — both with regards to the temperature-dependence and the $x$-dependence of the temperature-dependence of the experimental data. Apart from a subtraction of the values of $K_{\rm el}$ at $T=$ 280 K that are necessary to eliminate offsets between the experimental curves, no adjustment has been made to $K_{\rm el}$ calculated using Equation (\[bulkmodulusequation\]) — the parameters used are those determined elsewhere by fitting other thermodynamic quantities[@harrison2019] (tabulated in Table \[table1\]). The other compositions, $x=$ 0.2% and $x=$ 0%[@freibert2012; @migliori2016] in Fig. \[experimentalbulkmod\]a, exhibit trends relative to other compositions that are consistent with the model calculations of $K_S$. However, the lack of temperature-dependent data for these compositions means we cannot isolate the electronic component for these compositions. ![Calculated pressure derivative of the bulk modulus $K^\prime$ of Ga-stabilized $\delta$-Pu at 300 K (pink line and points) according to Equation (\[pressurederivative\]) versus atomic volume $V$ at 300 K for different concentrations $x$ of Ga as indicated. This is compared with experimental data for Ga and Am stabilized $\delta$-Pu for different concentrations $x$ as indicated (brown circles). Also shown, for comparison, are the value (light blue square) for $\alpha$-Pu (2.3% Ga), the calculated Ga-stabilized values $\delta$-Pu values at 600 K (cyan line and points) and the value expected for a normal metal (black dashed line). All of the experiments were performed on $^{239}$Pu with the exception of the sample with $x=$ 3.5%, for which $^{242}$Pu was measured.[]{data-label="pressurebulkmod"}](bulkmodpressure100319){width=".35\textwidth"} Turning once again to the effect of pressure, a fundamental question concerns whether prior measurements of a negative thermal expansion serve as a reliable predictor of a pressure-induced bulk modulus softening.[@lawson2006; @klosek2008] Existing experimental studies have focused only on the effect of hydrostatic pressure on the bulk modulus at ambient temperature ($\approx$ 300 K),[@klosek2008; @faure2005; @faure2010; @zhang2019] the results of which are compared with the model calculations at 300 K (from Fig. \[experimentalbulkmod\]b) in Fig. \[pressurebulkmod\]. Importantly, Equation (\[pressurederivative\]) is found to yield a pressure-induced change in $\delta$-Pu whose negative sign agrees with the softening obtained experimentally; on taking an average of the $\delta$-Pu data points in Fig. \[pressurebulkmod\] (including both Ga-stabilized and Am-stabilized $\delta$-Pu), we obtain $K^\prime=$ $-3$ $\pm$ $2$. The calculated value is clearly different from the positive (i.e. stiffening) $K^\prime=+9$ $\pm$ 2 value measured in $\alpha$-Pu (plotted for comparison),[@faure2010] and the stiffening $K^\prime=+4$ expected for a normal metal. However, the degree of scatter in the experimental data, the magnitude of the error bars and the lack of temperature-dependent data prevent firm conclusions from being reached concerning the absolute magnitude, the doping-dependence or the volume-dependence of $K^\prime$ (see Appendix). Figures \[electronicbulkmod\]b and \[pressurebulkmod\] show that a significant increase in the observable pressure-induced softening of Ga-stabilized $\delta$-Pu could in principle be achieved by increasing the experimental temperature, which would also enable a more robust verification of the role of excited electronic configurations to be made. Access to higher temperatures would also enable such experiments to be carried out on pure $\delta$-Pu. Discussion ========== By taking into consideration the compressible nature of the previously isolated multiple electronic configurations in $\delta$-Pu,[@harrison2019; @lawson2002; @lawson2006] we have discovered a previously unknown yet significant electronic contribution to the bulk modulus. We find this contribution to be primarily responsible for the excess softening of the bulk modulus in Pu, the dominant energy scale of which is the same as that previously associated with the invar effect.[@lawson2002; @lawson2006] The effect of the electronically driven softening is discernible in Ga composition-dependent experiments close to room temperature, but is shown to be strongly enhanced at temperatures substantially above room temperature in $\delta$-Pu$_{1-x}$Ga$_x$ samples with low concentrations of Ga. The bulk modulus is further shown to soften under pressure, as found experimentally,[@klosek2008; @faure2005; @faure2010] and is further predicted to undergo a collapse at low concentrations of Ga and high temperatures. Conversely, the electronic contribution to the softening is expected to be much smaller for samples with large concentrations of Ga at high temperatures, providing an opportunity for the phonon contribution $K_{\rm ph}$ to be more accurately isolated in future studies. What constitutes a significant advance in the present approach to modeling the bulk modulus is that the softening is based entirely on the same free energy that has been shown to accurately describe other thermodynamic quantities as a function of temperature and Ga composition. These include the experimentally observed thermal expansion,[@lawson2002; @lawson2006] magnetostriction[@harrison2019] and heat capacity.[@lashley2003; @lawson2019; @harrison2019] It also involves the same energy scale associated with the invar effect and detected in neutron scattering experiments.[@janoschek2015] Our results reveal the central role played by statistical thermodynamics in the equation of state in plutonium near ambient pressure. Appendix ======== Breakdown of conventional Grüneisen scaling ------------------------------------------- Whereas in ordinary solids the combination of the volume thermal expansivity $\alpha_v$, the molar heat capacity $C_v$, the isothermal bulk modulus $K_T$ and the molar volume $V_{\rm m}$ equates to a dimensionless Grüneisen coefficient[@mitra1957; @mitra1986] $$\label{gruneisen} \Gamma=\frac{\alpha_vV_{\rm m}K_T}{C_v},$$ simple scaling breaks down in Pu owing to multiple sources of entropy.[@lawson2002; @lawson2006; @harrison2019] This is demonstrated spectacularly by the bulk modulus continuing to fall unabated with increasing temperature[@suzuki2011; @freibert2012; @migliori2016; @soderlind2010] regardless of whether the thermal expansion coefficient is positive, as is the case for most solids and phases of Pu,[@lawson2002; @lawson2006] or whether it changes to negative, as occurs for the $\delta$ phase of Pu ($\delta$-Pu) at elevated temperatures and low concentrations of doped gallium (Ga).[@lawson2002; @lawson2006] Energy-versus-volume -------------------- The calculation of the electronic contribution to the bulk modulus from the second derivative of the free energy requires knowledge of the functional forms of $E_i(\nu)$, for which we turn to a generalized model of cohesion in metals. The total internal energy for a given electronic configuration is determined by a balance between Coulomb ($\propto\frac{1}{a}$) and kinetic ($\propto\frac{1}{a^2}$) energy terms, leading to an energy curve of the form[@ashcroft1976] $$\label{cohesion} E_i(a)=a_0-\frac{a_1}{a}+\frac{a_2}{a^2},$$ where $a$ is the lattice spacing and $a_0$, $a_1$ and $a_2$ are constants. Here, each $E_i(a)$ curve corresponds to different number of $5f$-electrons, $n_f=0,1,2\dots$, confined to the atomic core.[@eriksson1999; @svane2007] We can then proceed to obtain the functional form for each $E_i(\nu)$ curve in face-centered cubic $\delta$-Pu (see Appendix) by the substitution of $a=\sqrt[3]{4V}$ and $V=V_0~[1+\nu]$ into Equation (\[cohesion\]), from which we obtain $$\label{cohesionvolume} E_i(\nu)=E_{i,0}+\frac{9K_{i,0}}{2N}\big[1-2[1+\nu-\nu_i]^{-\frac{1}{3}}+[1+\nu-\nu_i]^{-\frac{2}{3}}\big].$$\ The parabolic approximation in Equation (\[cohesionsmall\]) is obtained by making a Taylor series expansion of Equation (\[cohesionvolume\]) about $\nu-\nu_i$. The energy minimum for each curve is given by $E_{i,0}=a_0-\frac{1}{4}\big[\frac{a^2_1}{a_2}\big]$ while the bulk modulus at the minimum is given by $K_{i,0}=\frac{N}{18}\big[\frac{a^2_1}{a_2}\big]$. Differentiation of Equation (\[cohesionvolume\]) yields $$\label{derivative} \frac{\partial E_i(\nu)}{\partial\nu}=\frac{3K_{i,0}}{N}\big[[1+\nu-\nu_i]^{-\frac{4}{3}}-[1+\nu-\nu_i]^{-\frac{5}{3}}\big],$$\ which can be considered as a (negative) partial pressure, while further differentiation yields $$\label{secondderivative} \frac{\partial^2E_i(\nu)}{\partial\nu^2}=\frac{K_{i,0}}{N}\big[5[1+\nu-\nu_i]^{-\frac{8}{3}}-4[1+\nu-\nu_i]^{-\frac{7}{3}}\big].$$\ For $\nu=0$, we obtain the relation $$\label{ambientmoduli} K_i=K_{i,0}\big[5[1-\nu_i]^{-\frac{8}{3}}-4[1-\nu_i]^{-\frac{7}{3}}\big].$$ Theoretical values of $K_i(\nu=0)$ are listed in Table \[table2\]. ------------- -------------------------------------- ----------------------------------- -- -- $n_{5f}$ $K_i$ (GPa), Ref.[@eriksson1999] $K_i$ (GPa), Ref.[@svane2007] \[0.5ex\] 0 - 26.0 1 - 28.0 2 21.8 31.0 3 34.5 24.4 4 36.4 20.6 5 28.5 33.9 6 - 25.6 \[1ex\] ------------- -------------------------------------- ----------------------------------- -- -- : [Theoretical bulk moduli parameters]{}. The theoretical contributions $K_i$ to the bulk modulus at $\nu=0$ for each of the electronic configurations calculated by Eriksson [et al.]{}[@eriksson1999] and Svane [et al.]{}.[@svane2007] Here, $n_{5f}$ refers to the number of $5f$-electrons artificially confined to the atomic core for each configuration.[@eriksson1999; @svane2007] Surprisingly, $K_i$ is not found to depend significantly on the calculation method or on $n_{5f}$. Their average is $\bar{K}_i(\nu=0)\approx$ 28.2 GPa while their standard deviation is $\sigma K_i=$ 5.0 GPa (i.e. $\approx\frac{1}{6}^{\rm th}$ of $\bar{K}_i$). \[table2\] The calculation of $K^\prime=NK_i\frac{\partial^3E_i}{\partial\nu^3}$ from the derivative of Equation (\[secondderivative\]) yields $K^\prime=4$ at $\nu=0$ for a normal metal. However, the absence of a discernible dependence of the Debye temperature on $x$[@harrison2019] suggests that $K^\prime$ is actually closer to zero for the ground state configuration of $\delta$-Pu, thereby further justifying the use of a parabolic approximation for small $\nu$ in the present study. A more accurate experimental determination of $K^\prime$ for the ground state configuration would require low temperature bulk modulus measurements to be performed under pressure. Thermodynamics of multiple configurations ----------------------------------------- For the thermal expansion, one differentiates the free energy once with respect to $\nu$ to obtain $$\label{thermalexpansionderivative} \frac{\partial F_{\rm el}}{\partial\nu}\bigg\|_T=-P=\sum_ip_i(\nu)K_i~[\nu-\nu_i],$$ where we have again made use of the parabolic approximation given by Equations (\[cohesionsmall\]) and (\[effectivepressure\]). Since the total pressure $P\approx0$ during experiments under ambient conditions, $\nu$ and $\nu_i$ can be separated in Equation (\[thermalexpansionderivative\]) to yield $$\label{thermalexpansion} \nu=\frac{\sum_ip_i(\nu)P_i}{\sum_ip_i(\nu)K_i}.$$ Here, the numerator is equivalent to a summation over partial pressures, where the total pressure is ambient pressure. Meanwhile, the denominator is equivalent to the first term of Equation (\[bulkmodulusequation\]), meaning that it is equivalent to the bulk modulus that one obtains on neglecting excitations. If we constrain the bulk moduli to be similar for all relevant configurations (i.e. $K_i=K_0$), then the denominator becomes $K_0$ and Equation (\[thermalexpansion\]) acquires the much simpler form: $$\label{thermalexpansionsimple} \nu\approx\sum_ip_i(\nu)\nu_i.$$ It is instructive to express Equation (\[bulkmodulusequation\]) in a similarly reduced form by setting $K_i=K_0$ for all configurations and defining $k_{\rm el}=\frac{K_{\rm el}}{K_0}-1$, whereupon we obtain $$\label{bulkmodulusequationsimple} k_{\rm el}\approx\frac{K_0}{Nk_{\rm B}T}\Big[-\sum_{i}p_i(\nu)\nu^2_i +\big[\sum_{i}p_i(\nu)\nu_i\big]^2\Big],$$ in which the negative term inside the parenthesis dominates. Grüneisen’s law, in its original form given by $\Gamma=\frac{\alpha_vV_{\rm m}K_T}{C_v}$ is violated because $\alpha_v$, which is the temperature derivative of Equation (\[thermalexpansionsimple\]), is proportional to the sum over $\nu_i$ contributions, whose individual values are both positive or negative, thereby giving rise to sign changes.[@harrison2019] By contrast, the bulk modulus softening depends on a summation over $\nu_i^2$ contributions, causing Equation (\[bulkmodulusequationsimple\]) always to have the same negative sign. The heat capacity, meanwhile, depends only on the energies $E_i$, which are positive (or zero) and indirectly related to $\nu_i$. In the case of the pressure derivative, the last term on the right-hand-side of Equation (\[pressurederivative\]) is given by $$\begin{aligned} \label{correction} \delta=3Nk_{\rm B}T\bigg[\big[\sum_ip_i(\nu)P_i(\nu)\big]\big[\sum_ip_i(\nu)K_i\big]~~~~~~~\nonumber\\ -\big[\sum_ip_i(\nu)P_i(\nu)K_i\big]\bigg],\end{aligned}$$ which vanishes when $K_i$ are the same for all configurations. Additive electronic and phonon bulk moduli ------------------------------------------ Since the isothermal bulk modulus is given by $\frac{\partial^2F}{\partial\nu^2}\big\|_T$ where the free energy $F=F_{\rm el}+F_{\rm ph}$ is the summation of electronic and phonon contributions, this same summation should generally carry over to the bulk modulus $K_T=K_{\rm el}+K_{\rm ph}$, with the electronic and phonon contributions therefore acting in parallel. The precise form of the dependence of $F_{\rm ph}$ leading to $K_{\rm ph}$ is still an area of active debate.[@digilov2019; @ida1969; @born1939; @varshni1970; @rose1984; @anderson1989; @ida1970] According to Ida[@ida1969] $$\label{phononbulkmodulus} K_{\rm ph}(T)=K_0\bigg[\frac{T}{T_{\rm 0}}\frac{1}{Q}-1\bigg],$$ where $T_{\rm 0}$ is the temperature scale associated with lattice vibrations (see Table \[table1\]), $(\Delta l/l)_{\rm ph}$ is the thermal expansion of the lattice attributable to phonons[@harrison2019] (see Fig. \[phononexpansion\]) and $Q$ is the vibrational elongation determined by solving $$\label{Qdetermination} Q=\frac{T}{T_{\rm 0}}{\rm e}^{2b\big[(\frac{\Delta l}{l})_{\rm ph}+Q\big]}.$$ ![The linear thermal expansion of $\delta$-Pu attributable to phonons, determined from a combined fit to thermal expansion and magnetostriction data, and its verification using heat capacity measurements.[@harrison2019][]{data-label="phononexpansion"}](phononexpansion090419.pdf){width=".45\textwidth"} Calculating the adiabatic bulk modulus -------------------------------------- According to basic thermodynamics $$\gamma=\frac{\alpha_v^2V_{\rm m}TK_T}{C_v}+1$$ where $\alpha_v=3\alpha_l$ and $C_v$ values have recently been calculated by Harrison [et al]{}[@harrison2019], from which we obtain the temperature-dependent $\gamma$ in Fig. \[adiabaticcorrection\]. Since $\gamma$ is close to unity, $K_S\approx K_T$. Significantly, for compositions $x\approx$ 2%, the difference between $K_S$ and $K_T$ becomes negligible. ![Calculated ratio $\gamma=K_S/K_T=C_p/C_T$, according to multiple electronic configurations,[@harrison2019] with the concentrations of Ga substituted into $\delta$-Pu indicated in different colors.[]{data-label="adiabaticcorrection"}](adiabaticcorrection092719.pdf){width=".45\textwidth"} Prior models of the bulk modulus softening ------------------------------------------ In a prior model of the bulk modulus based on the invar model, the Debye temperature, and consequently the bulk modulus, was assumed to be the probability-weighted sum of the ground state bulk modulus and an excited invar configuration bulk modulus,[@lawson2002; @lawson2006; @lawson2019] which is equivalent to the first term of Equation (\[bulkmodulusequation\]). In the absence of any other terms contributing to the bulk modulus, Lawson [et al.]{} were able to approximately account for the experimentally observed softening of the bulk modulus by setting $K_i=0$ for the excited invar configuration. Such a value is expected to occur only at a very large positive volume strain of $\nu=\frac{61}{64}+\nu_i$, which we obtain by equating Equation (\[secondderivative\]) to zero. A volume strain this large ($\nu\sim$ 100%) corresponds to an actual volume of order $V\sim$ 50 $\AA^3$, which is significantly larger than the value $V_i\approx$ 20 $\AA^3$ for the excited configuration of $\delta$-Pu obtained in fitting the invar model to the thermal expansion.[@lawson2002; @lawson2006] In another more recent model of the bulk modulus based on the disordered local moment model,[@migliori2016] the volume was assumed to remain approximately constant with the different configurations corresponding to a continuum of states with different degrees of orbital compensation of the local moment.[@soderlind2008] In this model, the softening of the bulk modulus occurs in response to a reduction in the moment with increasing temperature. However, since the degree of softening in this model is predicted to be reduced for samples with a larger negative contribution to the thermal expansion, which generally occurs for samples with lower concentrations $x$ of Ga,[@lawson2002] the predicted Ga-dependent trend is opposite to that found experimentally (plotted in Fig. \[experimentalbulkmod\]b).[@soderlind2010] Offsets in the measured bulk modulus ------------------------------------ Significant vertical displacements between bulk moduli values for different samples of the same composition together with the absence of a trend in $x$ near room temperature in Fig. \[experimentalbulkmod\]a, indicate that extrinsic factors unrelated to $K_{\rm el}$ and $K_{\rm ph}$ contribute random vertical offsets to the experimental data that are of order 1 or 2 GPa. Similar observations have been made in measuring control samples made of aluminum. In Fig. \[experimentalbulkmod\]b, the extrinsic and phonon contributions are removed from the analysis of the $x=$ 2.36%, 3.30% and 4.64% datasets by subtracting the values of the bulk moduli at $T=$ 280 K. Questions regarding pressure-dependent data ------------------------------------------- There are uncertainties in how some of the experimental compositions in Fig. \[pressurebulkmod\] can be compared with the model. For instance, the $x=$ 3.5% Ga-stabilized sample, for which $K^\prime\sim+4$ is positive in contrast to the other $\delta$-Pu samples,[@zhang2019] has a substantially lower volume ($V\approx$ 24.2 Å$^3$) than that ($V\approx$ 24.6 Å$^3$) previously found for samples of nominally the same composition (interpolating between $x=$ 2 and 4%).[@lawson2002] Also, the lowest error bar ($K^\prime=$ $-4$ $\pm$ $2$) is obtained for a heavily Am-stabilized $\delta$-Pu sample (likely due to the the wider range of accessible pressures), yet its thermal expansion and elastic properties remain largely unexplored.[@hecker2004]\ Acknowledgements ================ The work was performed under the Los Alamos National Laboratory LDRD program: project “20180025DR.” Measurements were performed at the National High Magnetic Field Laboratory, which is supported by the National Science Foundation, Florida State and the Department of Energy. N. H. thanks Albert Migliori, Boris Maiorov, Angus Lawson and Paul Tobash for insightful discussions. [99]{} Moore, K. T., van de Laan, G. Nature of the 5$f$ states in actinide metals. [Rev. Mod. Phys.]{} [81]{}, 235-298 (2009). Hecker, S. S., Plutonium – an element at odds with itself. [Los Alamos Science]{} [26]{}, 16-23 (2000); https://lib-www.lanl.gov/26-30.shtml Hecker, S. S., Harbur, D. R., Zocco, T. G. Phase stability and phase transformations in Pu-Ga alloys. [Prog. Mater. Science]{} [49]{}, 429-485 (2004). Smith, J. L., Kmetko, E. A., Magnetism or bonding: a nearly periodic table of the transition elements. [J. Less Common Metals]{} [90]{}, 83-88 (1983). Lashley, J. C. [et al..]{} Experimental electronic heat capacities of $\alpha$- and $\delta$-plutonium: heavy fermion physics in an element. [Phys. Rev. Lett.]{} [91]{}, 205901 (2003). Shim, J. H., Haule, K., Kotliar, G., Fluctuating valence in a correlated solid and the anomalous properties of $\delta$-plutonium. [Nature]{} [446]{}, 513-516 (2007). Zhu, J.-X. [et al..]{} Spectral properties of $\delta$-plutonium: Sensitivity to $5f$ occupancy, [Phys. Rev. B]{} [76]{}, 245118 (2007). N. Lanatà, Y. Yao, C.-Z. Wang, K.-M. Ho, G. Kotliar, Phase diagram and electronic structure of praseodymium and plutonium. [Phys. Rev. X]{} [5]{}, 011008 (2015); link.aps.org/doi/10.1103/PhysRevX.5.011008 A. C. Lawson , J. A. Roberts , B. Martinez, J. W. Richardson, Invar effect in Pu-Ga alloys. [Phil. Mag. B]{} [82]{} 1837 (2002). Lawson, A. C. [et al..]{} Invar model for $\delta$-phase Pu: thermal expansion, elastic and magnetic properties. [Phil. Mag.]{} [86]{}, 2713-2733 (2006). N. Harrison, J.B. Betts, M.R. Wartenbe, F.F. Balakirev, S. Richmond , M. Jaime, P.H. Tobash, Phase stabilization by electronic entropy in plutonium. [Nature Commun.]{} [10]{} 3159 (2019). P. Söderlind, A. Landa, J. E. Klepeis, Y. Suzuki, A. Migliori, Elastic properties of Pu metal and Pu-Ga alloys. [Phys. Rev. B]{} [82]{}, 224110 (2010). M.-H. Nadal, L. Bourgeois, Elastic moduli of Pu and Ga stabilized $\delta$-Pu: Experimental data and phenomenological behavior at high temperatures. [J. Appl. Phys.]{} [108]{}, 033512 (2010). Y. Suzuki, V. R. Fanelli, J. B. Betts, F. J. Freibert, C. H. Mielke, J. N. Mitchell, M. Ramos, T. A. Saleh, A. Migliori, Temperature dependence of elastic moduli of polycrystalline $\beta$ plutonium. [Phys. Rev. B]{} [84]{}, 064105 (2011). F. J. Freibert, J. N. Mitchell, T. A. Saleh, A. Migliori, Instability and anharmonicity in plutonium thermo-physical properties. [XIIth Fundamentals of Plutonium Workshop]{}, (Cambridge, United Kingdom, 2012); https://permalink.lanl.gov/object/tr?what=info:lanl-repo/lareport/LA-UR-12-22830; https://permalink.lanl.gov/object/tr?what=info:lanl-repo/lareport/LA-UR-12-21466. Migliori, A., Soderlind, P., Landa, A., Freibert, F. J., Maiorov, B., Ramshaw, B. J., Betts, J. B., Origin of the multiple configurations that drive the response of $\delta$-plutonium’s elastic moduli to temperature. [Proc. Nat. Acad. Sci. USA]{} [113]{}, 11158-11161 (2016). Born, M., Thermodynamics of crystals and melting. [J. Chem. Phys.]{} [7]{}, 591-603 (1939). Ida, Y., Theory of melting based on lattice instability. [Phys. Rev.]{} [187]{}, 951-958 (1969). Ida, Y., Anharmonic effect on heat capacity of solids up to the critical temperature of lattice instability. [Phys. Rev. B]{} [1]{}, 2488-2496 (1970). Varshni, Y. P., Temperature dependence of the elastic constants. [Phys. Rev. B]{} [2]{}, 3952-3958 (1970). Rose, J. H., Smith, J. R., Guinea, F., Ferrante, J., Universal features of the equations of state of metals. [29]{}, 2963-2969 (1984). Anderson, O. L., Isaak, D. G., Yamamoto, S., Anharmonicity and the equation of state for gold. [J. Applied Phys.]{} [65]{}, 1534-1543 (1989). Digilov, R. M., Abramovich, H., Temperature variation of the isothermal bulk modulus in solids: Thermo-elastic instability and melting. [J. Applied Phys.]{} [125]{}, 065104 (2019). Lawson, A. C., Thermodynamics of the bulk modulus of delta phase plutonium alloys. [Phil. Mag.]{} [99]{}, 1481-1498 (2019). Eriksson, O., Becker, J. N., Balatsky, A. V., Wills, J. M., Novel electronic configuration in $\delta$-Pu. [J. Alloy. & Comp.]{} [287]{}, 1 (1999). Wills, J. W. [et al..]{} A novel electronic configuration of the $5f$ states in $\delta$-plutonium as revealed by the photo-electron spectra. [Journal of Electron Spectroscopy and Related Phenomena]{} [135]{}, 163 (2004). Svane, A., Petit, L., Szotek, Z., Temmerman, W. M., Self-interaction-corrected local spin density theory of $5f$-electron localization in actinides. [Phys. Rev. B]{} [76]{}, 115116 (2007). Windorff, C. J., Chen, G. P., Cross, J. N., Evans, W. J., Furche, F., Gaunt, A. J., Janicke, M. T., Kozimor, S. A., Scott, B. L., Identification of the formal $+2$ oxidation state of plutonium: Synthesis and characterization of {Pu$^{\rm II}$\[C$_5$H$_3$(SiMe$_3$)$_2$\]$_3$}$^-$. [J. Am. Chem. Soc.]{}. [139]{}, 3970-3973 (2017). Booth, C. H., Jiang, Y., Wang, D. L., Mitchell, J. N., Tobash, P. H., Bauer, E. D., Wall, M. A., Allen, P. G., Sokaras, D., Nordlund, D., Weng, T. C., Torrez, M. A., Sarrao, J. L. Multiconfigurational nature of $5f$ orbitals in uranium and plutonium intermetallics. [Proc. Natl. Acad. Sci. U.S.A.]{} [109]{}, 10205 (2012). Janoschek, M. [et al..]{} The valence-fluctuating ground state of plutonium. [Sci. Adv.]{} [1]{}, e1500188 (2015); DOI: 10.1126/sciadv.1500188 Savrasov, S. Y., Kotliar, G., Abrahams, E., Correlated electrons in $\delta$-plutonium within a dynamical mean-field picture. [Nature]{} [410]{}, 793-795 (2001). Lawrence, J. M., Riseborough, P. S., Parks, R. D., Valence fluctuation phenomena. [Rep. Prog. Phys.]{} [44]{}, 1-84, (1981). Wohlleben, D., [Physics and chemistry of electrons and ions in condensed matter]{} ed. Acrivos, J. V., Mott, N. F., Yoffe, A. D. (Reidel, Dordrecht 1984) pp. 85-107. Faure, P., Klosek, C, Genestier, N., Baclet, S., Heathman, P., Normile, R., Haire, Structural investigation of $\delta$-stabilized plutonium alloys under pressure. [Materials Research Society Proceedings (eds. Sarrao, A.J. Schwartz, P. Burns, R.G. Haire, H. Nitsche, M. Antonio)]{}, vol. [893]{} , Materials Research Society, (Pittsburgh, PA, 2005). pp. 223. Klosek, V., Griveau, J. C., Faure1 P., Genestier, C., Baclet, N., Wastin, F. High pressure study of Pu$_{0.92}$Am$_{0.08}$ binary alloy. [J. Phys.: Condes. Matter]{} [20]{}, 275217 (2008). Faure, P., Genestier, C., X-ray diffraction study $\delta$-stabilized plutonium alloys under pressure. [J., Nucl. Mater.]{} [397]{}, 74 (2010). Zhang, J., Freibert, F. J., Clausen, B., Smith, A. I., Vogel, S. C., Brown, D. W., Equation of state and strain-induced stabilization $\delta$-phase stabilized plutonium alloys. [J., Nucl. Mater.]{} [524]{}, 54 (2019). Mitra, G. B., Mitra, S. K., Grüneisen’s rule and the thermal expansion of metals. [Nature]{} [179]{}, 1295 (1957). Mitra, G. B., Giri, A. K., Grüneisen’s law and extrapolated values of lattice constants of single-phase copper-aluminium alloys at absolute zero. [J. Phys. D: Appl. Phys.]{} [19]{}, 1065 (1986). Ashcroft, N. W., Mermin, N. D., [Solid state physics]{} (Saunders College Publishing, Orlando 1976). Söderlind, P., Quantifying the importance of orbital over spin correlations in $\delta$-Pu within density-functional theory. [Phys. Rev. B]{} [77]{}, 085101 (2008).	{ "pile_set_name": "ArXiv" }
--- abstract: 'The evolution of galaxies at Cosmic Noon (1$<$[z]{}$<$3) passed through a dust-obscured phase, during which most stars formed and black holes in galactic nuclei started to shine, which cannot be seen in the optical and UV, but it needs rest frame mid-to-far IR spectroscopy to be unveiled. At these frequencies, dust extinction is minimal and a variety of atomic and molecular transitions, tracing most astrophysical domains, occur. The future IR space telescope mission, SPICA, currently under evaluation for the 5th Medium Size ESA Cosmic Vision Mission, fully redesigned with its 2.5-m mirror cooled down to $T < 8\, \rm{K}$ will perform such observations. SPICA will provide for the first time a 3-dimensional spectroscopic view of the hidden side of star formation and black hole accretion in all environments, from voids to cluster cores over 90% of cosmic time. Here we outline what SPICA will do in galaxy evolution studies.' --- Introduction ============ Most of the activity in galaxy evolution, the formation of stars and supermassive black holes at the center of galaxies, took place more than six billion years ago, with a sharp drop to the present epoch [e.g., @ma_di14 Fig.\[fig\_madau\]]. Since most of the energy emitted by stars and accreting SMBHs is absorbed and re-emitted by dust, understanding the physics of galaxy evolution requires infrared (IR) observations of large, unbiased galaxy samples spanning a range in luminosity, redshift, environment, and nuclear activity. From [Spitzer]{} and [Herschel]{} photometric surveys the global Star Formation Rate (SFR) and Black Hole Accretion Rate (BHAR) density functions have been [estimated]{} through measurements of the bolometric luminosities of galaxies [@leflo05; @gru13; @del14]. However, such integrated measurements could not separate the contribution due to star formation from that due to BH accretion [see, e.g., @mul11]. This crucial separation has been attempted so far through modelling of the spectral energy distributions and relied on model-dependent assumptions and local templates, with large uncertainty and degeneracy. On the other hand, determinations from UV [e.g. @bou07] and optical spectroscopy [e.g., from the [Sloan]{} Digital Sky Survey, @eis11] track only marginally ($\sim$10%) the total integrated light (Fig.\[fig\_madau\]). X-ray analyses of the BHAR, in turn, are based on large extrapolations and possibly miss a large fraction of obscured objects. Furthermore, the SFR density at $z$ $>$ 2-3 is very uncertain, since it is derived from UV surveys, highly affected by dust extinction. As opposite, through IR emission lines, the contributions from stars and BH accretion can be separated. [SPICA]{} spectroscopy will allow us to directly measure redshifts, SFRs, BHARs, metallicities and dynamical properties of gas and dust in galaxies at lookback times down to about 12 Gyrs. [SPICA]{} spectroscopic observations will allow us for the first time to redraw the SFR rate and BHAR functions (Fig.\[fig\_madau\]) in terms of measurements directly linked to the physical properties of the galaxies. ![Estimated star-formation rate densities from the far-ultraviolet (blue points) and far-IR (red points) photometric surveys [figure adapted from @ma_di14]. The estimated BHAR density, scaled up by a factor of $3300$, is shown for comparison (in green shading from X-rays and light blue from the IR).[]{data-label="fig_madau"}](madau_plot.png){width="13cm"} The mid- to far-IR spectral range hosts a suite of atomic and ionic transitions, covering a wide range of excitation, density and metallicity, directly tracing the physical conditions in galaxies. Ionic fine structure lines (e.g. \[NeII\], \[SIII\], \[OIII\]) probe HII regions around hot young stars, providing a measure of the SFR and the gas density. Lines from highly ionized species (e.g. \[OIV\], \[NeV\]) trace the presence of AGN and can measure the BHAR [@sm92]. Through line ratio diagrams, like the [new IR BPT diagram]{} [@fer16 Fig.\[fig\_IR\_BPT\]], IR spectroscopy can separate the galaxies in terms of both the source of ionization –either young stars or AGN excitation– and the gas metallicity, during the dust-obscured era of galaxy evolution ($0.5 < z < 4$). ![ Observed line ratios of \[NeIII\]15.6$\mu$m/\[NeII\]12.8$\mu$m vs. \[OIV\]26$\mu$m/\[OIII\]88$\mu$m for AGN, LINER, starburst and dwarf galaxies in the local Universe. These ratios are powerful diagnostic tools: the star formation in galaxies increases from right to left, while the gas metallicity increases from the top to the bottom. For this reason we named this diagram the [new BPT diagram]{} in the IR [@fer16]. The active galaxies symbols have been color-coded from blue to red, according to the value of the ratio of the \[NeV14.3\]$\mu$m line to the FIR continuum as measured with [Herschel-PACS]{} spectra at $\sim$160$\mu$m, corresponding the AGN dominance, taken from @fer16. The solid and dashed lines show standard photoionization models for AGN, LINERS galaxies and dwarf [low-metallicity]{} galaxies. []{data-label="fig_IR_BPT"}](IR_btp.png){width="13cm"} SPICA observations of galaxy evolution ====================================== For a complete description of the SPace Infrared telescope for Cosmology and Astrophysics ([SPICA]{}), we refer to @roe18. We concentrate here on the work that [only]{} an IR observatory such as SPICA will be able to do to unveil galaxy evolution. 1. SPICA will provide for the first time a 3-dimensional spectroscopic view of the hidden side of star formation and black hole accretion in all environments, from voids to cluster cores over 90% of cosmic time. This will lead to a complete census – as a function of cosmic time – of: $\bullet$ the star formation in galaxies from low-mass dust-poor star-forming galaxies to high-mass heavily dust-obscured starbursts, setting a fundamental new benchmark for cosmological models, allowing unambiguous tests of the physics driving star, galaxy and large-scale structure formation. $\bullet$ the black hole accretion in the nuclei of galaxies, including also heavily obscured Compton thick nuclei. From these observations, we will be able to derive both the average cosmic star-formation and accretion histories of the Universe as well as the physical conditions of each individual galaxy through their spectra in the large cosmic time interval since about 12 billion years [@spi17]. SPICA will allow us to understand the physical origin of the dramatic change in the efficiency of star formation and accretion around cosmic noon, one of the major challenges for theoretical models in present-day cosmology. 2. SPICA will measure if and how energetic (radiative and mechanical) feedback from starburst and AGN influences galaxy evolution. SPICA will determine the role of active galactic nuclei in quenching star formation in galaxies and address the origin of the low stellar to dark matter halo mass ratio in massive galaxies [@gon17]. 3. SPICA will determine how galaxies build up their metals and dust during the last ten Gyrs of cosmic time and how are these metals recycled in the ISM and injected into the CGM [@fer17]. IR line based tracers will allow SPICA to peer through the heavily opaque medium of main-sequence galaxies at the cosmic noon and probe their true chemical ages. Galaxies at z $\sim$ 2-3 show extremely low metallicities in optically-based traces, down to 0.8 dex below solar [@ono16], which is in conflict with the large amounts of dust inferred from IR photometry (Fig. 1), suggesting much higher metallicities. SPICA will also probe the first production of dust and metals, assess the presence of hot dust in the most distant galaxies (z$>$6) and determine the physical properties of infrared bright galaxies and quasar hosts at the epoch of re-ionization. SPICA observational strategy to unveil galaxy evolution {#strategy} ------------------------------------------------------- We briefly show here some of the results of the scientific work that has been done in preparation of the galaxy evolution studies that will performed with the SPICA mission. In the context of the development of a coherent observing strategy, we have considered a step by step observational sequence, aimed at optimising the measure of basic astrophysical quantities in galaxies through their evolution in the last 12 Gyrs. This sequence can be summarised as follows: [(i)]{} Deep spectrophotometric and photometric surveys in the mid-IR, using the SPICA Mid-IR Instrument (SMI) of large enough fields (of order 1-10 deg$^2$) down to very faint flux limits (i.e. 3-12$\mu$Jy in the continuum at 34$\mu$m and 2-5$\times$10$^{-20}$ W/m$^2$ in the spectral lines). These observations will be complemented with the B-BOP camera at 70$\mu$m (at a 30$\mu$Jy depth). [This survey will provide a 3-dimensional view of galaxy evolution.]{} [(ii)]{} Identification of a sample of galaxies based on the outcome of the above surveys, in terms of redshift, mid-IR luminosity, classification (i.e. Starburst-dominated or AGN-dominated), characterisation of the Stellar Masses, Dust Masses and (where possible) Bolometric Luminosities using available multifrequency data. [(iii)]{} Deep follow-up observations with the grating spectrometers (both SAFARI and SMI at spectral medium resolution) of above defined sample. [This will provide a complete spectral atlas of galaxies as a function of redshift, luminosity and stellar mass.]{} From the spectral atlas of galaxies we will determine their physical properties as a function of cosmic time: the Star Formation Rate (SFR), the Black Hole Accretion Rate (BHAR), the metal abundances, the outflow/infall occurrence, etc. Predictions of the SMI ultra-deep survey ---------------------------------------- Following the work already presented in @kan17 and @gru17, we are planning to perform an ultra-deep survey of 1 deg$^2$ with the SMI imaging low-resolution spectrometer [@kan18] down to the photometric limit of 3$\mu$Jy in the continuum at 34$\mu$m in a typical observing time of about $\sim$ 600 hours. We show in Fig. \[fig\_SMI\] the predictions of how the SPICA SMI ultra-deep survey of 1 deg$^2$ will fill the luminosity-redshift plane for Star Forming Galaxies (left diagram) and for AGN (right diagram). The Star Forming galaxies are detected through the Polycyclic Aromatic Hydrocarbon (PAH) features at 6.2, 7.7, 8.6, 11.25 amd 17.0 $\mu$m, while the AGN can be detected through a set of high-ionization fine-structure lines, \[NeVI\]7.65$\mu$m, \[SIV\]10.5$\mu$m, \[NeV\]14.3 and 24.3$\mu$m, \[NeIII\]15.5$\mu$m, and \[OIV\]15.9$\mu$m. We have used the far-IR luminosity functions as derived from [Herschel]{} photometric observations [@gru13] and the IR linea and PAH features calibrations to the total IR luminosities from observations in the local Universe. The details of these simulation can be found in Mordini et al (2020, in preparation). ![[Left:]{} Redshift-Luminosity diagram with the simulation of the SMI ultra-deep survey of 1 deg$^2$ showing the Star Forming Galaxies detections in the various PAH spectral features. The green solid line shows the knee of the Luminosity Functions, as a function of redshift. [Right:]{} Same, but simulated with the AGN detections in high-ionization fine-structure lines.[]{data-label="fig_SMI"}](SMI3_SF.png "fig:"){width="7cm"} ![[Left:]{} Redshift-Luminosity diagram with the simulation of the SMI ultra-deep survey of 1 deg$^2$ showing the Star Forming Galaxies detections in the various PAH spectral features. The green solid line shows the knee of the Luminosity Functions, as a function of redshift. [Right:]{} Same, but simulated with the AGN detections in high-ionization fine-structure lines.[]{data-label="fig_SMI"}](SMI3_AGN.png "fig:"){width="6.7cm"} Predictions of SAFARI follow-up pointed spectroscopic observations ------------------------------------------------------------------ Following step [ii)]{} outlined in Section \[strategy\], i.e. the definition of a suitable sample of galaxies, of order 1,000 objects, we have made predictions of the observability of such sample with the SAFARI spectrometer [@roe18]. We show these predictions in Fig. \[fig\_SAFARI\], where, on the left, the luminosity-redshift plane for Star Forming Galaxies through the detection of the Polycyclic Aromatic Hydrocarbon (PAH) feature at 17$\mu$m is shown. Analogously, on the left diagram of Fig. \[fig\_SAFARI\], the luminosity-redshift plane for AGN is shown, where the detections by SPICA-SAFARI are made in the \[OIV\]25.9$\mu$m line. We can see from these predictions that SPICA will be able to measure the SFR and the BHAR of galaxies far below the knee of the luminosity functions at each redshift, in particular beyond z$\sim$4 for Star Forming galaxies and up to z$\sim$3.5 for AGN. ![[Left:]{} Redshift-Luminosity diagram with the simulation of the SAFARI follow-up pointed observations of Star Forming Galaxies with the PAH 17$\mu$m spectral features. The green solid line shows the knee of the Luminosity Functions, as a function of redshift. [Right:]{} Same, but simulated with the AGN detections of the \[OIV\]26$\mu$m fine-structure line.[]{data-label="fig_SAFARI"}](SAFARI_SF.png "fig:"){width="7cm"} ![[Left:]{} Redshift-Luminosity diagram with the simulation of the SAFARI follow-up pointed observations of Star Forming Galaxies with the PAH 17$\mu$m spectral features. The green solid line shows the knee of the Luminosity Functions, as a function of redshift. [Right:]{} Same, but simulated with the AGN detections of the \[OIV\]26$\mu$m fine-structure line.[]{data-label="fig_SAFARI"}](SAFARI_AGN.png "fig:"){width="6.7cm"} Acknowledgements {#acknowledgements .unnumbered} ================ This paper is dedicated to the memory of Bruce Swinyard, who initiated the SPICA project in Europe, but unfortunately died on 22 May 2015 at the age of 52. He was ISO-LWS calibration scientist, Herschel-SPIRE instrument scientist, first European PI of SPICA and first design lead of SAFARI. We acknowledge the whole SPICA Collaboration Team, as without its multi-year efforts and work this paper could not have been possible. We also thank the SPICA Science Study Team appointed by ESA and the SPICA Galaxy Evolution Working Group. LS and JAFO acknowledge financial support by the Agenzia Spaziale Italiana (ASI) under the research contract 2018-31-HH.0. [2]{} André, Ph. et al. 2019, PASA, 36, e029 Bouwens, R. R. et al. 2007, ApJ, 670, 928 Delvecchio, I. et al. 2014, MNRAS, 439, 2736 Eisenstein, D.J. et al. 2011, AJ, 142, 72 Fernández-Ontiveros, J.A. et al. 2016, ApJS, 226, 19 Fernández-Ontiveros, J.A. et al. 2017, PASA, 34, e053 González-Alfonso, E. et al.2017, PASA, 34, e054 Gruppioni, C. et al. 2013, MNRAS, 432, 23 Gruppioni, C. et al. 2017, PASA, 34, e055 Kaneda, H. et al. 2017, PASA, 34, e059 Kaneda, H. et al. 2018, Proceedings of the SPIE, Volume 10698, id. 106980C Le Floc’h, E. et al. 2005, ApJ, 632, 169 Madau, P. & Dickinson, M. 2014, ARA&A, 52, 415 Mullaney, J. R. et al. 2011, MNRAS, 414, 1082 Onodera, M. et al. 2016, ApJ, 822, 42 Roelfsema, P. R. et al. 2018, PASA, 35, e030 Spinoglio, L. & Malkan, M.A. 1992, ApJ, 399, 504 Spinoglio, L. et al. 2017, PASA, 34, e057	{ "pile_set_name": "ArXiv" }
--- abstract: 'By the assumption that the thermodynamics second law is valid, we study the possibility of $\omega=-1$ crossing in an interacting holographic dark energy model. Depending on the choice of the horizon and the interaction, the transition from quintessence to phantom regime and subsequently from phantom to quintessence phase may be possible. The second transition avoids the big rip singularity. We compute the dark energy density at transition time and show that by choosing appropriate parameters we can alleviate the coincidence problem.' author: - \| H. Mohseni Sadjadi$^1$[^1] and M. Honardoost$^2$\ $^1$ [School of Physics, University College of Science, University of Tehran,]{}\ [North Karegar Ave. Tehran, Iran.]{}\ $^2$ [Department of Physics, Shahid Beheshti University,]{}\ [Evin, Tehran 19839, Iran.]{} title: 'Thermodynamics second law and $\omega=-1$ crossing(s) in interacting holographic dark energy model ' --- Introduction ============ Recent observations suggest that the universe is undergoing an accelerated expansion [@acc]. This acceleration may be explained by the assumption that $70\%$ of the universe is filled by a perfect fluid with negative pressure, dubbed dark energy. Some present data seem to favor an evolving dark energy, corresponding to an equation of state (EOS) parameter less than $\omega=-1$ at present epoch from $\omega>-1$ in the near past [@Bo]. Many candidates for dark energy has been proposed such as the cosmological constant [@dark]: A constant quantum vacuum energy density which fills the space homogeneously, corresponding to a fluid with a constant EOS parameter $\omega=-1$; dynamical fields with a suitably chosen potential to make the vacuum energy vary with time [@field], and so on. Recently, using holographic principle, a new candidate for dark energy which is independent of any specific field has been suggested [@Li],[@hol]. Based on the holographic principle (which relates the number of degrees of freedom of a physical system to the area of its boundary), in order to allow the formation of black holes in local quantum field theory, Cohen et al [@cohen] proposed a relationship between UV and IR cutoff. This yields an upper bound on the zero-point energy density, which by a suitable choice of the infrared cutoff, can be viewed as the holographic dark energy density. In [@Li], three candidates for the infrared cutoff was proposed: the Hubble radius, the particle horizon and the future event horizon. There was shown that among these options only the future event horizon may be identified with the desired infrared cutoff. To study the coincidence problem and also to have other choices for the infrared cutoff, e.g. the Hubble radius, interaction between dark matter and dark energy [@inter] may be considered in the holographic dark energy model. As we have mentioned, based on astrophysical data, we may take into account the possibility of $\omega=-1$ (phantom divide line) crossing. Therefore dark energy models which can describe phantom divide line crossing, has been also studied vastly in the literature [@divide]. The phantom like behavior of interacting holographic dark energy was studied in [@abd], where it was claimed that by selecting appropriate interaction parameters the transition from the dark energy EOS parameter $\omega_D>-1$ to $\omega_D<-1$ is possible. Despite this, in [@antabd] it was shown that the dark energy effective EOS parameter cannot cross $\omega_D^{eff.}=-1$. In this paper we consider interacting holographic dark energy model and study the ability of the model to describe the transition from quintessence to phantom regime and vice versa. After preliminaries in section two, where we introduce the interacting holographic dark energy model and some of its general properties used in the subsequent sections, in section three we study the possibility of crossing $\omega=-1$. In section four we derive necessary conditions for existence of two transitions in our model. The first transition is from quintessence to phantom phase and the second the transition from phantom to quintessence regime. The importance of the second transition lies on the fact that it avoids the big rip singularity. We discuss also the behavior of Hubble parameter and dark energy density at transitions times. We use $\hbar=G=c=k_B=1$ units throughout the paper. preliminaries ============= We consider a spatially flat Friedmann–Lemaitre–Robertson–Walker (FRW) universe), with scale factor $a(t)$ $$\label{1} ds^2=dt^2-a^2(t)(dx^2+dy^2+dz^2).$$ We assume that this universe is filled with dark energy and pressureless dark matter fluids satisfying the following equations of state $$\begin{aligned} \label{2} \dot{\rho_D}+3H(1+\omega_D)\rho_D&=&-Q \nonumber \\ \dot{\rho_m}+3H\rho_m&=&Q.\end{aligned}$$ $H=\dot{a}/a$ is the Hubble parameter, $\rho_D$ is dark energy density and $\rho_m$ is the density of cold dark matter. “dot” denotes derivative with respect to the comoving time. $\omega_D$ is dark energy EOS parameter. $Q$ denotes the interaction of dark matter with dark energy. In this paper $Q$ is assumed to be $$\label{3} Q=(\lambda_m\rho_m+\lambda_D\rho_D)H,$$ where $\lambda_m$ and $\lambda_D$ are two real constants. Different choices such as $\lambda_m=0$, $\lambda_D=0$ and $\lambda_D=\lambda_m$ has been adopted in literature [@inter], [@abd]. The Hubble parameter satisfies $$\begin{aligned} \label{4} H^2&=&\frac{8\pi}{3}(\rho_m+\rho_D) \nonumber \\ &=&\frac{8\pi}{3}\rho,\end{aligned}$$ and $$\begin{aligned} \label{5} \dot{H}&=&-4\pi\left((1+\omega_D)\rho_D+\rho_m\right) \nonumber \\ &=&-4\pi(1+\omega)\rho.\end{aligned}$$ $\rho=\rho_m+\rho_D$ is the total energy density satisfying $$\label{6} \dot{\rho}+3H(1+\omega)\rho=0,$$ and $\omega$ is the parameter of the EOS of the universe. In terms of $\Omega_D=\rho_D/\rho$, we can write $$\label{7} \omega_D\Omega_D=\omega.$$ $\rho_D$ can be viewed as a holographic energy density $$\label{8} \rho_D=\frac{3}{8\pi}\frac{c^2}{L^2}.$$ The length scale $L$ is an infrared cutoff and $c>0$ is a positive numerical constant. We assume $$\label{9} L=\beta R_{FH}+\alpha R_{PH},$$ where $R_{FH}$ and $R_{PH}$ are the future and particle event horizons $$\begin{aligned} \label{10} R_{FH}&=&a\int_t^\infty\frac{dt}{a}\nonumber \\ R_{PH}&=&a\int_0^t\frac{dt}{a},\end{aligned}$$ and $\alpha$ and $\beta$ are two positive numerical constants. By taking $\beta=0$ and $\alpha=1$, we arrive at the holographic cosmology horizon adopted in [@sus]. For $\alpha=0$ and $\beta=1$, the infrared cutoff becomes the future event horizon [@Li]. Generally one can assume that $L$ is a function of $R_{FH}$ and $R_{PH}$ [@odin]. The time derivative of $L$ is obtained as $$\label{11} \dot{L}=HL+\alpha-\beta.$$ Using (\[4\]) and (\[8\]), $HL$ in the above equation can be written as $$\label{12} HL=c\Omega_D^{-\frac{1}{2}}.$$ By relating the entropy of the universe to the infrared cutoff $L$ via $$\label{13} S=\pi L^2,$$ the second law of thermodynamics results in $\dot{L}>0$, leading to $\dot{\rho_D}<0$. In a phantom dominated universe (identified by $\dot{H}>0$), using (\[4\]) and (\[8\]) we get $$\label{14} \dot{\Omega_D}=\frac{8\pi}{3H^4}\left(\dot{\rho_D}H^2-2H\dot{H}\rho_D\right)<0.$$ For $\dot{\Omega_D}>0$ we must have $(LH\dot{)}<0$ which leads to $\ddot{L}<0$. But if one requires $\dot{L}\geq 0$, then either $\lim_{t\to \infty}\dot{L}=0$ or $\ddot{L}$ becomes positive after a finite time. In terms of the Hubble parameter, $\omega$ is $$\label{15} \omega=-1-\frac{2\dot{H}}{3H^2}.$$ By substituting $\dot{H}=-H^2+(\beta-\alpha)H/L+(HL\dot{)}/L$ (which can be verified using (\[11\])), (\[15\]) becomes $$\label{16} \omega=-\frac{1}{3}-\frac{2}{3}(\beta-\alpha)\frac{\sqrt{\Omega_D}}{c}+\frac{\dot{\Omega_D}} {3H\Omega_D}.$$ Using (\[2\]) one can show that $$\label{17} \dot{r}=3rH\left[\omega_D +\frac{1}{3}\left(\frac{r+1}{r}\right)\left(\lambda_D+r\lambda_m\right)\right],$$ where we have defined $r=\rho_m/\rho_D$. Therefore, by considering $r=\Omega_D^{-1}-1$, we obtain $$\label{18} \omega_D=-\frac{1}{3H}\frac{\dot{\Omega_D}}{\Omega_D(1-\Omega_D)}-\frac{\lambda_D}{3(1-\Omega_D)} -\frac{\lambda_m}{3\Omega_D},$$ and consequently $$\label{19} \omega=-\frac{1}{3H}\frac{\dot{\Omega_D}}{(1-\Omega_D)}-\frac{\lambda_D\Omega_D}{3(1-\Omega_D)} -\frac{\lambda_m}{3}.$$ From (\[16\]) and (\[19\]) we can express $\omega$ and $\dot{\Omega_D}$ in terms of $\Omega_D$: $$\label{20} \omega=-\frac{1+\lambda_D-\lambda_m}{3}\Omega_D-\frac{2(\beta-\alpha)}{3c}\Omega_D^{\frac{3}{2}} -\frac{\lambda_m}{3},$$ $$\label{21} \frac{\dot{\Omega_D}}{H}=(\lambda_m-\lambda_D-1)\Omega_D^2+(1-\lambda_m)\Omega_D+ \frac{2(\beta-\alpha)}{c}\Omega_D^{\frac{3}{2}} (1-\Omega_D).$$ For an accelerated universe, $\ddot{a(t)}>0$ (which results in $\omega<-1/3$), we have $$\label{22} (1-\lambda_m)+(\lambda_m-\lambda_D-1)\Omega_D-\frac{2(\beta-\alpha)} {c}\Omega_D^{\frac{3}{2}}<0.$$ In the absence of interaction, acceleration is not possible for $\beta<\alpha$, as was pointed out by [@Li] for the case $\beta=0,\,\alpha=1$. This was the motivation of [@Li] to take the future event horizon (instead of the particle horizon) as the infrared cutoff. However the above calculation reveals that in the presence of interaction, inflation may be possible even when the particle horizon is taken as the infrared cutoff. Combining (\[21\]) and (\[22\]), yields $$\label{23} \dot{\Omega_D}<\frac{2(\beta-\alpha)}{c}H\Omega_D^{\frac{3}{2}}.$$ Hence like the phantom regime, for $\beta<\alpha$ and in an accelerating universe, $\Omega_D$ is decreasing. Applying the assumption $\dot{L}>0$, gives an upper bound for $\Omega_D$ which depends only on $H$ $$\label{24} \dot{\Omega_D}<2H\Omega_D<2H.$$ Crossing $\omega=-1$ in interacting holographic dark energy model ================================================================= At $\omega=-1$, $u=\Omega_D^{\frac{1}{2}}$ satisfies the cubic equation $$\label{25} \frac{2(\beta-\alpha)}{c}u^3+(1+\lambda_D-\lambda_m)u^2 +\lambda_m-3=0.$$ If $\omega=-1$ is allowed, the above equation must have, at least, one positive root which is less than one. Based on Descartes rule, we know that the above equation, at most, has two real positive roots. So, if $\omega=-1$ is crossed, two transitions may be possible, one from quintessence to phantom and the other from phantom to quintessence phase (by quintessence phase (regime) we mean $\omega<-1/3$ and $\dot{H}<0$). From (\[15\]) it is clear that at $\omega=-1$ we have $\dot{H}=0$. If a transition from quintessence to phantom phase occurs at time $t_1$, we must have $\dot{H}(t_1)=0$, and $\dot{H}(t<t_1)<0$ and $\dot{H}(t>t_1)>0$, therefore $H(t_1)$ must be a local minimum of $H$. In the same way, $H$ must have a local maximum at $t_2$, where the transition from phantom to quintessence era occurs. In the neighborhood of $t_1$, $\omega$ is a decreasing function while in the neighborhood of $t_2$, $\omega$ is an increasing function of time. So in order to see that if $\omega=-1$ crossing is permissible, we must also consider the behavior of the Hubble parameter near the roots of (\[25\]). In the following we assume that $H>0$ and $L>0$. We also assume that the Hubble parameter is a differentiable function of time [@sad]. Therefore following (\[11\]) and (\[12\]), $L$ and $u(=\Omega_D^{\frac{1}{2}})$ are also differentiable. Let us consider the Taylor expansion of $H$ at $t=t_i$, where $t_i$ is defined by $\dot{H}(t_i)=0$ (or $\omega(t_i)=-1$), $$\label{26} H=h_0+h_1\tau^a+O(\tau^{a+1}),\, a\geq 2,$$ where $\tau=t-t_i$, $h_1=\frac{1}{a!}\frac{d^aH}{dt^a}(t_i)$, and $a$ is the order of the first nonzero derivative of $H$ at $t=t_i$. If $a$ is an even integer and $h_1>0(<0)$, then $H$ has a minimum (maximum) at $t_i$ and the transition occurs at $t_i$. Using (\[15\]) we obtain $$\label{27} \omega=-1-\frac{2ah_1}{3h_0^2}\tau^{a-1}+O(\tau^a).$$ We consider the following expansion for $u$ at $t=t_i$, $$\label{28} u=u_0+u_1\tau^b+u_2\tau^{b+1}+O(\tau^{b+2}), \,\, u_1\neq 0, \,\, b\geq1$$ where $b$ is the order of the first nonzero derivative of $u$ at $t_i$. If the solution (\[26\]) is permissible, by inserting (\[27\]) and (\[28\]) in (\[20\]) and by comparing the powers of $\tau$ in both sides of (\[20\]), we obtain the following results: $$\label{29} (i)-\frac{1+\lambda_D-\lambda_m}{3}u_0^2-\frac{2(\beta-\alpha)}{3c} u_0^3=-1+\frac{\lambda_m}{3}.$$ The roots of this equation specify $u$ at transition time(s). In order that the transition occurs the above equation must have at least one real root in the interval $(0,1)$. (ii)For $$\label{30} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_0\neq 0,$$ we obtain $$\label{31} -\frac{ah_1}{3h_0^2}=\left(\frac{1+\lambda_D-\lambda_m}{-3}- \frac{\beta-\alpha}{c}u_0\right)u_0u_1, \,\, b=a-1.$$ In the case $$\label{32} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_0= 0,$$ and for $\beta\neq \alpha$, we find $$\label{33} \frac{2ah_1}{3h_0^2}=\frac{(\beta-\alpha)u_0u_1^2}{c},\,\, a=2b+1.$$ In this case $a$ is an odd number and transition does not occur. If $\beta=\alpha$, $1+\lambda_D-\lambda_m= 0$, and $\omega$ in (\[20\]), becomes a constant: $\omega(t)=-1$, and no transition occurs. To determine $a$ and the sign of $h_1$, (\[21\]) may be used : $$\label{34} 2\dot{u}=H\left[\frac{2(\alpha-\beta)}{c}u^4+(\lambda_m-\lambda_D-1)u^3+\frac{2(\beta-\alpha)}{c}u^2 +(1-\lambda_m)u\right].$$ By inserting (\[28\]) and (\[26\]) into the above equation, we see that the left hand side begins with $\tau^{b-1}$, while if the right hand side does not begin with $\tau^0$, it will begin by $\tau^{(\gamma>b)}$, which is inconsistent with the left hand side. Thereby $b=1$. Hence based on (\[31\]) and our previous discussion after (\[26\]), if the transition occurs, we must have $a=2$. Note that for the case (\[32\]), we obtain $a=3$ which may be corresponding to the inflection point of $H$. For $b=1$, by equalizing $\tau^0$’s coefficients in both sides of (\[34\]), we get $$\label{35} 2u_1=h_0\left[\frac{2(\alpha-\beta)}{c}u_0^4+(\lambda_m-\lambda_D-1)u_0^3 +\frac{2(\beta-\alpha)}{c}u_0^2 +(1-\lambda_m)u_0\right],$$ which using (\[29\]) reduces to $$\label{36} u_1=h_0u_0\left[\frac{\beta-\alpha}{c}u_0-1\right].$$ Now let us determine the sign of $h_1$. Combining (\[31\]) and (\[36\]) results in $$\label{37} -\frac{ah_1}{3h_0^2}=\left(\frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_0\right)\left( \frac{\beta-\alpha}{c}u_0-1\right)h_0u_0^2.$$ From (\[11\]), (\[12\]) and (\[13\]), by applying the thermodynamics second law we deduce $$\label{38} \frac{\beta-\alpha}{c}u_0-1<0,$$ which results in $u_1<0$. $\dot {S}=0$ is ruled out because $h_1\neq 0$. So we conclude that the sign of $h_1$ must be the same as the sign of $(1+\lambda_D-\lambda_m)/(-3)-(\beta-\alpha)u_0/c$. For transition from quintessence to phantom phase we must have $$\label{39} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_0>0,$$ while a transition from phantom to quintessence era requires $$\label{40} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_0<0.$$ Note that if, like [@abd], we take $\alpha=0$, $\beta=1$ and $\lambda_m=\lambda_D$, transition from quintessence to phantom is not allowed. It is also instructive to study the behavior of $L$ at $t_i$. Assuming that $\dot{L}>0$, by inserting the Taylor expansion of $L$ at $t=t_i$ up to the order $\tau^2$ in (\[11\]), we obtain $\dot{L}=[L_0+L_1\tau+O(\tau^2)][h_0+h_1\tau^a+O(\tau^{(a+1)})]+\alpha-\beta$, where $L_0=L(t_i)$ and $L_1=\dot{L}(t_i)$. Hence $L_1=h_0L_0+\alpha-\beta$. (\[12\]) results in $u=c/(LH)$, therefore $$\label{41} u_0=\frac{c}{L_0h_0},\,\,u_1=-\frac{c}{L_0}\left( 1+\frac{\alpha-\beta}{h_0L_0}\right),$$ which is consistent with (\[36\]). As a summary we have shown that in order that a transition phase occurs: (i) (\[29\]) must have at least a positive real root in the interval (0,1) (ii) At these roots the Hubble parameter (if it is differentiable) must have the Taylor expansion (\[26\]), with an even integer $a$. In interacting holographic dark energy model we obtained $a=2$ and verified that quintessence to phantom phase transition and vice versa occur provided (\[39\]) and (\[40\]) hold respectively. Two transitions in interacting holographic dark energy model ============================================================= In this part we try to investigate the ability of the system to return to the quintessence regime from the phantom phase. This may be interesting because it avoids the big rip singularity which may be encountered in phantom models. Let us write (\[29\]) as $$\label{42} pu_0^3+qu_0^2+1=0,$$ where $$\label{43} p=\frac{2(\beta-\alpha)}{c(\lambda_m-3)},\, q=\frac{1+\lambda_D-\lambda_m}{\lambda_m-3}.$$ In order to have two transitions, (\[42\]), must possess two real positive roots, which we denote by $u_{01}$, $u_{02}$, in the interval $(0,1)$. $p$ and $q$ are real numbers, hence the third root must be also real. Therefore the discriminant of (\[42\]), i.e. $-27p^2-4q^3$, must be positive $$\label{44} \left(\frac{p}{2}\right)^2+\left(\frac{q}{3}\right)^3<0.$$ From $\sum_{i\neq j}u_{0i}u_{0j}=0$, we find that the third root, $u_{03}$, is negative and $\|u_{03}\|<1$. So using $0<-u_{03}u_{02}u_{01}=1/p<1$, we deduce $p>1$. Also following Descartes rule of sign, having two positive roots is only possible when $p>0$ and $q<0$. The Sturm sequence constructed from (\[42\]) is $\mathcal{S}(x)=\{\mathcal{P}(x),\mathcal{P}_1(x),\mathcal{P}_2(x),\mathcal{P}_3(x)\}$, where $$\begin{aligned} \label{45} \mathcal{P}(x)&=&px^3+qx^2+1 \nonumber \\ \mathcal{P}_1(x)&=&3px^2+2qx \nonumber \\ \mathcal{P}_2(x)&=&\frac{2q^2x}{9p}-1\nonumber \\ \mathcal{P}_3(x)&=&-\frac{9p}{4q^4}(27p^2+4q^3).\end{aligned}$$ We have $\mathcal{S}(0)=\{1,0,-1,-9p(4q^3+27p^2)/(4q^4)\}$, $\mathcal{S}(1)=\{p+q+1,3p+2q,-1+2q^2/(9p),-9p(4q^3+27p^2)/(4q^4)\}$. If we expect to have two real roots in the interval $(0,1)$, using (\[44\]) and by applying the Sturm theorem, we find $$\begin{aligned} \label{46} &&p+q+1>0 \nonumber \\ &&\frac{p}{2}+\frac{q}{3}>0\nonumber \\ &&\left(\frac{q}{3}\right)^2-\frac{p}{2}>0.\end{aligned}$$ These conditions and (\[44\]), may be resumed as $$\label{47} -q-1<p<2\left( -\frac{q}{3}\right)^{\frac{3}{2}},\,\, q<-3.$$ Now let us consider (\[39\]) and (\[40\]). A transition from quintessence to phantom phase may be occurred at $t_1$ ($u(t=t_1)=u_{01}$), if $h_1>0$ $$\label{48} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_{01}>0.$$ For $t>t_1$ the system becomes phantom dominated until $t=t_2$ ($u(t=t_2)=u_{02}$), i.e. when the second transition occurs, provided $h_1<0$ at $t_2$ $$\label{49} \frac{1+\lambda_D-\lambda_m}{-3}-\frac{\beta-\alpha}{c}u_{02}<0.$$ From (\[14\]) we find that $u$ is a decreasing function of time in phantom dominated era, hence $u_{01}>u_{02}$. So in order that (\[48\]) and (\[49\]) become consistent we must have $$\label{50} \beta-\alpha<0, \,\, 1+\lambda_D-\lambda_m>0.$$ The existence of $u_{01}$ and $u_{02}$ which satisfy (\[48\]) and (\[49\]), may be verified as follows: Using the condition posed on $p$ and $q$, we obtain $$\label{51} 0<-\frac{2q}{3p}(=-\frac{(1+\lambda_D-\lambda_m)c}{3(\beta-\alpha)})<1.$$ The Sturm sequence at $-2q/(3p)$ is $$\label{100} \mathcal{S}\left( -2q/(3p)\right)=\{4q^3/(27p^2)+1,0,-4q^3/(27p^2)-1,-9p(4q^3+27p^2)/(4q^4)\}.$$ So by invoking the Sturm theorem, one can verify that one of the positive roots, which based on our discussion in previous paragraph we take $u_{02}$, is located in $(0, -2q/(3p))$ while the other, i.e. $u_{01}$ belongs to $(-2q/(3p),1)$. As we have previously mentioned (see discussion after (\[25\])), $\omega$ is a decreasing function of time in the neighborhood of $u_{01}$ and an increasing function of time in the vicinity of $u_{02}$, whence for a differentiable $\omega$, we must have $\dot{\omega}=0$ at a point in the phantom regime. It is interesting to note that for $t_1<t<t_2$, this happens at $u=-2q/(3p)$, as it can be verified using (\[20\]). Indeed from (\[20\]), it is obvious that $\dot{\omega(t)}=0$ has three solutions: $u=-2q/(3p)$, $\dot{u}=0$ and $u=0$. In our case, $\dot{u}=0$ is ruled out by (\[23\]). The solution $u=-2q/(3p)$, as we have seen lies in the phantom regime. The solution $u=0$, must be obtained in the limit $t\to \infty$ in the quintessence era (by the assumption that the inflation will be continued). To verify this claim, we note that in an accelerating universe $\dot{H}+H^2>0$, and for $\alpha>\beta$, $$\label{52} (HL\dot{)}=(\dot{H}+H^2)L+(\alpha-\beta)H>0.$$ Therefore $HL$ is an analytic differentiable increasing function of time. Thereby $u$, is a decreasing function of time, as we showed before (see (\[23\])) via another method. If $u=0$ occurs at a finite time $\tilde{t}$, then for $t>\tilde{t}$, $u(t>\tilde{t})<0$ which conflicts with definition of $u$. In addition $u(\tilde{t})=0$ leads to $H(\tilde{t})L(\tilde{t})=\infty$ which conflicts with continuity of $H$ and $L$. As a test of our results, we have plotted $\omega+1$, (\[20\]), in terms of $u$ for an interacting holographic dark energy model with parameters $\{\beta=0$, $\alpha=1$, $c=1$, $\lambda_D=3.9$, $\lambda_m=2.5\}$ (see fig.(\[fig1\])). In this example $\omega+1$ has two zero in the interval $(0,1)$, $u_{01}=0.86$ corresponding to $\Omega_D=0.75$ and $u_{02}=0.72$ corresponding to $\Omega_D=0.52$. For $u_{02}<u<u_{01}$, the system is in phantom phase, i.e. $\omega<-1$, and for $u>u_{01}$ and $u<u_{02}$ the system is in quintessence phase. Note that for an accelerating universe and when $\alpha>\beta$, $u$ is a decreasing function of time (see(\[23\]) and our discussion in the previous paragraph) and the directions of $t$ and $u$-axis are opposite. $\dot{\omega}=0$ occurs in the phantom regime: $u=0.8$, and at $u=0$. Summary ======= In this paper we considered the holographic dark energy model with a general interaction between dark matter and dark energy, see (\[3\]). We took the infrared cutoff as a linear combination of the future an particle horizon see (\[9\]). We derived an expression for EOS parameter of the universe, $\omega$, in terms of the ratio of dark energy density and total energy density of the flat FRW space-time, $\Omega_D$, see (\[20\]). Using (\[20\]) and the expression obtained for time derivative of $\Omega_D$ (\[21\]), and by assumption that the thermodynamics second law is still valid, we studied the possibility (and necessary conditions) for quintessence to phantom phase transition and vice versa with a differentiable Hubble parameter. Using some theorems about solutions of cubic equation satisfied by $\Omega_D^{1/2}$ see (\[25\]), we showed that such transitions occur provided we appropriately choose the parameters of the system see(\[47\]) and (\[50\]). [99]{} S. Perlmutter et al., Nature (London) [391]{}, 51 (1998); D. N. Spergel et al. \[WMAP Collaboration\], Astrophys. J. Suppl. [148]{}, 175 (2003); A. C. S. Readhead et al., Astrophys. J. [609]{}, 498 (2004); J. H. Goldstein et al., Astrophys. J. [599]{}, 773 (2003); M. Tegmark et al. \[SDSS Collaboration\], Phys. Rev. D [69]{}, 103501 (2004); A. G. Riess et al. \[Supernova Search Team Collaboration\], Astron. J. [116]{}, 1009 (1998); S. Perlmutter et al. \[Supernova Cosmology Project Collaboration\], Astrophys. J. [517]{}, 565 (1999). B. Feng, X. L. Wang and X. M. Zhang, Phys. Lett. B [607]{}, 35 (2005); D. Huterer and A. Cooray, Phys. Rev. D [71]{}, 023506 (2005); S. Nesserisa and L. Perivolaropoulos, Phys.Rev. D [72]{}, 123519 (2005); U. Seljak, A. Slosar and P. McDonald, JCAP, 0610, 014 (2006). T. Padmanabhan, Phys. Rept. [380]{}, 235 (2003) ; V. Sahni and A. A. Starobinsky, Int. J. Mod. Phys. D [9]{}, 373 (2000) ; S. M. Carroll, Living Rev. Rel. [4]{}, 1 (2001) ; S. Weinberg, Rev. Mod. Phys. [61]{}, 1 (1989). R. R. Caldwell, R. Dave and P. J. Steinhardt, Phys. Rev. Lett. [80]{}, 1582 (1998); P. J. E. Peebles and A. Vilenkin, Phys. Rev. D [59]{}, 063505 (1999) ; P. J. Steinhardt, L. M. Wang and I. Zlatev, Phys. Rev. D [59]{}, 123504 (1999); M. Doran and J. Jaeckel, Phys. Rev. D [66]{}, 043519 (2002); A. R. Liddle, P. Parson and J. D. Barrow, Phys. Rev. D [50]{}, 7222 (1994); R. R. Caldwell, Phys. Lett. B [545]{}, 23 (2002); R. R. Caldwell, M. Kamionkowski and N. N. Weinberg, Phys. Rev. Lett. [91]{}, 071301 (2003); S. M. Carroll, M. Hoffman and M. Trodden, Phys. Rev. D [68]{}, 023509 (2003); J. M. Cline, S. Y. Jeon and G. D. Moore, Phys. Rev. D [70]{}, 043543 (2004); H. Wei, R. G. Cai and D. F. Zeng, Class. Quant. Grav. [22]{}, 3189 (2005); H. Wei and R. G. Cai, Phys. Rev. D [72]{}, 123507 (2005); M. Alimohammadi and H. Mohseni Sadjadi, Phys. Rev. D [73]{}, 083527 (2006). M. Li, Phys. Lett. B [603]{}, 1 (2004). P. Horava and D. Minic, Phys. Rev. Lett. [85]{}, 1610 (2000); K. Enqvist and M. S. Sloth, Phys. Rev. Lett. [93]{}, 221302 (2004); Q-G. Huang and T. Gong, JCAP [08]{}, 006 (2004); Q-G. Huang and M. Li, JCAP [08]{}, 013 (2004); B. Guberina, R. Horvat and H. Nikolic, Phys. Lett. B [636]{}, 80 (2006); M. R. Setare and S. Shafei, JCAP [0609]{}, 011 (2006). A. G. Cohen, D. B. Kaplan and A. E. Nelson, Phys. Rev. Lett. [82]{}, 4971 (1999). L. Amendola, Phys. Rev. D 62, 043511 (2000); W. Zimdahl, D. Pavon and L. P. Chimento, Phys. Lett. B 521, 133 (2001); W. Zimdahl, Int. J. Mod. Phys. D [14]{}, 2319 (2005); L.P. Chimento, A.S. Jakubi, D. Pavon and W. Zimdahl, Phys. Rev. D 67, 083513 (2003); D. Pavon and W. Zimdahl, Phys. Lett. B [628]{}, 206 (2005); G. Mangano, G. Miele and V. Pettorino, Modern. Phys. Lett. A 18, 831 (2003); G. Farrar and P.J.E, Peebles, Astrophys. J. 604, 1 (2004); S. del Campo, R. Herrera, and D. Pavon, Phys. Rev. D 70, 043540 (2004); R-G. Cai and A. Wang, JCAP [0503]{}, 002 (2005); Micheal S. Berger and H. Shojaei, Phys. Rev. D [73]{}, 083528 (2006); Bo Hu and Y. Ling, Phys. Rev. D [73]{}, 123510 (2006); H. Li, Z-k Guo and Y-z Zhang, Int. J. Mod. Phys. D 15, 869 (2006); S. Campo, R. Herrera, G. Olivares and D. pavon, Phys. Rev. D 74, 023501 (2006); B.Gumjudpai, T. Naskar, M. Sami and S. Tsujikawa, JCAP [0506]{}, 007 (2005). Y. H. Wei and Y. Z. Zhang, Grav. Cosmol. [9]{}, 307 (2003); Y. H. Wei and Y. Tian, Class. Quant. Grav. [21]{}, 5347 (2004); R. G. Cai, H. S. Zhang and A. Wang, Commun. Theor. Phys. [44]{}, 948 (2005); V. Sahni and Y. Shtanov, JCAP [0311]{}, 014 (2003); I. Y. Arefeva, A. S. Koshelev and S. Y. Vernov, Phys. Rev. D [72]{}, 064017 (2005); H. Wei, R. G. Cai and D. F. Zeng, Class. Quant. Grav. [22]{}, 3189 (2005); H. Wei and R. G. Cai, Phys. Rev. D [72]{}, 123507 (2005); A. Vikman, Phys. Rev. D [71]{}, 023515 (2005); A. Anisimov, E. Babichev and A. Vikman, JCAP [0506]{}, 006 (2005); Wang, Y. G. Gong and E. Abdalla, Phys. Lett. B [624]{}, 141 (2005); C. Carvalho and A. Saa, Phys. Rev. D [70]{}, 087302 (2004); M. Li, B. Feng and X. Zhang, JCAP [0512]{}, 002 (2005) ; H. Mohseni Sadjadi, Phys. Rev. D [73]{}, 063525 (2006); V. K. Onemli and R. P. Woodard, Phys. Rev. D[70]{}, 107301 (2004); S. Nojiri, S. D. Odintsov and S. Tsujikawa, Phys. Rev. [D 71]{}, 063004 (2005); B. Wang, Y. Gong, E. Abdalla, Phys.Lett. B [624]{}, 141 (2005); B. Wang, C-Y Lin, E. Abdalla, Phys. Lett. B [637]{}, 357 (2006). H. Kim, HW Lee and YS Myung, Phys.Lett. B [632]{}, 605 (2006). W. Fischler and L. Susskind, hep-th/9806039. E. Elizalde, S. Nojiri, S. D. Odintsov and P. Wang, Phys. Rev. D [71]{}, 103504 (2005). H. Mohseni Sadjadi and M. Alimohammadi, Phys. Rev. D [74]{}, 043506 (2006). [^1]: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the deconstructed version of a topcolor-assisted technicolor model wherein the mechanism of top quark mass generation is separated from the rest of electroweak symmetry breaking. The minimal deconstructed version of this scenario is a “triangle moose” model, where the top quark gets its mass from coupling to a top-Higgs field, while the gauge boson masses are generated from a Higgsless sector. The spectrum of the model includes scalar (top-Higgs) and pseudoscalar (top-pion) states. In this paper, we study the properties of these particles, discuss their production mechanisms and decay modes, and suggest how best to search for them at the LHC.' author: - 'R. Sekhar Chivukula' - Baradhwaj Coleppa - 'Heather E. Logan' - Adam Martin - 'Elizabeth H. Simmons' title: 'Top-Higgs and Top-pion phenomenology in the Top Triangle Moose model' --- Introduction ============ Higgsless models, as the name implies, break the electroweak symmetry without invoking a fundamental scalar particle. Inspired by the idea that one could maintain perturbative unitarity in extra-dimensional models through heavy vector resonance exchanges in lieu of a Higgs [@SekharChivukula:2001hz; @Chivukula:2002ej; @Chivukula:2003kq], Higgsless models were intially introduced in an extra-dimensional context as $SU(2)\times SU(2)\times U(1)$ gauge theories living in a slice of $AdS_5$, with symmetry breaking codified in the boundary condition of the gauge fields [@Csaki:2009bb; @Cacciapaglia:2006gp; @Cacciapaglia:2004zv; @Csaki:2004sz; @Cacciapaglia:2004rb; @Cacciapaglia:2004jz; @Csaki:2003zu]. It emerged that the low energy dynamics of these extra-dimensional models can be understood in terms of a collection of 4-D theories, using the principle of “deconstruction” [@Hill:2000mu; @ArkaniHamed:2001ca]. Essentially, this involves latticizing the extra dimension, associating a 4-D gauge group with each lattice point and connecting them to one another by means of nonlinear sigma models. The five dimensional gauge field is now spread in this theory as four dimensional gauge fields residing at each lattice point, and the fifth scalar component residing as the eaten pion in the sigma fields. The picture that emerges is called a “Moose” diagram [@Georgi:1985hf]. The $AdS/CFT$ correspondence suggests that these models can be understood to be dual to strongly coupled technicolor models. The key features of these models [@Casalbuoni:1985kq; @Casalbuoni:1995qt; @Chivukula:2005ji; @Chivukula:2005xm; @Chivukula:2005bn; @Kurachi:2004rj; @Chivukula:2004af; @Chivukula:2004pk; @Lane:2009ct] are the following: - Spin-1 resonances created by the techni-dynamics are described as massive gauge bosons, following the Hidden-Local-Symmetry setup originally used for QCD [@Bando:1987br; @Bando:1985rf; @Bando:1984pw; @Bando:1984ej; @Bando:1987ym]. The mass of the resonances is roughly $\tilde g F$, where $F$ is around the weak scale and $\tilde g$ is a large coupling. Interactions of the resonances with other gauge bosons and fermions are calculated as a series in $1/\tilde g \ll 1$. - Standard model (SM) fermions reside primarily on the exterior sites – the sites approximately corresponding to $SU(2)_w$ and $U(1)_Y$ gauge groups. Fermions become massive through mixing with massive, vector-like fermions located on the interior, ‘hidden’ sites. - Precision electroweak parameters (S,T,U), perennially a thorn in the side of dynamical electroweak breaking models [@Peskin:1991sw], are accommodated by delicately spreading the SM fermions between sites. By adjusting the fermion distribution across sites to match the gauge boson distribution, S, T, U can all be reduced to acceptable levels. This is identical to the solution used in extra-dimensional Higgsless models, where the spreading of a fermion among sites becomes a continuous distribution, or profile, in the extra dimension [@Cacciapaglia:2006gp]. This adjustment is called “ideal delocalization” [@Chivukula:2005xm]. The most economical deconstructed Higgsless model constructed along these lines (a “three site” model) was presented in [@Chivukula:2006cg], and had, in addition to the SM spectrum, one heavy partner for each fermion and the $W$ and $Z$ bosons. Though providing an excellent ground for studying the low-energy properties of Higgsless models, the mass of the heavy Dirac partners of the SM fermions in this model was constrained to lie at $\sim$ 2 TeV, because of the tension between obtaining the correct value for the top-quark mass and keeping $\Delta \rho$ within experimental bounds. To alleviate this constraint, an extension of the three site model was constructed [@Chivukula:2009ck], whose goal was to separate the top-quark mass generation from the rest of electroweak symmetry breaking (EWSB), thus relaxing the aforementioned constraint. This idea of treating the top-quark mass as arising from a separate dynamics is not new - in fact, Topcolor-assisted technicolor models [@Hill:1991at; @Hill:1994hp; @Lane; @and; @Eichten; @-; @TC2; @Popovic:1998vb; @Hill; @and; @Simmons; @Braam:2007pm] employ precisely this idea. Topcolor-assisted technicolor is a scenario of dynamical electroweak symmetry breaking in which the strong dynamics is partitioned into two different sectors. One sector, the technicolor sector, is responsible for the bulk of electroweak symmetry breaking and is therefore characterized by a scale $F \sim v$, where $v=$ 246 GeV is the EWSB scale. Consequently, technicolor dynamics is responsible for the majority of gauge boson masses and, more indirectly, light fermion masses. The second strong sector, the topcolor sector, only communicates directly with the top quark. Its sole purpose is to generate a large mass for the top quark. In generating a top quark mass, this second sector also breaks the electroweak symmetry. If the characteristic scale of the topcolor sector is low, $f \ll F$, it plays only a minor role in electroweak breaking, but can still generate a sufficiently large top quark mass given a strong enough top-topcolor coupling. At low-energies, the top-color dynamics is summarized by the existence of a new dynamical top-Higgs which couples preferentially to the top-quark. The introduction of the top-Higgs field serves to alleviate the tension between obtaining the correct top quark mass and keeping $\Delta \rho$ small that exists in Higgsless models by separating the top quark mass generation from the rest of electroweak symmetry breaking. An important consequence of this separation is that the model permits heavy Dirac partners for the fermions that are potentially light enough to be seen at the LHC. Thus, the combination of two symmetry-breaking mechanisms can achieve both dynamical electroweak breaking and a realistic top quark mass. Because electroweak symmetry is effectively broken twice in this scenario, there are two sets of Goldstone bosons in the theory. One triplet of these Goldstones is eaten to become the longitudinal modes of the $W^{\pm}/Z^0$, while the other triplet remains in the spectrum. This remaining triplet, typically called the top-pions, and a singlet partner, the top-Higgs, are the focus of this paper. The top-pions and top-Higgs couple preferentially to the third generation of quarks, which makes them interesting for a number of reasons. First, the interactions of the top quark are the least constrained of all fermions, so new dynamics coupling preferentially to the top quark is not phenomenologically excluded. Second, the gluon fusion mechanism involves a top quark loop and is an efficient method for singly producing top-Higgses and neutral top-pions at the LHC. In fact, the strong top-quark–topcolor interaction, manifest in a top Yukawa of order $\sim$ few, significantly enhances the coupling of top-Higgs and top-pions relative to a SM Higgs of equal mass . Such a large cross section leads to exciting LHC signals which may be discoverable in the initial low-energy, low-luminosity run. Our goal in this paper is to lay the foundation for phenomenological studies of the top-pions and top-Higgs at the LHC. We begin in Sec. \[sec:model\] by setting out the relevant details of the Top Triangle Moose model. In Sec. \[sec:couplings\] we identify the physical top-pion states and summarize their couplings to other particles. Section \[sec:pheno\] contains the bulk of our phenomenological results. After identifying the existing experimental constraints on the top-pions and top-Higgs, we study their decay branching ratios, direct production cross sections in $pp$ collisions, and production in decays of the heavy vector-like top quark partners or the heavy gauge bosons in the model. We identify cases where the LHC has the clear ability to discover the new states and others where good potential for discovery exists and further detailed study is warranted. We summarize our findings and discuss their implications in Sec. \[sec:conclusions\]. The model {#sec:model} ========= The Top Triangle Moose model [@Chivukula:2009ck] is shown in moose notation in Fig. \[fig:Triangle\]. The circles represent global SU(2) symmetry groups; the full SU(2) at sites 0 and 1 are gauged with gauge couplings $g$ and $\tilde g$, respectively, while the $\tau^3$ generator of the global SU(2) at site 2 is gauged with U(1) gauge coupling $g^{\prime}$. The lines represent spin-zero link fields which transform as a fundamental (anti-fundamental) representation of the group at the tail (head) of the link. $\Sigma_{01}$ and $\Sigma_{12}$ are nonlinear sigma model fields, while $\Phi$ (the top-Higgs doublet) is a linear sigma model field. ![The gauge structure of the model in Moose notation. $g$ and $g^{\prime}$ are approximately the SM $SU(2)$ and hypercharge gauge couplings while $\tilde{g}$ represents the ‘bulk’ gauge coupling. The left (right) handed light fermions are mostly localized at site 0 (2) while their heavy counterparts are mostly at site 1. The links connecting sites 0 and 1 and sites 1 and 2 are nonlinear sigma model fields while the one connecting sites 0 and 2 is a linear sigma field. Site 2 is dotted to indicate that only the $\tau_3$ component is gauged.[]{data-label="fig:Triangle"}](triangle.pdf){width="2.4in"} The kinetic energy terms of the link fields corresponding to these charge assignments are: $$\mathcal{L}_{gauge}= \frac{F^{2}}{4} \textrm{Tr}[(D_{\mu}\Sigma_{01})^{\dagger}D^{\mu}\Sigma_{01}] + \frac{F^{2}}{4} \textrm{Tr}[(D_{\mu}\Sigma_{12})^{\dagger}D^{\mu}\Sigma_{12}] + (D_{\mu}\Phi)^{\dagger}D^{\mu}\Phi, \label{eqn:Gauge L}$$ where the covariant derivatives are: $$\begin{aligned} D_{\mu}\Sigma_{01} &=& \partial_{\mu}\Sigma_{01} + igW_{0\mu} \Sigma_{01} - i\tilde{g}\Sigma_{01}W_{1\mu}, \nonumber \\ D_{\mu}\Sigma_{12} &=& \partial_{\mu}\Sigma_{12} + i\tilde{g}W_{1\mu} \Sigma_{12} - ig^{\prime} \Sigma_{12}\tau^{3}B_{\mu}, \nonumber \\ D_{\mu}\Phi &=& \partial_{\mu}\Phi + igW_{0\mu} \Phi - \frac{ig^{\prime}}{2}B_{\mu}\Phi. \label{eqn:covariant}\end{aligned}$$ Here the gauge fields are represented by the matrices $W_{0\mu} = W_{0\mu}^{a}\tau^{a}$ and $W_{1\mu} = W_{1\mu}^{a}\tau^{a}$, where $\tau^a = \sigma^a/2$ are the generators of SU(2). The nonlinear sigma model fields $\Sigma_{01}$ and $\Sigma_{12}$ are 2$\times$2 special unitary matrix fields. To mimic the symmetry breaking caused by underlying technicolor and topcolor dynamics, we assume all link fields develop vacuum expectation values (vevs): $$\langle \Sigma_{01} \rangle = \langle \Sigma_{12} \rangle = \mathbf{1}_{2\times 2}, \qquad \qquad \langle \Phi \rangle = \left( \begin{array}{c} f/\sqrt{2} \\ 0 \end{array} \right).$$ In order to obtain the correct amplitude for muon decay, we parameterize the vevs in terms of a new parameter $\omega$, $$F=\sqrt{2}\,v\,\cos\,\omega, \qquad \qquad f=v\,\sin\,\omega, \label{eq:Ff}$$ where $v = 246$ GeV is the weak scale. As a consequence of the vacuum expectation values, the gauge symmetry is broken all the way down to electromagnetism and we are left with massive gauge bosons (analogous to techni-resonances), top-pions and a top-Higgs. To keep track of how the degrees of freedom are partitioned after we impose the symmetry breaking, we expand $\Sigma_{01}$, $\Sigma_{12}$ and $\Phi$ around their vevs. The coset degrees of freedom in the bi-fundamental link fields $\Sigma_{01}$ and $\Sigma_{12}$ can be described by nonlinear sigma fields: $$\Sigma_{01}=\textrm{exp}\left(\frac{2i\pi_{0}^{a}\tau^{a}}{F}\right), \qquad \qquad \Sigma_{12}=\textrm{exp}\left(\frac{2i\pi_{1}^{a}\tau^{a}}{F}\right), \label{eq:vevs}$$ while the degrees of freedom in $\Phi$ fill out a linear representation, $$\Phi= \left( \begin{array}{c} (f + H_t + i\pi_{t}^{0})/\sqrt{2} \\ i\pi^{-}_{t} \end{array} \right). \label{eqn:phi representation}$$ The gauge-kinetic terms in Eq. yield mass matrices for the charged and neutral gauge bosons. The photon remains massless and is given by the exact expression $$A_{\mu} = \frac{e}{g} W_{0 \mu}^3 + \frac{e}{\tilde g} W_{1 \mu}^3 + \frac{e}{g^{\prime}} B_{\mu},$$ where $e$ is the electromagnetic coupling. Normalizing the photon eigenvector, we get the relation between the coupling constants: $$\frac{1}{e^2}=\frac{1}{g^2}+\frac{1}{\tilde{g}^2}+\frac{1}{g'^2}.$$ This invites us to conveniently parametrize the gauge couplings in terms of $e$ by $$g = \frac{e}{\sin\theta \cos\phi} = \frac{g_0}{\cos\phi}, \qquad \qquad \tilde g = \frac{e}{\sin\theta \sin\phi} = \frac{g_0}{\sin\phi}, \qquad \qquad g^{\prime} = \frac{e}{\cos\theta}. \label{eqn:gauge couplings}$$ We will take $\tilde g \gg g$, which implies that $\sin\phi \equiv x$ is a small parameter. The result of the diagonalization of the gauge boson matrices perturbatively in $x$ is summarized in Appendix A. Counting the number of degrees of freedom, we see that there are six scalar degrees of freedom on the technicolor side ($\Sigma_{01}, \Sigma_{12}$) and four on the topcolor side ($\Phi$). Six of these will be eaten to form the longitudinal components of the $W^{\pm}$, $Z^0$, $W^{\prime \pm}$, and $Z^{\prime 0}$. This leaves one isospin triplet of scalars and the top-Higgs $H_t$ as physical states in the spectrum. While the interactions in Eq. are sufficient to give mass to the gauge bosons, the top-pions and top-Higgs remain massless at tree level. Quantum corrections will give the top-pions a mass, however this loop-level mass is far too small to be consistent with experimental constraints. To generate phenomenologically acceptable masses for the top-pions and top-Higgs, we add two additional interactions: $$\mathcal L_M = - \lambda\, {\rm Tr} \left( M^{\dagger}M - \frac{f^2}{2} \right)^2 - \kappa f^2\, {\rm Tr} \left\| M - \frac{f}{\sqrt 2} \Sigma_{01}\Sigma_{12} \right\|^2, \label{eq:Vhiggs}$$ where the first of these interactions arises from topcolor interactions, and the second from ETC-like interactions [@Eichten:1979ah]. Here $\lambda$ and $\kappa$ are two new parameters, $f$ is the same vacuum expectation value appearing in Eq. , and $M$ is the $\Phi$ field expressed as a matrix, schematically given by $M = (\tilde \Phi, \Phi)$ with $\tilde \Phi = -i \sigma_2 \Phi^$: $$M = \left( \begin{array}{cc} i \pi_t^+ & (f + H_t + i \pi_t^0)/\sqrt{2} \\ (f + H_t - i \pi_t^0)/\sqrt{2} & i \pi_t^- \end{array} \right),$$ where $\pi_t^+ = (\pi_t^-)^$. The first term in Eq. depends only on the modulus of $M$, and therefore contributes only to the mass of the top-Higgs. The second term gives mass to both the top-Higgs and the physical (uneaten) combination of pion fields, as we will show shortly. Because these masses depend on two parameters, $\lambda$ and $\kappa$, we can treat the mass of the top-Higgs and the common mass of the uneaten top-pions as two independent parameters. In addition to generating masses, the potential in Eq. also induces interactions between the top-Higgs and top-pions which can be important. Finally, we note that the mass terms for the light fermions arise from Yukawa couplings of the fermionic fields with the nonlinear sigma fields $$\begin{aligned} \mathcal{L} & = & M_{D}\left[\epsilon_{L}\bar{\psi}_{L0}\Sigma_{01}\psi_{R1}+\bar{\psi}_{R1}\psi_{L1}+\bar{\psi}_{L1}\Sigma_{12}\left(\begin{array}{cc} \epsilon_{uR} & 0\\ 0 & \epsilon_{dR}\end{array}\right)\left(\begin{array}{c} u_{R2}\\ d_{R2}\end{array}\right)\right]. \label{eqn:Light fermion mass}\end{aligned}$$ We have denoted the Dirac mass that sets the scale of the heavy fermion masses as $M_D$. Here, $\epsilon_{L}$ is a parameter that describes the degree of delocalization of the left handed fermions and is flavor universal. All the flavor violation for the light fermions is encoded in the last term; the delocalization parameters for the right handed fermions, $\epsilon_{fR}$, can be adjusted to realize the masses and mixings of the up and down type fermions. The mass of the top quark arises from similar terms with a unique left-handed delocalization parameter $\epsilon_{tL}$ and also from a unique Lagrangian term reflecting the coupling of the top-Higgs to the top quark: $$\mathcal{L}_{top}=-\lambda_{t}\bar{\psi}_{L0}\,\Phi\, t_{R}+h.c.\label{eq:top quark mass L}$$ Details of the fermion masses and mass eigenstates are given in Appendix A. Physical top-pions and their couplings {#sec:couplings} ====================================== The next step towards understanding top-pion phenomenology is to identify the combination of degrees of freedom which make up the physical (uneaten) top-pions. While the top-Higgs $H_t$ remains a mass eigenstate, the pions $\pi_0^a$, $\pi_1^a$ and $\pi_t^a$ mix. We can identify the physical top-pions as the linear combination of states that cannot be gauged away. We do this by isolating the Goldstone boson states that participate in interactions of the form $V_{\mu} \partial^{\mu} \phi$ in the Lagrangian. We start by expanding the nonlinear sigma fields to first order in $\pi/F$, $$\begin{aligned} \Sigma_{01} &=& 1 + \frac{2i\pi_{0}^{a}\tau^{a}}{F} + \mathcal{O}\left(\frac{\pi^2}{F^2}\right), \label{eqn:sigma01} \\ \Sigma_{12} &=& 1 + \frac{2i\pi_{1}^{a}\tau^{a}}{F} + \mathcal{O}\left(\frac{\pi^2}{F^2}\right). \label{eqn:sigma12} \end{aligned}$$ Plugging this in Eq. [^1], we can read off the various interaction terms. The complete details are given in Appendix B. Here, we concentrate on the gauge-Goldstone mixing terms: $$\mathcal{L}_{\textrm{mixing}} = \frac{g}{2} W_0^{a\mu} \partial_{\mu} \left[F\pi_0^a + f\pi_t^a \right] + \frac{\tilde{g}}{2} W_1^{a\mu} \partial_{\mu} \left[F \pi_1^a - F \pi_0^a \right] - \frac{g^{\prime}}{2} B_2^{\mu} \partial_{\mu} \left[F \pi_1^3 + f \pi_t^3 \right]. \label{eq:gauge-Goldstone}$$ Note that the pion combination in the third term can be written as a linear combination of those appearing in the first two terms: $$F \pi_1^3 + f \pi_t^3 = [F \pi_0^3 + f \pi_t^3] + [F \pi_1^3 - F \pi_0^3].$$ The two eaten triplets of pions span the linear combinations that appear in the first two terms of Eq. , leaving the third linear combination as the remaining physical top-pions, which we will denote $\Pi_t^a$: $$\Pi_t^a = - \sin\omega \left(\frac{\pi_0^a + \pi_1^a}{\sqrt{2}} \right) + \cos\omega \, \pi_t^a,$$ where we have normalized the state properly using the definitions of $F$ and $f$ in Eq. . The physical top-pions can also be identified by expanding the top-Higgs potential given in Eq. and collecting the mass terms. The masses of $H_t$ and $\Pi_t^a$ are given by, $$M^2_H = 2 v^2 (\kappa + 4 \lambda) \sin^2{\omega}, \qquad \qquad M^2_{\Pi_t} = 2 v^2 \kappa \tan^2{\omega}, \label{eq:pimass}$$ while the other two linear combinations of pions are massless, as true Goldstone bosons should be. Equation (\[eq:Vhiggs\]) also contains trilinear couplings between $H_t$ and two top-pions; we find $$\begin{aligned} \lambda_{H_t \Pi_t^0 \Pi_t^0} = \lambda_{H_t \Pi_t^+ \Pi_t^-} &=& 2 v \sin{\omega} \left( \kappa \sin^2{\omega} \tan^2{\omega} + 4 \lambda \cos^2{\omega} \right) \nonumber \\ &=& \frac{1}{2v \sin{\omega}} \left[(M^2_H - 2 M^2_{\Pi_t})\cos{2\omega} + M^2_H \right].\end{aligned}$$ These couplings are important for top-Higgs decays when $M_H > 2 M_{\Pi_t}$. Having worked out the physical top-pion combination, all that remains is to express the interactions in the Lagrangian in terms of mass eigenstates. The top-pion combination is given above, while the gauge boson and fermion mass eigenstates are given in Ref. [@Chivukula:2009ck] and are summarized in Appendix A. This conversion is straight-forward, but tedious, so we just summarize the results for the three-point couplings in Tables \[tab:pipiV\]–\[tab:H-couplings\]. We write the couplings in terms of $x$, $\sin\theta$, and $g_{0}$, with the latter two defined as in Eq. . The results in this section are given as an expansion in powers of $x$ and include terms up to order $x^2$. ------------------------------------------------------------------------------------------------ Vertex Strength ------------------------------------ ----------------------------------------------------------- $A_{\mu}\Pi_t^+\Pi_t^-$ $e\,(p_{\Pi^+}-p_{\Pi^-})_{\mu}$ $Z_{\mu}\Pi_t^+\Pi_t^-$ $\frac{g_0}{\cos\theta} \left[\left(\frac{1}{2} - \sin^2\theta \right) + \frac{x^2}{16} \sec^2\theta \left(2 + \cos2\theta \,\sec^2\omega \right) \right] (p_{\Pi^-}-p_{\Pi^+})_{\mu}$ $Z^{\prime}_{\mu}\Pi_t^+\Pi_t^-$ $\frac{g_0}{2x} \left[\sin^2\omega - \frac{x^2}{16}(7+\cos 2\omega) \sec^2\theta \right] (p_{\Pi^-}-p_{\Pi^+})_{\mu}$ $W_{\mu}^{+}\Pi_t^0\Pi_t^-$ $-\frac{g_{0}}{2} \left[1 + \frac{x^{2}}{8}(2+\cos 2\omega) \sec^2\omega \right] (p_{\Pi^0}-p_{\Pi^-})_{\mu}$ $W_{\mu}^{\prime +}\Pi_t^0\Pi_t^-$ $-\frac{g_0}{2x} \left[\sin^2\omega - \frac{x^2}{16}\left(7+\cos 2\omega \right) \right] (p_{\Pi^0}-p_{\Pi^-})_{\mu}$ ------------------------------------------------------------------------------------------------ : \[tab:pipiV\] Couplings of two top-pions to a vector boson. These have been calculated to $\mathcal{O}(x^{2})$. Here $p_{\Pi}$ is the outgoing momentum of particle $\Pi$. ---------------------------------------------------------------------------------------------------------------------- Vertex Strength --------------------------------------------- ------------------------------------------------------------------------ $\Pi_t^-A_{\mu}W_{\nu}^+$ $0$ $\Pi_t^-A_{\mu}W^{\prime +}_{\nu}$ $0$ $\Pi_t^-Z_{\mu}W_{\nu}^+$ $-\frac{i e^2 x^2}{16} v\,\sec^3\theta\,\tan\,\omega$ $\Pi_t^-Z_{\mu}W_{\nu}^{\prime +}$ $-\frac{i\,g_0^2}{4x} v\,\sec\theta\,\sin2\omega \left[1 + \frac{x^2}{16}\left(3+5\,\cos2\theta \right) \sec^2\theta \right]$ $\Pi_t^-Z^{\prime}_{\mu}W^+_{\nu}$ $\frac{i\,g_0^2}{4x} v \sin2\omega \left[1 + \frac{x^2}{16}\left(5+3\,\cos2\theta \right)\sec^2\theta \right]$ $\Pi_t^-Z^{\prime}_{\mu}W^{\prime +}_{\nu}$ $-\frac{i\,g_0^2}{8} v \, \tan^2\theta \left[\sin 2\omega + \frac{x^2}{32}\left(10-2\,\cos 2\theta + 3\,\cos [2(\theta-\omega)] + 2\,\cos 2\omega + 3\,\cos [2(\theta+\omega)] \right) \sec^2 \theta\, \tan\omega \right]$ $\Pi_t^0W^+_{\mu}W^{\prime -}_{\nu}$ $\frac{ig_0^2}{4x} v \, \sin 2\omega \left(1 + \frac{x^2}{2}\right)$ ---------------------------------------------------------------------------------------------------------------------- : \[tab:piVV\] Couplings of a top-pion to a pair of gauge bosons. The corresponding Feynman rule is obtained by inserting a $g_{\mu\nu}$ in each coupling. These have been calculated to $\mathcal{O}(x^{2})$. ---------------------------------------------------------------------------------------------------------------------- Vertex Strength ------------------------------------------- -------------------------------------------------------------------------- $H_tW^+_{\mu}W^-_{\nu}$ $\frac{g_{0}^{2}}{2} v \sin\omega \left(1+\frac{3\,x^2}{4} \right)$ $H_tW^{\prime +}_{\mu}W^-_{\nu}$ $-\frac{g_{0}^{2} x}{4} v \sin\omega$ $H_tW^{\prime +}_{\mu}W^{\prime -}_{\nu}$ $\frac{g_{0}^{2} x^{2}}{8} v \sin\omega$ $H_tZ_{\mu}Z_{\nu}$ $\frac{g_0^2}{4 \cos^2\theta} v \sin\omega \left[1 + \frac{x^2}{4}\left(1 + 2\cos 2\theta \right) \sec^2 \theta \right]$ $H_tZ^{\prime}_{\mu}Z_{\nu}$ $-\frac{g_{0}^{2} x}{4 \cos^2\theta} v \sec\theta \cos 2\theta \sin\omega$ $H_tZ^{\prime}_{\mu}Z^{\prime}_{\nu}$ $\frac{g_{0}^{2} x^2}{16 \cos^2\theta} v \sec^2\theta \cos^2 2\theta \sin\omega$ $H_t \Pi_t^- W^+_{\mu}$ $\frac{g_0}{2} \cos\omega \left(1+\frac{3 x^2}{8} \right)(p_H - p_{\Pi^-})_{\mu}$ $H_t \Pi_t^- W^{\prime +}_{\mu}$ $-\frac{g_{0} x}{4} \cos\omega (p_H - p_{\Pi^-})_{\mu}$ $H_t \Pi_t^0 Z_{\mu}$ $-\frac{g_{0}}{2 \cos\theta} \cos\omega \left[1 + \frac{x^2}{8} \left(1 + 2\cos 2\theta \right) \sec^2\theta \right](p_H - p_{\Pi^0})_{\mu}$ $H_t \Pi_t^0 Z^{\prime}_{\mu}$ $\frac{g_{0} x}{4 \cos\theta} \cos\omega \sec\theta \cos 2\theta (p_H - p_{\Pi^0})_{\mu}$ ---------------------------------------------------------------------------------------------------------------------- : \[tab:H-couplings\] Three-point couplings of the top-Higgs, again calculated to $\mathcal{O}(x^2)$. The Feynman rules for the couplings involving two gauge bosons are obtained by multiplying the coupling strength given here by $i g_{\mu\nu}$. Notice in particular that the couplings of the heavy gauge bosons $Z^{\prime}$ and $W^{\prime \pm}$ to two top-pions are proportional to the large gauge coupling $\tilde g = g_0/x$ associated with site 1. The leading term in these couplings is in fact $\tilde g \sin^2\omega/2$, with the two $\sin\omega$ factors reflecting the overlap of the $\Pi_t^a$ wavefunction with the combination of nonlinear sigma fields $(\pi_0^a + \pi_1^a)/\sqrt{2}$. The couplings of the top-Higgs and top pions to third generation fermions (and their heavy partners) can be likewise computed, by plugging in the mass eigenstates into the top quark mass term, Eq. , with $\Phi$ is given by Eq. ). The results are shown in Table IV, written in terms of the parameter $a= v\,\textrm{sin}\,\omega/ \sqrt{2}M_{D}$. We have also worked out the four-point interactions. While these are less important phenomenologically, we list the mass-basis couplings in Appendix C for completeness. Vertex Strength -------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $H_t\,t_{L}\bar{t}_R+h.c$ $\frac{\lambda_t}{\sqrt{2}}\left[-1+\frac{(1+a^2)(x^2+2 \epsilon_{tR}^2)+4\sqrt{2}ax\, \epsilon_{tR}}{4(a^2-1)^2} \right]$ $H_t\,T_{L}\bar{T}_R+h.c$ $-\frac{\lambda_t}{\sqrt{2}}\left[\frac{a (x^2+2 \epsilon_{tR}^2)+\sqrt{2}(1+a^2)x\,\epsilon_{tR}}{2(a^2-1)^2} \right]$ $H_t\,t_{L}\bar{T}_R+h.c$ $\lambda_t \left[\frac{a\,x+\sqrt{2}\epsilon_{tR}}{2(a^2-1)} \right]$ $H_t\,t_{R}\bar{T}_L+h.c$ $\lambda_t \left[\frac{x+\sqrt{2}a\,\epsilon_{tR}}{2(a^2-1)} \right]$ $\Pi_{t}^0\bar{t}_Lt_R-h.c.$ $\frac{i\lambda_t}{\sqrt{2}}\left[-\textrm{cos}\,\omega + \frac{(a^2+\textrm{cos}2\omega)\textrm{sec}\,\omega(x^2+2\,\epsilon_{tR}^2)+\frac{2\sqrt{2}}{a}x\,\epsilon_{tR}\left[2a^2\textrm{cos}\,\omega+(a^2-1)\textrm{sin}\,\omega\,\textrm{tan}\,\omega \right]}{4\sqrt{2}(a^2-1)^2} \right]$ $\Pi_{t}^0\bar{T}_LT_R-h.c.$ $\frac{i\lambda_t}{4\sqrt{2}}\left[\frac{ 2\sqrt{2}\,x\,\epsilon_{tR}(a^2+\textrm{cos}2\omega)+(2a^2\textrm{cos}\,\omega+(a^2-1)\textrm{sin}\,\omega\,\textrm{tan}\,\omega)\,(x^2+2\,\epsilon_{tR}^2)}{(a^2-1)^2} \right]$ $\Pi_{t}^0\bar{t}_LT_R-h.c.$ $\frac{i\lambda_t}{(a^2-1)}\left[\frac{x\textrm{sec}\,\omega(-1+3a^2+(1+a^2)\,\textrm{cos}2\omega)}{8a}+\frac{\epsilon_{tR}\textrm{cos}\,\omega}{\sqrt{2}} \right]$ $\Pi_{t}^0\bar{T}_Lt_R-h.c.$ $\frac{i\lambda_t}{(a^2-1)}\left[\frac{x\,\textrm{cos}\,\omega}{2}+\frac{\epsilon_{tR}\,\textrm{sec}\,\omega(-1+3a^2+(1+a^2)\,\textrm{cos}2\omega)}{4\sqrt{2}a} \right]$ $\Pi_{t}^{-}t_R\bar{b}_L-h.c.$ $i\lambda_t\left[\textrm{cos}\,\omega-\frac{x^2\left(a^4+(a^4-2a^2+2)\,\textrm{cos}2\omega \right)\textrm{cos}\,\omega}{8(a^2-1)^2}-\frac{x\,\epsilon_{tR}\,\textrm{sec}\,\omega \left(-2+5a^2-a^4+(a^4-a^2+2)\,\textrm{cos}2\omega \right)\textrm{cos}\,\omega}{4\sqrt{2}a(a^2-1)^2}-\frac{\epsilon_{tR}^2\,\textrm{cos}\,\omega}{2(a^2-1)^2} \right]$ $\Pi_{t}^{-}T_R\bar{B}_L-h.c.$ $\frac{i\lambda_t\,\textrm{sec}\,\omega}{(a^2-1)}\left[-\frac{x^2\,(-1+3a^2+(1+a^2)\,\textrm{cos}2\omega)}{8a}-\frac{x\,\epsilon_{tR}(1+3\textrm{cos}2\omega)}{4\sqrt{2}a}+\frac{\textrm{sin}^2\omega\,\epsilon_{tR}^2}{2a} \right]$ $\Pi_{t}^{-}T_R\bar{b}_L-h.c.$ $-\frac{i\lambda_t}{(a^2-1)}\left[\frac{x\left(-1+3a^2+(1+a^2)\,\textrm{cos}2\omega \right)\,\textrm{sec}\,\omega}{4\sqrt{2}a}+\epsilon_{tR}\,\textrm{cos}\,\omega \right]$ $\Pi_{t}^{-}t_R\bar{B}_L-h.c.$ $i\lambda_t\left[\frac{x\,\textrm{cos}\,\omega}{\sqrt{2}}-\frac{\epsilon_{tR}\,\textrm{sin}\,\omega\,\textrm{tan}\,\omega}{2a} \right]$ : The couplings of the top-Higgs and top pions to third generation fermions and their heavy partners, calculated to $\mathcal{O}(x^{2},\epsilon_{tR}^2)$. \[tab:H-fermion\] Top-Higgs and top-pion phenomenology {#sec:pheno} ==================================== We are now prepared to investigate the phenomenology of the new states related to the top quark: the top-Higgs $(H_t)$, the top-pions $(\Pi_t)$, and the heavy vector fermion partner of the top quark $(T)$. First, we will show how existing Tevatron data can be applied to place limits on the top triangle moose model. Essentially, rescaling to take altered coupling values into account allows limits derived for other models to be transformed into limits on our model’s top-pions and top-Higgs. Next, we study top-Higgs and top-pion production at LHC. As indicated in Figure (\[fig:higgs\_prod\]) below, the new scalars can be produced either directly, through gluon fusion via a top loop, or indirectly, via decays of the heavy $T$ quarks. ![\[fig:feynman1\] Feynman diagrams for the single production of the top-Higgs or top pion. Direct production proceeds via a top loop (left). Indirect production occurs as a decay product of heavy T quarks (right), where the pair production of the T quarks proceeds via gluon fusion and quark annihilation.[]{data-label="fig:higgs_prod"}](higgs-prod.pdf) The multiple production modes will make it possible to confirm that the scalar one has discovered is, in fact, the $H_t$ of this model, rather than some other scalar state. We examine the branching ratios of the $H_t, \Pi_t$, and $T$, in order to identify production channels at LHC that are likely to lead to discovery of these new particles. Then, we discuss how already-planned searches at LHC can be repurposed or extended to yield information about the individual states and the relationships between them. Current constraints on parameter values --------------------------------------- Before starting our phenomenological analysis, let us briefly recall some of the limits on the different parameters of our model. First, within the gauge sector there is the mass of the $W'$ boson, $M_{W'}$, and the ratio of gauge couplings $x$. Second, there are parameters related to fermion masses: the quantities $\epsilon_L$, $\epsilon_{fR}$, $\epsilon_{tL}$, $\epsilon_{tR}$, and $M_D$. In addition, from the top-Higgs sector, we have $\lambda_t$, $\textrm{sin}\,\omega$ and the masses of the top-pion and the top-Higgs, $M_{\Pi_t}$ and $M_{H_t}$. As discussed in [@Chivukula:2006cg], the value of the $W'$ mass is constrained to lie above 380 GeV by the LEP II measurement of the triple gauge boson vertex and to lie below 1.2 TeV by the need to maintain perturbative unitarity in $W_L W_L$ scattering. We will use the illustrative $M_{W'} = 500$ GeV in our calculations (except where noted otherwise) both for definiteness and because the value of the parameter $\epsilon_L$ is then derivable from $M_{W'}$ (and $x$) via ideal delocalization, as shown below. In principle, the values of the various $\epsilon_{fR}$ are proportional to the masses of the light fermions; since we will be working in the limit $m_f \rightarrow 0$ for fermions other than the top quark, we will set $\epsilon_{fR}=0$. Similarly, since the top quark mass depends very little on $\epsilon_{tR}$ we will set $\epsilon_{tR}=0$ as well for simplicity. In this limit $M_D$ corresponds to the (degenerate) masses of the heavy fermionic partners of the light ordinary fermions, and is closely related (as shown in Appendix A) to the mass of the heavy partner of the top quark. We will set $M_D$ to the illustrative value $M_D = 400$ GeV in calculations not depending strongly on the precise value, and will otherwise show how results vary with $M_D$. Within the top-pion sector we will set $M_{H_t}$ and $M_{\Pi_t}$ to the illustrative values $M_{H_t} = 250$ GeV and $M_{\Pi_t} = 200$ GeV when the precise value is not critical and will otherwise show how results vary with the values of these masses. Likewise, we will allow $\sin\omega$ to vary to show how various physical quantities depend on it; when the dependence is not critical, we will tend to use the illustrative value $\sin\omega = 0.5$. The remaining parameters, $x$ and the top Yukawa coupling $\lambda_t$, are now derived from the quantities above via: $$\begin{aligned} x &=& \sqrt{2} \, \epsilon_L = \frac{2 \cos\omega M_W}{M_{W^{\prime}}} \nonumber \\ \lambda_t &=& \frac{\sqrt{2} \, m_t}{v \sin\omega} \left[ \frac{M_D^2 (\epsilon_L^2 + 1) - m_t^2}{M_D^2 - m_t^2} \right], \label{eq:lambdat}\end{aligned}$$ where $m_t$ is the physical top quark mass and in the last expression we have set $\epsilon_{tR} = 0$. The relationship between $\epsilon_L$ and $x$ is imposed by ideal delocalization; similarly, as discussed below in subsection IV.A.3, flavor constraints tend to force $\epsilon_{tL} \simeq \epsilon_{L}$, so the value of this last parameter is set as well. ### Tevatron limits from Higgs searches {#sec:Higgs Limits} The Tevatron experiments analyzed the channel $gg \to H \to WW$ and set upper bounds on the cross-section as a function of $M_H$ in Ref. [@Aaltonen:2010sv]. We can adapt this data to our model by appropriately rescaling the couplings involved in the following way: Because of the $\sin\omega$ factor in the denominator of Eq. for $\lambda_t$ above, couplings of $H_t$ to top quarks are enhanced compared to those of the SM Higgs, particularly for small $\sin\omega$. This leads to an enhanced cross section for $H_t$ production in gluon fusion, scaling proportional to $(\lambda_t/\lambda_t^{\rm SM})^2$. Simultaneously, due to the absence of the decay mode $H_t \to b \bar b$ at low $H_t$ masses, the branching ratio for $H_t \to WW$ is larger than for the SM Higgs for masses below about 160 GeV. These two features lead to an enhancement of the predicted rate for $gg \to H_t \to WW$ compared to the corresponding SM process, which is already constrained by Tevatron data. We can now translate the Tevatron bounds on the cross-section [@Aaltonen:2010sv] into constraints on the $\sin\omega$–$M_{H_t}$ parameter space as follows. We compute the cross section for $gg \to H_t \to WW$ according to the approximation $$\begin{aligned} \sigma(gg \to H_t \to WW) &= \sigma^{\rm SM}(gg \to H) \frac{\Gamma(H_t \to gg)}{\Gamma^{\rm SM}(H \to gg)} {\rm BR}(H_t \to WW) \\ \nonumber &=\sigma^{\rm SM}(gg \to H) \frac{BR( H_t \to gg)\Gamma(H_t)}{BR( H_{\textrm{SM}} \to gg)\Gamma(H_{\textrm{SM}})}{\rm BR}(H_t \to WW),\end{aligned}$$ where $\Gamma^{\rm SM}(H \to gg)$ is the SM partial width of $H$ to gluons computed using HDECAY [@HDECAY], $\Gamma(H_t \to gg)$ and BR($H_t \to WW$) are the partial width of $H_t$ to gluons and the branching ratio of $H_t$ to $WW$, respectively, computed using our modified version of HDECAY, and $\sigma^{\rm SM}(gg \to H)$ is the SM Higgs gluon fusion cross section, which we take from Table 2 of Ref. [@Anastasiou:2008tj] for $M_H\leq 200$ GeV and compute using the public code RGHIGGS [@Ahrens:2008qu; @Ahrens:2008nc; @Ahrens:2010rs] for $M_H> 200$ GeV. For each value of $\sin\omega$, a specific range of masses for the top-Higgs is excluded by the Tevatron data. For example, for the illustrative value $\sin\omega = 0.5$, the data implies that the mass range $140\ {\rm GeV} < M_{H_t} < 195\ {\rm GeV}$ is excluded. We present this in Fig. \[fig:limit-sin.pdf\] - as can clearly be seen, the scaling of the Yukawa enhances the production cross-section in our model. As we move to larger $M_{H_t}$ values, $\sigma\, \cdot\, \text{BR}$ declines toward zero as the parton distribution function of the gluon falls rapidly. ![Constraints on $\sigma\cdot BR$ for top-Higgs production at the Tevatron as a function of $M_{H_t}$ and $\sin\omega$. One can clearly see how $\sin\omega < 1$ enhances the production cross-section. For a given value of $\sin\omega$, the range of masses for which the corresponding curve lies above the “Tevatron upper limit" curve is excluded. Each curve for fixed $\sin\omega$ falls off at small $M_{H_t}$ due to a decline in $BR(H_t \to WW)$ and drops off at high $M_{H_t}$ because the falling gluon pdf reduces the production cross-section $\sigma(gg \to H_t)$.[]{data-label="fig:limit-sin.pdf"}](limit-sin.pdf) Turning to the top-pion, we find that there are more important constraints on the $\Pi_t$ masses than those derived from the Tevatron Higgs search limit. These constraints come from limits on rare top decays, from $Zb\bar{b}$ couplings, and from B-physics. We discuss these in turn below. ### Lower bound on the top-pion mass If the charged top-pion $\Pi_t^+$ is lighter than the top quark, it can appear in top decays, $t \to \Pi_t^+ b$. The Tevatron experiments have searched for this process in the context of two-Higgs-doublet models and set upper bounds of about 10–20% on the branching fraction of $t \to H^+ b$, with $H^+$ decaying to $\tau \nu$ or $c \bar s$ [@Aaltonen:2009ke; @Abazov:2009zh] - we can use this to set a lower bound on the top-pion mass. In our model, below the $t \bar b$ threshold $\Pi_t^+$ decays via its mixing with $\pi_0^+$ to lighter SM fermions, with couplings controlled by the fermions’ SM Yukawa couplings. The branching fraction of $\Pi_t^+$ to $\tau\nu$ is therefore about 70%, with the remainder of decays to $c \bar s$. The Tevatron studies can then be applied directly to the top-pion. The relevant limit is BR($t \to \Pi_t^+ b) \lesssim 0.2$ based on D0 data [@Abazov:2009zh]. In our model, the branching fraction of $t \to \Pi_t^+ b$ is controlled by the top-pion mass, the pion mixing angle $\omega$, and the coupling $\lambda_t$: $$\Gamma(t \to \Pi_t^+b) = \frac{G_F m_t}{8 \sqrt{2} \pi} \left[m_t^2 R^2 + \mathcal{O}(m_b^2)\right] \left[1 - \frac{M_{\Pi_t^+}^2}{m_t^2} \right]^2, \label{eqn:gamma_tb}$$ with $$R \equiv \cos\omega \frac{\lambda_t}{\lambda_t^{\rm SM}} = \cot\omega \left[ \frac{M_D^2 (\epsilon_L^2 + 1) - m_t^2}{M_D^2 - m_t^2} \right].$$ To evaluate this expression numerically, we choose the illustrative values: $\sin\omega = 0.5$, $M_D=400$ GeV. Plugging in these values in Eq. leads to the constraint $M_{\Pi_t^+} \gtrsim 150$ GeV from top quark decay[^2]. Because $\Pi_t^0$ and $\Pi_t^+$ are degenerate, this sets a lower bound on both particles’ masses. Having established that $M_{\Pi_t}$ cannot be much lighter than $m_t$, we will assume in the rest of the paper that $M_{\Pi_t} > m_t$, so that decays of $\Pi_t^+ \to t \bar b$ dominate [^3]. ### $\textrm{R}_{\textrm{b}}$ and non-ideal delocalization of the left-handed top quark Next, we consider how data on the $Zb\bar{b}$ coupling constrains the allowed values of $\sin\omega$ and $M_{H_t}$. At tree level, the $Z$ coupling to left-handed bottom quarks is generically modified due to the profile of the $b_L$ wavefunction at the three sites, yielding [@Chivukula:2009ck], $$g_L^{Zbb} = - \frac{e}{s_Wc_W} \left[ \left(1 + \frac{x^2}{4} - \frac{\epsilon_{tL}^2}{2} \right) T_3 - Q s_W^2 \right].$$ The tree-level shift in $R_b$ can be eliminated by imposing ideal delocalization on the left-handed fermions [@Chivukula:2009ck], which means setting: $$\epsilon^2_{tL} = (\epsilon^{ideal}_{tL})^2 \equiv \frac{x^2}{2}~.$$ This has the additional benefit of decoupling the SM fermions from the $W^{\prime}$, $Z^{\prime}$ gauge bosons, eliminating a potentially dangerous source of 4-fermion operators. At one loop, $R_b$ receives additional contributions which we parameterize as $\delta g_L$ according to $$g_L^{Zbb} = - \frac{e}{s_Wc_W} \left[ \left( 1 + \frac{(\epsilon^{ideal}_{tL})^2}{2} - \frac{\epsilon_{tL}^2}{2} \right) T_3 + \delta g_L - Q s_W^2 \right]. \label{eq:glzbb}$$ The one-loop corrections come from: - Loops involving $W$, $W^{\prime}$, SM fermions, and/or heavy vector-like fermions - these were computed for the three-site model [@Chivukula:2006cg] in Ref. [@Abe:2009ni]. - Loops involving the charged top-pion and at least one vector-like heavy fermion. Note that the couplings of $\Pi_t^-$ to one (two) vector-like heavy fermions are suppressed by one (two) powers of $x$ or $\epsilon_{tR}$. - Loops involving the charged top-pion and SM fermions, including contributions from the Goldstone boson eaten by the $W$ in the Top Triangle Moose model, which contains an admixture of the original top-Higgs doublet. These contributions were studied in Ref. [@Burdman:1997pf] for a generic topcolor model based on the calculation of the contribution to $R_b$ in the two-Higgs-doublet model done in Ref. [@Denner:1991ie]. In the Top Triangle Moose model that we consider in this paper, most of the top quark mass comes from the topcolor mechanism and the contribution from $\epsilon_{tR}$ is small, no more than a few GeV. Therefore the contributions to $\delta g_L$ given by the first two sources will be negligible, and the dominant correction comes from the charged top-pion loops. These give a $\lambda_t^2$–enhanced correction to $R_b$ given by $$\delta g_L^{\Pi^-_t} = \frac{m_t^2}{16 \pi^2 v^2} \cot^2 \omega \left[ \frac{R}{R-1} - \frac{R \ln R}{(R-1)^2} \right], \label{eq:dglpi}$$ where $R \equiv m_t^2/M_{\Pi^-_t}^2$; note that $M_Z$ and $m_b$ have been neglected in the loop calculation relative to $m_t$ and $M_{\Pi_t}$. We now consider the size of the dominant new-physics correction $\delta g_L^{\Pi_t^-}$ and compare it to the experimental constraints on $R_b$. We can express the new physics contribution to $R_b$ in terms of $\delta g_L^{\rm new}$ according to [@Oliver:2002up], $$\delta R_b = 2 R_b (1 - R_b) \frac{g_L}{g_L^2 + g_R^2} \delta g_L^{\rm new}, \label{eq:rbdefnn}$$ where $g_L$ and $g_R$ are the SM $Zbb$ couplings at leading order, $$g_L = -\frac{1}{2} + \frac{s_W^2}{3}, \qquad \qquad g_R = \frac{s_W^2}{3}, \qquad \qquad {\rm with} \ \ s_W^2 \simeq 0.23.$$ To leading order we can insert the SM prediction for $R_b$ [@Amsler:2008zzb], $$R_b^{\rm SM} = 0.215 \, 84 \pm 0.000 \, 06\,, \label{eq:rbsmm}$$ in the right-hand side of Eq. (\[eq:rbdefnn\]). This yields the convenient numerical expression, $$\delta R_b \simeq -0.774 \, \delta g_L^{\rm new}. \label{eq:numexprrb}$$ The current experimental value of $R_b$ is [@Amsler:2008zzb], $$R_b^{\rm expt} = 0.216 \, 29 \pm 0.000 \, 66.$$ Subtracting this from the SM prediction Eq. (\[eq:rbsmm\]) gives us a value for the left-hand side of Eq. (\[eq:numexprrb\]), yielding a constraint on the new physics contribution, $$\delta g_L^{\rm new} = (-5.8 \pm 8.6) \times 10^{-4}. \label{eq:deltagLlimits}$$ This, in turn, implies a 2$\sigma$ (3$\sigma$) upper bound on $\delta g_L^{\rm new}$ of $11.4 \times 10^{-4}$ ($20.0 \times 10^{-4}$). Now let us see what we can deduce about constraints on the parameter space of our model. Let us first consider ideal delocalization (i.e., no tree-level contribution to $R_b$ from the distribution of the light fermion wavefunction among the sites). The coefficient in Eq. is numerically, $$\frac{m_t^2}{16 \pi^2 v^2} = 32 \times 10^{-4}.$$ We take $\sin\omega = 0.5$, which yields $\cot^2 \omega = 3$. When $M_{\Pi_t} = m_t$ (i.e., $R=1$), the function of $R$ in square brackets in Eq. is equal to $1/2$. At this parameter point we thus have, $$\delta g_L^{\Pi_t^-} = 48 \times 10^{-4} \qquad \qquad (\cot^2\omega = 3, \ \ M_{\Pi_t} = m_t),$$ which is forbidden. The function of $R$ in square brackets in Eq. falls with increasing $M_{\Pi_t}$. This allows us to put a lower bound on $M_{\Pi_t}$ assuming ideal fermion delocalization and taking $\sin\omega = 0.5$: $$M_{\Pi_t} \gtrsim 760 \ (480) \ {\rm GeV} \qquad {\rm at} \ 2\sigma \ (3\sigma).$$ A lighter top-pion can be allowed if we shift the left-handed third generation quarks away from ideal delocalization. A positive one-loop $\delta g_L^{\rm new}$ can be compensated by choosing a smaller $\epsilon_{tL}$. At our parameter point we have chosen $M_{W^{\prime}} = 500$ GeV, which (with $\sin\omega = 0.5$) yields, $$x = \frac{2 \cos\omega M_W}{M_{W^{\prime}}} \simeq 0.28, \qquad \qquad {\rm or} \qquad\, (\epsilon_{tL}^{ideal})^2\,\simeq 0.039.$$ Returning to Eq. , the combined tree-level and one-loop new physics contribution can be eliminated by choosing $\epsilon_{tL}$ to satisfy $$\frac14 \left(\epsilon^2_{tL} - (\epsilon_{tL}^{ideal})^2 \right) + \delta g_L^{\Pi_t^-} = 0 \,.$$ More generally, if we define $$\Delta \epsilon^2_{tL} = \epsilon^2_{tL} - (\epsilon_{tL}^{ideal})^2$$ then we can deduce from Eq. that the value of $\epsilon_{tL}$ must satisfy $$\frac14 \Delta \epsilon^2_{tL} + \delta g_L^{\Pi_t^-} < 11.4 \times 10^{-4}\ \ (20.0 \times 10^{-4})$$ in order for the predicted value of $R_b$ to agree with experiment at the $2\sigma$ ($3\sigma$) level. Figure \[fig:one\] shows a contour plot of the fractional deviation $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$ from ideal fermion delocalization required in order for top quark delocalization to compensate for top-pion corrections to $R_b$ (meaning agreement at the 90% CL level). Note that for a fractional deviation of order 1, essentially the entire $\sin\omega$ vs $M_{\Pi_t}$ plane is allowed. The illustrative value $M_{W'} = 500$ GeV was used in making this plot; since $\epsilon_{tL}^{ideal} \propto M_{W'}^{-1}$, for heavier $W'$ bosons the contours would retain their shape and label but correspond to a larger value of $\epsilon_{tL}$. ![A contour plot of the deviation $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$ required to compensate for the top-pion contribution to $R_b$, for $M_{W'}=500\, \textrm{GeV}$. The contour boundaries correspond to $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$ equal to 0 to 1.0 in increments of 0.2. As discussed in the text, for larger values of $M_{W'}$, each contour would correspond to a larger value of $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$. Note that, for a deviation $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$ of order 1, essentially the entire plane of values is allowed. \[fig:one\] ](mpi-sin.pdf){width="60.00000%"} Finally, one may worry that changing the value of $\epsilon_{tL}$ from its ideal value might cause problems with flavor changing neutral current constraints. We demonstrate in Appendix D that, in the case of “next-to-minimal" flavor violation [@Agashe:2005hk], these do not rule out compensating for the deviation in $R_b$ resulting from top-pion exchange by modifying the delocalization of the third-generation quarks. Top Higgs production and decay ------------------------------ Having derived the relevant interactions between matter and the top-Higgs/top-pions and understood current constraints on the Top-Triangle moose parameter space, we are ready to move on to phenomenology. In this section and the following, we present the dominant production and decay rates for the top-Higgs and top-pions respectively in a viable region of parameter space. For both the $H_t$ and $\Pi_t$ we consider both direct production $pp \rightarrow H_t,\Pi_t$ and indirect production – top-Higgses/top-pions which arise from the decays of $T$ quarks. ### Decay Branching Ratios The major two-body decay modes of the top-Higgs are the $t\bar{t}$ channel, gauge boson pair modes, and (when kinematically allowed), the $W\Pi_{t}$ and $\Pi_t\Pi_t$ modes. In Fig. \[fig:br-higgs-linear\], we present a plot of the branching ratios of the top-Higgs including only the dominant decay modes for the illustrative set of parameter values: $$\begin{aligned} M_{\Pi_t}&=\,200\, \textrm{GeV},\ \quad M_{D} = \,400\, \textrm{GeV} \nonumber \\ M_{W'}&=\,500\, \textrm{GeV},\ \quad \textrm{sin}\,\omega =\,0.5. \end{aligned}$$ Note that for $M_{H_t}$ below the $WW$ threshold, the top-Higgs tends to decay to $gg$ (through a top loop), or to virtual $W$’s and $Z$’s, as shown in the left-hand pane of Fig. \[fig:br-higgs-linear\] (computed using a modified version of HDECAY [@HDECAY]). ![\[fig:br-higgs-linear\]Branching ratios for the dominant decay modes of the top-Higgs $H_t$. At left, the key shows (from top to bottom) the order of the curves at low $M_{H_t}$. At right, the key shows (from top to bottom) the order of the curves for $M_{H_t}$ greater than the $WW$ threshold. We fixed the following: $M_D$= 400 GeV, $M_{W'}$= 500 GeV, $M_{\Pi_t}$= 200 GeV, and $\sin\omega$= 0.5.](br-higgs-threshold.pdf "fig:"){width="3.39in"} ![\[fig:br-higgs-linear\]Branching ratios for the dominant decay modes of the top-Higgs $H_t$. At left, the key shows (from top to bottom) the order of the curves at low $M_{H_t}$. At right, the key shows (from top to bottom) the order of the curves for $M_{H_t}$ greater than the $WW$ threshold. We fixed the following: $M_D$= 400 GeV, $M_{W'}$= 500 GeV, $M_{\Pi_t}$= 200 GeV, and $\sin\omega$= 0.5.](br-higgs-linear.pdf "fig:"){width="3.39in"} ### Top Higgs production: Direct The direct production of the top-Higgs, $pp \to H_t$ occurs at the LHC via gluon fusion (Fig. \[fig:feynman1\], left) just as for its SM counterpart. Within the Top-Triangle Moose model this process is completely dominated by loops of top quarks; the heavy top contribution is negligibly small. The production cross sections at the LHC are presented in Fig. \[fig:ggH\] for two different center of mass energies, $\sqrt s = 7\ \text{TeV}$ and $\sqrt s = 14\ \text{TeV}$. ![\[fig:ggH\] The production cross-section for $pp\rightarrow H$ at the LHC. The dashed lines indicate the LO production cross sections for a standard model Higgs at $7\ \text{TeV}$ and $14\ \text{TeV}$ at the LHC and are taken from [@HDECAY], while the solid lines are the leading order (LO) top-Higgs production cross sections at the same energy values. The top-Higgs cross sections were calculated using CTEQ5L parton distribution functions, with factorization scale $\mu_F = m_{H_t}/2$ and renormalization scale $\mu_R = m_{H_t}$.[]{data-label="fig:ggh"}](tophiggs_production.pdf){width="4.5in"} As expected, the top-Higgs cross section is significantly larger than that for a standard model Higgs of equivalent mass. The enhancement is roughly a factor of four for our current parameter choice, though the actual value does depend somewhat on the width, and hence the mass, of the top-Higgs. Once the top-Higgs is sufficiently heavy that it can decay into a pair of top pions it becomes considerably wider than its SM equivalent, bringing down the cross section slightly. ### Top Higgs production: Indirect Since the top-Higgs has a non-zero off-diagonal coupling to a light and a heavy top, we could look for it in the decays of the heavy top. To see when this strategy might be useful, we examine the decays of the heavy top, shown in Fig. \[fig:br-T\]. ![\[fig:br-T\]The decay branching ratios of the heavy top for $M_{H_t}$= 250 GeV, $M_{\Pi_t}$= 200 GeV and $M_{W'}$= 500 GeV. The key shows (from top to bottom) the order of the curves in the middle of the plot. We see that the decay modes involving the top higgs and top pion are comparable, while the $Wb$ mode dominates for a wide range of Dirac masses.](br-T.pdf){width="4.5in"} We see that the $Wb$ mode dominates for Dirac masses up to about a TeV. This suggests that one could look at the pair production of the heavy tops, with one of them decaying to $Wb$, and the other decaying to a top-Higgs, i.e., $pp\rightarrow T\bar{T}\rightarrow WbH_tt$, as shown in Fig. \[fig:feynman1\]. This strategy is identical to the indirect Higgs-production mechanism proposed previously in the context of vector-like fermion extensions of the standard model [@delAguila:1989ba; @delAguila:1989rq; @AguilarSaavedra:2006gw; @Kribs:2010ii]. To get an idea for the size of indirect top-Higgs production in the Top-Triangle Moose model, we present the rate for $pp \rightarrow Wb\,H_tt$ at the LHC (14 TeV) in Fig. \[fig:ppHH\] below. In this plot, we have fixed $M_D\, =\, 650$ GeV, $M_{\Pi_t}\,=\, 200$ GeV, and scanned over top-Higgs mass values from $100\ \text{GeV}$ up to[^4] $600\ \text{GeV}$. We will discuss the implications of top-Higgs production and decay modes in more detail in sub-section (\[sec:LHCdisc\]). ![$\sigma\, \cdot\, \text{BR}$ for the process $pp\rightarrow T\bar{T}\rightarrow WbH_tt$ for $M_{\Pi_t}\, =\, 200\ \text{GeV}$, $M_{D}\, =\, 650\ \text{GeV}$, $M_{W'}\, =\, 500\ \text{GeV}$ and $\sin\omega=0.5$. This choice of final state takes advantage of the high BR for $T\rightarrow Wb$. The key shows (from top to bottom) the order of the curves in the middle of the plot. The cross section were calculated using an implementation of the Top-Triangle Moose model in CalcHEP [@Pukhov-Calchep] and assumed a LHC center of mass energy of 14 TeV.[]{data-label="fig:ppHH"}](pp_TT_HtWb.pdf){width="4.5in"} Top Pion production and decay ----------------------------- We now turn to the top-pions. Before discussing their production channels, which are similar to the ones discussed for the top-Higgs, we will first work out the decay branching ratios of the charged and neutral top-pion. ### Decay Branching Ratios The charged (neutral) top pion, when produced, decays to $tb$ ($tt$), $Wh$ ($Zh$), or $Tb$ ($tT$). The decays involving heavy gauge bosons or two heavy fermions are suppressed. We show the plot of branching ratios of the $\Pi_{t}^{-}$ and $\Pi^0_{t}$ in Fig. \[fig:br-pi\] for the following illustrative set of parameter values: $$\begin{aligned} M_{H_t}&=\,250\, \textrm{GeV},\ \quad M_{D} = \,400\, \textrm{GeV} \nonumber \\ M_{W'}&=\,500\, \textrm{GeV},\ \quad \textrm{sin}\,\omega =\,0.5. \end{aligned}$$ ![\[fig:br-pi\]The decay branching ratios of the charged top-pions (left) and the neutral top-pion (right). The key shows (from top to bottom) the order of the curves in the middle of the plot. The dominant decay modes are to $tb, tB, WH_t, ZW', \textrm{and} WZ'$ for the $\Pi^{\pm}_t$, and $tt, tT, WW', \textrm{and} \, ZH_t$ for the $\Pi^0_t$. Below the $t \bar t $ threshold, the $\Pi^0_t$ will decay almost exclusively into a pair of gluons. ](br-piminus.pdf "fig:"){width="3.4in"} ![\[fig:br-pi\]The decay branching ratios of the charged top-pions (left) and the neutral top-pion (right). The key shows (from top to bottom) the order of the curves in the middle of the plot. The dominant decay modes are to $tb, tB, WH_t, ZW', \textrm{and} WZ'$ for the $\Pi^{\pm}_t$, and $tt, tT, WW', \textrm{and} \, ZH_t$ for the $\Pi^0_t$. Below the $t \bar t $ threshold, the $\Pi^0_t$ will decay almost exclusively into a pair of gluons. ](br-pizero.pdf "fig:"){width="3.4in"} As $\Pi_t$ is a pseudoscalar, it cannot decay into longitudinally polarized gauge bosons. With the longitudinal $W/Z$ modes forbidden, the dominant decay mode of $\Pi_t$ below the top-pair threshold is $\Pi_t \rightarrow gg$. Decays to pairs of (transversely polarized) electroweak bosons are present but suppressed by small coupling. Similarly, phase space suppresses three and four-body decay modes like $\Pi_t \rightarrow \bar t t^$. As a result, the neutral top-pion is quite narrow below the top-pair threshold. ### Top pion: Direct, indirect and associated single production The neutral top pion, by analogy with the top-Higgs, can either be produced directly via $gg\rightarrow \Pi_t$, or could show up as a decay product of the heavy top quarks. The production cross section for the first process $gg\rightarrow \Pi_{t}$ is shown in Fig. \[fig:ggP\] for two different LHC energies. As with top-Higgs production, the top-quark loop contribution is dominant. We see that there is a small sharp peak at $M_{\Pi}\sim 350\ \text{GeV}$ - this is due to the effect of the $t\bar{t}$ in the loop going on-shell. In Fig. \[fig:ppPP\], we present $\sigma \cdot$ BR for indirect production, again looking at the case where one of the heavy tops decays to $Wb$ and the other decays to $\Pi_{t}t$. Here, we fixed $M_D\, =\,650$ GeV, and $M_{H_t}\, =\, 250$ GeV. ![\[fig:ggP\] The production cross-section for $pp\rightarrow \Pi_{t}$ at the LHC at 14 TeV (top curve) and 7 TeV (bottom curve). Parton distribution functions and parameters are the same as in Fig. \[fig:ggh\].](toppi_production.pdf){width="4.5in"} ![$\sigma\cdot$BR for the process $pp\rightarrow T\bar{T}\rightarrow \Pi_{t}tWb$. The final state was chosen to take advantage of the high BR for $T\rightarrow Wb$. The key shows (from top to bottom) the order of the curves. This plot was made using the same tools and assumptions as Fig. \[fig:ppHH\].[]{data-label="fig:ppPP"}](pp_TT_PtWb.pdf){width="4.5in"} In addition, the top-pion can also be produced in association with a top-quark - see Fig. \[fig:tP-Feynman\]. We present the cross-section for this process in Fig. \[fig:tP\] as a function of the top-pion mass, summing over $\Pi_{t}^+$ and $\Pi_{t}^-$ production. ![Feynman diagrams for the process $pp\to \Pi^{-}_t t$ at the LHC ($\sqrt s = 14$ TeV).[]{data-label="fig:tP-Feynman"}](tP.pdf) ![The cross-section for the process $pp\to t\Pi_t^{\pm}$ at the LHC ($\sqrt s = 14$ TeV) as a function of the top-pion mass.[]{data-label="fig:tP"}](pp_tP.pdf){width="4.5in"} ### Pair production of $H_t$ and $\Pi_t$ In addition to the processes considered above, one could also look at pair production of two top-pions or production of one top-pion and one top-Higgs at the LHC. The latter occurs via a $W^{}$ exchange, e.g. the process $pp \rightarrow W^{}\rightarrow \Pi_{t}^{\pm}H_t$; see Fig. \[fig:higgs-pion\]. We present the cross-section for this process on the left-hand side in Fig. \[fig:pp-HP\] as a function of the top-pion mass (keeping $M_{H_t}=$250 GeV) and summing over the $\Pi_{t}^+$ and $\Pi_{t}^-$ production. We also show the pair production of a neutral and a charged top-pion in the same plot. We have isolated the pair production of charged top-pions (the right-hand pane in Fig. \[fig:pp-HP\]) - one can see that the cross-section for this process is higher than the rest. This is because of the contribution of additional $t$-channel diagrams involving the top-quark (and its heavy partner) when we include the bottom quark parton distribution function. ![\[fig:higgs-pion\] Feynman diagram for the production of $H_t+\Pi_{t}$](higgs-pion.pdf){width="2.0in"} - the process occurs through an $s$ channel $W^$. ![\[fig:pp-HP\]Left: The production cross-section for the processes $pp \to \Pi_t H_t$ and $pp \to \Pi^{0}_t\Pi^{\pm}_t$ for $M_{H_t}$= 250 GeV. The key shows (from top to bottom) the order of the curves. Right: Cross-section for the pair productions of charged top-pions - this is seen to be signigicantly higher than the values in the plot on the left because of the contribution of additional diagrams involving the top-quark (and its heavy partner) in the $t$ channel once we include bottom quark pdf’s. Both plots were made assuming a $\sqrt s = 14$ TeV LHC.](pp_HP.pdf "fig:"){width="3.5in"} ![\[fig:pp-HP\]Left: The production cross-section for the processes $pp \to \Pi_t H_t$ and $pp \to \Pi^{0}_t\Pi^{\pm}_t$ for $M_{H_t}$= 250 GeV. The key shows (from top to bottom) the order of the curves. Right: Cross-section for the pair productions of charged top-pions - this is seen to be signigicantly higher than the values in the plot on the left because of the contribution of additional diagrams involving the top-quark (and its heavy partner) in the $t$ channel once we include bottom quark pdf’s. Both plots were made assuming a $\sqrt s = 14$ TeV LHC.](pp_PP.pdf "fig:"){width="3.5in"} Discovery prospects at the LHC {#sec:LHCdisc} ------------------------------ Now that we have discussed the production and decay of the top-Higgs, top-pion and the heavy $T$-quark in the model, we survey their discovery prospects at the LHC. We identify channels with clear discovery prospects and estimate their LHC reach. We also point out which channels are promising enough to warrant detailed investigation in future work. Since heavy scalars can be produced indirectly, through the decay of the heavy $T$-quark, we start by commenting on the visibility of this heavy fermion at the LHC.\ *Heavy $T$-quark: The LHC phenomenology of the heavy partners of the first and second generation quarks in this model was already discussed in [@Chivukula:2009ck] - the essential conclusion of that analysis is that, by considering both single and pair productions and subsequently letting the quarks decay to SM gauge bosons, we can discover them at the 5$\sigma$ level for a 14 TeV LHC with $\approx$ 300 $\text{fb}^{-1}$ luminosity for masses up to $\sim$ 1 TeV. Light masses, naturally, require less integrated luminosity. Here, we discuss the prospects for discovering the heavy partner of the top-quark. The $T$ state decays predominantly to $Wb$ for a wide range of $M_D$. Thus, for a wide range of the heavy-T mass, the best possible discovery channel, based on branching fraction considerations, seems to be $pp\to TT \to WbWb$, with the $W$’s decaying to either leptons or quarks. If both $W$’s decay leptonically, we would have two sources of missing energy, and reconstructing the heavy-$T$ mass would be problematic. Hence, the best bet[^5] seems to be $pp\to TT \to WbWb \to l\nu +4j$. But in order to facilitate comparison with [@Chivukula:2009ck], we first consider the process $pp\to \bar TT \to WbWb \to l\nu l\nu jj$, ignoring for the moment the complication arising due to the presence of two neutrinos in the final state. In order to make definitive statements regarding discovery prospects, we would have to calculate the complete SM background. But it is conceivable that once we impose hard $p_T$ cuts on the jets, the SM background reduces to almost zero, as was the case in [@Chivukula:2009ck]. In this case, one could translate the results of that analysis by scaling the couplings. Thus, comparing the process of interest to one that was analyzed, we see that the particular ratio we are after is: $$\frac{pp\to T\bar{T} \to WbW\bar{b} \to l\nu l\nu jj}{pp\to Q\bar{Q} \to WZjj \to l\nu lljj}=\frac{BR(T\to Wb)^2}{BR(Q\to Wj) BR(Q\to Zj)}\frac{BR(W\to l\nu)}{BR(Z \to ll)}.$$ The branching ratios of the heavy quarks depend on the Dirac mass, but we can still make rough estimates. Comparing the branching ratio plot Fig. \[fig:br-T\] to the one for the heavy-$U$ in [@Chivukula:2009ck], we see that the branching ratio to $Wj$ is enhanced for the heavy-$T$, while that to $Zj$ is suppressed by roughly the same amount. Also, $BR(Q\to Wj)\approx 2 BR(Q\to Zj)$, as can be readily verified from Fig. 3 in [@Chivukula:2009ck]. These two facts mean that the first ratio in the above equation is $\geq$2. The second ratio is approximately 3.2 (using the SM values: $BR(W\to l\nu=$0.108, $BR(Z \to ll=$0.033)). Thus, we see that the reach for the heavy-$T$ is roughly enhanced by a factor of 6. But in the analysis for the heavy-$U$ quarks, there is a factor of 4 included (for the heavy partners of the first two generations), and thus in our comparison, we have to divide out by the same factor. This gives an enhancement of 1.5. Considering all this, it is conservative to estimate that the reach for the heavy-$T$ quarks, via pair production at the LHC, would be comparable to that of the heavy-$U$, and that the analysis of the pair production scenario in [@Chivukula:2009ck] applies here. Thus, referring to Fig. 12 in [@Chivukula:2009ck], we conclude that, for a fixed $M_{W'}$= 500 GeV, the heavy-$T$ is discoverable at the LHC with a luminosity of 1 $\text{fb}^{-1}$ for masses up to 450 GeV. This reach is extended to about 650 (850) GeV for 10 (100) $\text{fb}^{-1}$. This indicates that it would be worth doing a thorough analysis of the signal and background for the search for the $T$ states; we plan to present this in forthcoming work.\ Top-Higgs: Much as with the standard model Higgs, the detection prospects of the top-Higgs depend on its mass. Top-Higgses lighter than $\sim 150\ \text{GeV}$ decay dominantly into two gluons and will be impossible to see unless produced in association with a vector boson. Even when produced with a $W/Z$, the immense SM $W/Z + \text{jet}$ backgrounds would make detection difficult, especially for lighter top-Higgses[^6]. Above $160\ \text{GeV}$, top-Higgses produced via gluon fusion are detectable through leptonic $WW/ZZ$ modes. Gluon fusion to top-Higgses is enhanced by $1/\sin^2{\omega} \sim 4 $ over a SM Higgs of equivalent mass, making the discovery prospects excellent. To get some idea of the accessible parameter range we can rescale SM Higgs discovery projections to account for the altered production rate and decay of the top-Higgs. This is most easily done for a 14 TeV collider, where many studies have been done for all Higgs masses (see, for example [@Djouadi:2005gi]). As an example, we can concentrate on top-Higgses heavier than $200\ \text{GeV}$ where the 4-lepton ‘golden’ mode will be dominant. The $h \rightarrow ZZ$ significances found in [@Cranmer:2004ys; @Djouadi:2005gi] are rescaled, then translated into a luminosity required for $S/\sqrt B = 5.0$ at a given top-Higgs mass. This gives us the top-Higgs discovery luminosity curve, which we show in Fig. (\[fig:hzzdisc\]). ![Luminosity (in $\text{fb}^{-1}$) necessary for the discovery the top-Higgs via the fully leptonic mode $pp \rightarrow H_t \rightarrow Z\,Z \rightarrow 4\ell$ at a $14\ \text{TeV}$ LHC.[]{data-label="fig:hzzdisc"}](higgs_ldisc_rescaled){width="3.5in"} Discovery of top-Higgses ligher than $350\ \text{GeV}$ using the leptonic mode alone is possible over a wide range of masses; top-Higgses with mass $ < 350\ \text{GeV}\ (< 400\ \text{GeV})$ would be seen within the first $\text{fb}^{-1}$ ($10\ \text{fb}^{-1}$). Using the leptonic $H_t \rightarrow WW \rightarrow 2\ell\, 2\nu$ mode, we expect similar number for discovery prospects extending down to $m_{H_t} \sim 160\ \text{GeV}$. While the discovery prospects of a $\sim160 - 400\ \text{GeV}$ top-Higgs at a full-powered LHC are excellent, one may ask what the discovery prospects are during the initial, low-energy LHC run. Few phenomenology studies have been carried out for SM Higgs at this lower energy, however Ref. [@Berger:2010nc] has studied the leptonic $WW$ mode for a 7 TeV collider for Higgses lighter than $200\ \text{GeV}$ - however, as can be seen from Fig. \[fig:limit-sin.pdf\], $145\,\textrm{GeV}<M_{H_t}<195\,\textrm{GeV}$ has already been excluded by the Tevatron for $\sin\omega\leq0.5$. A more thorough investigation of light top-Higgses including other modes would be interesting, however from this simple rescaling alone we can say that light top-Higgses are certainly detectable even during the initial LHC run. For heavier top-Higgses, discovery becomes more challenging because the $t \bar t$ mode opens up. In the SM the $t \bar t$ mode is never discussed as a discovery mode since the Higgs branching fraction to $t \bar t$ never gets bigger than 10% and cannot compete with the cleaner leptonic $WW/ZZ$ channels. In contrast, the branching fraction to $t \bar t$ for the top-Higgs can be much bigger than 10% because of the enhanced top - $H_t$ coupling and, consequently, the $WW/ZZ$ branching fractions drop at high $M_{H_t}$ much more sharply than in the SM [@Djouadi:2005gi]. This, unfortunately, has a net negative effect on the discovery potential: $H_t \rightarrow t \bar t$ is unlikely to be a discovery mode due to large backgrounds and the signal size in the cleaner di-boson channels is reduced. A better option for the discovery of heavy $H_t$ is $H_t \rightarrow \Pi_t^{0} Z$. Provided that $M_{\Pi_t} < 2 m_t$, this decay mode yields the final state $\ell\ell jj$, where the jets are quite energetic and can be reconstructed to $M_{\Pi_t}$. Rejecting events with $\gtrsim 2 $ jets or heavy flavor and by exploiting the kinematics of the $\Pi_t \rightarrow jj$ system, it may be possible to suppress SM $Z + \text{jets}$ and $t \bar t$ backgrounds to the point that the top-Higgs is discoverable. While this channel can arise any time there are multiple sectors which break EWS (such as 2HDM), we are not aware of any phenomenological studies. We plan to address this in forthcoming work. Note that the $Z\Pi_{t}^{0}$ mode is only potentially useful for top-Higgses lighter than $2 M_{\Pi_t}$; above $2 M_{\Pi_t}$, top-Higgses will decay primarily to $H_t \rightarrow \Pi_t\,\Pi_t$, where the top-pions can be charged or neutral. In either case, this final state will be extremely challenging to discover [@Chivukula:1990di; @Chivukula:1991zk; @Kilic:2008ub].\ Charged top-pion: The charged top-pion is phenomenologically similar to a charged Higgs boson in a two-Higgs-doublet model with low $\tan\beta$ (i.e., enhanced top quark Yukawa coupling). Discovery prospects for a charged Higgs boson have previously been studied for the 14 TeV LHC in the context of supersymmetric models. The charged top-pion can be produced in association with a top quark through bottom-gluon fusion, $gb \to t \Pi_t^-$, and through gluon-gluon fusion, $gg \to \bar b t \Pi_t^-$. The cross section has been computed to next-to-leading order in QCD [@Plehn:2002vy]; it grows proportional to $\cot^2\omega$ (analogous to $\cot^2\beta$ in the usual two-Higgs-doublet notation). Due to the popularity of supersymmetric models, studies of $tH^-$, $H^- \to \bar t b$ at ATLAS [@Aad:2009wy] and CMS [@CMSnote] have focused entirely on the large $\tan\beta$ regime. The major background comes from $t \bar t$ plus jets; the systematic uncertainty from the background normalization presents the biggest challenge to this search. The CMS study [@CMSnote] gives values of $\sigma(pp \to t H^{\pm}) \times {\rm BR}(H^{\pm} \to tb)$ required for $5\sigma$ discovery in this channel as a function of the mass of the pseudoscalar $A^0$ in the MSSM, which is nearly degenerate with $H^-$ in that model - the sensitivity depends very strongly on the systematic uncertainty on the $t\bar{t}$ background. We present a plot of the cross-section for the process $pp\to t\Pi^{-}_t$ in Fig. \[fig:tP\]. For $M_{\Pi_t}\leq600\,\textrm{GeV}$, the charged top-pion decays to $tb$ 100% of the time (see Fig. \[fig:br-pi\]), and so we can make a direct comparison between Fig. \[fig:tP\] and the CMS study. Doing so, we find that one can discover a charged top-pion at the 5$\sigma$ level at a luminosity of 30 $\text{fb}^{-1}$ for $M_{\Pi_t}\leq450\,\textrm{GeV}$, assuming a 0% uncertainty on the $t\bar{t}$ background[^7]. For 1% (3%) systematic uncertainty, the reach goes down to 350 GeV (250 GeV)[^8]. Even when decays to other final states ($WH_{t}$, $tB$, $ZW^{\prime}$, $Z^{\prime} W$) are kinematically accessible, the branching fraction to $\bar t b$ remains high - see Fig. \[fig:br-pi\]. Studies of this channel done in the context of the MSSM can thus still be applied, with the caveat that angular correlations among the final-state particles in the event may be different. In the MSSM at large $\tan\beta$, $H^-$ decays to $\bar t_L b_R$ through the bottom Yukawa coupling. For the top-pion, however, $\Pi_t^- \to \bar t_R b_L$ through the top Yukawa coupling. This difference may affect details of the experimental acceptance for the signal. However, it is reasonable to conclude that this channel is quite promising for $\Pi_t$ discovery and warrants further study. Finally we note that ATLAS [@Aad:2009wy] combines this $tH^-$, $H^- \to \bar t b$ channel with $tH^-$, $H^- \to \tau \nu$ in the MSSM to present combined discovery reach contours at large $\tan\beta$, but the $H^- \to \bar t b$ contribution improves the reach only marginally [@Aad:2009wy]. We emphasize that while BR($H^- \to \tau \nu) \simeq$ 10% above the $tb$ threshold in the MSSM at large $\tan\beta$, for the charged top-pion this decay mode is absent. Discussion and conclusions {#sec:conclusions} ========================== This paper has explored the collider physics of the new heavy fermionic top-quark partner ($T$), the top-Higgs boson ($H_t$) and the top-pion ($\Pi_t$) states in the deconstructed topcolor-assisted technicolor theory known as the Top Triangle Moose. After establishing the spectrum and the couplings of the new states to each other and to standard model particles in Section III, we turned to phenomenology. We showed in Section IV.A how existing Fermilab Tevatron data constrains $M_{H_t}$ and $M_{\Pi_t}$ as a function of the mixing angle $\sin\omega$ between the linear and nonlinear sigma model symmetry-breaking sectors: $\Pi_t$ lighter than 150 GeV would likely have been seen already in $t \to \Pi_t b$ while $H_t$ in the 150-200 GeV range would likely have been visible in $WW$ decays. We also established that the presence of relative light $\Pi_t$ does not present insurmountable challenges related to third-generation flavor physics, as one might have feared. In particular, allowing the delocalization of the left-handed top quark to deviate from the value suggested by ideal delocalization can cancel contributions from one-loop diagrams involving $\Pi_t$ exchange that would otherwise have shifted $R_b$ from agreement with experiment. As shown in Figure 3, nearly the full $\sin\omega$ vs. $M_{\Pi_t}$ parameter space can be accommodated in this way. Moreover, limits from third-generation FCNCs are consistent with this finding, as shown in Appendix D. In Sections IV.B and IV.C, we laid the foundation for studies of LHC phenomenology by calculating the decay branching ratios of $H_t$ and $\Pi_t$, as well as their production cross-sections, for a variety of key processes. This information allowed us to determine which channels are most promising for discovery of $T$, $H_t$ and $\Pi_t$. Adapting previous work on the heavy partners $Q$ of the first and second generation quarks enabled us to demonstrate that the $T$ should be visible at the LHC for $M_T \leq 900$ GeV in the pair-production channel $pp \to T \bar{T} \to Wb Wb \to \ell\nu\ \ell\nu\ jj$. Hence, a full study of the detailed background processes and optimal cuts for this process is indicated (and is now underway). The alternative channel in which one $W$ decays hadronically, so that the final state is $\ell\nu\ 4j$, should offer a larger signal along with the welcome possibility for full reconstruction of the top quark’s heavy partner; we are also planning to study this channel. In the case of a moderately light $H_t$, we found that the situation resembles that of the standard model Higgs. For $M_{H_t} \leq 160$ GeV, the top-Higgs will be invisible because it decays almost exclusively to dijets, for which the background is overwhelming. For $160\ {\rm GeV} \leq M_{H_t} \leq 400$ GeV, the top-Higgs should actually be easier to find than the standard model Higgs, because the “golden" all-leptonic decay modes open up and the signal rate is enhanced by a factor of $1 / \sin\omega$. In fact, for $H_t$ in the lighter end of this mass range, discovery in the first $1 \text{fb}^{-1}$ of LHC data would be possible. Top-Higgs bosons heavier than 400 GeV will be more challenging to find since $BR(H_t \to WW)$ will be below even the already-reduced diboson branching ratio for the standard model Higgs. The most promising decay channel for top-Higgs discovery in the window $2 m_t \leq M_{H_t} \leq 2 M_{\Pi_t}$ would be $H_t \to W \Pi_t$ and we plan to study this in detail in forthcoming work. Once $M_{H_t} > 2 M_{\Pi_t}$, the primary decay mode is $H_t \to \Pi_t \Pi_t \to 4 g$ and the large multijet background will make discovery difficult (though the methods advocated in [@Chivukula:1990di; @Chivukula:1991zk; @Kilic:2008ub] can be of help). Adapting existing work on charged-Higgs search protocols suggests that $\Pi_t^\pm$ with masses below 400 GeV should be visible in 30 $\text{fb}^{-1}$ of LHC data through the process $pp \to t \Pi_t^\pm \to ttb$. Further studies of final state particle angular correlations and the dependence on $\sin\omega$ are needed. In particular, most studies of charged-Higgs searches have focused on the case of large mixing angle (essentially, large $\tan\beta$) whereas the case of small $\sin\omega$ is of greatest interest in the Top Triangle Moose. Finally, it is interesting to reflect on how one would know that the new states one had discovered, whether $T$, $H_t$ or $\Pi_t$, were those of the Top Triangle Moose rather than some other model. The answer will surely lie in the overall pattern of observable relationships among these three states. Consider, for instance, a top-Higgs boson of moderate mass. One would first find this state in single production followed by diboson decays, $pp \to H_t \to WW$; the fact that the signal rate noticeably exceeded the standard model prediction would show that one had found an exotic rather than a standard model Higgs state. As the LHC integrated luminosity grew, the $T$ state would eventually be found in $pp \to T\bar{T} \to WbWb$ channels. Once the existence of that state is confirmed, it would be possible to measure the rarer $T \to H_t t$ decay path and confirm that the $H_t$ found in $T$ decays is the same particle that one had already discovered in $H_t \to WW$. This would show that the $H_t$ was both part of the electroweak sector (as witnessed by its diboson coupling) and strongly coupled to the top quark sector. In the case of top-pions, one might begin by establishing their presence in associated production with top quarks; this would help show that they were strongly coupled to the top sector, which a measurement of $T \to \Pi_t t$ could also confirm. Then finding either joint production of $H_t$ and $\Pi_t$ through an off-shell $W$ boson ($pp \to W^ \to H_t \Pi_t$) or one of the decay paths $H_t \to \Pi_t W$ or $\Pi_t \to H_t W$ would demonstrate the relationship of $\Pi_t$ to the electroweak sector, including the top-Higgs. As the LHC data set grows, it will be interesting to watch for signs of these new states, heralding the presence of new strong dynamics in the top quark sector. BC and HEL were supported, in part, by the Natural Sciences and Engineering Research Council of Canada. RSC and EHS were supported, in part, by the US National Science Foundation under grant PHY-0854889. AM is supported by Fermilab operated by Fermi Research Alliance, LLC under contract number DE-AC02-07CH11359 with the US Department of Energy. Masses and Eigenstates ====================== Gauge Bosons ------------ The neutral gauge boson mass matrix is given by: $$\begin{aligned} M_{Z}^{2}=\frac{e^{2}\, v^{2}}{4\, x^{2}\,\textrm{sin}^{2}\,\theta}\left(\begin{array}{ccc} \frac{x^{2}}{1-x^{2}}(1+\textrm{cos}^{2}\,\omega) & -\frac{2 x}{\sqrt{1-x^{2}}}\textrm{cos}^{2}\,\omega & -\frac{x^{2}}{\sqrt{1-x^{2}}}\textrm{sin}^{2}\omega\,\textrm{tan}\,\theta \\ -\frac{2 x}{\sqrt{1-x^{2}}}\textrm{cos}^{2}\,\omega & 4\,\textrm{cos}^{2}\omega & -2\, x\textrm{\, cos}^{2}\,\omega\textrm{\, tan}\,\theta\\ -\frac{x^{2}}{\sqrt{1-x^{2}}}\textrm{sin}^{2}\omega\,\textrm{tan}\,\theta & -2\, x\,\textrm{cos}^{2}\,\omega\,\textrm{tan}\,\theta & x^{2}(1+\textrm{cos}^{2}\,\omega)\textrm{tan}^{2}\,\theta\end{array}\right). \label{eqn:Z mass matrix}\end{aligned}$$ Diagonalizing perturbatively in the small parameter $x$ yields the following masses for the $Z$ and the $Z'$ [@Chivukula:2009ck]: $$\begin{aligned} M_{Z}^{2}&=\frac{e^{2}\, v^{2}}{4\,\textrm{sin}^{2}\,\theta\,\textrm{ cos}^{2}\,\theta}\left(1+x^{2}\left(1-\frac{\textrm{sec}^{2}\,\theta}{4}\right)\right) \label{Z mass} \\ M_{Z'}^{2}&=\frac{e^{2}\, v^{2}\,\textrm{cos}^{2}\,\omega}{4\,\textrm{sin}^{2}\,\theta\, x^{2}}\left(4+x^{2}\textrm{sec}^{2}\,\theta\right), \label{Z' mass}\end{aligned}$$ while the photon remains massless. The eigenvector of the $Z$ is given by: $$Z^{\mu} = v_{z}^{0}W_{0}^{\mu}+v_{z}^{1}W_{1}^{\mu}+v_{z}^{2}B^{\mu}, \label{Z eigenvector}$$ where $$v_{z}^{0} = \frac{1}{8}\,(4(-2+x^{2})\textrm{cos\,}\theta-3\, x^{2}\textrm{sec}\,\theta), \,\,\,\, v_{z}^{1} = \frac{1}{2}\, x(-2\textrm{cos}^{2}\,\theta+1)\textrm{sec}\,\theta, \,\, \,\,\,\, v_{z}^{2} = \textrm{sin}\,\theta-\frac{1}{2}x^{2}\,\textrm{sec}\,\theta\,\textrm{tan}\,\theta. \nonumber$$ The eigenvector of $Z'$ is the orthogonal combination. The charged gauge boson mass matrix is the upper $2\times 2$ block of Eq. . The masses of the physical gauge bosons are given by: $$\begin{aligned} M_{W}^{2}&=\frac{e^{2}v^{2}}{4\,\textrm{sin}^{2}\,\theta}\left(1+\frac{3x^{2}}{4}\right) \\ M_{W'}^{2}&=\frac{e^{2}\, v^{2}\textrm{cos}^{2}\,\omega}{4\,\textrm{sin}^{2}\,\theta\, x^{2}}\left(4+x^{2}\right),\end{aligned}$$ with the respective eigenvectors: $$\begin{aligned} W^{\mu} & =\left(1-\frac{x^{2}}{8}\right)W_{0}^{\mu}+\frac{1}{2}xW_{1}^{\mu}\ \textrm{and} \\ W'^{\mu} & =-\frac{1}{2}xW_{0}^{\mu}+\left(1-\frac{x^{2}}{8}\right)W_{1}^{\mu}. \end{aligned}$$ We are now in a position to define the weak mixing angle, $1 - \sin^2\theta_W \equiv M_W^2/M_Z^2$. Including corrections up to $\mathcal{O}(x^2)$, we obtain, $$\sin\theta_W = \left( 1 - \frac{x^2}{8} \right) \sin\theta.$$ Fermions -------- The light fermion mass matrix is derived from the Lagrangian: $$\begin{aligned} \mathcal{L} & = & M_{D}\left[\epsilon_{L}\bar{\psi}_{L0}\Sigma_{01}\psi_{R1}+\bar{\psi}_{R1}\psi_{L1}+\bar{\psi}_{L1}\Sigma_{12}\left(\begin{array}{cc} \epsilon_{uR} & 0\\ 0 & \epsilon_{dR}\end{array}\right)\left(\begin{array}{c} u_{R2}\\ d_{R2}\end{array}\right)\right],\end{aligned}$$ and is given by: $$\begin{aligned} M_{u,d}=M_{D}\left(\begin{array}{cc} \epsilon_{L} & 0\\ 1 & \epsilon_{uR,dR}\end{array}\right). \label{fermion mass matrix}\end{aligned}$$ This can be diagonalized in the small parameters $\epsilon_L$ and $\epsilon_{fR}$ to yield the masses of the light fermion and its heavy Dirac partner: $$\begin{aligned} m_{f}&=\frac{M_{D}\epsilon_{L}\epsilon_{fR}}{\sqrt{1+\epsilon_{fR}^{2}}}\left[1-\frac{\epsilon_{L}^{2}}{2(1+\epsilon_{fR}^{2})}+...\right] \label{eqn:light quark mass}\\ m_{F}&=M_{D}\sqrt{1+\epsilon_{fR}^{2}}\left[1+\frac{\epsilon_{L}^{2}}{2(1+\epsilon_{fR}^{2})^{2}}+....\right].\end{aligned}$$ The left- and right-handed eigenstates of the light fermion can be derived to be: $$\begin{aligned} u_{L} & = \left(-1+\frac{\epsilon_{L}^{2}}{2(1+\epsilon_{uR}^{2})^{2}}\right)\psi_{L0}+\left(\frac{\epsilon_{L}}{1+\epsilon_{uR}^{2}}\right)\psi_{L1}, \\ u_{R} & = \left(-\frac{\epsilon_{uR}}{\sqrt{1+\epsilon_{uR}^{2}}}+\frac{\epsilon_{L}^{2}\epsilon_{uR}}{(1+\epsilon_{uR}^{2})^{5/2}}\right)\psi_{R1}+\left(\frac{1}{\sqrt{1+\epsilon_{uR}^{2}}}+\frac{\epsilon_{L}^{2}\epsilon_{uR}^{2}}{(1+\epsilon_{uR}^{2})^{5/2}}\right)u_{R2}. \label{uR vector}\end{aligned}$$ The eigenvector of the left- and right-handed heavy quark are the orthogonal combinations. For the top, the mass term is dominated by the top-Higgs contribution. The mass matrix is given by: $$\begin{aligned} M_{t}=M_{D}\left(\begin{array}{cc} \epsilon_{tL} & a\\ 1 & \epsilon_{tR} \end{array}\right), \label{eqn:top mass matrix}\end{aligned}$$ where the parameter $a$ is defined as $a \equiv v \sin\omega/\sqrt{2}\,M_D$. Diagonalizing Eq. perturbatively in $\epsilon_{tL}$ and $\epsilon_{tR}$, we get the mass of the SM top-quark: $$m_{t}= \lambda_{t}v\,\textrm{sin}\,\omega\left[1+\frac{\epsilon_{tL}^{2}+\epsilon_{tR}^{2}+\frac{2}{a}\epsilon_{tL}\epsilon_{tR}}{2(-1+a^{2})}\right]. \label{eqn:top mass}$$ Thus, we see that $m_{t}$ depends only slightly on $\epsilon_{tR}$, in contrast to the light fermion mass, Eq. , where the dominant term is $\epsilon_{fR}$ dependent. The mass of the heavy partner of the top is given by: $$m_{T}= M_{D}\left[1-\frac{\epsilon_{tL}^{2}+\epsilon_{tR}^{2}+2a\epsilon_{tL}\epsilon_{tR}}{2(-1+a^{2})}\right]. \label{eqn:mass of TOP}$$ The left- and right-handed eigenvectors of the SM top are given by: $$\begin{aligned} t_{L} & = \left(1-\frac{\epsilon_{tL}^{2}+a^{2}\epsilon_{tR}^{2}+2a\epsilon_{tL}\epsilon_{tR}}{2(-1+a^{2})^{2}}\right)\psi_{L0}^{t}+\left(\frac{\epsilon_{tL}+a\epsilon_{tR}}{-1+a^{2}}\right)\psi_{L1}^{t}\ \textrm{and}\\ t_{R} & =\left(1-\frac{a^{2}\epsilon_{tL}^{2}+\epsilon_{tR}^{2}+2a\epsilon_{tL}\epsilon_{tR}}{2(-1+a^{2})^{2}}\right)\psi_{R1}^{t}+\left(\frac{a\epsilon_{tL}+\epsilon_{tR}}{-1+a^{2}}\right)t_{R2}.\end{aligned}$$ The Lagrangian ============== In order to derive the terms in the Lagrangian describing the interaction of the top-Higgs and the top pions with the gauge bosons, we start by plugging Eq. in Eq. , and writing the covariant derivative of $\Sigma_{01}$ as $$D_{\mu}\Sigma_{01}=\frac{i}{F}\partial_{\mu}\pi_0 +igW_{0\mu}-\frac{g}{F}W_{0\mu}\pi_{0}-i\tilde{g}W_{1\mu}-\frac{\tilde{g}}{F}\pi_{0} W_{1\mu},$$ where we have denoted $\pi_{0}=\pi_{0}^{a}\sigma^{a}$. The product can be evaluated to be: $$\begin{aligned} (D_{\mu}\Sigma_{01})^{\dagger}(D_{\mu}\Sigma_{01} )&=\frac{1}{F^2}\left(\partial_{\mu}\pi_{0}\right)^2+\left[g^2 W_{0\mu}^{2}+\tilde{g}^2 W_{1\mu}^2-g\tilde{g}\, W_{0}^{\mu}W_{1\mu}-g\tilde{g}\, W_{1}^{\mu}W_{0\mu}\right] \nonumber\\ &+\frac{1}{F}\left(\partial^{\mu}\pi_{0}\right)\left[gW_{0\mu}-\tilde{g}W_{1\mu}\right]+\left[\frac{g}{F}W^{0\mu}\left(\partial_{\mu}\pi_{0}\right)-\frac{\tilde{g}}{F}W^{1\mu}\left(\partial_{\mu}\pi_{0}\right)\right] \nonumber \\ &-\frac{i}{F^2}\left(\partial^{\mu}\pi_{0}\right)\left[-gW_{0\mu}\pi_{0}+\tilde{g}\pi_{0}W_{1\mu}\right]-\left[\frac{ig}{F^2}\pi_{0}W^{\mu}_{0} \left(\partial_{\mu}\pi_{0}\right)-\frac{i\tilde{g}}{F^2}W^{\mu}_{1}\pi_{0} \left(\partial_{\mu}\pi_{0}\right)\right] \nonumber \\ &+\frac{i}{F}\left[-g W^{\mu}_{0}+\tilde{g}W^{\mu}_{1}\right]\left[-g W_{0\mu}\pi_{0}+\tilde{g}\pi_{0}W_{1\mu}\right]-\frac{1}{F}\left[g\pi_{0}W^{\mu}_{0}-\tilde{g}W^{\mu}_{1}\pi_{0}\right]\left[igW_{0\mu}-i\tilde{g}W_{1\mu}\right] \nonumber \\ &+\frac{1}{F^2}\left[-g\pi_{0}W^{\mu}_{0}+\tilde{g}W^{\mu}_{1}\pi_{0}\right] \left[-gW_{0\mu}\pi_{0}+\tilde{g}\pi_{0}W_{1\mu}\right]. \label{eqn:sigma0}\end{aligned}$$ The first line gives the kinetic energy term for the pions and the gauge bosons masses. The second line gives the mixing between the gauge and the Goldstone bosons. The third and fourth lines give the $\pi\pi V$ and $\pi V V$ couplings respectively, while the last line gives the four point coupling, $\pi \pi V V$. Plugging in the matrix definitions of the fields and taking the trace, we get, $$\addtolength{\fboxsep}{10pt} \boxed{ \begin{split} &\mathcal{L}^{(2)}_{\pi KE}=\frac{1}{2}\left(\partial_{\mu}\pi_{1}^{a}\right)^2\\ &\mathcal{L}^{(2)}_{\textrm{mixing}}=\frac{F}{2}\left[\tilde{g}W_{1\mu}^{a}(\partial_{\mu}\pi_{1}^{a})-g'B_{2\mu}(\partial_{\mu}\pi_{1}^{3})\right]\\ &\mathcal{L}^{(2)}_{\pi \pi V}=-\frac{\tilde{g}}{2}\epsilon_{abc}(\partial^{\mu}\pi_{1}^{a})W_{1\mu}^{b}\pi_{1}^{c}+\frac{g'}{2}\epsilon_{ab3}(\partial^{\mu}\pi_{1}^{a})\pi_{1}^{b}B_{2\mu}\\ &\mathcal{L}^{(2)}_{\pi VV}=\frac{g'\tilde{g}F}{2}\epsilon_{ab3}W_{1}^{a\mu}B_{2 \mu}\pi_{1}^{b}\\ &\mathcal{L}^{(2)}_{\pi \pi VV}=\mathcal{M}_{abcd}\tilde{g}^{2}\pi_{1}^{a}W_{1}^{b\mu}W_{1\mu}^{c}\pi_{1}^{d}-2\tilde{g}g'\mathcal{M}_{abc3}\pi_{1}^{a}W_{1}^{b\mu}\pi_{1}^{c}B_{2\mu}+\frac{g'^{2}}{8}B_{2\mu}^2(\pi_{1}^{a})^{2}\\ \end{split} }$$ where $\mathcal{M}_{abcd}=\frac{1}{8}\left(\delta_{ab}\delta_{cd}-\delta_{ac}\delta_{bd}+\delta_{ad}\delta_{bc}\right)$. The corresponding terms from the kinetic term of the other nonlinear sigma model field can be read off by relabeling the fields and couplings as follows: $$gW_{0\mu}\rightarrow \tilde{g}W_{1\mu};\, \tilde{g}W_{1\mu}\rightarrow g'B_{2\mu};\, \pi_{0}\rightarrow \pi_{1}.$$ We summarize the results for the sake of completeness: $$\addtolength{\fboxsep}{10pt} \boxed{ \begin{split} &\mathcal{L}^{(2)}_{\pi KE}=\frac{1}{2}\left(\partial_{\mu}\pi_{1}^{a}\right)^2\\ &\mathcal{L}^{(2)}_{\textrm{mixing}}=\frac{F}{2}\left[\tilde{g}W_{1\mu}^{a}(\partial_{\mu}\pi_{1}^{a})-g'B_{2\mu}(\partial_{\mu}\pi_{1}^{3})\right]\\ &\mathcal{L}^{(2)}_{\pi \pi V}=-\frac{\tilde{g}}{2}\epsilon_{abc}(\partial^{\mu}\pi_{1}^{a})W_{1\mu}^{b}\pi_{1}^{c}+\frac{g'}{2}\epsilon_{ab3}(\partial^{\mu}\pi_{1}^{a})\pi_{1}^{b}B_{2\mu}\\ &\mathcal{L}^{(2)}_{\pi VV}=\frac{g'\tilde{g}F}{2}\epsilon_{ab3}W_{1}^{a\mu}B_{2 \mu}\pi_{1}^{b}\\ &\mathcal{L}^{(2)}_{\pi \pi VV}=\mathcal{M}_{abcd}\tilde{g}^{2}\pi_{1}^{a}W_{1}^{b\mu}W_{1\mu}^{c}\pi_{1}^{d}-2\tilde{g}g'\mathcal{M}_{abc3}\pi_{1}^{a}W_{1}^{b\mu}\pi_{1}^{c}B_{2\mu}+\frac{g'^{2}}{8}B_{2\mu}^2(\pi_{1}^{a})^{2}\\ \end{split} }$$ where $M_{abcd}$ is given as before. Turning to the kinetic energy term of $\Phi$, we see that its covariant derivative $$D_{\mu}\Phi=\partial _{\mu}\Phi+igW_{0\mu}^1 \Phi-\frac{ig'}{2}B_{2\mu}\Phi$$ can be expanded by plugging in Eq. : $$\Phi= \left( \begin{array}{c} \frac{1}{\sqrt{2}}(\partial_{\mu}H+i\partial_{\mu}\pi_{t}^{0}) \\ i\partial_{\mu}\pi_{t}^{-} \end{array} \right)+ \frac{ig}{2} \left( \begin{array}{c} \frac{W_{0\mu}^{3}}{\sqrt{2}}(f+H+i\pi_{t}^{0})+\sqrt{2}iW_{0\mu}^{+}\pi_{t}^{-} \\ W_{0}^{-}(f+H+i\pi_{t}^{0})-iW_{0\mu}\pi_{t}^{-} \end{array} \right)-\frac{ig'}{\sqrt{2}} \left( \begin{array}{c} \frac{B_{2\mu}}{\sqrt{2}}(f+H+i\pi_{t}^{0}) \\ iB_{2\mu}\pi_{t}^{-} \end{array} \right).$$ In order to make the expressions more compact, we will introduce the following notation: $$\begin{aligned} Z_{\mu}&=gW_{3\mu}-g'B_{2\mu}, \\ A_{\mu}&=gW_{3\mu}+g'B_{2\mu}.\end{aligned}$$ The $Z$ and $A$ appearing in the above formulas are convenient aids to make the expressions look simple, and are not the physical $Z_{\mu}$ and $A_{\mu}$. Using this, the product can be evaluated to be: $$\begin{aligned} D_{\mu}\Phi^{\dagger}D_{\mu}\Phi=&\frac{1}{2}(\partial_{\mu}H)^{2}+\frac{1}{2}(\partial_{\mu}\pi_{t}^{0})^{2}+(\partial^{\mu}\pi_{t}^{+})(\partial_{\mu}\pi_{t}^{-}) +\frac{Z_{\mu}}{2}\left[(f+H)(\partial_{\mu}\pi_{t}^{0})-\pi_{t}^{0}\partial_{\mu}H\right] \nonumber \\ &-\frac{g}{2}(\partial^{\mu}H) (W_{0\mu}^{+}\pi_{t}^{-}+W_{0\mu}^{-}\pi_{t}^{+})+\frac{ig}{2}(\partial^{\mu}\pi_{t}^{0})( W_{0\mu}^{+}\pi_{t}^{-}-W_{0\mu}^{-}\pi_{t}^{+}) +\frac{(Z_{\mu})^2}{8}\left[(f+H)^2+(\pi_{t}^0)^2\right] \nonumber \\ &+\frac{g}{2}\left[(f+H)[W_{0}^{-\mu}(\partial_{\mu}\pi_{t}^{+}) + W_{0}^{+\mu}(\partial_{\mu}\pi_{t}^{-})\right]+\frac{ig}{2}\pi_{t}^{0}\left[W_{0}^{-\mu}(\partial_{\mu}\pi_{t}^{+}) - W_{0}^{+\mu}(\partial_{\mu}\pi_{t}^{-})]\right] \nonumber \\ &+\frac{ig}{4}Z_{\mu}\left[(f+H)(W_{0}^{+\mu}\pi_{t}^{-}-W_{0}^{-\mu}\pi_{t}^{+})-i\pi_{t}^{0}(W_{0}^{+\mu}\pi_{t}^{-}+ W_{0}^{-\mu}\pi_{t}^{+})\right] \nonumber \\ &-\frac{iA_{\mu}}{2}\left[(\partial^{\mu}\pi_{t}^{+})\pi_{t}^{-}-(\partial^{\mu}\pi_{t}^{-})\pi_{t}^{+}\right]+\frac{g^2}{4}W^{-\mu}_{0}W_{0\mu}^{+}\left[(f+H)^2+(\pi_{t}^0)^2\right] \nonumber \\ &-\frac{ig}{4}A^{\mu}\left[(f+H)(W_{0\mu}^{+}\pi_{t}^{-}-W_{0\mu}^{-}\pi_{t}^{+})-i\pi_{t}^{0}(W_{0\mu}^{+}\pi_{t}^{-}+W_{0\mu}^{-}\pi_{t}^{+})\right] \nonumber \\ &+\frac{1}{2}\pi_{t}^{+}\pi_{t}^{-}(A_{\mu})^{2}+\frac{1}{2}\pi_{t}^{+}\pi_{t}^{-}W_{0}^{+\mu}W_{0\mu}^{-}, \label{eqn:phi lagrangian}\end{aligned}$$ where $W^{\pm}=(W^{1}\mp iW^{2})/\sqrt{2}$, and similarly for the $\pi_{t}^{\pm}$. Eq. gives us the coupling of the top-Higgs and the pions to the gauge bosons, and the gauge-Goldstone mixing terms. Let us pick the latter contribution to the Lagrangian. $$\mathcal{L}^{(3)}_{\textrm{mixing}} =\frac{gf}{2}\left[W_{0}^{-\mu}(\partial_{\mu}\pi_{t}^{+}) + W_{0}^{+\mu}(\partial_{\mu}\pi_{t}^{-}) \right]+\frac{f}{2}Z_{\mu}(\partial_{\mu}\pi_{t}^{0}).$$ Plugging in the definitions of the fields, this becomes: $$\mathcal{L}^{(3)}_{\textrm{mixing}}=\frac{f}{2}\left[g(\partial_{\mu}\pi_{t}^{a})W_{0}^{a\mu}-g'(\partial_{\mu}\pi_{t}^{3})B_{2}^{\mu}\right].$$ Four point couplings ==================== We present the four point couplings involving two gauge bosons and top-pions/top-Higgs in Table \[tab:H-couplings4\]. Vertex Strength -------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $H_tH_tWW$ $\frac{g_{0}^{2}}{4}\left(1+\frac{3\,x^2}{4}\right)$ $H_tH_tW'W$ $-\frac{g_{0}^{2}x}{8}$ $H_tH_tW'W'$ $\frac{g_{0}^{2}x^2}{16}$ $H_tH_tZZ$ $\frac{e^{2}}{2}\left(\textrm{cosec}^{2}2\theta+\frac{x^2}{16}\left[1+2\,\textrm{cos}\,2\theta \right] \textrm{cosec}^{2}\theta\,\textrm{sec}^{4}\theta \right)$ $H_tH_tZ'Z$ $-\frac{g_{0}^{2}x}{8}\textrm{sec}^{3}\theta\,\textrm{cos}\,2\theta$ $H_tH_tZ'Z'$ $\frac{g_{0}^{2}x^2}{32}\textrm{sec}^{4}\theta\,\textrm{cos}^{2} 2\theta$ $H_tZW^{-}\Pi_{t}^{+}$ $-\frac{i g_{0}^{2}}{2}\,\textrm{cos}\,\omega\,\textrm{tan}\,\theta\left(\textrm{sin}\,\theta+\frac{x^2}{16}\left[1+3\,\textrm{cos}\,2\theta \right]\textrm{sec}\,\theta\,\textrm{tan}\,\theta \right)$ $H_tZW^{'-}\Pi_{t}^{+}$ $\frac{i g_{0}^{2}}{4}x\,\textrm{cos}\,\omega\,\textrm{sin}\,\theta\,\textrm{tan}\,\theta$ $H_tZ'W^{-}\Pi_{t}^{+}$ $-\frac{i g_{0}^{2}}{4}x\,\textrm{cos}\,\omega \,\textrm{tan}^{2}\theta$ $H_tZ'W^{'-}\Pi_{t}^{+}$ $\frac{i g_{0}^{2}}{8}x^2\,\textrm{cos}\,\omega \,\textrm{tan}^{2}\theta$ $H_tAW^{-}\Pi_{t}^{+}$ $-\frac{i g_{0}^{2}}{2}\,\textrm{cos}\,\omega \,\textrm{sin}\,\theta\left(1+\frac{3\,x^2}{8} \right)$ $H_tAW^{'-}\Pi_{t}^{+}$ $\frac{i g_{0}^{2}}{4}x\,\textrm{cos}\,\omega \,\textrm{sin}\,\theta$ : \[tab:H-couplings4\] Four point couplings involving the top-Higgs, again calculated to $\mathcal{O}(x^{2})$. FCNC constraints and ideal delocalization ========================================= Limits on $\Delta \epsilon^2_{tL}$: $\Delta F=2$ ------------------------------------------------ Limits on the deviation of $\epsilon_{tL}$ from ideal comes from the minimal size of tree-level flavor-changing neutral currents from $Z$-exchange. Consider re-writing Eq. for the left-handed quarks of the $i$th family (where $i=u,c,t$) as $$g_L^{Zii} = - \frac{e}{s_Wc_W} \left[ \left(1 - \frac{\Delta \epsilon^2_{iL}}{2} \right) T_3 - Q s_W^2 \right]~,$$ where $\Delta \epsilon^2_{iL}$ denotes the deviation from ideal delocalization of the $i$th family in the top-quark mass-eigenstate basis $$\Delta \epsilon^2_{iL} = \epsilon^2_{iL} - \frac{x^2}{2}~.$$ In this notation $t_L$ is the left-handed quark (in the top-quark mass eigenstate basis) whose “down" component receives a large correction from top-pion exchange. In general, this “down" component may be written $$d^t_{tL} = U_{3j} d_{jL}~,$$ where $d^t_{iL}$ represent the “down" components of the left-handed doublet fields in the top-quark mass eigenstate basis, and $d_{jL}$ are the same fields in the down-quark mass-eigenstate basis, and the $U_{3j}$ are the third row of a unitary matrix. The minimal size of the $U_{3j}$, corresponding to “next-to-minimal" flavor violation [@Agashe:2005hk], is $$U_{3j} = {\cal O}\left(V^{CKM}_{tj})\right),$$ where $V^{CKM}$ is the usual CKM flavor-mixing matrix in the standard model. Since GIM cancellation is exact when $\Delta \epsilon^2_{tL}=0$, we find the tree-level flavor changing $Z$-boson couplings to down-quarks $$\begin{aligned} g^{Zb_L s_L}_Z = \frac{e \Delta \epsilon^2_{tL} V^{CKM}_{ts}}{4 s_W c_W}\label{eq:Zbs}\\ g^{Zb_L d_L}_Z = \frac{e \Delta \epsilon^2_{tL} V^{CKM}_{td}}{4 s_W c_W}\,.\end{aligned}$$ $Z$-exchange then produces the $\Delta F=2$ effective operators $$\begin{aligned} C^1_K (\bar{s}_L \gamma^\mu d_L) (\bar{s}_L \gamma_\mu d_L) \label{eq:cdeffs}\\ C^1_{B_d} (\bar{b}_L \gamma^\mu d_L) (\bar{b}_L \gamma_\mu d_L) \\ \! C^1_{B_s} (\bar{b}_L \gamma^\mu s_L) (\bar{b}_L \gamma_\mu s_L) \,,\!\end{aligned}$$ where, since GIM cancellation is exact when $\Delta \epsilon^2_{tL}=0$, we find $$\begin{aligned} \|\Re (C^1_K)\| = \left\| \Re\left(\frac{e^2 (\Delta \epsilon^2_{tL}V^{CKM}_{ts} V^{CKM}_{td})^2}{8(s_W c_W)^2 M^2_Z}\right)\right\| <\frac{1}{(1.0\times 10^6\, {\rm GeV})^2}\\ \|\Im (C^1_K)\| = \left\| \Im\left(\frac{e^2 (\Delta \epsilon^2_{tL}V^{CKM}_{ts} V^{CKM}_{td})^2}{8(s_W c_W)^2 M^2_Z}\right) \right\| <\frac{1}{(1.5\times 10^7\, {\rm GeV})^2}\\ \|C^1_{B_d}\| =\left\| \frac{e^2 (\Delta \epsilon^2_{tL}V^{CKM}_{tb} V^{CKM}_{td})^2}{8(s_W c_W)^2 M^2_Z}\right\| < \frac{1}{(2.1\times 10^5\, {\rm GeV})^2}\\ \|C^1_{B_s}\| =\left\| \frac{e^2 (\Delta \epsilon^2_{tL}V^{CKM}_{tb} V^{CKM}_{td})^2}{8(s_W c_W)^2 M^2_Z}\right\|< \frac{1}{(3 \times 10^4\, {\rm GeV})^2} \,,\!\end{aligned}$$ where the bounds given by the last inequality in each expression come from Ref. [@Bona:2007vi]. The strongest constraint arises from limits on extra contributions to CP-violation in K-meson mixing, for which we find $$\Delta \epsilon^2_{tL} < 7.2 \times 10^{-2}~. \label{eq:CPlimit}$$ We plot this bound as a limit on $\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2$ in the upper curve in Fig. \[fig:two\], as a function of $\sin\omega$ for $M_{W'}=500$ GeV. ![The upper bounds in the deviation $\|\Delta \epsilon^2_{tL}/(\epsilon^{ideal}_{tL})^2\|$ arising from limits on extra contributions to CP-violation in $K$-meson mixing (upper curve) and from bounds on the rare decay $B_s \to \mu^+\mu^-$ (lower curve), as a function of $\sin \omega$ for $M_{W'}=500$ GeV. \[fig:two\] ](FCNCconstraints.pdf){width="60.00000%"} Limits on $\Delta \epsilon^2_{tL}$: $\Delta F=1$ ------------------------------------------------ The strongest limits from $\Delta F=1$ processes come from constraints on the $B$-meson decays $B_{d,s} \to \mu^+\mu^-$. The limits arising from experimental constraints have been summarized in [@Carpentier:2010ue]. In Table 2 of that reference, we find the strongest constraint arising from Tevatron limits on $B_s \to \mu^+ \mu^-$ and, for the operator $$\frac{2\varepsilon}{v^2} (\bar{b}_L \gamma^\nu s_L)(\bar{\mu}_L \gamma_\nu \mu_L)~, \label{eq:varepsilon}$$ where $\sqrt{2} G_F = v^{-2}$ and $v \approx 246$ GeV is the weak scale. In our case, using Eq. and $M_Z = e v/(2 s_W c_W)$, we find $$\varepsilon = \frac{1}{2}\Delta \epsilon^2_{tL} V^{CKM}_{ts} V^{CKM}_{tb} \approx \frac{\lambda^2 \Delta\epsilon^2_{tL}}{2}~.$$ Using the limit $BR(B_s \to \mu^+ \mu^-)< 4.3 \times 10^{-8}$ [@CDFnew] and the techniques[^9] of ref. [@Carpentier:2010ue], we find the bound $\varepsilon < 7.6 \times 10^{-4}$. From eqn. (\[eq:varepsilon\]), we then obtain $$\Delta \epsilon^2_{tL} < 3.8 \times 10^{-2}~,$$ a constraint roughly twice as small as that given by limiting contributions to CP-violation in $K$-meson mixing in Eq. . Comparing Figs. \[fig:one\] and \[fig:two\], we see that compensating for the deviation in $R_b$ resulting from top-pion exchange by modifying the delocalization of the third-generation quarks is not, in the context of “next-to-minimal" flavor violation [@Agashe:2005hk], ruled out by flavor changing neutral current constraints. [100]{} R. S. Chivukula, D. A. Dicus and H. J. He, Phys. Lett. B [525]{}, 175 (2002) \[arXiv:hep-ph/0111016\]. R. S. Chivukula and H. J. He, Phys. Lett. B [532]{}, 121 (2002) \[arXiv:hep-ph/0201164\]. R. S. Chivukula, D. A. Dicus, H. J. He and S. Nandi, Phys. Lett. B [562]{}, 109 (2003) \[arXiv:hep-ph/0302263\]. C. Csaki and D. Curtin, Phys. Rev. D [80]{}, 015027 (2009) \[arXiv:0904.2137 \[hep-ph\]\]. G. Cacciapaglia, C. Csaki, G. Marandella and J. Terning, Phys. Rev. D [75]{}, 015003 (2007) \[arXiv:hep-ph/0607146\]. G. Cacciapaglia, C. Csaki, C. Grojean and J. Terning, eConf [C040802]{}, FRT004 (2004) \[Czech. J. Phys. [55]{}, B613 (2005)\]. C. Csaki, arXiv:hep-ph/0412339. G. Cacciapaglia, C. Csaki, C. Grojean and J. Terning, Phys. Rev. D [71]{}, 035015 (2005) \[arXiv:hep-ph/0409126\]. G. Cacciapaglia, C. Csaki, C. Grojean and J. Terning, Phys. Rev. D [70]{}, 075014 (2004) \[arXiv:hep-ph/0401160\]. C. Csaki, C. Grojean, L. Pilo and J. Terning, Phys. Rev. Lett. [92]{}, 101802 (2004) \[arXiv:hep-ph/0308038\]. N. Arkani-Hamed, A. G. Cohen and H. Georgi, Phys. Rev. Lett. [86]{}, 4757 (2001) \[arXiv:hep-th/0104005\]. C. T. Hill, S. Pokorski and J. Wang, Phys. Rev. D [64]{}, 105005 (2001) \[arXiv:hep-th/0104035\]. H. Georgi, Nucl. Phys. [B266]{}, 274 (1986). R. Casalbuoni, S. De Curtis, D. Dominici [et al.]{}, Phys. Lett. [B155]{}, 95 (1985). R. Casalbuoni, A. Deandrea, S. De Curtis, D. Dominici, R. Gatto and M. Grazzini, Phys. Rev. D [53]{}, 5201 (1996) \[arXiv:hep-ph/9510431\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [72]{}, 075012 (2005) \[arXiv:hep-ph/0508147\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [72]{}, 015008 (2005) \[arXiv:hep-ph/0504114\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [71]{}, 115001 (2005) \[arXiv:hep-ph/0502162\]. M. Kurachi, R. S. Chivukula, E. H. Simmons, H. J. He and M. Tanabashi, arXiv:hep-ph/0409134. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Lett. B [603]{}, 210 (2004) \[arXiv:hep-ph/0408262\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [70]{}, 075008 (2004) \[arXiv:hep-ph/0406077\]. K. Lane, A. Martin, Phys. Rev. [D80]{}, 115001 (2009). \[arXiv:0907.3737 \[hep-ph\]\]. M. Bando, T. Kugo and K. Yamawaki, Phys. Rept. [164]{}, 217 (1988). M. Bando, T. Kugo and K. Yamawaki, Nucl. Phys. B [259]{}, 493 (1985). M. Bando, T. Kugo and K. Yamawaki, Prog. Theor. Phys. [73]{}, 1541 (1985). M. Bando, T. Kugo, S. Uehara, K. Yamawaki and T. Yanagida, Phys. Rev. Lett. [54]{}, 1215 (1985). M. Bando, T. Fujiwara and K. Yamawaki, Prog. Theor. Phys. [79]{}, 1140 (1988). M. E. Peskin and T. Takeuchi, Phys. Rev. D [46]{}, 381 (1992). R. S. Chivukula, B. Coleppa, S. Di Chiara, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [74]{}, 075011 (2006) \[arXiv:hep-ph/0607124\]. R. Sekhar Chivukula, N. D. Christensen, B. Coleppa and E. H. Simmons, Phys. Rev. D [80]{}, 035011 (2009) \[arXiv:0906.5567 \[hep-ph\]\]. C. T. Hill, Phys. Lett. B [266]{}, 419 (1991). C. T. Hill, Phys. Lett. B [345]{}, 483 (1995) \[arXiv:hep-ph/9411426\]. K. Lane and E. Eichten, Phys. Lett. B 352: 382-387 (1995). M. B. Popovic and E. H. Simmons, Phys. Rev. D [58]{}, 095007 (1998) \[arXiv:hep-ph/9806287\]. C. T. Hill and E. H. Simmons, Phys. Rept. 381: 235-402 (2003), Erratum-ibid. 390: 553-554 (2004) [\[]{}arXiv: hep-ph:0203079[\]]{}. F. Braam, M. Flossdorf, R. S. Chivukula, S. Di Chiara and E. H. Simmons, Phys. Rev. [77]{}, 055005 (2008) \[arXiv:0711.1127 \[hep-ph\]\]. E. Eichten, K. D. Lane, Phys. Lett. [B90]{}, 125-130 (1980). T. Aaltonen [et al.]{} \[CDF and D0 Collaborations\], arXiv:1005.3216 \[hep-ex\]. A. Djouadi, J. Kalinowski and M. Spira, Comput. Phys. Commun. [108]{}, 56 (1998) \[arXiv:hep-ph/9704448\]. C. Anastasiou, R. Boughezal and F. Petriello, JHEP [0904]{}, 003 (2009) \[arXiv:0811.3458 \[hep-ph\]\]. V. Ahrens, T. Becher, M. Neubert and L. L. Yang, Phys. Rev. D [79]{}, 033013 (2009) \[arXiv:0808.3008 \[hep-ph\]\]. V. Ahrens, T. Becher, M. Neubert and L. L. Yang, Eur. Phys. J. C [62]{}, 333 (2009) \[arXiv:0809.4283 \[hep-ph\]\]. V. Ahrens, T. Becher, M. Neubert and L. L. Yang, arXiv:1008.3162 \[hep-ph\]. T. Aaltonen [et al.]{} \[CDF Collaboration\], Phys. Rev. Lett. [103]{}, 101803 (2009) \[arXiv:0907.1269 \[hep-ex\]\]. V. M. Abazov [et al.]{} \[D0 Collaboration\], Phys. Lett. B [682]{}, 278 (2009) \[arXiv:0908.1811 \[hep-ex\]\]. G. Aad [et al.]{} \[The ATLAS Collaboration\], arXiv:0901.0512 \[hep-ex\]. M. Baarmand, M. Hashemi and A. Nikitenko, J. Phys. G [32]{}, N21 (2006). T. Abe, R. S. Chivukula, N. D. Christensen, K. Hsieh, S. Matsuzaki, E. H. Simmons and M. Tanabashi, Phys. Rev. D [79]{}, 075016 (2009) \[arXiv:0902.3910 \[hep-ph\]\]. G. Burdman and D. Kominis, Phys. Lett. B [403]{}, 101 (1997) \[arXiv:hep-ph/9702265\]. A. Denner, R. J. Guth, W. Hollik and J. H. Kuhn, Z. Phys. C [51]{}, 695 (1991). J. F. Oliver, J. Papavassiliou and A. Santamaria, Phys. Rev. D [67]{}, 056002 (2003) \[arXiv:hep-ph/0212391\]. C. Amsler [et al.]{} \[Particle Data Group\], Phys. Lett. B [667]{}, 1 (2008). K. Agashe, M. Papucci, G. Perez and D. Pirjol, arXiv:hep-ph/0509117. F. del Aguila, G. L. Kane, M. Quiros, Phys. Rev. Lett. [63]{}, 942 (1989). F. del Aguila, L. Ametller, G. L. Kane [et al.]{}, Nucl. Phys. [B334]{}, 1 (1990). J. A. Aguilar-Saavedra, JHEP [0612]{}, 033 (2006). \[hep-ph/0603200\]. G. D. Kribs, A. Martin, T. S. Roy, \[arXiv:1012.2866 \[hep-ph\]\]. A. Pukhov, arXiv: hep-ph/0412191. See http://www-cdf.fnal.gov/physics/ewk/2010/WW\_WZ/index.html A. Djouadi, Phys. Rept. [457]{}, 1-216 (2008). \[hep-ph/0503172\]. K. Cranmer, Y. Q. Fang, B. Mellado [et al.]{}, \[hep-ph/0401148\]. E. L. Berger, Q. -H. Cao, C. B. Jackson [et al.]{}, Phys. Rev. [D82]{}, 053003 (2010). \[arXiv:1003.3875 \[hep-ph\]\]. R. S. Chivukula, M. Golden, E. H. Simmons, Phys. Lett. [B257]{}, 403-408 (1991). R. S. Chivukula, M. Golden, E. H. Simmons, Nucl. Phys. [B363]{}, 83-96 (1991). C. Kilic, S. Schumann, M. Son, JHEP [0904]{}, 128 (2009). \[arXiv:0810.5542 \[hep-ph\]\]. T. Plehn, Phys. Rev. D [67]{}, 014018 (2003) \[arXiv:hep-ph/0206121\]; S. Lowette, J. D’Hondt and P. Vanlaer, Report No. CERN-CMS-NOTE-2006-109, 2006, available from `http://cdsweb.cern.ch`. M. Bona [et al.]{} \[UTfit Collaboration\], JHEP [0803]{}, 049 (2008) \[arXiv:0707.0636 \[hep-ph\]\]. M. Carpentier and S. Davidson, Eur. Phys. J. C [70]{}, 1071 (2010) \[arXiv:1008.0280 \[hep-ph\]\]. “Search for $B^0_s \to \mu^+\mu^-$ and $B^0_d \to \mu^+\mu^-$ with 3.7 pb$^{-1}$ of $p\bar{p}$ Collisions with CDF II", CDF Public Note 9892, CDF Collaboration. [^1]: Here and in the Appendices, the subscripts appearing in the fields will refer to the “site” numbers and the superscripts will be reserved for $SU(2)$ indices. [^2]: This bound gets stronger as $\sin\omega$ becomes smaller. [^3]: LHC experiments should be able to reduce the upper limit on BR($t \to \Pi_t^+ b$) to about $10^{-2}$ [@Aad:2009wy; @Baarmand:2006dm], which would push the lower bound on the top-pion mass above 170 GeV for the parameter point considered here. However, the studies of the LHC reach have been done only for charged Higgs masses below 150 GeV; for higher masses, off-shell decays to $t^*b$ should also be considered. [^4]: We choose 600 GeV as the upper limit because the top-Higgs becomes a broad resonance beyond this point. [^5]: We could also consider one or both of the heavy quarks decaying to a $Z$, but this would introduce extra top decays in the final state, and is not likely to compete with the charged current channel. [^6]: Amusingly, the CDF collaboration does see a slight excess in the di-jet invariant mass distribution of $W + \text{jets}$ events at $\sim 150\ \text{GeV}$ [@CDFwjj]. Though it is unlikely that the top-Higgs can be produced with sufficient rate to explain this excess, further study may be warranted. [^7]: Note that $\sigma(pp\to\Pi^{-}_t t)\propto \cot^2\omega$, and hence the reach becomes higher for lower $\sin\omega$. [^8]: The CMS study only looks at charged Higgs of mass at least 250 GeV. [^9]: We disagree with the numerical extraction of the bound on $\varepsilon$ presented in [@Carpentier:2010ue], though we agree with their method.	{ "pile_set_name": "ArXiv" }
--- author: - \| R. Cowsik$^1$ & B.V.Sreekantan$^2$\ $^1$Indian Institute of Astrophysics, Bangalore 560 034\ $^2$ National Institute of Advanced Studies, Bangalore 560 012 title: A Bound on Violations of Lorentz Invariance --- To test by experiments the limits of validity of Lorentz invariance or indeed any of the fundamental principles of physics we need a theoretical model which assumes a specific form for the violation and makes predictions of physical phenomena which can be searched for by the experiments \[2-5\]. The recent model of Coleman and Glashow incorporates tiny departures from Lorentz invariance which do not also respect flavor conservation \[1\]. One of the signatures of such a flavor non conservation is the transition $\mu \rightarrow e+ \gamma$ whose rate increases rapidly with the energy of the muon as measured in a preferred frame such as the one in which the 2.7$^o$K universal microwave background does not have any dipole anisotropy. Following their suggestion we calculate the possible contributions of such a process to the flux of “horizontal air showers” and $\mu$-less showers which provide useful estimates for the possible strength of such an interaction and also provide a good bound on such violations. The idea on which the bound on flavor violating interactions is derived becomes clear by noting that the primary cosmic rays consist mainly of nuclei which interact strongly when they are incident on the top of the earth’s atmosphere. The amount of shielding provided by the atmosphere in the vertical direction above the earth is about 1000 g cm$^{-2}$ and increases as the secant of the zenith angle $\theta$ upto $\sim 80^o$. The total grammage in the horizontal direction is about 36500 g cm$^{-2}$. The primary cosmic rays interact in the atmosphere and create a ‘nuclear active’ cascade. Since the atmosphere is tenuous with a scale height h $\approx 7 \times 10^5$ cm pions and kaons in the cascade decay producing the cosmic-ray muonic component. Nuclear interactions of pions and kaons with the atmosphere compete with their decay and become increasingly dominant as the particle energy increases, so that the spectrum of the muonic component at high energies is steeper than that of the nuclear active component by a factor E$^{-1}$. Also the muon component at high energies increases as $\sim$ sec $\theta$, as the scale height of the atmosphere also has this dependence. Since the interaction mean free path of the hadronic components is after reaching their maximum development, they are absorbed with an absorption mean free path of $\sim$ 100 g cm$^{-2}$. In contrast the muons suffer only electromagnetic interactions and propagate with hardly any reduction in flux. Now note that as we move away from the vertical towards the horizontal direction, with increasing sec $\theta$ the nuclear active components get severely absorbed but the high energy muonic component increases as $\sim$ sec $\theta$! Thus at large angles we have a nearly pure beam of high energy muons, traversing distances of the order of few times the scale height $h_\theta \sim h sec \theta$. Now should the muon decay radiatively the decay products e and $\gamma$ will induce an electromagnetic cascade which can easily be observed signalling the violation of flavor conservation, as described in the model of Glashow and Coleman. Indeed as the energy of the muon increases the observability of the e$\gamma$-cascade increases as it penetrates deeper, spreads wider and produces more observable electrons and photons. The electromagnetic cascade has a very broad peak at about 500 g cm $^{-2}$ from the point of initiation for an electron or $\gamma$ of energy E $\sim 10^4 $ GeV and the depth of maximum increases logarithmically with energy. The total number of electrons at the peak of an electromagnetic cascade is approximately equal to the energy of the initiating electron or gamma ray in GeV units. Thus any array of particle detectors deployed to detect extensive air showers will be able to detect such showers generated by the radiative decay of the muon. There will be negligible amount of nuclear active particles and muons in these showers. The background due to showers induced by the primary cosmic ray nuclei become negligible as we go to large zenith angles. Thus ‘$\mu$-less’ showers appearing in near horizontal directions constitute a signal of the new process described by Coleman and Glashow. To quantify these ideas we note that the spectrum of muons at high energies near the earth may be parametrized as $$\mu (E) = \frac{\kappa_i sec \theta}{E^{\beta +1}_i} cm^{-2} s^{-1} sr^{-1} GeV^{-1} \eqno (1)$$ with $$\kappa_1 = 10 , \beta_1 = 2.7 \,\,\,\, for \,\,\,\, 1000 GeV <E< 10^5\,\,\, GeV \eqno (2)$$ and $$\kappa_2 = 10^4, \beta_2 = 3.3 \,\,\,\,\, for \,\,\,\, 10^5 GeV <E< 3 \times 10^7 GeV \eqno (3)$$ Here $\beta_1$ and $\beta_2$ are the power law exponents of the primary cosmic ray spectrum at energies of 10 to 30 times the energy of the muon. According to Coleman and Glashow\[1\] the total decay probability per unit time, $\Gamma$, of a muon of Lorentz factor $\gamma$ is given by: $$\Gamma = \Gamma_w + \Gamma_r = \frac{1 + b \gamma^4}{\gamma \tau_o} = \frac{1}{\gamma \tau_o} + \frac{b \gamma^3}{\tau_o} \eqno (4)$$ Here $\tau_o \approx 2.2 \times 10^{-6}s$ is the life-time of the muon and b is a very small parameter describing the violation of Lorentz invariance and flavor conservation. For a muon decay close to the earth say within a distance d of about 5 km ($\sim 700 g cm^{-2}$ from the air shower array), it has to survive decay during its flight though the atmosphere beyond this i.e. a distance of few times $h_\theta$, the scale height in that direction. Thus the number of muons decaying in the 5 km stretch is given by $$s(E) \approx \kappa sec \theta E^{- \beta -1} exp \left\{ - j h_\theta \Gamma/c \right\} \Gamma d/c \eqno (5)$$ where j is a number of the order of 2 to 3. Noting that $\Gamma$ is a small number and that at high energies $\Gamma \sim \Gamma_{r}$, the exponential in eq. 5 may be set to unity and eq. 5 is rewritten as $$s(E) \sim \kappa sec \theta E^{- \beta - 1} . \Gamma_r d/c \approx \frac{\kappa sec \theta b d m^{-3}_\mu}{c \tau_o} E^{2-\beta} \eqno (6)$$ $$\equiv \kappa b \eta \,\, E^{2-\beta}$$ where $\eta = d m^{-3}_\mu sec\theta/c\tau_o \,\, GeV^{-3} \approx 5 \times 10^4 \,\, GeV^{-3}$. The products of the radiative decay of the muon generate an extensive air shower which contains a large number of electrons near the maximum, n, related to the muon energy through the simple relation $$n_e \approx E/\epsilon \eqno (7)$$ where $\epsilon \approx $ 1 GeV for an electromagnetic shower of primary energy in the range $10^4$ GeV - $10^6$ GeV. The number spectrum of particles that will be seen by an air shower array is given by $$f(n) \approx \epsilon^{3-\beta}. \kappa b \eta \, n^{2-\beta}$$ Or the number of showers F, of size larger than n is given by $$F(n) = \int ^\infty_n \, f(n^\prime)dn^\prime \eqno (9)$$ $$F_2(n) = \frac{\epsilon^{3-\beta} \kappa_2 b \eta}{\beta_2-3} n^{3-\beta} \,\,\,\, for \, n \geq 10^5 \eqno(10)$$ $$F_1(n) = \frac{\epsilon^{3-\beta} \kappa_1 b \eta }{3-\beta_1} \left[10^{5(3-\beta)} - n^{3-\beta} \right ] + F_2(10^5) \,\,\,\, for \,\, n < 10^5 \eqno (11)$$ We compare the integral number spectrum of horizontal air showers obtained by Nagano et al \[6\] with the Akeno array in Fig. 1 for $b=10^{-23}$ (curve a) and $b=10^{-25}$ (curve b). Note that $b\sim 10^{-23}$ excluded even by the lower energy data at $n_e \sim 10^5$ and the bound $$b<10^{-25} \eqno (12)$$ obtains when we consider the fluxes of horizontal air showers quoted by Nagano et al for $n_e \sim 5 \times 10^6$. Clearly these bounds are considerably more stringent than those derived by looking at the depth intensity curves for muons and as such small values of branching ratio for radiative decay will not have any detrimental effects on the functioning of muon colliders (Coleman and Glashow 1998). It is interesting to note that in the Coleman Glashow model this limit translates to $$\mid 1-c \mid \leq 6 \times 10^{-21} \eqno (13)$$ This limit is of course several others weaker than those reviewed in their paper. [99]{} S.Coleman and S.L.Glashow, Phys. Lett. B , 249 (1997); Harvard University Theoretical Physics Preprint 98/AO76 (prt. comm). C.M. Will, Theory and Experiment in Gravitational Physics (Cambridge University Press, Cambridge, 1993) M.I. Haugan and C.M.Will, Physics Today, (May 1987) E.Fischbach, M.P.Haugan, D.Tadic and H.Y.Chang, Phys. Rev. (1985) 154. G.L.Greene, M.S.Dewey, E.G.Kessler, Jr and E.Fischbach, Phys. Rev. (1991) 2216. M.Nagano, H.Yoshii, T.Hara, N.Hayashida, M.Honda, K.Kamata, S.Kawaguchi, T.Kifune, Y.Matsubara, G.Tanahashi and M.Teshima, J.Phys. G: Nucl. Phys. (1986) 69-84. Fig. 1 The integral flux of horizontal air showers given by Nagano et al. is compared with the expectation from the Coleman-Glashow process for the two values of b, $10^{-23}$ and $10^{-25}$ respectively.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Combining the usual energy functional with a higher-order conserved quantity originating from integrability theory, we show that the black soliton is a local minimizer of a quantity that is conserved along the flow of the cubic defocusing NLS equation in one space dimension. This unconstrained variational characterization gives an elementary proof of the orbital stability of the black soliton with respect to perturbations in $H^2({\mathbb{R}})$.' author: - \| Thierry Gallay$^{1}$ and Dmitry Pelinovsky$^{2}$\ [$^{1}$ Institut Fourier, Université de Grenoble 1, 38402 Saint-Martin-d’Hères, France]{}\ [$^{2}$ Department of Mathematics, McMaster University, Hamilton, Ontario, Canada, L8S 4K1]{} title: 'Orbital stability in the cubic defocusing NLS equation: II. The black soliton' --- Introduction {#sec:intro} ============ In this work we show how the techniques developed in the companion paper [@GP] to investigate the stability properties of the cnoidal periodic waves of the cubic defocusing nonlinear Schrödinger equation in one space dimension can be extended to provide a new and rather elementary proof of orbital stability in the limiting case of the black soliton. We thus consider the cubic defocusing NLS equation $$\label{nls} i \psi_t(x,t) + \psi_{xx}(x,t) - \|\psi(x,t)\|^2 \psi(x,t) \,=\, 0,$$ where $\psi$ is a complex-valued function of $(x,t) \in {\mathbb{R}}\times {\mathbb{R}}$. The black soliton is the particular solution of given by $\psi(x,t) = e^{-it}u_0(x)$, where $$\label{black-soliton} u_0(x) \,=\, \tanh\Bigl(\frac{x}{\sqrt{2}}\Bigr), \qquad x \in {\mathbb{R}}.$$ For later use, we note that the soliton profile $u_0 : {\mathbb{R}}\to {\mathbb{R}}$ satisfies the differential equations $$\label{wave} u_0' \,=\, \frac{1}{\sqrt{2}}\Bigl(1-u_0^2\Bigr), \qquad \hbox{hence}\qquad u_0'' + u_0 - u_0^3 \,=\, 0.$$ The NLS equation has many symmetries and conserved quantities, which play a crucial role in the dynamics of the system. In particular, the gauge invariance $\psi \mapsto e^{i\theta}\psi$ and the translation invariance $\psi \mapsto \psi(\cdot-\xi)$ give rise to the conservation of the charge $Q$ and the momentum $M$, respectively, where $$\label{QMdef} Q(\psi) \,=\, \int_{\mathbb{R}}\Bigl(\|\psi\|^2 -1\Bigr){\,\mathrm{d}}x, \qquad M(\psi) \,=\, \frac{i}{2} \int_{\mathbb{R}}\Bigl(\bar{\psi} \psi_x - \psi \bar{\psi}_x\Bigr) {\,\mathrm{d}}x.$$ Since the NLS equation is an autonomous Hamiltonian system, we also have the conservation of the energy $$\label{energy} E(\psi) \,=\, \int_{\mathbb{R}}\left(\|\psi_x\|^2 + \frac{1}{2} (1 - \|\psi\|^2)^2 \right) {\,\mathrm{d}}x.$$ In what follows, our goal is to study the stability of the black soliton , and we shall therefore restrict ourselves to solutions of for which $\|\psi\| \to 1$ as $\|x\| \to \infty$. This is why we defined the conserved quantities , in such a way that the integrands vanish when $\|\psi\| = 1$ and $\psi_x = 0$. The nonlinear stability of the black soliton has been studied in several recent works. In [@BGSS] the authors apply the variational method of Cazenave and Lions [@CL], which relies on the fact that the black soliton is a global minimizer of the energy $E$ for a fixed value of the momentum $M$. The difficulty with this approach is that the momentum is not defined for all finite-energy solutions, so that the integral defining $M$ in has to be renormalized and properly interpreted. A slightly different proof was subsequently given in [@GS], in the spirit of the work by Weinstein [@We] and Grillakis, Shatah, and Strauss [@GSS]. The main idea is to show that the energy functional becomes coercive in a neighborhood of the black soliton if the conservation of the momentum is used to get rid of one unstable direction. Both results in [@BGSS; @GS] are variational in nature and establish orbital stability of the black soliton in the energy space. Note that asymptotic stability of the black soliton is also proved in [@GS], using ideas and techniques developed by Martel and Merle for the generalized Korteweg-de Vries equation [@MM]. In a different direction, a more precise orbital stability result was obtained in [@GZ] for sufficiently smooth and localized perturbations, using the inverse scattering transform method which relies on the integrability of the cubic defocusing NLS equation . Similarly, asymptotic stability of the black soliton and several dark solitons was recently proved in [@Cuccagna-private]. As as consequence of integrability, the NLS equation has many conserved quantities in addition to the charge, the momentum, and the energy. In the present work, we introduce a new variational approach based on the higher-order functional $$\label{Sdef} S(\psi) \,=\, \int_{{\mathbb{R}}} \left( \|\psi_{xx}\|^2 + 3 \|\psi\|^2 \|\psi_x\|^2 + \frac{1}{2} (\bar{\psi} \psi_x + \psi \bar{\psi}_x)^2 + (1 - \|\psi\|^2)^2 \Bigl( 1 + \frac{1}{2} \|\psi\|^2 \Bigr) \right) {\,\mathrm{d}}x,$$ which is also conserved under the evolution defined by . The latter claim can be proved by a straightforward but cumbersome calculation, or by more educated techniques as described, e.g., in [@Yang Section 2.3]. The natural domain of definition for the functional is the $H^2$ energy space defined by $$\label{Xdef} X \,=\, \Bigl\{\psi \in H^2_{\rm loc}({\mathbb{R}})\,: \quad \psi_x \in H^1({\mathbb{R}}), ~1 - \|\psi\|^2 \in L^2({\mathbb{R}}) \Bigr\}.$$ Indeed, if $\psi \in X$, then $\zeta := 1 - \|\psi\|$ belongs to $H^1({\mathbb{R}})$, because $\|\zeta\| \le \|1 - \|\psi\|^2\| \in L^2({\mathbb{R}})$ and $\zeta_x = -\|\psi\|_x \in L^2({\mathbb{R}})$. By Sobolev’s embedding of $H^1({\mathbb{R}})$ into $L^{\infty}({\mathbb{R}})$, we thus have $\|\psi\| = 1 - \zeta \in L^\infty({\mathbb{R}})$, and from the definitions (\[Sdef\]) and (\[Xdef\]), it follows easily that $S(\psi) < \infty$. Since $u_0'$, $u_0''$, and $1 - u_0^2$ decay exponentially to zero as $\|x\| \to \infty$, it is clear that $u_0 + H^2({\mathbb{R}}) \subset X$, so that the functional is well defined for $H^2$ perturbations of the soliton profile $u_0$. This allows us to define the differential of $S$ at $u_0$, and a direct calculation using the differential equations reveals that $u_0$ is a [critical point]{} of $S$, in the sense that $S'(u_0) = 0$. Unfortunately, the second variation $S''(u_0)$ has no definite sign [@GP], hence it is not possible to prove orbital stability of the black soliton using the functional $S$ alone. As is explained in the companion paper [@GP], which is devoted to the stability of periodic waves for the NLS equation , it is possible to cure that problem by subtracting from $S$ an appropriate multiple of the energy $E$, which is well defined on $X$ and also satisfies $E'(u_0) = 0$. The optimal choice is $$\label{Lamdef} \Lambda(\psi) \,=\, S(\psi) - 2 E(\psi), \qquad \psi \in X.$$ We then have $\Lambda'(u_0) = 0$, and the starting point of our approach is the following result, which asserts that the second variation $\Lambda''(u_0)$ is nonnegative. \[prop-main\] The second variation of the functional at the black soliton is nonnegative for perturbations in $H^2({\mathbb{R}})$. It is important to realize that Proposition \[prop-main\] gives an [unconstrained]{} variational characterization of the black soliton $u_0$, which is our main motivation for introducing the higher-order conserved quantity . In contrast, the approach in [@BGSS; @GS] relies on the fact that $u_0$ is a minimum of the energy $E(\psi)$ subject to the constraint ${\mathcal{M}}(\psi) = {\mathcal{M}}(u_0)$, where ${\mathcal{M}}$ is a suitably renormalized version of the momentum $M$ defined in . The proof of Proposition \[prop-main\] developed in Section \[sec:positive\] actually shows that the second variation $\Lambda''(u_0)$ is positive except for degeneracies due to symmetries: the nonnegative self-adjoint operator associated with $\Lambda''(u_0)$ has a simple zero eigenvalue which is due to translation invariance, and the essential spectrum extends all the way to the origin due to gauge invariance. As a consequence, perturbations in $H^2({\mathbb{R}})$ can include slow modulations of the phase of the black soliton far away from the origin, which hardly increase the functional $\Lambda$. This means that the second variation $\Lambda''(u_0)$ is not coercive in $H^2({\mathbb{R}})$, even if modulation parameters are used to remove the zero modes due to the symmetries. For that reason, we are not able to control the perturbations of the black soliton in the topology of $H^2({\mathbb{R}})$, but only in a weaker sense that allows for a slow drift of the phase at infinity, see Section \[sec:modulation\] below for a more detailed discussion. To formulate our main result, we equip the space $X$ with the distance $$\label{distance} d_R(\psi_1,\psi_2) \,=\, \\|(\psi_1 - \psi_2)_x\\|_{H^1({\mathbb{R}})} + \\| \|\psi_1\|^2 - \|\psi_2\|^2\\|_{L^2({\mathbb{R}})} + \\| \psi_1-\psi_2\\|_{L^2(-R,R)},$$ where $R \ge 1$ is a parameter. Note that $d_R$ is the exact analogue, at the $H^2$ level, of the distance that is used in previous variational studies of the black soliton, including [@BGSS; @Gerard; @GS]. As is easily verified, a function $\psi \in H^2_{\rm loc}({\mathbb{R}})$ belongs to $X$ if and only if $d_R(\psi,u_0) < \infty$; moreover, different choices of $R$ give equivalent distances on $X$. To prove orbital stability of the black soliton with profile $u_0$, the idea is to consider solutions $\psi$ of the NLS equation for which $d_R(\psi,u_0)$ is small. This is certainly the case if $\\|\psi-u_0\\|_{H^2}$ is small, but the converse is not true because $d_R(\psi,u_0)$ does not control the $L^2$ norm of the difference $\psi - u_0$ on the whole real line. We shall prove in Section \[sec:stability\] that the distance $d_R$ is well adapted to the functional $\Lambda$ near $u_0$, in the sense that $$\label{Lamcoer} \Lambda(\psi) - \Lambda(u_0) \,\ge\, C d_R(\psi,u_0)^2 \qquad \hbox{when}\quad d_R(\psi,u_0) \ll 1,$$ provided the perturbation $\psi-u_0$ satisfies a pair of orthogonality conditions. As is usual in orbital stability theory, these orthogonality conditions can be fulfilled if we replace $\psi$ by $e^{i\theta} \psi(\cdot+\xi)$ for some appropriate modulation parameters $\theta,\xi \in {\mathbb{R}}$, see Section \[sec:modulation\] below. It is then easy to deduce from that solutions of the NLS equation with initial data $\psi$ satisfying $d_R(\psi_0,u_0) \ll 1$ will stay close for all times to the orbit of the black soliton under the group of translations and phase rotations. The precise statement is: \[theorem-soliton\] Fix $R \ge 1$ and let $u_0 \in X$ be the black soliton . Given any $\epsilon > 0$, there exists $\delta > 0$ such that, for any $\psi_0 \in X$ satisfying $$\label{bound-initial} d_R(\psi_0,u_0) \,\le\, \delta,$$ the global solution $\psi(\cdot,t)$ of the NLS equation with initial data $u_0$ has the following property. For any $t \in {\mathbb{R}}$, there exist $\xi(t) \in {\mathbb{R}}$ and $\theta(t) \in {\mathbb{R}}/(2\pi{\mathbb{Z}})$ such that $$\label{bound-final} d_R\Bigl(e^{i (t + \theta(t))} \psi(\cdot + \xi(t),t)\,,u_0\Bigr) \,\le\, \epsilon.$$ Moreover $\xi$ and $\theta$ are continuously differentiable functions of $t$ which satisfy $$\label{bound-time-per} \|\dot \xi(t)\| + \|\dot \theta(t)\| \,\le\, C \epsilon, \quad t \in {\mathbb{R}},$$ for some positive constant $C$. It is known from the work of Zhidkov [@Zhidkov] that the Cauchy problem for the NLS equation is globally well-posed in $X$. This is the functional framework that is used to define solutions of in Theorem \[theorem-soliton\]. Except for the use of a different distance $d_R$, which controls the perturbations in the topology of $H^2_{\rm loc}({\mathbb{R}})$, Theorem \[theorem-soliton\] is the exact analogue of the orbital stability results obtained in [@BGSS; @GS]. However the proof is quite different, and in some sense simpler, because the profile $u_0$ of the black soliton is an unconstrained local minimizer of the higher-order functional $\Lambda$. It is also possible to prove asymptotic stability results for the black soliton of the cubic NLS equation . In that perspective, it is useful to consider the black soliton as a member of the one-parameter family of traveling dark solitons, given by the exact expression $$\label{darksoliton} e^{it}\psi_{\nu}(x + \nu t,t) \,=\, \sqrt{{{\textstyle}1 - \frac{1}{2} \nu^2}} \,\tanh\left(\sqrt{{{\textstyle}\frac{1}{2} - \frac{1}{4} \nu^2}}\,x\right) + \frac{i\nu}{\sqrt{2}},$$ where $\nu \in (-\sqrt{2},\sqrt{2})$. Asymptotic stability of the family of dark solitons with nonzero speed $\nu$ was proved in [@Bethuel], using the Madelung transformation and the hydrodynamic formulation of the NLS equation. This approach applies to solutions whose modulus is strictly positive, and therefore excludes the case of the black soliton. Very recently, the asymptotic stability of the black soliton (within the one-parameter family of all dark solitons) has been established in [@Cuccagna-private; @GS]. The rest of this article is organized as follows. In Section \[sec:positive\] we establish positivity and coercivity properties for the quadratic form associated with the second variation of the functional at $u_0$. In Section \[sec:modulation\], we introduce modulation parameters in a neighborhood of the soliton profile to eliminate the zero modes of the second variation $\Lambda''(u_0)$. Combining these results and using a new variable borrowed from [@GS], we prove in Section \[sec:stability\] the orbital stability of the black soliton in the space $X$. Positivity and coercivity of the second variation {#sec:positive} ================================================= Let $u_0$ be the soliton profile and $\Lambda = S - 2E$ be the functional defined by , , and . In this section, we prove that the second variation $\Lambda''(u_0)$ is nonnegative, as stated in Proposition \[prop-main\], and we deduce some coercivity properties that will be used in the proof of Theorem \[theorem-soliton\]. We consider perturbations of $u_0$ of the form $\psi = u_0 + u + i v$, where $u,v \in H^2({\mathbb{R}})$ are real-valued. As in [@GP], the second variations at $u_0$ of the functionals $E$ and $S$ satisfy $$\begin{aligned} {\textstyle\frac12}\langle E''(u_0)[u,v], [u,v]\rangle \,&=\, \langle L_+ u,u\rangle_{L^2} + \langle L_- v,v\rangle_{L^2}, \\[1mm] {\textstyle\frac12}\langle S''(u_0)[u,v], [u,v]\rangle \,&=\, \langle M_+ u,u\rangle_{L^2} + \langle M_- v,v\rangle_{L^2},\end{aligned}$$ where $\langle\cdot\,,\cdot\rangle_{L^2}$ denotes the usual scalar product in $L^2({\mathbb{R}})$. The self-adjoint operators $L_\pm$ and $M_\pm$ have the following expressions: $$\label{operatorsdef} \begin{array}{l} L_+ \,=\, -\partial_x^2 + 3 u_0^2 - 1, \\[1mm] L_- \,=\, -\partial_x^2 + u_0^2 - 1, \end{array} \qquad \begin{array}{lcl} M_+ \,=\, \partial_x^4 - 5 \partial_x u_0^2 \partial_x -5 u_0^4 + 15 u_0^2 - 4, \\[1mm] M_- \,=\, \partial_x^4 - 3 \partial_x u_0^2 \partial_x + u_0^2 - 1. \end{array}$$ In view of , it follows that $$\label{Lambdasecond} {\textstyle\frac12}\langle \Lambda''(u_0)[u,v], [u,v]\rangle \,=\, \langle K_+ u,u\rangle_{L^2} + \langle K_- v,v\rangle_{L^2},$$ where $K_{\pm} = M_{\pm} - 2L_{\pm}$. More explicitly, the quadratic forms associated with $K_\pm$ are given by $$\begin{aligned} \label{Kquad+} \langle K_+ u,u\rangle_{L^2} \,&=\, \int_{\mathbb{R}}\Bigl(u_{xx}^2 + (5u_0^2-2)u_x^2 + (9u_0^2 -5u_0^4-2)u^2\Bigr){\,\mathrm{d}}x, \\ \label{Kquad-} \langle K_- v,v\rangle_{L^2} \,&=\, \int_{\mathbb{R}}\Bigl(v_{xx}^2 + (3u_0^2-2)v_x^2 + (1-u_0^2)v^2\Bigr){\,\mathrm{d}}x.\end{aligned}$$ Our first task is to show that the quadratic forms , are nonnegative on $H^2({\mathbb{R}})$. Due to translation invariance of the NLS equation , we have $L_+ u_0' = M_+ u_0' = 0$, hence also $K_+ u_0' = 0$. As $u_0' \in H^2({\mathbb{R}})$, this shows that the quadratic form associated with $K_+$ has a neutral direction, hence is not strictly positive, see Lemma \[lemma-K-plus\] below. The situation is slightly different for $K_-$: due to gauge invariance, we have $L_- u_0 = M_- u_0 = 0$, hence $K_- u_0 = 0$, but of course $u_0 \not\in H^2({\mathbb{R}})$. In fact, the result of Lemma \[lemma-K-minus\] below shows that the quadratic form associated with $K_-$ is strictly positive on $H^2({\mathbb{R}})$. We first prove that the quadratic form is nonnegative, see also [@GP Corollary 4.5]. \[lemma-K-plus\] For any $u \in H^2({\mathbb{R}})$, we have $$\label{operator-K-plus-soliton} \langle K_+ u, u \rangle_{L^2} \,=\, \\|w_x\\|_{L^2}^2 + \\|w\\|_{L^2}^2 \,\ge\, 0,$$ where $w = u_x + \sqrt{2} u_0 u$. Integrating by parts and using the differential equations satisfied by $u_0$, we easily obtain $$\label{LemK+1} \int_{\mathbb{R}}w^2 {\,\mathrm{d}}x \,=\, \int_{\mathbb{R}}\Bigl(u_x^2 + 2\sqrt{2}u_0 u u_x + 2u_0^2 u^2\Bigr) {\,\mathrm{d}}x \,=\, \int_{\mathbb{R}}\Bigl(u_x^2 + (3u_0^2-1)u^2\Bigr){\,\mathrm{d}}x.$$ Similarly, as $w_x = u_{xx} + \sqrt{2}u_0u_x + \sqrt{2}u_0'u$, we find $$\begin{aligned} \nonumber \int_{\mathbb{R}}w_x^2 {\,\mathrm{d}}x \,&=\, \int_{\mathbb{R}}\Bigl(u_{xx}^2 + 2\sqrt{2}u_0 u_x u_{xx}+ 2u_0^2 u_x^2 + 2\sqrt{2}u_0' uu_{xx} + 4u_0 u_0' uu_x + 2u_0'^2 u^2\Bigr){\,\mathrm{d}}x \\ \nonumber \,&=\, \int_{\mathbb{R}}\Bigl(u_{xx}^2 + (5u_0^2-3)u_x^2 + 8u_0 u_0' uu_x + 2u_0'^2 u^2\Bigr){\,\mathrm{d}}x \\ \label{LemK+2} \,&=\, \int_{\mathbb{R}}\Bigl(u_{xx}^2 + (5u_0^2-3)u_x^2 + (1-u_0^2)(5u_0^2-1) u^2\Bigr){\,\mathrm{d}}x,\end{aligned}$$ because $2u_0'^2 - 4(u_0u_0')' = (1-u_0^2)(5u_0^2-1)$. If we now combine and , we see that $\\|w_x\\|_{L^2}^2 + \\|w\\|_{L^2}^2$ is equal to the right-hand side of , which is the desired conclusion. \[remark-black-plus\] The right-hand side of vanishes if and only if $w = 0$, which is equivalent to $u = C u_0'$ for some constant $C$. Thus zero is a simple eigenvalue of $K_+$ in $L^2({\mathbb{R}})$. Moreover, since $u_0(x) \to \pm 1$ as $x \to \pm \infty$, it is clear from that the essential spectrum of $K_+$ is the interval $[2,\infty)$. Thus if we restrict ourselves to the orthogonal complement of $u_0'$ with respect to the scalar product $\langle\cdot\,, \cdot\rangle_{L^2}$, the spectrum of $K_+$ is bounded from below by a strictly positive constant, and the corresponding quadratic form is thus coercive in the topology of $H^2({\mathbb{R}})$, see Remark \[oldcond\] below. We next prove the positivity of the quadratic form , see also [@GP Lemma 4.1]. \[lemma-K-minus\] For any $v \in H^2({\mathbb{R}})$, we have $$\label{operator-K-minus} \langle K_- v, v \rangle_{L^2} \,=\, \\| L_- v \\|_{L^2}^2 + \\| u_0 v_x - u_0' v \\|_{L^2}^2 \,\ge\, 0,$$ where $L_- = -\partial_x^2 + u_0^2 - 1$. Integrating by parts we obtain $$\begin{aligned} \int_{\mathbb{R}}(L_-v)^2 {\,\mathrm{d}}x \,&=\, \int_{{\mathbb{R}}} \Bigl(v_{xx}^2 + 2(1 - u_0^2) v v_{xx} + (1-u_0^2)^2 v^2 \Bigr) {\,\mathrm{d}}x \\ \,&=\, \int_{{\mathbb{R}}} \Bigl(v_{xx}^2 + 2(u_0^2-1) v_x^2 -2(u_0 u_0')'v^2 + (1-u_0^2)^2v^2 \Bigr) {\,\mathrm{d}}x.\end{aligned}$$ Similarly, we have $$\int_{\mathbb{R}}\Bigl(u_0 v_x - u_0' v\Bigr)^2{\,\mathrm{d}}x \,=\, \int_{{\mathbb{R}}} \Bigl(u_0^2 v_x^2 + (u_0u_0')'v^2 + u_0'^2 v^2\Bigr) {\,\mathrm{d}}x.$$ It follows that $$\\|L_- v\\|_{L^2}^2 + \\|u_0 v_x - u_0' v\\|_{L^2}^2 \,=\, \int_{{\mathbb{R}}} \Bigl(v_{xx}^2 + (3 u_0^2 - 2) v_x^2 + [(1-u_0^2)^2 - u_0 u_0''] v^2 \Bigl){\,\mathrm{d}}x,$$ and that expression coincides with the right-hand side of since $(1-u_0^2)^2 - u_0 u_0'' = 1 - u_0^2$ by . This proves . The right-hand side of vanishes if and only if $L_-v = 0$ and $u_0 v_x - u_0' v = 0$, namely if $v = C u_0$ for some constant $C$. As $u_0 \notin H^2({\mathbb{R}})$, this shows that $\langle K_- v, v \rangle_{L^2} > 0$ for any nonzero $v \in H^2({\mathbb{R}})$. However, since $\|u_0(x)\| \to 1$ as $\|x\| \to \infty$, it is clear from the representation that zero belongs to the essential spectrum of the operator $K_-$, hence the associated quadratic form is not coercive in the topology of $H^2({\mathbb{R}})$. Some weaker coercivity property will nevertheless be established below, see Remark \[newcond\]. In view of the decomposition , Proposition \[prop-main\] is an immediate consequence of Lemmas \[lemma-K-plus\] and \[lemma-K-minus\]. In the rest of this section, we show that the quadratic forms , are not only positive, but also coercive in some appropriate sense. \[lemma-soliton-1\] Let $u_0$ be the black soliton . There exists a positive constant $C$ such that, for any $u \in H^2({\mathbb{R}})$ satisfying $\langle u_0', u \rangle_{L^2} = 0$, we have the estimate $$\label{bound-u} \\| u \\|_{H^2} \,\le\, C \\| w \\|_{H^1},$$ where $w = u_x + \sqrt{2} u_0 u$. Solving the linear differential equation $u_x + \sqrt{2} u_0 u = w$ by Duhamel’s formula, we find $u = A u_0' + W$ for some $A \in {\mathbb{R}}$, where $$\label{variation-u} W(x) \,=\, \int_0^x K(x,y) w(y) {\,\mathrm{d}}y, \qquad K(x,y) \,=\, \frac{\cosh^2(y/\sqrt{2})}{\cosh^2(x/\sqrt{2})}.$$ The constant $A$ is uniquely determined by the orthogonality condition $\langle u_0', u \rangle_{L^2} = 0$, which implies that $A \\|u_0'\\|_{L^2}^2 + \langle u_0', W \rangle_{L^2} = 0$. Using , we easily obtain $$\begin{aligned} \nonumber \langle u_0', W \rangle_{L^2} \,&=\, \int_{-\infty}^\infty \biggl\{\int_0^x K(x,y) w(y) {\,\mathrm{d}}y\biggr\}u_0'(x) {\,\mathrm{d}}x \\ \nonumber \,&=\, \int_0^\infty \biggl\{\int_y^\infty K(x,y) u_0'(x){\,\mathrm{d}}x\biggr\}\Bigl(w(y)-w(-y)\Bigr) {\,\mathrm{d}}y \\ \label{intex} \,&=\, \frac{1}{3} \int_0^\infty e^{-\sqrt{2}y}\,\frac{3+e^{-\sqrt{2}y}}{ 1+e^{-\sqrt{2}y}}\Bigl(w(y)-w(-y)\Bigr) {\,\mathrm{d}}y,\end{aligned}$$ hence $\|\langle u_0', W \rangle_{L^2}\| \le 2^{-1/4}\\|w\\|_{L^2}$. It follows that $\|A\| \le C\\|w\\|_{L^2}$ for some $C > 0$. On the other hand, if we introduce the operator notation $W = \hat{K}(w)$ for the representation , we note that $\hat{K}$ is a bounded operator from $L^{\infty}({\mathbb{R}})$ to $L^{\infty}({\mathbb{R}})$ with norm $$K_{\infty} \,=\, \sup_{x \in {\mathbb{R}}} \int_0^{\|x\|} K(x,y) {\,\mathrm{d}}y \,=\, \frac{1}{\sqrt{2}} \sup_{x \in {\mathbb{R}}} \frac{1 + 2 \sqrt{2} \|x\| e^{-\sqrt{2}\|x\|} - e^{-2 \sqrt{2}\|x\|}}{1 + 2 e^{-\sqrt{2}\|x\|} + e^{-2 \sqrt{2}\|x\|}} \,<\, \infty,$$ as well as a bounded operator from $L^1({\mathbb{R}})$ to $L^1({\mathbb{R}})$ with norm $$K_1 \,=\, \sup_{y \in {\mathbb{R}}} \int_{\|y\|}^{\infty} K(x,y) {\,\mathrm{d}}x \,=\, \frac{1}{\sqrt{2}} \sup_{y \in {\mathbb{R}}} \Bigl(1 + e^{-\sqrt{2}\|y\|}\Bigr) = \sqrt{2}.$$ By the Riesz-Thorin interpolation theorem, it follows that $\hat{K}$ is a bounded operator from $L^2({\mathbb{R}})$ to $L^2({\mathbb{R}})$, and we have the estimate $\\|W\\|_{L^2} = \\|\hat{K}(w) \\|_{L^2} \leq (K_1 K_{\infty})^{1/2} \\|w\\|_{L^2}$. Summarizing, we have shown that $\\|u\\|_{L^2} \le \|A\| \\|u_0'\\|_{L^2} + \\|W\\|_{L^2} \le C \\|w\\|_{L^2}$ for some $C > 0$. Since $w = u_x + \sqrt{2} u_0 u$, we also have $\\|u_x\\|_{L^2} \le \\|w\\|_{L^2} + \sqrt{2}\\|u\\|_{L^2}$ and (after differentiating) $\\|u_{xx}\\|_{L^2} \le \\|w_x\\|_{L^2} + \sqrt{2} \\|u_x\\|_{L^2} + \\| u \\|_{L^2}$. This proves the bound . \[oldcond\] Combining and , we conclude that there exists a constant $C_+ > 0$ such that $$\label{K+coercive} \langle K_+ u, u \rangle_{L^2} \,\ge\, C_+ \\|u\\|_{H^2}^2,$$ for all $u \in H^2({\mathbb{R}})$ satisfying $\langle u_0', u \rangle_{L^2} = 0$. \[lemma-soliton-2\] Let $u_0$ be the black soliton . There exists a positive constant $C$ such that, for any $v \in H^2_{\rm loc}({\mathbb{R}})$ satisfying $v_x \in H^1({\mathbb{R}})$ and $\langle u_0'', v \rangle_{L^2} = 0$, we have the estimate $$\label{bound-v} \\| v_{xx} \\|_{L^2} + \\| v_x \\|_{L^2} + \|v(0)\| \,\le\, C(\\| p \\|_{L^2} + \\| q \\|_{L^2}),$$ where $p = u_0 v_x - u_0' v$ and $q = -L_- v = v_{xx} + (1-u_0^2) v$. Any solution of the linear differential equation $u_0 v_x - u_0' v = p$ has the form $v = B u_0 + Z$ for some $B \in {\mathbb{R}}$, where $$\label{variation-v} Z(x) \,=\, u_0(x)\int_0^x \Bigl(p(y) + \sqrt{2}q(y)\Bigr){\,\mathrm{d}}y - \sqrt{2}p(x).$$ Indeed, we observe that $p_x = u_0 v_{xx} - u_0'' v = u_0 (v_{xx} + (1-u_0^2) v) = u_0 q$. Thus, if $v = B u_0 + Z$, we have $$\label{variation-v2} v_x(x) \,=\, u_0'(x)\left(B + \int_0^x \Bigl(p(y) + \sqrt{2}q(y) \Bigr){\,\mathrm{d}}y\right) + u_0(x)p(x),$$ hence $u_0 v_x - u_0' v = (u_0^2 + \sqrt{2}u_0')p = p$. The constant $B$ is uniquely determined by the orthogonality condition $\langle u_0'', v \rangle_{L^2} = 0$, which implies that $B \\|u_0'\\|_{L^2}^2 = \langle u_0'',Z\rangle$. Since $p \in L^2({\mathbb{R}})$ and $p_x = u_0 q \in L^2({\mathbb{R}})$, we have $p \in L^\infty({\mathbb{R}})$ by Sobolev’s embedding resulting in the bound $\\|p\\|_{L^\infty}^2 \le \\|p\\|_{L^2} \\|p_x\\|_{L^2} \le \\|p\\|_{L^2}\\|q\\|_{L^2}$. Thus, using and Hölder’s inequality, we deduce that $$\|Z(x)\| \,\le\, \sqrt{2} (\|x\|^{1/2} + 1)(\\|p\\|_{L^2} + \\|q\\|_{L^2}), \quad x \in {\mathbb{R}}.$$ This moderate growth of $Z$ is compensated for by the exponential decay of $u_0''$ to zero at infinity, and we obtain $\|\langle u_0'',Z\rangle\| \le C(\\|p\\|_{L^2} + \\|q\\|_{L^2})$ for some $C > 0$, hence also $\|B\| \le C(\\|p\\|_{L^2} + \\|q\\|_{L^2})$. In the same way, it follows from that $\\|v_x\\|_{L^2} \le C(\\|p\\|_{L^2} + \\|q\\|_{L^2})$. A similar estimate holds for $\\|v_{xx}\\|_{L^2}$ because $v_{xx} = q - (1 - u_0^2) v$ and $1 - u_0^2$ has the exponential decay to zero at infinity. Finally, since $v(0) = -\sqrt{2}p(0)$, we also have $\|v(0)\| \le C(\\|p\\|_{L^2} + \\|q\\|_{L^2})$. This proves the bound . \[newcond\] Combining and , we conclude that there exists a constant $C_- > 0$ such that $$\label{K-coercive} \langle K_- v, v \rangle_{L^2} \,\ge\, C_- \Bigl(\\|v_x\\|_{H^1}^2 + \|v(0)\|^2\Bigr),$$ for all $v \in H^2_{\rm loc}({\mathbb{R}})$ satisfying $v_x \in H^1({\mathbb{R}})$ and $\langle u_0'', v \rangle_{L^2} = 0$. As is clear from the proof of Lemma \[lemma-soliton-2\], we need some orthogonality condition on $v$ to prove estimate , and since $u_0 \notin L^2({\mathbb{R}})$ we cannot impose $\langle u_0, v \rangle_{L^2} = 0$. Thus we use $u_0'' = u_0(u_0^2-1)$ instead of $u_0$. Although $u_0''$ is only an approximate eigenfunction of $K_-$, the orthogonality condition $\langle u_0'', v \rangle_{L^2} = 0$ is good enough for our purposes, as we shall see in Section \[sec:modulation\]. Modulation parameters near the black soliton {#sec:modulation} ============================================ This section contains some important preliminary steps in the proof of Theorem \[theorem-soliton\]. To establish the orbital stability of the black soliton with profile $u_0$, our general strategy is to consider solutions $\psi(x,t)$ of the cubic NLS equation of the form $$\label{decomposition2} e^{i(t + \theta(t))} \psi(x + \xi(t),t) \,=\, u_0(x) + u(x,t) + i v(x,t), \qquad (x,t) \in {\mathbb{R}}\times {\mathbb{R}},$$ where the perturbations $u,v$ are real-valued and satisfy the orthogonality conditions $$\label{projections2} \langle u_0', u(\cdot,t) \rangle_{L^2} \,=\, 0, \qquad \langle u_0'', v(\cdot,t) \rangle_{L^2} \,=\, 0, \qquad t \in {\mathbb{R}}.$$ As was discussed in Remarks \[oldcond\] and \[newcond\], these conditions are needed to exploit the coercivity properties of the second variation $\Lambda''(u_0)$, where $\Lambda$ is the conserved quantity . They also allow us to determine uniquely the “modulation parameters”, namely the translation $\xi(t)$ and the phase $\theta(t)$, at least for solutions $\psi(x,t)$ in a small neighborhood of the black soliton. To make these considerations rigorous, we first need to specify in which topology that neighborhood is understood; in other words, we need to choose an appropriate perturbation space. Next we have to verify that the modulation parameters exist and depend smoothly on the solution $\psi(x,t)$ in the vicinity of the black soliton. Concerning the first point, we observe that the functional which serves as a basis for our analysis is invariant under translations and gauge transformations, and we recall that $\Lambda'(u_0) = 0$. Thus, if $\psi(x,t)$ is a solution of the NLS equation of the form with $u,v \in H^2({\mathbb{R}})$, we have for each fixed $t \in {\mathbb{R}}$ the following expansion $$\label{DeltaLambda2} \Lambda(\psi) - \Lambda(u_0) \,=\, \langle K_+ u, u \rangle_{L^2} + \langle K_- v, v \rangle_{L^2} + N(u,v),$$ where $N(u,v)$ collects all terms that are at least cubic in $u$ and $v$. However, unlike in the periodic case considered in the companion paper [@GP], the decomposition is not sufficient to prove the orbital stability of the black soliton. Indeed, the quadratic terms in are nonnegative, but they are degenerate in the sense that they do not control the $L^2(\mathbb{R})$ norm of $v$, as can be seen from the lower bound . This is due to the fact that the operator $K_-$ has essential spectrum touching the origin, with generalized eigenfunctions corresponding to slow modulations of the phase of the black soliton. As is clear from the proof of Lemma \[lemma-soliton-2\], one cannot even prove that $v \in L^\infty({\mathbb{R}})$ if we only know that $\langle K_- v, v \rangle_{L^2} < \infty$. This in turn makes it impossible to control the nonlinearity $N(u,v)$ in in terms of the quadratic part $\langle K_+ u, u \rangle_{L^2} + \langle K_- v, v \rangle_{L^2}$. There are good reasons to believe that the above problem is not just a technical one, and that the $H^2$ topology for the perturbations $u,v$ is not appropriate to prove orbital stability of the black soliton. Indeed, as is well known, the cubic NLS equation has a family of travelling dark solitons $\psi_{\nu}(x,t)$ given by . Rigorous results [@GS] and numerical simulations indicate that a small, localized perturbation of the black soliton $\psi_0$ can lead to the formation of a dark soliton $\psi_{\nu}$ with a small nonzero speed $\nu$. If this happens, the functions $u,v$ defined in cannot stay bounded in $L^2({\mathbb{R}})$ for all times, because $\psi_{\nu} - \psi_0 \notin L^2({\mathbb{R}})$ if $\nu \neq 0$. Note, however, that the quantity $\|\psi_{\nu}\| - \|\psi_0\|$ does belong to $L^2({\mathbb{R}})$ and decays exponentially at infinity. This suggests that a particular combination of $u,v$ may be controlled in $L^2({\mathbb{R}})$ for all times. Following [@GS], we introduce the auxiliary variable $$\label{etadef} \eta \,=\, \|u_0 + u + iv\|^2 - \|u_0\|^2 \,=\, 2 u_0 u + u^2 + v^2,$$ which allows us to control the perturbations of the modulus of the black soliton $u_0$. The idea is now to consider perturbations $u,v$ for which $u_x, v_x \in H^1({\mathbb{R}})$, $\eta \in L^2({\mathbb{R}})$, and $u,v \in L^2(-R,R)$ for some fixed $R \ge 1$. If $\psi = u_0 + u + iv$, this is equivalent to requiring that $\psi \in X$, where $X$ is the function space , or that $d_R(\psi,u_0) < \infty$, where $d_R$ is the distance . Indeed, we have by definition $$\label{distance2} d_R(\psi,u_0) \,=\, \\|u_x + iv_x\\|_{H^1({\mathbb{R}})} + \\|\eta\\|_{L^2({\mathbb{R}})} + \\|u + iv\\|_{L^2(-R,R)}.$$ Note, however, that we do not assume any longer that $u,v$ are square integrable at infinity. In particular, the perturbed solutions we consider include dark solitons $\psi_{\nu}$ with nonzero speed $\nu$. Now that we have defined a precise perturbation space, we can state our first result showing the existence and the continuity of the modulation parameters $\xi$ and $\theta$ in a neighborhood of the orbit of the soliton profile $u_0$. The following statement is very close in spirit to Proposition 2 in [@GS] or Lemma 6.1 in [@GP]. \[lemma-xith\] Fix any $R \ge 1$. There exists $\epsilon_0 > 0$ such that, for any $\psi \in X$ satisfying $$\label{inf2} \inf_{\xi, \theta \in {\mathbb{R}}} d_R\Bigl(e^{i \theta} \psi(\cdot + \xi), u_0\Bigr) \,\le\, \epsilon_0,$$ there exist $\xi \in {\mathbb{R}}$ and $\theta \in {\mathbb{R}}/(2\pi{\mathbb{Z}})$ such that $$\label{decomp2} e^{i \theta} \psi(x + \xi) \,=\, u_0(x) + u(x) + i v(x), \quad x \in {\mathbb{R}},$$ where the real-valued functions $u$ and $v$ satisfy the orthogonality conditions . Moreover, the modulation parameters $\xi \in {\mathbb{R}}$ and $\theta \in {\mathbb{R}}/(2\pi{\mathbb{Z}})$ depend continuously on $\psi$ in the topology defined by the distance . It is sufficient to prove for all $\psi \in X$ such that $\epsilon := d_R(\psi,u_0)$ is sufficiently small. Given such a $\psi \in X$, we consider the smooth function ${\bf f} : {\mathbb{R}}^2 \to {\mathbb{R}}^2$ defined by $${\bf f}(\xi,\theta) \,=\, \begin{pmatrix} \langle u_0'(\cdot - \xi), {\rm Re}(e^{i \theta} \psi) \rangle_{L^2} \\[1mm] \langle u_0''(\cdot - \xi), {\rm Im}(e^{i \theta} \psi) \rangle_{L^2} \end{pmatrix}, \qquad (\xi,\theta) \in {\mathbb{R}}^2.$$ By construction, we have ${\bf f}(\xi,\theta) = {\bf 0}$ if and only if $\psi$ can be represented as in for some real-valued functions $u,v$ satisfying the orthogonality conditions . If we decompose $\psi = u_0 + u + iv$ where $u,v$ are real-valued, we have $\langle u_0', {\rm Re}(\psi)\rangle_{L^2} = \langle u_0', u\rangle_{L^2}$ because $\langle u_0',u_0\rangle_{L^2} = 0$. As in the proof of Lemma \[lemma-soliton-2\], we observe that $$\|u(x)\| \,\le\, C\Bigl(\\|u\\|_{L^2(-1,1)} + (1+\|x\|^{1/2})\\|u_x\\|_{L^2({\mathbb{R}})} \Bigr) \,\le\, C(1+\|x\|^{1/2})d_R(\psi,u_0),$$ where in the last inequality we have used . Thus $\|\langle u_0', {\rm Re}(\psi)\rangle_{L^2}\| \le C d_R(\psi,u_0)$, and a similar argument gives $\|\langle u_0'', {\rm Im}(\psi) \rangle_{L^2}\| \le C d_R(\psi,u_0)$. This shows that $\\|{\bf f} (0,0) \\| \le C \epsilon$ for some positive constant $C$ independent of $\epsilon$. On the other hand, the Jacobian matrix of the function ${\bf f}$ at the origin $(0,0)$ is given by $$D {\bf f}(0,0) \,=\, \begin{pmatrix} \\|u_0'\\|_{L^2}^2 & 0 \\ 0 & -\\|u_0'\\|_{L^2}^2 \end{pmatrix} \,+\, \begin{pmatrix} -\langle u_0'',{\rm Re}(\psi - u_0) \rangle_{L^2} & -\langle u_0', {\rm Im}(\psi - u_0)\rangle_{L^2} \\[.5mm] -\langle u_0''', {\rm Im}(\psi- u_0)\rangle_{L^2} & \langle u_0'', {\rm Re}(\psi - u_0)\rangle_{L^2} \end{pmatrix}.$$ The first term in the right-hand side is a fixed invertible matrix and the second term is bounded in norm by $C\epsilon$, hence $D {\bf f}(0,0)$ is invertible if $\epsilon$ is small enough. In addition, the norm of the inverse of $D {\bf f}(0,0)$ is bounded by a constant independent of $\epsilon$. Finally, it is straightforward to verify that the second-order derivatives of ${\bf f}$ are uniformly bounded when $\epsilon \le 1$. These observations together imply that there exists a unique pair $(\xi,\theta)$, in a neighborhood of size $\mathcal{O}(\epsilon)$ of the origin, such that ${\bf f}(\xi,\theta) = {\bf 0}$. Thus the decomposition holds for these values of $(\xi,\theta)$. In addition, the above argument shows that the modulation parameters $\xi,\theta$ depend continuously on $\psi \in X$ in the topology defined by the distance . This concludes the proof. As was already mentioned, the Cauchy problem for the NLS equation is globally well-posed in the space $X$ [@Zhidkov]. If $\psi(\cdot,t)$ is a solution of in $X$ which stays for all times in a neighborhood of the orbit of the black soliton, the modulation parameters $\xi(t)$, $\theta(t)$ given by the decomposition (\[decomposition2\]) subject to the orthogonality conditions (\[projections2\]) are continuous functions of time. In fact, as in [@GP Lemma 6.3], we have the following stronger conclusion: \[difflem\] If $\epsilon > 0$ is sufficiently small, and if $\psi(\cdot,t)$ is any solution of the NLS equation satisfying estimate for all $t \in {\mathbb{R}}$, then the modulation parameters $\xi(t),\theta(t)$ in the decomposition (\[decomposition2\]) subject to (\[projections2\]) are continuously differentiable functions of $t$ satisfying . If $\psi(\cdot,t)$ is any solution of the NLS equation in $X$, we know from [@Gerard; @Zhidkov] that $t \mapsto \psi(\cdot,t)$ is continuous in the topology defined by the distance . Thus, if estimate holds for all $t \in \mathbb{R}$, Lemma \[lemma-xith\] shows that $\psi(\cdot,t)$ can be decomposed as in with modulation parameters $\xi(t),\theta(t)$ that depend continuously on time. To prove differentiability, we first consider more regular solutions for which $\psi(\cdot,t) \in Y$, where $$Y \,=\, \Bigl\{\psi \in H^4_{\rm loc}({\mathbb{R}})\,: \quad \psi_x \in H^3({\mathbb{R}}), ~1 - \|\psi\|^2 \in L^2({\mathbb{R}}) \Bigr\}.$$ For such solutions, it is not difficult to verify (by inspecting the proof of Lemma \[lemma-xith\]) that the modulation parameters are $C^1$ functions of time, so that we can differentiate both sides of and obtain from (\[nls\]) the evolution system $$\left\{\!\!\begin{array}{l} ~\,\,u_t \,=\, L_- v + \dot{\xi} (u_0' + u_x) - \dot{\theta} v + (2 u_0 u + u^2 + v^2) v, \\ -v_t \,=\, L_+ u -\dot{\xi} v_x - \dot{\theta} (u_0 + u) + (3 u_0 u + u^2 + v^2) u + u_0 v^2, \end{array} \right.$$ where the operators $L_\pm$ are defined in . Using the orthogonality conditions , we eliminate the time derivatives $u_t, v_t$ by taking the scalar product of the first line with $u_0'$ and of the second line with $u_0''$. This gives the following linear system for the derivatives $\dot{\xi}$ and $\dot{\theta}$: $$\label{Bsys} B \begin{pmatrix} \dot{\xi} \\[.5mm] \dot{\theta} \end{pmatrix} \,=\, \begin{pmatrix} \langle L_- u_0', v \rangle_{L^2} \\[.5mm] \langle L_+ u_0'', u \rangle_{L^2} \end{pmatrix} \,+\, \begin{pmatrix} \langle u_0', (2 u_0 u + u^2 + v^2) v \rangle_{L^2} \\[.5mm] \langle u_0'', (3 u_0 u + u^2 + v^2) u + u_0 v^2 \rangle_{L^2} \end{pmatrix},$$ where $$\label{BBdef} B \,=\, \begin{pmatrix} -\\| u_0' \\|^2_{L^2} & 0 \\ 0 & -\\|u_0'\\|^2_{L^2}\end{pmatrix} \,+\, \begin{pmatrix} -\langle u_0', u_x \rangle_{L^2} & \langle u_0', v \rangle_{L^2} \\[.5mm] \langle u_0'', v_x \rangle_{L^2} & \langle u_0'', u\rangle_{L^2} \end{pmatrix}.$$ As in the proof of Lemma \[lemma-xith\], it is easy to verify using that the second term in the right-hand side of is bounded by $C \epsilon$ for some positive constant $C$, hence the matrix $B$ is invertible if $\epsilon$ is small enough. Inverting $B$ in , we obtain a formula for the derivatives $\dot{\xi},\dot{\theta}$ in which the right-hand side makes sense (and is a continuous function of time) for any solution $\psi(\cdot,t) \in X$ of satisfying for all times. Since $Y$ is dense in $X$, we conclude by a standard approximation argument that the modulation parameters $\xi(t),\theta(t)$ are $C^1$ functions of time in the general case, and that their derivatives satisfy . Finally, the first term in the right-hand side of is of size $\mathcal{O}(\epsilon)$, whereas the second term is $\mathcal{O}(\epsilon^2)$, hence $\|\dot{\xi}(t)\| + \|\dot{\theta}(t)\| \le C\epsilon$ for all $t \in {\mathbb{R}}$, where the positive constant $C$ is independent of $t$. This concludes the proof. Proof of orbital stability of the black soliton {#sec:stability} =============================================== This final section is entirely devoted to the proof of Theorem \[theorem-soliton\]. As in the previous section, we consider solutions of the NLS equation of the form , where the real-valued perturbations $u,v$ satisfy the orthogonality conditions . Our main task is a detailed analysis of the functional in a neighborhood of the orbit of the soliton profile $u_0$. Instead of using the straightforward decomposition , the main idea is to express the difference $\Lambda(\psi) - \Lambda(u_0)$ in terms of the variables $u$, $v$, and $\eta$, where $\eta$ is defined in . \[lemma-soliton-3\] If $\psi = u_0 + u + iv$ satisfies $d_R(\psi,u_0) < \infty$, then $$\begin{aligned} \nonumber \Lambda(\psi) - \Lambda(u_0) \,=\, \int_{{\mathbb{R}}} \Bigl(&u_{xx}^2 + v_{xx}^2 + (3u_0^2-2) (u_x^2 + v_x^2) + (1-u_0^2)(u^2+v^2) \\ \label{Lamexp} &-3(1-u_0^2)(1-3u_0^2)u^2 + \frac12 \eta_x^2 + \frac12(3u_0^2-2) \eta^2\\ \nonumber &+\frac12 \eta^3 + 3\eta(u_x^2 + v_x^2) + 6u_0'(u^2+v^2)u_x\Bigr){\,\mathrm{d}}x~.\end{aligned}$$ We observe that $\|\psi\|^2 = u_0^2 + \eta$ and $\bar \psi \psi_x + \psi \bar \psi_x = 2u_0 u_0' + \eta_x$. Thus, if $$A(\psi) \,=\, \|\psi_{xx}\|^2 + \|\psi_x\|^2 (3 \|\psi\|^2 -2) + \frac{1}{2} (\bar{\psi} \psi_x + \psi \bar{\psi}_x)^2 + \frac{1}{2} \|\psi\|^2 (1 - \|\psi\|^2)^2$$ denotes the integrand in the functional $\Lambda = S - 2 E$, a direct calculation shows that $$\begin{aligned} \nonumber A(\psi) - A(u_0) \,=~ &\mathcal{L}(u,\eta) + 6\eta u_0' u_x + u_{xx}^2 + v_{xx}^2 + (3u_0^2-2)(u_x^2 + v_x^2) \\ \label{Aexp} &+ \frac12 \eta_x^2 + \frac12(3u_0^2-2)\eta^2 + \frac12 \eta^3 + 3\eta(u_x^2 + v_x^2),\end{aligned}$$ where $\mathcal{L}(u,\eta) = 2u_0''u_{xx} + 2(3u_0^2-2)u_0'u_x + 2u_0u_0'\eta_x + \eta(1-u_0^2)(2-3u_0^2)$. We now integrate the right-hand side of over $x \in {\mathbb{R}}$, starting with the terms $\mathcal{L}(u,\eta)$ which are linear in $u$ and $\eta$. Using the identities $u_0'' + u_0 - u_0^3 = 0$ and $u_0'''' + (1-3u_0^2)u_0'' -6u_0 u_0'^2 = 0$, we find $$\begin{aligned} 2\int_{\mathbb{R}}\Bigl(u_0''u_{xx} + (3u_0^2-2)u_0'u_x\Bigr){\,\mathrm{d}}x \,&=\, 2\int_{\mathbb{R}}\Bigl(u_0'''' - (3u_0^2-2)u_0'' - 6u_0 u_0'^2\Bigr)u{\,\mathrm{d}}x \\ \,&=\, 2\int_{\mathbb{R}}u_0'' u {\,\mathrm{d}}x \,=\, -2\int_{\mathbb{R}}(1-u_0^2)u_0 u {\,\mathrm{d}}x.\end{aligned}$$ Similarly, as $2(u_0 u_0')' = (1-u_0^2)(1-3u_0^2)$, we have $$2\int_{\mathbb{R}}u_0 u_0' \eta_x {\,\mathrm{d}}x = -2\int_R (u_0 u_0')'\eta {\,\mathrm{d}}x = - \int_{\mathbb{R}}(1-u_0^2)(1-3u_0^2)\eta d x.$$ We conclude that $$\label{Lterms} \int_{\mathbb{R}}\mathcal{L}(u,\eta){\,\mathrm{d}}x \,=\, \int_{\mathbb{R}}(1-u_0^2) (\eta - 2u_0 u){\,\mathrm{d}}x \,=\, \int_{\mathbb{R}}(1-u_0^2)(u^2+v^2){\,\mathrm{d}}x.$$ Note that is now quadratic in $u$ and $v$, which could be expected since $u_0$ is a critical point of the functional $\Lambda$. We next consider the quadratic term $6\eta u_0' u_x$ in , which has no definite sign. Using the representation , we find $6\eta u_0' u_x = 12 u_0 u_0' u u_x + 6 u_0' (u^2+v^2)u_x$, and integrating by parts, we obtain $$\label{Qterm} 6\int_{\mathbb{R}}\eta u_0' u_x{\,\mathrm{d}}x \,=\, -3\int_{\mathbb{R}}(1-u_0^2)(1-3u_0^2)u^2 {\,\mathrm{d}}x + 6 \int_{\mathbb{R}}u_0' (u^2+v^2)u_x {\,\mathrm{d}}x.$$ Now, combining , , and , we arrive at . To simplify the notations, we define $$\begin{aligned} \nonumber B_0(u) \,&=\, u_{xx}^2 + (5u_0^2-2)u_x^2 - (1-3u_0^2)u^2 - (1-u_0^2)(1-5u_0^2)u^2\\ \label{Bdef} B_1(u) \,&=\, u_{xx}^2 + (3u_0^2-2)u_x^2 + (1-u_0^2)u^2 - 3 (1-u_0^2)(1-3u_0^2)u^2\\ \nonumber B_2(v) \,&=\, v_{xx}^2 + (3u_0^2-2)v_x^2 + (1-u_0^2)v^2\\ \nonumber B_3(\eta) \,&=\, {{\textstyle}\frac12\eta_x^2 + \frac12(3u_0^2-2)\eta^2}.\end{aligned}$$ The quadratic terms in the right-hand side of can be written in the compact form $$\label{Qdef} Q(u,v,\eta) = \int_{\mathbb{R}}\Bigl(B_1(u) + B_2(v) + B_3(\eta)\Bigr){\,\mathrm{d}}x.$$ We see that $Q(u,v,\eta)$ contains $\langle K_-v,v\rangle \equiv \int_{\mathbb{R}}B_2(v){\,\mathrm{d}}x$, but not $\langle K_+u,u\rangle \equiv \int_{\mathbb{R}}B_0(u){\,\mathrm{d}}x$. Instead, it only contains $\int_{\mathbb{R}}B_1(u){\,\mathrm{d}}x$ and $\int_{\mathbb{R}}B_3(\eta){\,\mathrm{d}}x$. This discrepancy is due to that fact that the variables $u$ and $\eta$ are not independent. As $\eta = 2 u_0 u + u^2 + v^2$, the quantity $\int_{\mathbb{R}}B_3(\eta){\,\mathrm{d}}x$ also contains quadratic terms in $u$ and $u_x$, which should be added to $\int_{{\mathbb{R}}} B_1(u) dx$ to obtain $\int_{{\mathbb{R}}} B_0(u) dx$. Due to the relation between $u$ and $\eta$, it is not obvious that each quadratic term in is positive independently of the others. To avoid that difficulty, we fix some $R \ge 1$ (which will be chosen large enough below) and we split the integration domain into two regions. When $\|x\| \le R$, we replace $\eta$ by $2 u_0 u + u^2 + v^2$, and we use extensions of Lemmas \[lemma-soliton-1\] and \[lemma-soliton-2\] to prove positivity of the quadratic terms in . In the outer region $\|x\| > R$, the analysis is much simpler, because the expressions $B_1(u)$, $B_2(v)$, and $B_3(\eta)$ are obviously positive if $R$ is large enough. Since $\eta$ is a nonlinear function of $u$ and $v$, the analysis of the quadratic expression will produce higher-order terms, which will be controlled using a smallness assumption on the distance $d_R(\psi,u_0)$. To that purpose, we find it convenient to introduce the quantity $$\label{rhodef} \rho^2(u,v,\eta) \,=\, \int_{\mathbb{R}}\Bigl(u_{xx}^2 + v_{xx}^2 + u_x^2 + v_x^2 \Bigr){\,\mathrm{d}}x + \int_{\|x\|\le R} \Bigl(u^2 + R^{-2}v^2\Bigr) {\,\mathrm{d}}x + \int_{\|x\|\ge R} \Bigl(\eta_x^2 + \eta^2\Bigr){\,\mathrm{d}}x,$$ which is equivalent to the squared distance in a neighborhood of $u_0$. Indeed, we have the following elementary result: \[auxlem\] Fix $R \ge 1$, and assume that $\psi = u_0 + u + iv$, where $u,v \in H^2_{\rm loc}({\mathbb{R}})$ are real-valued. Let $d_R(\psi,u_0)$ be given by and $\rho(u,v,\eta)$ by .\ [a)]{} One has $d_R(\psi,u_0) < \infty$ if and only if $\rho(u,v,\eta) < \infty$.\ [b)]{} There exists a constant $C_0 \ge 1$ (independent of $R$) such that, if $d_R(\psi,u_0) \le 1$ or if\ $R^{1/2}\rho(u,v,\eta) \le 1$, then $$\label{rhoequiv} C_0^{-1} \rho(u,v,\eta) \le d_R(\psi,u_0) \le C_0 R \rho(u,v,\eta).$$ Throughout the proof, we denote $d_R(\psi,u_0)$ by $d_R$ and $\rho(u,v,\eta)$ simply by $\rho$. We proceed in three steps. [Step 1:]{} Assume first that $d_R < \infty$, so that $u_x,v_x \in H^1({\mathbb{R}})$, $u,v \in L^2(-R,R)$, and $\eta \in L^2({\mathbb{R}})$, where $\eta = \|\psi\|^2 - \|u_0\|^2 = 2 u_0 u + u^2 + v^2$. We claim that $u,v \in L^\infty({\mathbb{R}})$ and that $$\label{Kbound1} K \,:=\, \\|u\\|_{L^\infty({\mathbb{R}})} + \\|v\\|_{L^\infty({\mathbb{R}})} \,\le\, C(1 + d_R),$$ for some universal constant $C > 0$. Indeed, if $f = \|\psi\| - \|u_0\|$, we observe that $$d_R^2 \,\ge\, \int_{\mathbb{R}}\eta^2 {\,\mathrm{d}}x \,\ge\, \int_{\|x\|\ge 1} (\|\psi\| - \|u_0\|)^2(\|\psi\| + \|u_0\|)^2{\,\mathrm{d}}x \ge C \int_{\|x\|\ge 1} f^2 {\,\mathrm{d}}x,$$ hence $f \in L^2(I)$, where $I = \{x \in {\mathbb{R}}: \|x\| \ge 1\}$, and $\\|f\\|_{L^2(I)} \le C d_R$. Moreover, we have $\|f_x\| \le 2 u_0' + \|u_x\| + \|v_x\|$ almost everywhere, hence $f_x \in L^2({\mathbb{R}})$ and $\\|f_x\\|_{L^2({\mathbb{R}})} \le C (1+d_R)$. By Sobolev embedding, this implies that $f \in L^\infty(I)$, hence also $u,v \in L^\infty(I)$, and we have the bound $\\|u\\|_{L^\infty(I)} + \\|v\\|_{L^\infty(I)} \le C (1+d_R)$. Finally, since $\\|u_x\\|_{L^2({\mathbb{R}})} + \\|v_x\\|_{L^2({\mathbb{R}})} \le C d_R$, we conclude that $u,v \in L^\infty({\mathbb{R}})$ and that holds. [Step 2:]{} Next, we assume that $\rho < \infty$, so that $u_x,v_x \in H^1({\mathbb{R}})$, $u,v \in L^2(-R,R)$, and $\eta \in H^1(I_R)$, where $I_R = \{x \in {\mathbb{R}}: \|x\| \ge R\}$. We claim that $u,v \in L^\infty({\mathbb{R}})$ and that $$\label{Kbound2} K \,:=\, \\|u\\|_{L^\infty({\mathbb{R}})} + \\|v\\|_{L^\infty({\mathbb{R}})} \,\le\, C(1 + R^{1/2}\rho),$$ for some universal constant $C > 0$. Indeed, we know that $\eta \in L^\infty(I_R)$ with $\\|\eta\\|_{L^\infty(I_R)} \le C\rho$. This implies that $\psi \in L^\infty(I_R)$, hence also $u,v \in L^\infty(I_R)$, and that $\\|u\\|_{L^\infty(I_R)}+\\|v\\|_{L^\infty(I_R)} \le C(1+\rho)^{1/2}$. On the other hand, we know that $\\|u\\|_{L^\infty(-R,R)} \le C \\|u\\|_{H^1(-R,R)} \le C\rho$ and that $$\\|v\\|_{L^\infty(-R,R)} \,\le\, C\biggl(\frac{\\|v\\|_{L^2(-R,R)}}{R^{1/2}} + \\|v\\|_{L^2(-R,R)}^{1/2}\\|v_x\\|_{L^2(-R,R)}^{1/2}\biggr) \,\le\, CR^{1/2}\rho,$$ because $\\|v\\|_{L^2(-R,R)} \le R\rho$ and $\\|v_x\\|_{L^2(-R,R)} \le \rho$. Thus we conclude that $u,v \in L^\infty({\mathbb{R}})$ and that holds. [Step 3:]{} Finally we assume that $K = \\|u\\|_{L^\infty({\mathbb{R}})} + \\|v\\|_{L^\infty({\mathbb{R}})} < \infty$, which is the case if $d_R < \infty$ or if $\rho < \infty$. As $\eta = 2u_0u + u^2 + v^2$, we find $$\\|\eta\\|_{L^2(-R,R)} \,\le\, C(1+K) \Bigl(\\|u\\|_{L^2(-R,R)} + \\|v\\|_{L^2(-R,R)}\Bigr) \,\le\, C(1+K)R\rho,$$ because $\\|u\\|_{L^2(-R,R)} \le \rho$ and $\\|v\\|_{L^2(-R,R)} \le R\rho$. This shows that, if $\rho < \infty$, then $\eta \in L^2({\mathbb{R}})$, so that $d_R < \infty$, and we have the bound $d_R \le C(1+K)R\rho$. Conversely, since $\eta_x = 2(u_0' u + u_0 u_x + uu_x + vv_x)$, we obtain $$\\|\eta_x\\|_{L^2({\mathbb{R}})} \,\le\, C(1+K) \Bigl(\\|u\\|_{L^2(-1,1)} + \\|u_x\\|_{L^2({\mathbb{R}})} + \\|v_x\\|_{L^2({\mathbb{R}})}\Bigr) \,\le\, C(1+K)d_R,$$ where to estimate $u_0' u$ we used the fact that $\|u(x)\| \le C(\\|u\\|_{L^2(-1,1)} + (1+\|x\|)^{1/2}\\|u_x\\|_{L^2({\mathbb{R}})})$. This shows that, if $d_R < \infty$, then $\eta_x \in L^2({\mathbb{R}})$, so that $\rho < \infty$, and we have the bound $\rho \le C(1+K) d_R$. This concludes the proof. In the calculations below, to avoid boundary terms when integrating by parts in expressions such as , it is technically convenient to split the integration domain using a smooth partition of unity. Let $\chi : {\mathbb{R}}\to [0,1]$ be a smooth cut-off function such that $$\chi(x) \,=\, 1 \quad \mbox{\rm for} \quad \|x\| \le \frac{1}{2}\,, \qquad {\rm and} \qquad \chi(x) \,=\, 0 \quad \mbox{\rm for} \quad \|x\| \ge \frac{3}{2}\,.$$ We further assume that $\chi$ is even, that $\chi'(x) \le 0$ for $x \ge 0$, and that $\chi(1) = \frac{1}{2}$. Given $R \ge 1$, we denote $\chi_R(x) = \chi(x/R)$. The following estimates will be useful to control the functions $u,v$ on the support of $\chi_R'$. \[auxlem2\] Fix $R \ge 1$, and assume that $\psi = u_0 + u + iv$ satisfies $d_R(\psi,u_0) < \infty$. Then there exists a constant $C_1 > 0$ (independent of $R$) such that $$\begin{aligned} \label{locbound1} \\|u\\|_{L^2(-2R,2R)} \,&\le\, C_1 (\rho(u,v,\eta) + R^{3/2}\rho(u,v,\eta)^2), \\ \label{locbound2} \\|u\\|_{L^\infty(-2R,2R)} + \\|v\\|_{L^\infty(-2R,2R)} \,&\le\, C_1 R^{1/2}\rho(u,v,\eta),\end{aligned}$$ where $\rho(u,v,\eta)$ is given by . If $f$ is either $u$ or $v$, then $\|f(x)\| \le C(R^{-1/2}\\|f\\|_{L^2(-R,R)} + (\|x\|+R)^{1/2}\\|f_x\\|_{L^2({\mathbb{R}})})$, and this gives the bound . To prove estimate , we recall that $\\|u\\|_{L^2(-R,R)} \le \rho(u,v,\eta)$, so we only need to control $u(x)$ for $R \le \|x\| \le 2R$. In that region we have $\|u\| \le C(\|\eta\| + u^2 + v^2)$, hence using the bound and the fact that $\\|\eta\\|_{L^2(\|x\| \ge R)} \le \rho(u,v,\eta)$ we obtain the desired result. We now analyze the quadratic terms in the representation . \[lemma-soliton-4\] Under the assumptions of Lemma \[auxlem\], if $d_R(\psi,u_0) \le 1$, we have $$\label{Bident} \int_{\mathbb{R}}\Bigl(B_1(u) + B_3(\eta)\Bigr)\chi_R(x){\,\mathrm{d}}x \,=\, \int_{\mathbb{R}}B_0(u)\chi_R(x){\,\mathrm{d}}x + \mathcal{O}(R^3\rho(u,v,\eta)^3 + e^{-R}\rho(u,v,\eta)^2),$$ where the estimate in the big O term holds uniformly for $R \ge 1$. Since $\eta = 2u_0u + u^2 + v^2$, we find by a direct calculation $$B_3(\eta) \,=\, 2u_0'^2u^2 + 2 u_0^2 u_x^2 + 4 u_0 u_0' u u_x + 2(3u_0^2-2) u_0^2 u^2 + \tilde N(u,v),$$ where $$\begin{aligned} \tilde N(u,v) \,&=\, 4(uu_x + vv_x)(u_0'u + u_0u_x) + 2(uu_x + vv_x)^2 \\ &\quad + 4(3u_0^2-2)u_0u(u^2+v^2) + 2(3u_0^2-2)(u^2+v^2)^2.\end{aligned}$$ In view of the definitions , this implies that $$B_1(u) + B_3(\eta) \,=\, B_0(u) + (2u_0 u_0' u^2)_x + \tilde N(u,v).$$ If we now multiply both sides by $\chi_R(x)$ and integrate over $x \in {\mathbb{R}}$, we arrive at , because it is straightforward to verify using , and that $$-2\int_{\mathbb{R}}u_0 u_0' u^2 \chi_R'(x) {\,\mathrm{d}}x \,=\, \mathcal{O}(e^{-R} \rho(u,v,\eta)^2), \quad \hbox{and}\quad \int_{\mathbb{R}}\tilde N(u,v) \chi_R(x) {\,\mathrm{d}}x \,=\, \mathcal{O}(R^3\rho(u,v,\eta)^3).$$ This concludes the proof of the lemma. Using Lemma \[lemma-soliton-4\], we are able to derive the desired lower bound on the difference $\Lambda(\psi) - \Lambda(u_0)$ in terms of the quantity $\rho(u,v,\eta)$. \[prop-soliton\] If $R \ge 1$ is sufficiently large, there exists a constant $C_2 > 0$ such that, if $\psi = u_0 + u + iv$ satisfies $d_R(\psi,u_0) \le 1$ and if $\langle u_0',u\rangle_{L^2} = \langle u_0'',v\rangle_{L^2} = 0$, then $$\label{Lamlower} \Lambda(\psi) - \Lambda(u_0) \,\ge\, C_2 \rho(u,v,\eta)^2 + \mathcal{O}(R^3\rho(u,v,\eta)^3),$$ where the estimate in the big O term is uniform in $R$. Proceeding as in the proof of Lemma \[auxlem\], it is easy to estimate the cubic terms in in terms of $\rho(u,v,\eta)$ using, in particular, the uniform bound and the estimate . We thus find $$\label{lowbd0} \Lambda(\psi) - \Lambda(u_0) \,=\, Q(u,v,\eta) + \mathcal{O}(R^3\rho(u,v,\eta)^3),$$ where $Q(u,v,\eta)$ is given by and . Then, in the definition , we split the integral using the partition of unity $1 = \chi_R + (1-\chi_R)$ and we use Lemma \[lemma-soliton-4\]. This gives $$\begin{aligned} \nonumber Q(u,v,\eta) \,&=\, \int_{\mathbb{R}}B_2(v){\,\mathrm{d}}x + \int_{\mathbb{R}}B_0(u)\chi_R(x){\,\mathrm{d}}x \\ \label{lowbd1} &\quad + \int_{\mathbb{R}}\Bigl(B_1(u) + B_3(\eta)\Bigr)(1-\chi_R(x)){\,\mathrm{d}}x + \mathcal{O}(R^3\rho(u,v,\eta)^3 + e^{-R}\rho(u,v,\eta)^2).\end{aligned}$$ As $\langle u_0'',v\rangle = 0$, we know from Lemmas \[operator-K-minus\] and \[lemma-soliton-2\] that $$\label{lowbd2} \int_{\mathbb{R}}B_2(v){\,\mathrm{d}}x \,\ge\, C \int_{\mathbb{R}}(v_{xx}^2 + v_x^2){\,\mathrm{d}}x + \frac{C}{R^2}\int_{\|x\| \le R} v^2 {\,\mathrm{d}}x,$$ where the last term in the right-hand side follows from the bound $\|v(x)\| \le \|v(0)\| + \|x\|^{1/2}\\|v_x\\|_{L^2}$, which implies $$\int_{\|x\| \le R} v^2 {\,\mathrm{d}}x \,\le\, 4R\|v(0)\|^2 + 2R^2 \int_{\mathbb{R}}v_x^2 {\,\mathrm{d}}x \,\le\, C R^2 \int_{\mathbb{R}}B_2(v){\,\mathrm{d}}x.$$ On the other hand, if $R \ge 1$ is large enough so that $3u_0^2 - 2 \ge \frac{1}{2}$ for $\|x\| \ge R$, it is clear from that $$\label{lowbd3} \int_{\mathbb{R}}\Bigl(B_1(u) + B_3(\eta)\Bigr)(1-\chi_R(x)){\,\mathrm{d}}x \,\ge\, C \int_{\|x\| \ge R} (u_{xx}^2 + u_x^2 + \eta_x^2 + \eta^2){\,\mathrm{d}}x.$$ Finally, we estimate from below the term $\int_{\mathbb{R}}B_0(u)\chi_R(x){\,\mathrm{d}}x$ under the orthogonality assumption $\langle u_0',u\rangle_{L^2} = 0$. Arguing as in Lemma \[lemma-K-plus\] and Corollary \[lemma-K-minus\], we introduce the auxiliary variable $w = u_x + \sqrt{2}u_0 u$. After integrating by parts, we obtain the identity $$\int_{\mathbb{R}}B_0(u) \chi_R(x) {\,\mathrm{d}}x = \int_{\mathbb{R}}\Bigl(w_x^2 + w^2 \Bigr)\chi_R(x) {\,\mathrm{d}}x + J_R,$$ where $$J_R \,=\, \int_{\mathbb{R}}\Bigl(\sqrt{2}u_0 u_x^2 + 2\sqrt{2}u_0' u u_x + (2u_0 u_0'- \sqrt{2}u_0'')u^2 + \sqrt{2}u_0^2 u^2\Bigr) \chi_R'(x){\,\mathrm{d}}x.$$ Since $\chi_R'(x) = R^{-1} \chi'(x/R)$, we have using the estimate $$\|J_R\| \,\le\, \frac{C}{R} \int_{\|x\| \le 3R/2} \Bigl(u_x^2 + u^2\Bigr){\,\mathrm{d}}x \,\le\, \frac{C_3 \rho(u,v,\eta)^2}{R} + \mathcal{O}(R^2\rho(u,v,\eta)^4),$$ where $C_3 > 0$ is independent of $R$. Moreover, proceeding as in the proof of Lemma \[lemma-soliton-1\], we find $$\label{lowbdaux} \int_{\|x\| \le R} \Bigl(u_{xx}^2 + u_x^2 + u^2\Bigr){\,\mathrm{d}}x \,\le\, C \int_{\|x\| \le R} \Bigl(w_x^2 + w^2\Bigr){\,\mathrm{d}}x + \mathcal{O}(e^{-R}\rho(u,v,\eta)^2).$$ Indeed, we have the representation $u = A u_0' + W$, where the function $W$ is defined in and the constant $A$ is fixed by the orthogonality condition $\langle u_0',u\rangle_{L^2} = 0$. The proof of Lemma \[lemma-soliton-1\] shows that $\\|W\\|_{L^2(\|x\|\le R)} \le C \\|w\\|_{L^2(\|x\|\le R)}$. From the orthogonality relation $$0 \,=\, \int_{\|x\|\le R} u_0'(x)\Bigl(A u_0'(x) + W(x)\Bigr){\,\mathrm{d}}x + \int_{\|x\|\ge R} u_0'(x) u(x){\,\mathrm{d}}x,$$ we easily obtain the bound $\|A\| \le C\\|W\\|_{L^2(\|x\|\le R)} + \mathcal{O}(e^{-R} \rho(u,v,\eta))$. This shows that $$\\|u\\|_{L^2(\|x\|\le R)} \,\le\, C\\|w\\|_{L^2(\|x\|\le R)} + \mathcal{O}(e^{-R}\rho(u,v,\eta)),$$ and since $u_x = w - \sqrt{2}u_0 u$ we obtain similar estimates for the derivatives $u_x$ and $u_{xx}$, which altogether give . Summarizing, we have shown $$\begin{aligned} \nonumber \int_{\mathbb{R}}B_0(u) \chi_R(x) {\,\mathrm{d}}x \,&\ge\, C \int_{\|x\| \le R} \Bigl(u_{xx}^2 + u_x^2 + u^2\Bigr){\,\mathrm{d}}x - \frac{C_3 \rho(u,v,\eta)^2}{R} \\ \label{lowbd4} &\quad\, + \mathcal{O}(R^2\rho(u,v,\eta)^3 + e^{-R}\rho(u,v,\eta)^2),\end{aligned}$$ where in the big O term we replaced $R^2\rho(u,v,\eta)^4$ with $R^2\rho(u,v,\eta)^3$ using the fact that $\rho(u,v,\eta) \le C_0 d_R(\psi,u_0) \le C_0$ by . Now, combining , , , , , and taking $R \ge 1$ sufficiently large, we arrive at . \[Lambdafinal\] Fix any $R \ge 1$. There exist $\epsilon_1 \in (0,1)$ and $C_4 \ge 1$ such that, if $\psi = u_0 + u + iv$ satisfies $d_R(\psi,u_0) \le \epsilon_1$ and if $\langle u_0',u\rangle_{L^2} = \langle u_0'',v\rangle_{L^2} = 0$, then $$\label{Lambdaest} C_4^{-1}d_R(\psi,u_0)^2 \le \Lambda(\psi) - \Lambda(u_0) \,\le\, C_4 d_R(\psi,u_0)^2.$$ Choose $R \ge 1$ large enough so that the conclusion of Proposition \[prop-soliton\] holds, and $\rho_0 > 0$ small enough so that $R^3 \rho_0 \ll C_2$, where $C_2$ is as in . Take $\epsilon_1 \le 1$ such that $C_0\epsilon_1 \le \rho_0$, where $C_0$ is as in . If $\psi = u_0 + u + iv$ satisfies $d_R(\psi,u_0) \le \epsilon_1$ and $\langle u_0',u\rangle_{L^2} = \langle u_0'',v\rangle_{L^2} = 0$, it follows from that the quantity $\rho(u,v,\eta)$ defined in satisfies $\rho(u,v,\eta) \le \rho_0$. By Proposition \[prop-soliton\], we thus have $$\frac12 C_2 \rho(u,v,\eta)^2 \,\le\, \Lambda(\psi) - \Lambda(u_0) \le C_2' \rho(u,v,\eta)^2,$$ where the lower bound follows from , and the upper bound can be established by a much simpler argument (which does not use any orthogonality condition). Since $\rho(u,v,\eta)$ is equivalent to $d_R(\psi,u_0)$ by Lemma \[auxlem\], we obtain . Finally, Corollary \[Lambdafinal\] holds for any $R \ge 1$ because different values of $R$ give equivalent distances $d_R$ on $X$. It is now easy to conclude the proof of Theorem \[theorem-soliton\]. Fix any $R \ge 1$. Given any $\epsilon > 0$, we take $$\delta \,=\, \frac{1}{2C_4}\,\min(2\epsilon,\epsilon_0,\epsilon_1),$$ where $C_4 \ge 1$ and $\epsilon_1 > 0$ are as in Corollary \[Lambdafinal\] and $\epsilon_0 > 0$ is as in Lemma \[lemma-xith\]. If $\psi_0 \in X$ satisfies $d_R(\psi_0,u_0) \le \delta$, then $\Lambda(\psi_0) - \Lambda(u_0) \le C_4 \delta^2$ by the upper bound in , which does not require any orthogonality condition. Since $\Lambda$ is a conserved quantity, we deduce that the solution $\psi(\cdot,t)$ of the cubic NLS equation with initial data $\psi_0$ satisfies $\Lambda(\psi(\cdot,t)) - \Lambda(u_0) \le C_4 \delta^2$ for all $t \in {\mathbb{R}}$. We claim that, for all $t \in {\mathbb{R}}$, we have $$\label{inf3} \inf_{\xi, \theta \in {\mathbb{R}}} d_R\Bigl(e^{i \theta} \psi(\cdot + \xi,t), u_0\Bigr) \,\le\, 2C_4\delta \le \epsilon_0.$$ Indeed, the bound holds for $t = 0$ by assumption. Let $\mathcal{J} \subset {\mathbb{R}}$ be the largest time interval containing the origin such that the bound holds for all $t \in \mathcal{J}$. As is well-known [@Gerard; @Zhidkov], the solutions of the cubic NLS equation with initial data in $X$ depend continuously on time with respect to the distance $d_R(\psi,u_0)$. This implies that the left-hand side of the bound is a continuous function of $t$, so that $\mathcal{J}$ is closed. On the other hand, if $t \in \mathcal{J}$, then by Lemma \[lemma-xith\] we can find $\xi,\theta \in {\mathbb{R}}$ such that the function $\tilde \psi(x) = e^{i(\theta+t)}\psi(x+\xi,t)$ can be decomposed as in with $u,v$ satisfying the orthogonality conditions . Applying Corollary \[Lambdafinal\] to $\tilde \psi$, we deduce that $$C_4^{-1}d_R(\tilde\psi,u_0)^2 \le \Lambda(\tilde\psi) - \Lambda(u_0) \,=\, \Lambda(\psi_0) - \Lambda(u_0) \le C_4 \delta^2,$$ so that $d_R(\tilde\psi,u_0) \le C_4\delta$. Using again a continuity argument, we conclude that $\mathcal{J}$ contains a neighborhood of $t$. Thus $\mathcal{J}$ is open, hence finally $\mathcal{J} = {\mathbb{R}}$, so that the bound holds for all $t \in {\mathbb{R}}$. Using Lemma \[lemma-xith\], we thus obtain modulations parameters $\xi(t)$, $\theta(t)$ such that $$d_R\Bigl(e^{i(\theta(t)+t)} \psi(\cdot + \xi(t),t)\,,u_0\Bigr) \,\le\, C_4 \delta \le \epsilon, \qquad t \in {\mathbb{R}}.$$ Finally, Lemma \[difflem\] shows that the functions $\xi : {\mathbb{R}}\to {\mathbb{R}}$ and $\theta : {\mathbb{R}}\to {\mathbb{R}}/(2\pi{\mathbb{Z}})$ are continuously differentiable and satisfy the bounds . The proof of Theorem \[theorem-soliton\] is now complete. Instead of introducing the auxiliary variable $\eta$ to cure the imperfect decomposition , it would be advantageous to find a parametrization of the perturbations that fully takes into account the geometry of the functional $\Lambda$, and in particular the degeneracy of $\Lambda''(u_0)$. Near the constant solution $u_1 \equiv 1$, it is most natural to write $\psi(x,t) = (1 + r(x,t))e^{i{\varphi}(x,t)}$, where $r$ and ${\varphi}$ are real-valued functions. In that case, the usual energy function allows us to control $r$ in $H^1({\mathbb{R}})$ and ${\varphi}_x$ in $L^2({\mathbb{R}})$. In the same spirit, it is tempting to consider perturbations of the black soliton of the form $$\label{decomposition3} \psi(x,t) \,=\, (u_0(x) + r(x,t))e^{i{\varphi}(x,t)}, \quad x \in {\mathbb{R}},$$ where $r,{\varphi}$ are again real-valued functions. With this representation, we find $$\label{Lambdatry} \Lambda(\psi) - \Lambda(u_0) \,=\, \langle K_+r,r\rangle + \int_{\mathbb{R}}\Bigl(u_0^2 {\varphi}_{xx}^2 + {\varphi}_x^2)\Bigr){\,\mathrm{d}}x + \tilde N(r,{\varphi}_x),$$ where $\tilde N(r,{\varphi}_x)$ collects the higher order terms. This formula is interesting, because it is not difficult to verify that $\tilde N(r,{\varphi}_x)$ can be controlled by the quadratic terms in if $r$ is small in $H^2({\mathbb{R}})$ and ${\varphi}_x$ small in $H^1({\mathbb{R}})$. However, not all perturbations of the black soliton can be written in the form with $r,{\varphi}$ satisfying such smallness conditions, because $u_0$ vanishes at $x = 0$ in . [Acknowledgement.]{} D.P. is supported by the Chaire d’excellence ENSL/UJF. He thanks members of Institut Fourier, Université Grenoble for hospitality and support during his visit (January-June, 2014). [99]{} F. Bethuel, P. Gravejat, J.C. Saut, and D. Smets, “Orbital stability of the black soliton for the Gross–Pitaevskii equation", Indiana Univ. Math. J. [57]{} (2008), 2611–2642. F. Bethuel, P. Gravejat, and D. Smets, “Asymptotic stability in the energy space for dark solitons of the Gross–Pitaevskii equation", arXiv: 1212.5027v1 (2012), preprint. Th. Cazenave and P.-L. Lions, “Orbital stability of standing waves for some nonlinear Schrödinger equations”, Comm. Math. Phys. [85]{} (1982), 549–561. S. Cuccagna and R. Jenkins, “On asymptotic stability of $N$-solitons of the Gross–Pitaevskii equation", arXiv:1410.6887 (2014). Th. Gallay and D.E. Pelinovsky, “Orbital stability in the cubic defocusing NLS equation: I. Cnoidal periodic waves”, arXiv:1409.6453 (2014). P. Gérard, “The Gross-Pitaevskii equation in the energy space”, [Stationary and time dependent Gross-Pitaevskii equations]{}, Contemp. Math. [473]{} (AMS, Providence, RI, 2008), 129–148. P. Gérard and Z. Zhang, “Orbital stability of traveling waves for the one-dimensional Gross-Pitaevskii equation”, J. Math. Pures Appl. [91]{} (2009), 178–210. Ph. Gravejat and D. Smets, “Asymptotic stability of the black soliton for the Gross-Pitaevskii equation”, hal-01002094 (2014), preprint. M. Grillakis, J. Shatah, and W. Strauss. “Stability theory of solitary waves in the presence of symmetry. I”, J. Funct. Anal. [74]{} (1987), 160–197. Y. Martel and F. Merle, “Asymptotic stability of solitons of the gKdV equations with general nonlinearity", Math. Ann. [341]{} (2008), 391–427. J. Yang, Nonlinear Waves in Integrable and Nonintegrable Systems (SIAM, Philadelphia, 2010). M. Weinstein, “Lyapunov stability of ground states of nonlinear dispersive evolution equations”, Comm. Pure Appl. Math. [39]{} (1986), 51–67. P. Zhidkov, [Korteweg-de Vries and nonlinear Schrödinger equations]{}, Lecture Notes in Mathematics [1756]{} (Springer-Verlag, Berlin, 2001).	{ "pile_set_name": "ArXiv" }
--- abstract: 'By considering the $x$-dependence of $\pi^+$, $\pi^-$, $K^+$, $K^-$, $\Lambda$, $\bar{\Lambda}$, $p$, $\bar{p}$ hadron productions in charged lepton semi-inclusive deep inelastic scattering off nuclear target (using Fe as an example) and deuteron D target, we find that $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ are ideal to figure out the nuclear sea content, which is predicted to be different by different models accounting for the nuclear EMC effect.' author: - Baogui Lu - 'Bo-Qiang Ma' title: 'EMC effect in semi-inclusive deep-inelastic scattering process' --- \[sec:level1\]Introduction ========================== In 1982, the European Muon Collaboration (EMC) at CERN found that the structure function ratio of bound nucleon to free nucleon, in the form of $F^A_2(x,Q^2)/F^D_2(x,Q^2)$, is not consistent with the expectation by assuming that a nuclei is composed by almost free nucleons with Fermi motion correction taken into account [@EMCA; @EMC], and such phenomenon was confirmed by E139 collaboration at SLAC [[@SLAC]]{}. This discovery, which is called the nuclear EMC effect, has received extensive attention by the nuclear and hadronic physics society. Many nuclear models, such as the pion excess model [@LEST; @ME], the quark cluster model [@Jaffe; @Carlson; @Vary] and the rescaling model [@Jaffe; @FEC; @RLJ; @FECB; @ONachtmann] [et al.]{}, have been proposed to explain the data, and all these models can qualitatively describe the data in the mediate $x$ region. The inclusive deep inelastic scattering (DIS) data are expressed as ${F^A_2(x,Q^2)}/{F^D_2(x,Q^2)}$, which can be written in the naive parton model as: $$\frac{F^A_2(x,Q^2)}{F^D_2(x,Q^2)} =\frac{\Sigma_i{e_i^2\left[q_i(x,Q^2,A)+\bar{q}_i(x,Q^2,A)\right]}}{\Sigma_i{e_i^2\left[q_i(x,Q^2)+\bar{q}_i(x,Q^2)\right]}},$$ where $e_i$ denotes the charge of the partons with flavor $i$, and $q(x,Q^2)$ is the parton distribution function of a nucleon. Fig. \[f2\] shows the $F_2^A(x,Q^2)/F_2^D(x,Q^2)$ results of the cluster model, the pion excess model and the rescaling model respectively at $Q^2=5$ GeV$^2$ in the mediate $x$ region. All these models, as can be seen from Fig. \[f2\], predict similar behavior of $F_2^A(x,Q^2)/F_2^D(x,Q^2)$ at mediate $x$ region. However, the sea quark of the nuclei is differently described by the three models. In the cluster model, all sea quarks are enhanced. In the pion excess model, the sea quarks $\bar{u}$ and $\bar{d}$ are enhanced while the other quarks are reduced. However, in the rescaling model, all sea quarks are reduced in the nuclei compared with those in the free nucleon (Fig. \[sea\]). ![[]{data-label="f2"}](f2.EPS){width="300pt"} ![. The target nuclei assumed here is Fe.[]{data-label="sea"}](sea.EPS){width="350pt"} The Fermilab experiment 772 [@DMA] measured the dimuon yield in Drell-Yan process induced by 800 GeV proton off various nucleus and compared the data with the theoretical predictions of the three models accounting for the EMC effect [@DMA]. The data were explained in Ref. [@DMA] to favor a conclusion that the sea quark in the nuclei is not enhanced, by neglecting the energy loss effect of the incident quark, which is not precisely determined yet [@Mav; @Mbj; @Gtg]. To avoid the uncertainties concerning the sea quarks in the nuclei by the dimuon yield in Drell-Yan process solely, the sea content can be also measured in other experiments. The purpose of this work is to show that the semi-inclusive hadron productions in charged lepton deep inelastic scattering are sensitive to the sea quark content of the nuclei. The pion, kaon, proton and antiproton productions in the charged lepton semi-inclusive deep inelastic scattering off nuclei have been checked by the HERMES collaboration [@Hermes]. However, while the HERMES data, which are expressed as multiplicity ratios and also with $z$-dependence, are convenient to study the modification of the fragmentation functions in nuclear environment [@Hermes], they are not ideal to provide much information about the sea quarks of the nuclei. We will show that the $\bar{\Lambda}/{\Lambda}$ production, especially the $x$ dependent behavior, is ideal to distinguish between different predictions on the sea content of the nuclei in the different models of the EMC effect. Nuclear models and the sea quark distributions ============================================== The semi-inclusive hadron productions in charged lepton deep inelastic scattering can be related with the quark distribution functions as: $$\frac{d^3\sigma^{h} }{dxdydz}=\frac{4\pi\alpha s}{Q^4}(1+(1-y)^2) \sum_ie_i^2\left[q_i(x,Q^2)D_{q_i}^{h}(z,Q^2)+\bar{q}_i(x,Q^2)D_{\bar{q}_i}^{h}(z,Q^2) \right ],$$ where $q_i(x,Q^2)$ is the parton distribution for quarks with flavor $i$, and $D_{q_i}^h(z,Q^2)$ is the fragmentation function of quark $q_i$ to hadron $h$. The formula is also applicable to the nuclei, with the parton distributions and fragmentation functions replaced by $q(x,Q^2){\rightarrow}q(x,Q^2,A)$ and $D(z,Q^2){\rightarrow}D(z,Q^2,A)$ respectively. The inclusive production itself can not offer enough information about the sea quark enhancement, while the ratio $$\frac{d\sigma^{h}_A/dx}{d{\sigma}^{h}_D/dx}=\frac{\int_{a}^{b} dz{\sum}_ie_i^2(q_i^A(x,Q^2)D_{q_i}^{h}(z,Q^2,A)+{\bar{q}}_i^A(x,Q^2) D_{{\bar{q}}_i}^{h}(z,Q^2,A))}{\int_{a}^{b}dz{\sum}_ie_i^2(q_i^D(x,Q^2) D_{q_i}^{h}(z,Q^2)+{\bar{q}}_i^D(x,Q^2)D_{{\bar{q}}_i}^{h}(z,Q^2))},$$ is useful to reveal the difference between the sea quark behavior in the nuclei and that in the nucleon. In the following, we will consider the ratios $(d\sigma^{h}_A/dx)/(d{\sigma}^{h}_D/dx)$ for various hadrons. We use the pion excess model, the rescaling model and the pion excess model to calculate the sea quark content of the nuclei. For the sea quark distributions in the deuteron D, we use the result offered by the model itself [@Carlson] for the cluster model, and for the other two models we adopt the CTEQ5L parametrization [@cteq] of parton distributions for free nucleons by considering the isospin symmetry between proton and neutron. In the pion excess model, the quark distribution in the nuclei is modified by the extra pions caused by the interaction between the nucleons in nuclei [@LEST]. The quark distribution of nuclei is: $$q_i^A(x)=\int_x^1{\frac{dy}{y}f_\pi^A(y)q_i^\pi(\frac{x}{y})}+\int_x^1{\frac{dy}{y}f_N^A(y)q_i^N(\frac{x}{y})},$$ in which $q_i^\pi(x)$ and $q_i^N(x)$ are the parton distributions in the free pion and in the free nucleon respectively, and $f_\pi(y)$ is the probability to find extra pions in the nuclei [@ME]. For simplicity, we adopt the parametrization in a toy model [@ELB], in which the proton is supposed to be partially in the nucleon-pion subsystem state and the parton distributions in the nucleon and in the pion are assumed to be the same as those in the free nucleon and in the free pion. Thus, the excess pion and the nucleon probabilities per nucleon are given as [@ELB]: $$\begin{aligned} f_{\pi}^A(y) &=& \langle{n_{\pi}}\rangle\frac{\Gamma(a+b+2)}{\Gamma(a+1)\Gamma(b+1)}y^a(1-y)^b, \\ f_N^A(z) &=& (1-\langle{n_{\pi}}\rangle)\delta(z-1)+f_{\pi}^A(1-z),\end{aligned}$$ where $n_{\pi}=0.22$, $a=1$ and $b=3$. The CTEQ5L [@cteq] parametrization of the parton distribution of the nucleon and MRS [@MRS] parametrization of the parton distribution of the pion are adopted to obtain $F_2^A(x)$ and the parton distribution of the nuclei. In the quark cluster model, six or more quark cluster is supposed to exist in the nuclei to account for the EMC effect. For the sake of simplicity, only six quark cluster is considered here. $q(x)$ in a six-quark cluster can not be measured directly from experiment, but Carlson and Havens [@Carlson] estimated quark distribution q(x) per nucleon in the six quark cluster based on QCD counting rules: $$\begin{aligned} v_6(x) &=& N_vz^{(-1/2)}(1-z)^{10}, \\ \bar{u}_6(x) &=& (N_{sea}/4)z^{(-1)}(1-z)^{14}, \end{aligned}$$ where $N_v=1.3875$ and $N_{sea}=0.2521$ are the coefficients to warrant momentum conservation and the quark number. In the six quark cluster, $x$, the variable defined by $Q^2/(2M_N\nu)$, equals to $2z$ because $z$ is defined by $Q^2/(2M_6\nu)$ [@Carlson]. Therefore, $q^A(x)$ and $F_2^A(x,Q^2)/F_2^N(x,Q^2)$ can be given as: $$\begin{aligned} q^A(x) &=& (1-f)q^N(x)+fq^6(x), \\ \frac{F_2^A(x,Q^2)}{F_2^N(x,Q^2)}&=& (1-f)+f\frac{F_2^6(x,Q^2)}{F_2^N(x,Q^2)},\end{aligned}$$ where $f$ is the probability to find the six-quark cluster in the nuclei and is adjusted to fit the inclusive deep inelastic $e(\mu)$-$A$ scattering data. Its value is given as 0.30 by Carlson and Havens [@Carlson]. For the rescaling model, the quark in the nuclear medium is considered to have different confinement size compared with that of the quark in the free nucleon. $q^A(x,Q^2)$ is related with $q^N(x,Q^2)$ (the parton distribution in the free nucleon) by the relation: $$q^A(x,Q^2)=q^N(x,\xi(Q^2)Q^2),$$ where $\xi(Q^2)$ varies with $A$ and $Q^2$. $\xi$ equals to $1.83$ [@FECB] while $Q^2=5$ GeV$^2$ and A=56 (Fe). For $q^N(x,Q^2)$, the parton distribution per nucleon, CTEQ5L parametrization [@cteq] is adopted. Given the above analysis, $F^A_2(x,Q^2)/F^D_2(x,Q^2)$ and sea quark enhancement are checked in Fig. \[f2\] and Fig. \[sea\]. Fragmentation function and parton energy loss model =================================================== Due to the non-perturbative nature of the fragmentation process, the fragmentation function can not be calculated from first principle, thus models are used to obtain the fragmentation function. Experimentally, the process of $e^+ + e^- \rightarrow h+X$ can offer much information about the fragmentation [@pdg]. Based on the experimental data and theoretical analysis, Kretzer [@Kre] gave a parametrization of the fragmentation in the form: $$D_a^h(x,Q_0^2)=Nx^\alpha(1-x)^\beta,$$ where $N$, $\alpha$ and $\beta$ are the constants chosen to fit the experimental data. For specified hadron, $\alpha$, which determines the low $z$ region behavior of the fragmentation function, is the same for all light flavor quarks while $\beta$, related with the high $z$ region behavior of the fragmentation function, is different for various quarks. Therefore, in low $z$ region, all quark fragmentation functions have the same shape, and in the large $z$ region, the favored quark fragmentation function is larger than the unfavored quark fragmentation function. In addition, because of the strange dominance at large $x$ for $s$ in $K^-$, the fragmentation function such as $s\rightarrow K^-$ is larger than that of $\bar{u}\rightarrow K^-$. As for $\Lambda,\bar{\Lambda},p,\bar{p}$, there exists a phenomenological parametrization [@Ma] of their fragmentation functions based on the assumption that the fragmentation function of quark $q$ to hadron $h$ is proportional to the $q$ quark distribution in the hadron $h$: $$D_q^h(z)\propto q^h(z).$$ In general, the fragmentation functions can be written as: $$\begin{aligned} D_V^h(z) &=& C_V(z)z^\alpha q_V^h(z), \\ D_S^h(z) &=& C_S(z)z^\alpha q_S^h(z),\end{aligned}$$ where $D_S^h(z)$ means unfavored fragmentation function. There are three options for the favored and unfavored quark fragmentations: (1). $C_V=1$ and $C_S=0$ for $\alpha=0$; (2). $C_V=1$ and $C_S=1$ for $\alpha=0.5$; (3). $C_V=1$ and $C_S=3$ for $\alpha=1$. The parton distributions of $\Lambda$ and $\bar{\Lambda}$ are essential to get the fragmentation functions of quark to $\Lambda$ and $\bar{\Lambda}$. Based on the fact that there is no direct parton distribution of $\Lambda$ and $\bar{\Lambda}$, SU(3) symmetry between the proton and the $\Lambda$ is adopted to get the parton distribution in $\Lambda$ [@Ma]. Although there are three models, SU(3) symmetry model, quark-diquark model and pQCD based analysis, the difference of the fragmentation functions will not affect the qualitative predictions by the fact that only the flavor structure of the parton distributions of the proton are different in the three models in the high $x$ region [@MaB]. The common feature of the above parameterizations of fragmentation functions is that the favored quark fragmentation is larger than the unfavored quark fragmentation in the large $z$ region. Thus the favored quark fragmentation process is able to obtain sea quark information of the nuclei if the produced hadron is from the favored fragmentation of sea quarks in the nuclei. Then the hadron events of produced particles in the large $z$ region can be chosen to get the $x$ dependence of the production ratio, in which the unfavored fragmentation contribution from the valence quark can be largely suppressed. Fragmentation function in the nuclei is important for producing the hadron production from the nuclei. HERMES collaboration has measured the hadron production from the nuclei and found that the production is reduced compared with that from the free nucleon [@Hermes], and many effects such as nuclear absorption [@bialas], parton energy loss [@XNW; @XG], gluon bremsstrahlung [@kopeli] and partial deconfinement [@RLJ; @FECB; @ONachtmann; @partial] have been developed to account for the data. In this paper, the parton energy loss model is adopted to get the fragmentation function in the nuclei. In Refs. [@XNW; @XG], the modification of the fragmentation function is caused by the interaction between the hard quark and the debris of the nuclei. Given that the original parton energy loss model is complicated to apply in the real process, an effective model suggested in Ref. [@Xwn] is used in Ref. [@Fran] to get the modification of the fragmentation function. In effective parton energy loss model, modified fragmentation function is expressed in the form: $$zD^h_q(z,Q^2,A)=\int_0^{(\nu-E_h)}d\epsilon{D(\epsilon,\nu)}z^D^h_q(z^,Q^2),$$ where $E_h=\mu-\epsilon$. $E_h$ is the measured hadron energy and $\epsilon$ is the energy loss of the hard quark going through the nuclei. $z^$ is the rescaled momentum fraction caused by the quark energy shift in presence of QCD medium: $$z^=\frac{E_h}{1-\left( \frac{\epsilon}{\nu}\right ) }.$$ $D(\epsilon,Q^2)$, the probability for a quark with energy $E=\nu$ to lose energy $\epsilon$, is parameterized by Arleo [@Arleo]: $$D(\epsilon)=\frac{1}{\sqrt{2\pi}\sigma\epsilon}{\mathrm{exp}}[-\frac{(\mathrm{log}(\epsilon/\omega_c)-\mu)^2}{2\sigma^2}],$$ where $\mu$, $\sigma$ are two parameters with $\sigma=0.73$ and $\mu=-1.5$, as the energy of the quark, which has absorbed the virtual photon, is much higher than the energy loss when it passes through the nuclei environment. And $\omega_c$ is the relevant scale of the typical gluon energy and denotes the energy loss scale of the hard quark, $$\omega_c=\frac{1}{2}{\hat{q}}L^2.$$ Although $\hat{q}$ is not precisely determined yet, in the next section we will show that the result is not sensitive on $\hat{q}$ when $Q^2$ is large enough. Here we set $\hat q=0.72\ {\rm GeV}/{\rm fm}^2$ and $L=3/4R$, where $R$ is the nuclear radius [@Fran]. From another point of view, the modification of the fragmentation function in the parton energy loss model is from assumption. Its conformation with the theoretical framework of factorization and renormalization is not fully justified, as the definition of the fragmentation functions are vacuum matrix elements with no relation to the target material. As we will find, the qualitative conclusion of our paper on the ratio $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ will not be influenced by including the modification of the fragmentation function in the nuclear environment. In order to get rid of the nuclear absorption or the energy loss process, it would be fine to consider hadron production at larger energy (say larger than 20 GeV), hence at higher virtuality, where we know from the present HERMES data that these two effects prove negligible. Results ======= Fig. \[pion\] and Fig. \[kplus\] present the results of $\pi^+(u\bar{d})$, $\pi^-(\bar{u}d)$ and $K^+$ respectively. They show that there is no large difference between various model predictions in the large $z$ region and in the low $z$ region, no matter by including the favored quark fragmentation process only or by including all favored and unfavored fragmentation processes. The reason is that $\pi^+(u\bar{d})$, $\pi^-(\bar{u}d)$ and $K^+$ are contributed mainly by the favored fragmentation processes $u\rightarrow \pi^+$, $d\rightarrow \pi^-$ and $u\rightarrow K^+$ because that the valence quarks $u$ and $d$ are predominant over the sea quarks in the mediate $x$ region and that production of those hadrons are dominated by the behavior of the valence quark in the nuclei (Fig. \[uv\]). ![[]{data-label="pion"}](pion.EPS){width="350pt"} ![[]{data-label="kplus"}](kplus.EPS){width="350pt"} ![[]{data-label="uv"}](uv.EPS){width="300pt"} ![[]{data-label="kminus"}](kminus.EPS){width="350pt"} ![[]{data-label="lambda"}](lambda.EPS){width="350pt"} ![[]{data-label="lambdabar"}](lambdabar.EPS){width="400pt"} ![[]{data-label="ratio"}](ratio.EPS){width="350pt"} ![[]{data-label="antiproton"}](antiproton.eps){width="350pt"} ![[]{data-label="antiproton25"}](antiproton25.eps){width="350pt"} ![[]{data-label="ratiop"}](ratiop.eps){width="350pt"} ![[]{data-label="lambda25"}](lambda25.eps){width="350pt"} ![[]{data-label="lambdabar25"}](lamdabar25.eps){width="400pt"} ![[]{data-label="ratio25"}](ratio25.eps){width="350pt"} When focused on $K^-$, the result with only favored quark fragmentation functions is different from that when all fragmentation processes are considered (Fig. \[kminus\]), which indicates that the unfavored quark fragmentation function, $u$ and $d$ to $K^-$, can not be neglected both at low and high $z$ region for the predominance of the valence quarks in the $x$ region we considered. In the high $z$ region, due to parton energy loss, the $K^-$ is largely suppressed and we can hardly see any difference in the three model predictions with all fragmentation processes being considered. $\Lambda$ and $\bar{\Lambda}$ production ratios in different $z$ regions are examined and the results are plotted in Fig. \[lambda\] and Fig. \[lambdabar\]. For the same reason as $K^+$, three models predict almost the same $x$-dependence of $\Lambda^A/\Lambda^D$. From Fig. \[lambdabar\], difference are generated among various model predictions on the $x$-dependence of $\bar{\Lambda}^A/\bar{\Lambda}^D$, and the these difference are not sensitive to the three options of the fragmentation functions [@Ma]. The reason is that the dominant production of $\bar{\Lambda}$ is through the favored fragmentation of anti-quarks inside the targets, so that the $x$-dependence of the production ratio is sensitive to the sea quark behaviors of the nuclei. But in the large $z$ region, due to the parton energy loss effect, the large difference between three models do not manifest themselves significantly as expected. The reason is, at large $z$ region, the fragmentation function is largely modified by the parton energy loss. And such a phenomenon also happens on $\Lambda^A/\Lambda^D$. Fig. \[lambdabar\] shows that $\bar{\Lambda}^A/\bar{\Lambda}^D$ is not an ideal variable to figure out the nuclear sea quark content by including the modification of the fragmentation function in the nuclear environment. Fortunately, largely difference between three models appears for the quantity $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$, which is more accessible in experiment than $\bar{\Lambda}^A/\bar{\Lambda}^D$. When the integral upper limit is fixed with unity and the lower limit varies from 0.4 to 0.7, $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ is still model dependent (Fig. \[ratio\]). Thus it is plausible to conclude that $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ can offer information about the sea content of the nuclei. Similar to $\bar{\Lambda}^A/\bar{ \Lambda}^D$, the antiproton and proton production ratio $({\bar{p}}^A)/({\bar{p}}^D)$ can not offer much information while ${\bar{p}}^A/{p}^A/{\bar{p}}^A/p^A$ do generate large difference with different nuclear model (Figs. \[antiproton\]-\[ratiop\]). So, the ratio $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ is another choice to check the sea content of the nuclei in experiment. Attention should be paid to extract possible background contribution as a large number of protons and antiprotons might be produced from the decays of other baryons. $\hat{q}$ is a sensitive parameter that could largely affect the modification of fragmentation function and is not determined clearly yet. We also calculate $\Lambda^{A}/\Lambda^{D}$, $\bar{\Lambda}^A/\bar{\Lambda}^D$ and $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ when $\hat q=0.25\ {\rm GeV}/{\rm fm}^2$ (Figs. \[lambda25\]-\[ratio25\]). From the figures we can conclude that $\Lambda^{A}/\Lambda^{D}$ and $\bar{\Lambda}^A/\bar{\Lambda}^D$ are largely affected by different $\hat{q}$, while $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ is almost $\hat{q}$ independent. And $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ also has such properties(Figs. \[antiproton25\]-\[ratiop\]). So, the ratio $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ are not sensitive to $\hat{q}$. Besides, we should mention that the fixed order calculation used here is not appropriate to describe hadron production at high $z$ where large logarithms need to be resumed. Therefore we should consider the results here as qualitative predictions rather than quantitative ones. For more convinced quantitative predictions, we would need better constrained fits of the nuclear parton distributions, rather than the earlier EMC model results adopted in this paper. Summary ======= In this paper, we adopted three models of the nuclear EMC effect: the cluster model, the rescaling model and the pion excess model, to calculate their predictions on the hadron production ratio in charged lepton semi-inclusive deep inelastic scattering off nuclei in the large $z$ region. Our purpose is to find hadrons which are produced mainly from the sea quarks of nucleus, so that we can distinguish between different predictions on the sea content of the nuclei. For completeness, we considered the production ratios of $\pi^{+A}/\pi^{+D}$, ${\pi^{-A}/\pi^{-D}}$, $K^{+A}/K^{+D}$, $K^{-A}/K^{-D}$, $\Lambda^{A}/\Lambda^{D}$, $\bar{\Lambda}^A/\bar{\Lambda}^D$,$(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$, $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ and found that the ratios of $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ are ideal to figure out the sea content of the nuclei. More significantly, $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ are accessible in experiment and the behaviors of $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ are different for different models. According to Fig. \[ratio\] and Fig. \[ratiop\] , we conclude that the various models about the EMC effect with different sea behaviors can be distinguished by the future data of the $x$-dependence of $(\bar{\Lambda}^A/\Lambda^A)/(\bar{\Lambda}^D/\Lambda^D)$ and $({\bar{p}}^A/{p}^A)/({\bar{p}}^A/p^A)$ in semi-inclusive deep inelastic scattering process. The difference between the pion excess model and cluster model is not good enough to be checked out in experiment, but whether the sea quark is enhanced or not is clear to be distinguished according to the result. [Acknowledgements]{} This work is partially supported by National Natural Science Foundation of China (Nos. 10421503, 10575003, 10528510), by the Key Grant Project of Chinese Ministry of Education (No. 305001), and by the Research Fund for the Doctoral Program of Higher Education (China). [99]{} J.J. Aubert, et al., Phys. Lett. B 105,322 (1982). CERN NA2/EMC, J.J. Aubert, et al., Phys. Lett. B 123, 275 (1983). E139, R.G. Arnold, et.al., Phys. Rev. Lett 52, 727 (1984). C.H. Llewellyn Smith, Phys. Lett. B 128, 107 (1983). M. Ericson and A.W. Thomas, Phys. Lett. B 128, 112 (1983). R.L Jaffe, Phys. Rev. Lett. 50, 228 (1983). C.E Carlson, T.J. Havens, Phys. Rev. Lett. 51, 261 (1983). H.J. Pirner and J.P. Vary, Phys. Rev. Lett. 46, 1376-1379 (1981) F.E. Close, R.G. Roberts, and G.C. Ross, Phys. Lett. B 129, 346 (1983). R.L. Jaffe, et al., Phys. Lett. B 134, 449 (1984). F.E. Close, R.L. Jaffe, and R.G. Roberts, and G. G. Ross, Phys. Rev. D 31, 1004 (1985). O. Nachtmann and H. J. Pirner, Zeit. Phys. C 21, 277 (1984). D.M. Alde, et. al., Phys. Rev. Lett. 64, 2479 (1990). M.A. Vasiliev, et. al., Phys. Rev. Lett. 83, 2304 (1999). M.B. Johnson, et. al., Phys. Rev. Lett. 86, 4483 (2001). G.T. Garvey and J.C. Peng, Phys. Rev. Lett. 90, 092302 (2003). HERMES Collaboration, A. Airapetian, et. al., Phys. Lett. B 577, 37 (2003). H.L. Lai, et. al., Eur. Phys. J. C 12, 375 (2000). E.L. Berger, F. Coester, and R.B Wiringa, Phys. Rev. D 29, 398 (1983). P.J. Sutton, A.D. Martin, R.G. Roberts, and W.J. Stirling, Phys. Rev. D 45, 2349 (1992). O. Biebel, P. Nason, and B.R. Webbler, hep-ph/0109282. S. Kretzer, Phys. Rev. D 62, 054001 (2000). B.-Q. Ma, I. Schmidt, and J.-J. Yang, Phys. Lett. B 574, 35 (2003); B.-Q. Ma, I. Schmidt, J. Soffer, and J.-J. Yang, Phys. Lett. B 547, 245 (2002). B.-Q. Ma, I. Schmidt, and J.-J. Yang, Phys. Lett. B 477, 107 (2000). A. Bialas and T. Chmaj, Phys. Lett. B 133, 241 (1983); A. Bialas and M. Gyulassy, Nucl. Phys. B291, 793 (1987). E. Wang and X.N. Wang, Phys. Rev. Lett. 89, 162301 (2002). X. Guo and X.N. Wang, Phys. Rev. Lett. 85, 3591 (2000); X.N. Wang and X. Guo, Nucl. Phys. A 696, 788 (2001); J.A. Osborne, E. Wang, and X.N. Wang, Phys. Rev. D 67, 094022 (2003). B. Kopeliovich, J. Nemchik and E. Predazzi, nucl-th/9607036 A. Accardi and H.J. Pirner, Nucl. Phys. A 711, 264 (2002). X.N. Wang, Z. Huang, and I. Sarcevic, Phys. Rev. Lett. 77, 231 (1996). F. Arleo, Eur. Phys. J. C 30, 213 (2003). F. Arleo, JHEP 0211, 044 (2002).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study two methods for computing network features with topological underpinnings: the Rips and Dowker persistent homology diagrams. Our formulations work for general networks, which may be asymmetric and may have any real number as an edge weight. We study the sensitivity of Dowker persistence diagrams to asymmetry via numerous theoretical examples, including a family of highly asymmetric cycle networks that have interesting connections to the existing literature. In particular, we characterize the Dowker persistence diagrams arising from asymmetric cycle networks. We investigate the stability properties of both the Dowker and Rips persistence diagrams, and use these observations to run a classification task on a dataset comprising simulated hippocampal networks. Our theoretical and experimental results suggest that Dowker persistence diagrams are particularly suitable for studying asymmetric networks. As a stepping stone for our constructions, we prove a functorial generalization of a theorem of Dowker, after whom our constructions are named.' address: - 'Department of Mathematics, The Ohio State University. Phone: (614) 292-6805.' - 'Department of Mathematics and Department of Computer Science and Engineering, The Ohio State University. Phone: (614) 292-4975, Fax: (614) 292-1479.' author: - Samir Chowdhury and Facundo Mémoli bibliography: - 'biblio.bib' title: A functorial Dowker theorem and persistent homology of asymmetric networks --- Introduction ============ Networks are used throughout the sciences for representing the complex relations that exist between the objects of a dataset [@newman2003structure; @kleinberg-book]. Network data arises from applications in social science [@kumar2010structure; @kleinberg-book], commerce and economy [@elliott2014financial; @kleinberg-book; @acemoglu2015systemic], neuroscience [@sporns2011networks; @sporns2012discovering; @sporns2004motifs; @rubinov2010complex; @pessoa2014understanding], biology [@barabasi2004network; @huson2010phylogenetic], and defence [@masys2014networks], to name a few sources. Networks are often directed, in the sense that weights attached to edges do not satisfy any symmetry property, and this asymmetry often precludes the applicability of many standard methods for data analysis. Network analysis problems come in a wide range of flavors. One problem is in exploratory data analysis: given a network representing a dataset of societal, economic, or scientific value, the goal is to obtain insights that are meaningful to the interested party and can help uncover interesting phenomena. Another problem is network classification: given a “bag" of networks representing multiple instances of different phenomena, one wants to obtain a clustering which groups the networks together according to the different phenomena they represent. Because networks are often too complex to deal with directly, one typically extracts certain invariants of networks, and infers structural properties of the networks from properties of these invariants. While there are numerous such network invariants in the existing literature, there is growing interest in adopting a particular invariant arising from persistent homology [@frosini1992measuring; @robins1999towards; @edelsbrunner2002topological; @zomorodian2005computing], known as a persistence diagram, to the setting of networks. Persistence diagrams are used in the context of finite metric space or point cloud data to pick out features of significance while rejecting random noise [@edelsbrunner2002topological; @carlsson2009topology]. Since a network on $n$ nodes is regarded, in the most general setting, as an $n\times n$ matrix of real numbers, i.e. as a generalized metric space, it is conceivable that one should be able to describe persistence diagrams for networks as well. The motivation for computing persistence diagrams of networks is at least two-fold: (1) comparing persistence diagrams has been shown to be a viable method for shape matching applications [@frosini1992measuring; @frosini1999size; @collins2004barcode; @bot-stab; @carlsson2005persistence; @dgh-pers], analogous to the network classification problem described above, and (2) persistence diagrams have been successfully applied to feature detection, e.g. in detecting the structure of protein molecules (see [@krishnamoorthy2007topological; @xia2014persistent] and [@edelsbrunner2002topological §6]) and solid materials (see [@hiraoka2016hierarchical]) and might thus be a useful tool for exploratory analysis of network datasets. We point the reader to [@ghrist2008barcodes; @edelsbrunner2008persistent; @carlsson2009topology; @zigzag; @weinberger2011persistent; @burghelea2013topological; @dey2014computing] for surveys of persistent homology and its applications, and some recent extensions. Some extant approaches that obtain persistence diagrams from networks assume that the underlying network data actually satisfies metric properties [@lee2011computing; @khalid2014tracing]. A more general approach for obtaining persistence diagrams from networks is followed in [@horak2009persistent; @carstens2013persistent; @giusti2015clique; @petri2013topological], albeit with the restriction that the input data sets are required to be symmetric matrices. Our chief goal is to devise notions of persistent homology that are directly applicable to asymmetric networks in the most general sense, and are furthermore capable of absorbing structural information contained in the asymmetry. Contributions and an overview of our approach --------------------------------------------- \(N) at (0,0)[${\mathcal{N}}$]{}; (F) at (3,0)[${\mathcal}{F}$]{}; (D) at (6,0)[${\operatorname{Dgm}}$]{}; \(N) edge\[loop above, out=150, in=90, looseness=8, ->\] node\[left\][${\mathfrak}{s}$]{} (N); (N) edge\[loop above, out=210, in=270, looseness=8, ->\] node\[left\][${\mathfrak}{t}$]{} (N); \(N) edge\[->, bend left, looseness =1.5\] node\[above\][${\mathfrak}{R}$]{} (F); (N) edge\[->\] node\[above\][${{\mathfrak}{D}^{\operatorname{si}}}$]{} (F); (N) edge\[->, bend right, looseness=1.5\] node\[above\][${{\mathfrak}{D}^{\operatorname{so}}}$]{} (F); \(F) edge\[->\] node\[above\][$H_k$]{} (D); In this paper, we study two types of persistence diagrams: the Rips and Dowker diagrams. We define both invariants in the setting of asymmetric networks with real-valued weights, without assuming any metric properties at all (not symmetry and not even that the matrix representing the networks weights vanishes on the diagonal). As a key step in defining the Dowker persistence diagram, we first define two dual constructions, each of which can be referred to as a Dowker persistence diagram, and then prove a functorial Dowker theorem which implies that these two possible diagrams are equivalent. Following the line of work in [@dgh-pers], where stability of Rips persistence diagrams arising from finite metric spaces was first established, we formulate similar stability results for the Rips and Dowker persistence diagrams of a network. Through various examples, in particular a family of cycle networks, we espouse the idea that Dowker persistence diagrams are more appropriate than Rips persistence diagrams for studying asymmetric networks. We test our methods by solving a network classification problem on a database of simulated hippocampal networks. The first step in constructing a persistence diagram from a network is to construct a nested sequence of simplicial complexes, i.e. a simplicial filtration, which, in our work, will be the Rips or Dowker filtrations associated to a network. Rips and Dowker simplicial complexes and their associated filtrations are classically defined for metric spaces [@de2004topological; @ghrist-eat], and the generalization to networks that we use is a natural extension of the metric versions. After producing the simplicial filtrations, the standard framework of persistent homology takes over, and we obtain the Rips or Dowker persistence diagrams. Practitioners of persistent homology might recall that there are two Dowker complexes [@ghrist-eat p. 73], which we describe as the source and sink Dowker complexes. A subtle point to note here is that each of these Dowker complexes can be used to construct a persistence diagram. A folklore result in the literature about persistent homology of metric spaces, known as Dowker duality, is that the two persistence diagrams arising this way are equal [@chazal2014persistence Remark 4.8]. In this paper we prove a stronger result—a functorial Dowker theorem—from which the duality follows easily. Furthermore, the context of this result is strictly more general than that of metric spaces (see below for a more thorough description of the functorial version of Dowker’s theorem). Providing a construction of Rips and Dowker persistence diagrams is not enough: in order for these invariants to be useful in practice, one must verify that the diagrams are stable. In this context, stability means the following: the dissimilarity between two Rips (resp. Dowker) persistence diagrams obtained from two networks should be bounded above by a function of the dissimilarity between the two networks. To our knowledge, stability is not addressed in the existing literature on producing persistence diagrams from networks. In our work, we provide stability results for both the Rips and Dowker persistence diagrams (Propositions \[prop:rips-stab\] and \[prop:dowker-stab\]). One key ingredient in our proof of this result is a notion of network distance that follows previous work in [@clust-net; @nets-allerton; @nets-icassp]. This network distance is analogous to the Gromov-Hausdorff distance between metric spaces, which has previously been used to prove stability results for hierarchical clustering [@carlsson2008persistent; @clust-um] and Rips persistence diagrams obtained from finite metric spaces [@dgh-pers Theorem 3.1]. The Gromov-Hausdorff distance was later used in conjunction with the Algebraic Stability Theorem of [@chazal2009proximity] to provide alternative proofs of stability results for Rips and Dowker persistence diagrams arising from metric spaces [@chazal2014persistence]. Our proofs also involve this Algebraic Stability Theorem, but the novelty of our approach lies in a reformulation of the network distance (Proposition \[prop:dn-ko\]) that yields direct maps between two networks, thus passing naturally into the machinery of the Algebraic Stability Theorem (without having to define auxiliary constructions such as multivalued maps, as in [@chazal2014persistence]). A crucial issue that we point out in this paper is that even though we can construct both Rips and Dowker persistence diagrams out of asymmetric networks, Rips persistence diagrams appear to be blind to asymmetry, whereas Dowker persistence diagrams do exhibit sensitivity to asymmetry. In the case of Rips complexes, this purported insensitivity to asymmetry can be immediately seen from its definition. In the case of Dowker complexes, we argue about its sensitivity to asymmetry in two different ways. Firstly, we do so by explicitly computing Dowker persistence diagrams of multiple examples of asymmetric networks. In particular, we consider a family of highly asymmetric networks, the cycle networks, and by bulding upon results from [@adamaszek2015vietoris; @adamaszek2016nerve] we prove a complete characterization result for the Dowker persistence diagrams—across all dimensions—of any network belonging to this family. These networks constitute directed analogues of circles and may be motifs of interest in different applications related to network data analysis. More specifically, appearance of nontrivial 1-dimensional persistence in the Dowker persistence diagram of asymmetric network data may suggest the presence of directed cycles in the data. Some of our experimental results suggest that the Rips persistence diagrams of this family of networks are pathological, in the sense that they do not represent the signatures one would expect from the underlying dataset, which is a directed circle. Dowker persistence diagrams, on the other hand, are well-behaved in this respect in that they succeed at capturing relevant features. Secondly, we study the degree to which Dowker persistence diagrams are insensitive to changes (such as edge flips, or transposition) in the network structure. An overview of this thread of work is provided in Figure \[fig:overview\].\ #### Dowker’s theorem and a functorial generalization Let $X,Y$ be two totally ordered sets, and let $R\subseteq X\times Y$ be a nonempty relation. Then one can define two simplicial complexes $E_R$ and $F_R$ as follows. A finite subset ${\sigma}\subseteq X$ belongs to $E_R$ whenever there exists $y \in Y$ such that $(x,y) \in R$ for each $x\in {\sigma}$. Similarly a finite subset ${\tau}\subseteq Y$ belongs to $F_R$ whenever there exists $x\in X$ such that $(x,y) \in R$ for each $y\in {\tau}$. These constructions can be traced back to [@dowker1952homology], who proved the following result that we refer to as Dowker’s theorem: \[thm:dowker\] Let $X,Y$ be two totally ordered sets, let $R\subseteq X\times Y$ be a nonempty relation, and let $E_R, F_R$ be as above. Then for each $k \in {\mathbb{Z}}_+$, $$H_k(E_R) \cong H_k(F_R).$$ There is also a strong form of Dowker’s theorem that Björner proves via the classical nerve theorem [@bjorner-book Theorems 10.6, 10.9]: \[thm:dowker-strong\] Under the assumptions of Theorem \[thm:dowker\], we in fact have $\|E_R\| \simeq \|F_R\|$. The Functorial Dowker Theorem is the following generalization of the strong form of Dowker’s theorem: instead of a single nonempty relation $R \subseteq X\times Y$, consider any pair of nested, nonempty relations $R\subseteq R' \subseteq X\times Y$. Then there exist homotopy equivalences between the geometric realizations of the corresponding complexes that commute with the canonical inclusions, up to homotopy. We formalize this statement below. \[thm:dowker-functorial\] Let $X,Y$ be two totally ordered sets, let $R\subseteq R' \subseteq X\times Y$ be two nonempty relations, and let $E_R, F_R, E_{R'}, F_{R'}$ be their associated simplicial complexes. Then there exist homotopy equivalences $\Gamma_{\|E_R\|}:\|F_R\| {\rightarrow}\|E_R\|$ and $\Gamma_{\|E_{R'}\|}: \|F_{R'}\| {\rightarrow}\|E_{R'}\|$ such that the following diagram commutes up to homotopy: \(1) at (0,0)[$\|F_R\|$]{}; (2) at (3,0)[$\|F_{R'}\|$]{}; \(3) at (0,-2)[$\|E_{R}\|$]{}; (4) at (3,-2)[$\|E_{R'}\|$]{}; \(1) edge\[->\] node\[above\][$\|\iota_E\|$]{} (2); (3) edge\[->\] node\[above\][$\|\iota_F\|$]{}(4); (1) edge\[->\] node\[left\][$\Gamma_{\|E_R\|}$]{} node\[right\] [$\simeq$]{}(3); (2) edge\[->\] node\[right\][$\Gamma_{\|E_{R'}\|}$]{} node\[left\] [$\simeq$]{} (4); In other words, we have $\|\iota_F\|\circ \Gamma_{\|E_R\|} \simeq \Gamma_{\|E_{R'}\|} \circ \|\iota_E\|$, where $\iota_E,\iota_F$ are the canonical inclusions. From Theorem \[thm:dowker-functorial\] we automatically obtain Theorem \[thm:dowker-strong\] (the strong form of Dowker’s theorem) as an immediate corollary. The strong form does not appear in Dowker’s original paper [@dowker1952homology], but Björner has given a proof using the nerve theorem [@bjorner-book Theorems 10.6, 10.9]. Moreover, Björner writes in a remark following [@bjorner-book Theorem 10.9] that the nerve theorem and the strong form of Dowker’s theorem are equivalent, in the sense that one implies the other. We were not able to find an elementary proof of the strong form of Dowker’s theorem in the existing literature. However, such an elementary proof is provided by our proof of Theorem \[thm:dowker-functorial\] (given in Section \[sec:dowker-dual\]), which we obtained by extending ideas in Dowker’s original proof of Theorem \[thm:dowker\].[^1] Whereas the Functorial Dowker Theorem and our elementary proof are of independent interest, it has been suggested in [@chazal2014persistence Remark 4.8] that such a functorial version of Dowker’s theorem could also be proved using a functorial nerve theorem [@chazal2008towards Lemma 3.4]. Despite being an interesting possibility, we were not able to find a detailed proof of this claim in the literature. In addition, Björner’s remark regarding the equivalence between the nerve theorem and the strong form of Dowker’s theorem suggests the following question: \[q:f-nerve-f-dowker\] Are the Functorial Nerve Theorem (FNT) of [@chazal2008towards] and the Functorial Dowker Theorem (FDT, Theorem \[thm:dowker-functorial\]) equivalent? This question is of fundamental importance because the Nerve Theorem is a crucial tool in the applied topology literature and its functorial generalizations are equally important in persistent homology. In general, the answer is no, and moreover, one (of the FNT and FDT) is not stronger than the other. The FNT of [@chazal2008towards] is stated for paracompact spaces, which are more general than the simplicial complexes of the FDT. However, the FNT of [@chazal2008towards] is stated for spaces with finitely-indexed covers, so the associated nerve complexes are necessarily finite. All the complexes involved in the statement of the FDT are allowed to be infinite, so the FDT is more general than the FNT in this sense. To clarify these connections, we formulate a simplicial Functorial Nerve Theorem (Theorem \[thm:nerve-functorial-II\]) and prove it via a finite formulation of the FDT (Theorem \[thm:dowker-functorial-finite\]). In turn, we show that the simplicial FNT implies the finite FDT, thus proving the equivalence of these formulations (Theorem \[thm:dowker-nerve-eq\]). Dowker complexes are also known to researchers who use Q-analysis to study social networks [@johnson2013hypernetworks; @atkin1975mathematical; @atkin1972cohomology]. We perceive that viewing Dowker complexes through the modern lens of persistence will enrich the classical framework of Q-analysis by incorporating additional information about the meaningfulness of features, thus potentially opening new avenues in the social sciences. An announcement of part of our work has appeared in [@dowker-asilo]. Implementations --------------- Following work in [@curto2008cell; @dabaghian2012topological], we implement our methods in the setting of classifying simulated hippocampal networks. We simulate the activity pattern of hippocampal cells in an animal as it moves around arenas with a number of obstacles, and compile this data into a network which can be interpreted as the transition matrix for the time-reversal of a Markov process. The motivating idea is to ascertain whether, by just observing hippocampal activity and not using any higher reasoning ability, one might be able to determine the number of obstacles in the arena that the animal has just finished traversing. The results of computing Dowker persistence diagrams suggest that the hippocampal activity is indeed sufficient to accurately count the number of obstacles in each arena. Our datasets and software are available on <https://research.math.osu.edu/networks/Datasets.html> as part of the `PersNet` software package. Organization of the paper ------------------------- Notation used globally is defined directly below. §\[sec:background\] contains the necessary background on persistent homology. §\[sec:nets\] contains our formulations for networks, as well as some key ingredients of our stability results. §\[sec:rips\] contains details about the Rips persistence diagram. The first part of §\[sec:dowker\] contains details about the Dowker persistence diagram. §\[sec:dowker-dual\] contains the Functorial Dowker Theorem. The connection between the simplicial Functorial Nerve Theorem and the finite Functorial Dowker Theorem is detailed in §\[sec:dowker-nerve-equiv\]. In §\[sec:symmetry\] we show that Dowker complexes are sensitive to asymmetry. §\[sec:cycle\] contains a family of asymmetric networks, the cycle networks, and a full characterization of their Dowker persistence diagrams. In §\[sec:exp\] we provide details on an implementation of our methods. Finally, proofs of statements not contained in the main body of the paper are relegated to Appendix \[app:proofs\], whereas details about the characterization results for Dowker persistence diagrams of cycle networks are given in Appendix \[sec:cycle-addendum\]. Notation -------- We will write $\mathbb{K}$ to denote a field, which we will fix and use throughout the paper. We will write ${\mathbb{Z}}_+$ and ${\mathbb{R}}_+$ to denote the nonnegative integers and reals, respectively. The extended real numbers ${\mathbb{R}}\cup {\left\{\infty, -\infty\right\}}$ will be denoted $\overline{{\mathbb{R}}}$. The cardinality of a set $X$ will be denoted ${\operatorname{card}}(X)$. The collection of nonempty subsets of a set $X$ will be denoted ${\operatorname{pow}}(X)$. The natural numbers ${\left\{1,2,3,\ldots\right\}}$ will be denoted by ${\mathbb{N}}$. The dimension of a vector space $V$ will be denoted $\dim(V)$. The rank of a linear transformation $f$ will be denoted ${\operatorname{rank}}(f)$. An isomorphism between vector spaces $V$ and $W$ will be denoted $V\cong W$. A homotopy relation for two maps $f,g:X{\rightarrow}Y$ between topological spaces will be denoted $f\simeq g$. Occasionally we will need to take about multisets, i.e. sets where elements can have multiplicity greater than 1. We will use square bracket notation $[\ldots]$ to denote multisets. Identity maps will be denoted by the notation ${\operatorname{id}}_\bullet$. Given a simplicial complex ${\Sigma}$, we will often write $V({\Sigma})$ to denote the vertex set of ${\Sigma}$. We will write ${\operatorname{Bd}}({\sigma})$ to denote the boundary of a simplex ${\sigma}$. Background on persistent homology {#sec:background} ================================= We assume that the reader is familiar with terms and concepts related to simplicial homology, and refer to [@munkres-book] for details. Here we describe our choices of notation. Whenever we have a simplicial complex over a set $X$ and a $k$-simplex ${\left\{x_0,x_1,\ldots, x_k\right\}}$, $k\in {\mathbb{Z}}_+$, we will assume that the simplex is oriented by the ordering $x_0< x_1 < \ldots < x_k$. We will write $[x_0,x_1,\ldots,x_k]$ to denote the equivalence class of the even permutations of this chosen ordering, and $-[x_0,x_1,\ldots,x_k]$ to denote the equivalence class of the odd permutations of this ordering. Given a simplicial complex ${\Sigma}$, we will denote its geometric realization by $\|{\Sigma}\|$. The weak topology on $\|{\Sigma}\|$ is defined by requiring that a subset $A \subseteq \|{\Sigma}\|$ is closed if and only if $A \cap \|{\sigma}\|$ is closed in $\|{\sigma}\|$ for each ${\sigma}\in {\Sigma}$. A simplicial map $f: {\Sigma}{\rightarrow}\Xi$ between two simplicial complexes induces a map $\|f\|: \|{\Sigma}\| {\rightarrow}\|\Xi\|$ between the geometric realizations, defined as $\|f\|(\sum_{v\in {\Sigma}}a_v v):= \sum_{v\in {\Sigma}}a_v f(v)$. These induced maps satisfy the usual composition identity: given simpicial maps $f:{\Sigma}{\rightarrow}\Xi$ and $g:\Xi {\rightarrow}\Upsilon$, we have $\|g\circ f\| = \|g\| \circ \|f\|$. To see this, observe the following: $$\label{eq:htpy-func} \|g\circ f\|(\sum_{v\in {\Sigma}}a_v v) = \sum_{v\in {\Sigma}} a_vg(f(v)) = \|g\|(\sum_{v\in {\Sigma}}a_v f(v)) = \|g\|\circ\|f\|(\sum_{v\in {\Sigma}}a_v v).$$ A filtration of a simplicial complex ${\Sigma}$ (also called a filtered simplicial complex) is defined to be a nested sequence $\{{\Sigma}^{{\delta}}\subseteq {\Sigma}^{{\delta}'}\}_{{\delta}\leq {\delta}' \in {\mathbb{R}}}$ of simplicial complexes satisfying the condition that there exist ${\delta}_I,\, {\delta}_F \in {\mathbb{R}}$ such that ${\Sigma}^{{\delta}} = {\varnothing}$ for all ${\delta}\leq {\delta}_I$, and ${\Sigma}^{{\delta}} = {\Sigma}\text{ for all }{\delta}\geq {\delta}_F$. Fix a field ${\mathbb}{K}$. Given a finite simplicial complex ${\Sigma}$ and a dimension $k \in {\mathbb{Z}}_+$, we will denote a $k$-chain in ${\Sigma}$ as $\sum_ia_i{\sigma}_i$, where each $a_i \in \mathbb{K}$ and ${\sigma}_i \in {\Sigma}$. We write $C_k({\Sigma})$ or just $C_k$ to denote the $\mathbb{K}$-vector space of all $k$-chains. We will write ${\partial}_k$ to denote the associated boundary map ${\partial}_k : C_k {\rightarrow}C_{k-1}$: $${\partial}_k[x_0,\ldots,x_k]:=\sum_i(-1)^i[x_0,\ldots,\hat{x}_i,\ldots, x_k], \text{ where $\hat{x}_i$ denotes omission of $x_i$ from the sequence.}$$ We will write ${\mathcal{C}}=(C_k,{\partial}_k)_{k\in {\mathbb{Z}}_+}$ to denote a chain complex, i.e. a sequence of vector spaces with boundary maps such that ${\partial}_{k-1}\circ {\partial}_k =0$. Given a chain complex ${\mathcal{C}}$ and any $k\in {\mathbb{Z}}_+$, the $k$-th homology of the chain complex ${\mathcal{C}}$ is denoted $H_k({\mathcal{C}}) :=\ker({\partial}_k)/{\operatorname{im}}({\partial}_{k+1})$. The $k$-th Betti number of ${\mathcal{C}}$ is denoted ${\beta}_k({\mathcal{C}})$. Given a simplicial map $f$ between simplicial complexes, we write $f_$ to denote the induced chain map between the corresponding chain complexes [@munkres-book §1.12], and $(f_k)_{\#}$ to denote the linear map on $k$th homology vector spaces induced for each $k \in {\mathbb{Z}}_+$. The operations of passing from simplicial complexes and simplicial maps to chain complexes and induced chain maps, and then to homology vector spaces with induced linear maps, will be referred to as passing to homology. Recall the following useful fact, often referred to as functoriality of homology* [@munkres-book Theorem 12.2]: given a composition $g\circ f$ of simplicial maps, we have $$(g_k\circ f_k)_\# = (g_k)_\#\circ (f_k)_\# \qquad\text{ for each } k\in {\mathbb{Z}}_+. \label{eq:functoriality}$$ A persistence vector space is defined to be a family of vector spaces $\{U^{\delta}{\xrightarrow}{\mu_{{\delta},{\delta}'}} U^{{\delta}'}\}_{{\delta}\leq {\delta}'\in {\mathbb{R}}}$ such that: (1) $\mu_{{\delta},{\delta}}$ is the identity for each ${\delta}\in {\mathbb{R}}$, and (2) $\mu_{{\delta},{\delta}''} = \mu_{{\delta}',{\delta}''}\circ \mu_{{\delta},{\delta}'}$ for each ${\delta}\leq {\delta}' \leq {\delta}'' \in {\mathbb{R}}$. The persistence vector spaces that we consider in this work also satisfy the following conditions: (1) $\dim(U^{\delta}) <\infty$ at each ${\delta}\in {\mathbb{R}}$, (2) there exist ${\delta}_I,\, {\delta}_F \in {\mathbb{R}}$ such that all maps $\mu_{{\delta},{\delta}'}$ are isomorphisms for ${\delta},{\delta}' \geq {\delta}_F$ and for ${\delta},{\delta}' \leq {\delta}_I$, and (3) there are only finitely many values of ${\delta}\in {\mathbb{R}}$ such that $U^{{\delta}-{\varepsilon}} \not\cong U^{{\delta}}$ for each ${\varepsilon}>0$. Here ${\delta}$ is referred to as a resolution parameter, and such a persistence vector space is described as being ${\mathbb{R}}$-indexed. The collection of all such persistence vector spaces is denoted ${{\operatorname{\mathbf{PVec}}}}({\mathbb{R}})$. Observe that by fixing $k \in {\mathbb{Z}}_+$ and passing to the $k$th homology vector space at each step ${\Sigma}^{{\delta}}$ of a filtered simplicial complex $({\Sigma}^{{\delta}})_{{\delta}\in {\mathbb{R}}}$, the functoriality of homology gives us the $k$th persistence vector space associated to $({\Sigma}^{{\delta}})_{{\delta}\in {\mathbb{R}}}$, denoted $${\mathcal{H}}_k({\Sigma}) := \{H_k({\mathcal{C}}^{{\delta}}){\xrightarrow}{(\iota_{{\delta},{\delta}'})_\#} H_k({\mathcal{C}}^{{\delta}'})\}_{{\delta}\leq {\delta}' \in {\mathbb{R}}}.$$ The elements of ${{\operatorname{\mathbf{PVec}}}}({\mathbb{R}})$ contain only a finite number of vector spaces, up to isomorphism. By the classification results in [@carlsson2005persistence §5.2], it is possible to associate a full invariant, called a persistence barcode or persistence diagram, to each element of ${{\operatorname{\mathbf{PVec}}}}({\mathbb{R}})$. This barcode is a multiset of persistence intervals, and is represented as a set of lines over a single axis. The barcode of a persistence vector space ${\mathcal{V}}$ is denoted ${{\operatorname{\mathbf{Pers}}}}({\mathcal{V}})$. The intervals in ${{\operatorname{\mathbf{Pers}}}}({\mathcal{V}})$ can be represented as the persistence diagram of ${\mathcal{V}}$, which is as a multiset of points lying on or above the diagonal in $\overline{{\mathbb{R}}}^2$, counted with multiplicity. More specifically, $${\operatorname{Dgm}}({\mathcal{V}}):=\big[({\delta}_i,{\delta}_{j+1}) \in \overline{{\mathbb{R}}}^2 : [{\delta}_i,{\delta}_{j+1}) \in {{\operatorname{\mathbf{Pers}}}}({\mathcal{V}}) \big],$$ where the multiplicity of $({\delta}_i,{\delta}_{j+1})\in \overline{{\mathbb{R}}}^2$ is given by the multiplicity of $[{\delta}_i,{\delta}_{j+1}) \in {{\operatorname{\mathbf{Pers}}}}({\mathcal{V}})$. Persistence diagrams can be compared using the bottleneck distance, which we denote by ${d_{\operatorname{B}}}$. Details about this distance, as well as the other material related to persistent homology, can be found in [@chazal2012structure]. Numerous other formulations of the material presented above can be found in [@edelsbrunner2002topological; @zomorodian2005computing; @zigzag; @edelsbrunner2010computational; @edelsbrunner2014persistent; @bauer-isom; @ph-self]. \[rem:trivial-diag\] Whenever we describe a persistence diagram as being trivial, we mean that either it is empty, or it does not have any off-diagonal points. Interleaving distance and stability of persistence vector spaces. {#sec:background-int} ----------------------------------------------------------------- In what follows, we will consider ${\mathbb{R}}$-indexed persistence vector spaces ${{\operatorname{\mathbf{PVec}}}}({\mathbb{R}})$. Given ${\varepsilon}\geq 0$, two ${\mathbb{R}}$-indexed persistence vector spaces ${\mathcal{V}}=\{V^{\delta}{\xrightarrow}{\nu_{{\delta},{\delta}'}} V^{{\delta}'}\}_{{\delta}\leq {\delta}'}$ and ${\mathcal{U}}=\{U^{\delta}{\xrightarrow}{\mu_{{\delta},{\delta}'}} U^{{\delta}'}\}_{{\delta}\leq {\delta}'}$ are said to be ${\varepsilon}$-interleaved [@chazal2009proximity; @bauer-isom] if there exist two families of linear maps $$\begin{aligned} \{{\varphi}_{{\delta},{\delta}+{\varepsilon}}&:V^{\delta}{\rightarrow}V^{{\delta}+ {\varepsilon}}\}_{{\delta}\in {\mathbb{R}}},\\ \{\psi_{{\delta},{\delta}+{\varepsilon}}&:U^{\delta}{\rightarrow}U^{{\delta}+ {\varepsilon}}\}_{{\delta}\in {\mathbb{R}}}\end{aligned}$$ such that the following diagrams commute for all ${\delta}' \geq {\delta}\in {\mathbb{R}}$: $$\begin{tikzcd}[column sep=large] V^{\delta}\arrow{r}{\nu_{{\delta},{\delta}'}} \arrow[swap]{dr}{{\varphi}_{\delta}} & V^{{\delta}'}\arrow{dr}{{\varphi}_{{\delta}'}} & {} & {} & V^{{\delta}+{\varepsilon}} \arrow{r}{\nu_{{\delta}+\eta,{\delta}'+\eta}} & V^{{\delta}'+{\varepsilon}} \\ {} & U^{{\delta}+{\varepsilon}} \arrow{r}{\mu_{{\delta}+\eta,{\delta}'+\eta}} & U^{{\delta}'+{\varepsilon}} & U^{{\delta}} \arrow{ur}{\psi_{\delta}} \arrow{r}{\mu_{{\delta},{\delta}'}} & U^{{\delta}'} \arrow[swap]{ur}{\psi_{{\delta}'}} & {} \end{tikzcd}$$ $$\begin{tikzcd}[column sep=large] V^{\delta}\arrow{rr}{\nu_{{\delta},{\delta}+2{\varepsilon}}} \arrow[swap]{dr}{{\varphi}_{\delta}} & {} & V^{{\delta}+2{\varepsilon}} & {}& V^{{\delta}+{\varepsilon}} \arrow{dr}{\psi_{{\delta}+{\varepsilon}}}\\ {} & U^{{\delta}+{\varepsilon}} \arrow[swap]{ur}{{\varphi}_{{\delta}+{\varepsilon}}} & {}& U^{{\delta}} \arrow{ur}{\psi_{{\delta}}} \arrow{rr}{\mu_{{\delta},{\delta}+2{\varepsilon}}} & {} & U^{{\delta}+2\eta} \end{tikzcd}$$ The purpose of introducing ${\varepsilon}$-interleavings is to define a pseudometric on the collection of persistence vector spaces. The interleaving distance between two ${\mathbb{R}}$-indexed persistence vector spaces ${\mathcal{V}},{\mathcal{U}}$ is given by: $${d_{\operatorname{I}}}({\mathcal{U}},{\mathcal{V}}) := \inf \{{\varepsilon}\geq 0 : \text{${\mathcal{U}}$ and ${\mathcal{V}}$ are ${\varepsilon}$-interleaved}\}.$$ One can verify that this definition induces a pseudometric on the collection of persistence vector spaces [@chazal2009proximity; @bauer-isom]. The interleaving distance can then be related to the bottleneck distance as follows: Let ${\mathcal{U}}, {\mathcal{V}}$ be two ${\mathbb{R}}$-indexed persistence vector spaces. Then, $${d_{\operatorname{B}}}({\operatorname{Dgm}}({\mathcal{U}}),{\operatorname{Dgm}}({\mathcal{V}}))\leq {d_{\operatorname{I}}}({\mathcal{U}},{\mathcal{V}}).$$ Stability results are at the core of persistent homology, beginning with the classical bottleneck stability result in [@bot-stab]. One of our key contributions is to use the Algebraic Stability Theorem stated above, along with Lemma §\[sec:nets\] stated below, to prove stability results for methods of computing persistent homology of a network. Before stating the following lemma, recall that two simplicial maps $f,g: {\Sigma}{\rightarrow}\Xi$ are contiguous if for any simplex ${\sigma}\in {\Sigma}$, $f({\sigma}) \cup g({\sigma})$ is a simplex of $\Xi$. Contiguous maps satisfy the following useful properties: \[prop:contigo-props\] Let $f,g: {\Sigma}{\rightarrow}\Xi$ be two contiguous simplicial maps. Then, 1. $\|f\|,\|g\|:\|{\Sigma}\| {\rightarrow}\|\Xi\|$ are homotopic [@spanier-book §3.5], and 2. The chain maps induced by $f$ and $g$ are chain homotopic, and as a result, the induced maps $f_\#$ and $g_\#$ for homology are equal [@munkres-book Theorem 12.5]. \[lem:stab\] Let ${\mathfrak}{F}, {\mathfrak}{G}$ be two filtered simplicial complexes written as $$\{{\mathfrak}{F}^{\delta}{\xrightarrow}{s_{{\delta},{\delta}'}} {\mathfrak}{F}^{{\delta}'}\}_{{\delta}'\geq {\delta}\in {\mathbb{R}}} \text{ and } \{{\mathfrak}{G}^{\delta}{\xrightarrow}{t_{{\delta},{\delta}'}} {\mathfrak}{G}^{{\delta}'}\}_{{\delta}'\geq {\delta}\in {\mathbb{R}}},$$ where $s_{{\delta},{\delta}'}$ and $t_{{\delta},{\delta}'}$ denote the natural inclusion maps. Suppose $\eta\geq 0$ is such that there exist families of simplicial maps ${\left\{{\varphi}_{\delta}:{\mathfrak}{F}^{\delta}{\rightarrow}{\mathfrak}{G}^{{\delta}+\eta}\right\}}_{{\delta}\in {\mathbb{R}}}$ and ${\left\{\psi_{\delta}:{\mathfrak}{G}^{\delta}{\rightarrow}{\mathfrak}{F}^{{\delta}+\eta}\right\}}_{{\delta}\in {\mathbb{R}}}$ such that the following are satisfied for any ${\delta}' \geq {\delta}$: 1. $t_{{\delta}+\eta,{\delta}'+\eta}\circ {\varphi}_{\delta}$ and ${\varphi}_{{\delta}'}\circ s_{{\delta},{\delta}'}$ are contiguous 2. $s_{{\delta}+\eta,{\delta}'+\eta}\circ \psi_{\delta}$ and $\psi_{{\delta}'}\circ t_{{\delta},{\delta}'}$ are contiguous 3. $\psi_{{\delta}+\eta}\circ {\varphi}_{\delta}$ and $s_{{\delta},{\delta}+2\eta}$ are contiguous 4. ${\varphi}_{{\delta}+\eta}\circ \psi_{\delta}$ and $t_{{\delta},{\delta}+2\eta}$ are contiguous. All the diagrams are as below: $$\begin{tikzcd}[column sep=large] {\mathfrak}{F}^{\delta}\arrow{r}{s_{{\delta},{\delta}'}} \arrow[swap]{dr}{{\varphi}_{\delta}} & {\mathfrak}{F}^{{\delta}'}\arrow{dr}{{\varphi}_{{\delta}'}} & {} & {} & {\mathfrak}{F}^{{\delta}+\eta} \arrow{r}{s_{{\delta}+\eta,{\delta}'+\eta}} & {\mathfrak}{F}^{{\delta}'+\eta} \\ {} & {\mathfrak}{G}^{{\delta}+\eta} \arrow{r}{t_{{\delta}+\eta,{\delta}'+\eta}} & {\mathfrak}{G}^{{\delta}'+\eta} & {\mathfrak}{G}^{{\delta}} \arrow{ur}{\psi_{\delta}} \arrow{r}{t_{{\delta},{\delta}'}} & {\mathfrak}{G}^{{\delta}'} \arrow[swap]{ur}{\psi_{{\delta}'}} & {} \end{tikzcd}$$ $$\begin{tikzcd}[column sep=large] {\mathfrak}{F}^{\delta}\arrow{rr}{s_{{\delta},{\delta}+2\eta}} \arrow[swap]{dr}{{\varphi}_{\delta}} & {} & {\mathfrak}{F}^{{\delta}+2\eta} & {} & {}& {\mathfrak}{F}^{{\delta}+\eta} \arrow{dr}{{\varphi}_{{\delta}+\eta}}\\ {} & {\mathfrak}{G}^{{\delta}+\eta} \arrow[swap]{ur}{\psi_{{\delta}+\eta}} & {} & {}& {\mathfrak}{G}^{{\delta}} \arrow{ur}{\psi_{\delta}} \arrow{rr}{t_{{\delta},{\delta}+2\eta}} & {} & {\mathfrak}{G}^{{\delta}+2\eta} \end{tikzcd}$$ For each $k\in {\mathbb{Z}}_+$, let ${\mathcal{H}}_k({\mathfrak}{F}), {\mathcal{H}}_k({\mathfrak}{G})$ denote the $k$-dimensional persistence vector spaces associated to ${\mathfrak}{F}$ and ${\mathfrak}{G}$. Then for each $k\in {\mathbb{Z}}_+$, $${d_{\operatorname{B}}}({\operatorname{Dgm}}_k({\mathcal{H}}_k({\mathfrak}{F})),{\operatorname{Dgm}}_k({\mathcal{H}}_k({\mathfrak}{G}))) \leq {d_{\operatorname{I}}}({\mathcal{H}}_k({\mathfrak}{F}),{\mathcal{H}}_k({\mathfrak}{G})) \leq \eta.$$ Background on networks and our network distance {#sec:nets} =============================================== A network is a pair $(X,{\omega}_X)$ where $X$ is a finite set and ${\omega}_X: X\times X {\rightarrow}{\mathbb{R}}$ is a weight function. Note that ${\omega}_X$ need not satisfy the triangle inequality, any symmetry condition, or even the requirement that ${\omega}_X(x,x) = 0$ for all $x\in X$. The weights are even allowed to be negative. The collection of all such networks is denoted ${\mathcal{N}}$. When comparing networks, one needs a way to correlate points in one network with points in the other. To see how this can be done, let $(X,{\omega}_X), (Y,{\omega}_Y) \in {\mathcal{N}}$. Let $R$ be any nonempty relation between $X$ and $Y$, i.e. a nonempty subset of $X \times Y$. The distortion of the relation $R$ is given by: $${\operatorname{dis}}(R):=\max_{(x,y),(x',y')\in R}\|{\omega}_X(x,x')-{\omega}_Y(y,y')\|.$$ A correspondence between $X$ and $Y$ is a relation $R$ between $X$ and $Y$ such that $\pi_X(R)=X$ and $\pi_Y(R)=Y$, where $\pi_X:X\times Y {\rightarrow}X$ and $\pi_Y:X\times Y {\rightarrow}Y$ denote the natural projections. The collection of all correspondences between $X$ and $Y$ will be denoted ${\mathscr{R}}(X,Y)$. Following previous work in [@clust-net; @nets-allerton; @nets-icassp] the network distance ${d_{\mathcal{N}}}:{\mathcal{N}}\times {\mathcal{N}}{\rightarrow}{\mathbb{R}}_+$ is then defined as: $${d_{\mathcal{N}}}(X,Y):=\frac{1}{2}\min_{R\in{\mathscr{R}}}{\operatorname{dis}}(R).$$ It can be verified that ${d_{\mathcal{N}}}$ as defined above is a pseudometric, and that the networks at 0-distance can be completely characterized [@nets-allerton]. Next we wish to prove the reformulation in Proposition \[prop:dn-ko\]. First we define the distortion of a map between two networks. Given any $(X,{\omega}_X),(Y,{\omega}_Y)\in {\mathcal{N}}$ and a map ${\varphi}:(X,{\omega}_X) {\rightarrow}(Y,{\omega}_Y)$, the distortion of ${\varphi}$ is defined as: $${\operatorname{dis}}({\varphi}):= \max_{x,x'\in X}\|{\omega}_X(x,x')-{\omega}_Y({\varphi}(x),{\varphi}(x'))\|.$$ Next, given maps ${\varphi}:(X,{\omega}_X){\rightarrow}(Y,{\omega}_Y)$ and $\psi:(Y,{\omega}_Y){\rightarrow}(X,{\omega}_X)$, we define two co-distortion terms: $$\begin{aligned} C_{X,Y}({\varphi},\psi) &:= \max_{(x,y)\in X\times Y}\|{\omega}_X(x,\psi(y)) - {\omega}_Y({\varphi}(x),y)\|,\\ C_{Y,X}(\psi,{\varphi}) &:= \max_{(y,x)\in Y\times X}\|{\omega}_Y(y,{\varphi}(x)) - {\omega}_X(\psi(y),x)\|.\end{aligned}$$ \[prop:dn-ko\] Let $(X,{\omega}_X), (Y,{\omega}_Y)\in {\mathcal{N}}$. Then, $${d_{\mathcal{N}}}(X,Y) = \tfrac{1}{2}\min\{\max({\operatorname{dis}}({\varphi}),{\operatorname{dis}}(\psi),C_{X,Y}({\varphi},\psi), C_{Y,X}(\psi,{\varphi})) : {\varphi}:X {\rightarrow}Y, \psi:Y {\rightarrow}X \text{ any maps}\}.$$ Proposition \[prop:dn-ko\] is analogous to a result of Kalton and Ostrovskii [@kalton1997distances Theorem 2.1] where—instead of ${d_{\mathcal{N}}}$—one has the Gromov-Hausdorff distance between metric spaces. We remark that when restricted to the special case of networks that are also metric spaces, the network distance ${d_{\mathcal{N}}}$ agrees with the Gromov-Hausdorff distance. Details on the Gromov-Hausdorff distance can be found in [@burago]. An important remark is that in the Kalton-Ostrovskii formulation, there is only one co-distortion term. When Proposition \[prop:dn-ko\] is applied to metric spaces, the two co-distortion terms become equal by symmetry, and thus the Kalton-Ostrovskii formulation is recovered. But a priori, the lack of symmetry in the network setting requires us to consider both terms. In the following sections, we propose methods for computing persistent homology of networks, and prove that they are stable via Lemma \[lem:stab\]. Note that similar results, valid in the setting of metric spaces, have appeared in [@dgh-pers; @chazal2014persistence]. Whereas the proofs in [@chazal2014persistence] invoke an auxiliary construction of multivalued maps arising from correspondences, our proofs simply use the maps ${\varphi}, \psi$ arising directly from the reformulation of ${d_{\mathcal{N}}}$ (Proposition \[prop:dn-ko\]), thus streamlining the treatment. When studying the effect of asymmetry on persistent homology, it will be useful to consider the network transformations that we define next. \[defn:sym-trans\] Define the max-symmetrization map ${\mathfrak}{s}:{\mathcal{N}}{\rightarrow}{\mathcal{N}}$ by $(X,{\omega}_X)\mapsto (X,\widehat{{\omega}_X})$, where for any network $(X,{\omega}_X)$, we define $\widehat{{\omega}_X}:X\times X {\rightarrow}{\mathbb{R}}$ as follows: $$\widehat{{\omega}_X}(x,x'):= \max({\omega}_X(x,x'),{\omega}_X(x',x)), \text{ for } x,x'\in X.$$ Also define the transposition map ${\mathfrak}{t}:{\mathcal{N}}{\rightarrow}{\mathcal{N}}$ by $(X,{\omega}_X) \mapsto (X,{\omega}_X^\top)$, where for any $(X,{\omega}_X) \in {\mathcal{N}}$, we define ${\omega}_X^\top(x,x'):= {\omega}_X(x',x)$ for $x,x'\in X$. For convenience, we denote $X^\top:={\mathfrak}{t}(X)$ for any network $X$. We are now ready to formulate our two methods for computing persistent homology of networks. The Rips filtration is the “workhorse” of persistent homology of metric spaces so it is natural to consider its generalization to general asymmetric networks. The Rips filtration of a network {#sec:rips} ================================ Recall that for a metric space $(X,d_X)$, the Rips complex is defined for each ${\delta}\geq 0$ as follows: $${\mathfrak}{R}^{\delta}_X := {\left\{{\sigma}\in {\operatorname{pow}}(X): {\operatorname{diam}}({\sigma}) \leq {\delta}\right\}}, \text{ where } {\operatorname{diam}}({\sigma}) := \max_{x,x'\in {\sigma}}d_X(x,x').$$ Following this definition, we can define the Rips complex for a network $(X,{\omega}_X)$ as follows: $${\mathfrak}{R}^{\delta}_X:=\{{\sigma}\in {\operatorname{pow}}(X) : \max_{x,x'\in {\sigma}}{\omega}_X(x,x') \leq {\delta}\}.$$ To any network $(X,{\omega}_X)$, we may associate the Rips filtration $\{{\mathfrak}{R}^{\delta}_X{\hookrightarrow}{\mathfrak}{R}^{{\delta}'}_X\}_{{\delta}\leq {\delta}'}$. We denote the $k$-dimensional persistence vector space associated to this filtration by ${\mathcal{H}}_k^{{\mathfrak}{R}}(X)$, and the corresponding persistence diagram by ${\operatorname{Dgm}}_k^{{\mathfrak}{R}}(X)$. The Rips filtration is stable to small perturbations of the input data:9 \[prop:rips-stab\] Let $(X,{\omega}_X), (Y,{\omega}_Y) \in {\mathcal{N}}$. Then ${d_{\operatorname{B}}}({\operatorname{Dgm}}_k^{{\mathfrak}{R}}(X),{\operatorname{Dgm}}_k^{{\mathfrak}{R}}(Y)) \leq 2{d_{\mathcal{N}}}(X,Y).$ We omit the proof because it is similar to that of Proposition \[prop:dowker-stab\], which we will prove in detail. \[rem:rips-benefits\] The preceding proposition serves a dual purpose: (1) it shows that the Rips persistence diagram is robust to noise in input data, and (2) it shows that instead of computing the network distance between two networks, one can compute the bottleneck distance between their Rips persistence diagrams as a suitable proxy. The advantage to computing bottleneck distance is that it can be done in polynomial time (see [@efrat2001geometry]), whereas computing ${d_{\mathcal{N}}}$ is NP-hard in general. This follows from the fact that the problem of computing ${d_{\mathcal{N}}}$ includes the problem of computing the Gromov-Hausdorff distance between finite metric spaces, which is an NP-hard problem [@schmiedl]. We remark that the idea of computing Rips persistence diagrams to compare finite metric spaces first appeared in [@dgh-pers], and moreover, that Proposition \[prop:rips-stab\] is an extension of Theorem 3.1 in [@dgh-pers]. The Rips filtration in the setting of symmetric networks has been used in [@horak2009persistent; @carstens2013persistent; @giusti2015clique; @petri2013topological], albeit without addressing stability results. To our knowledge, Proposition \[prop:rips-stab\] is the first quantitative result justifying the constructions in these prior works. \[rem:rips-symm\] A critical weakness of the Rips complex construction is that it is not sensitive to asymmetry. To see this, recall the symmetrization map defined in Definition \[defn:sym-trans\], and let $(X,{\omega}_X) \in {\mathcal{N}}$. Now for any ${\sigma}\in {\operatorname{pow}}(X)$, we have $\max_{x,x' \in {\sigma}}{\omega}_X(x,x') = \max_{x,x'\in {\sigma}}\widehat{{\omega}_X}(x,x').$ It follows that for each ${\delta}\geq 0$, the Rips complexes of $(X,{\omega}_X)$ and $(X,\widehat{{\omega}_X})={\mathfrak}{s}(X,{\omega}_X)$ are equal, i.e. ${\mathfrak}{R} = {\mathfrak}{R} \circ {\mathfrak}{s}$. Thus the Rips persistence diagrams of the original and max-symmetrized networks are equal. The Dowker filtration of a network {#sec:dowker} ================================== Given $(X,\omega_X)\in \mathcal{N}$, and for any ${\delta}\in {\mathbb{R}}$, consider the following relation on $X$: $$R_{{\delta},X}:={\left\{(x,x') : {\omega}_X(x,x') \leq {\delta}\right\}}. \label{eq:relation}$$ Then $R_{{\delta},X} \subseteq X \times X$, and $R_{{\delta}_F,X} = X \times X$ for some sufficiently large ${\delta}_F$. Furthermore, for any ${\delta}' \geq {\delta}$, we have $R_{{\delta},X} \subseteq R_{{\delta}',X}$. Using $R_{{\delta},X}$, we build a simplicial complex ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ as follows: $$\label{eq:d-sink} {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}:={\left\{{\sigma}=[x_0,\ldots, x_n] : \text{ there exists } x'\in X \text{ such that } (x_i,x')\in R_{{\delta},X} \text{ for each } x_i\right\}}.$$ If ${\sigma}\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$, it is clear that any face of ${\sigma}$ also belongs to ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$. We call ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ the Dowker ${\delta}$-sink simplicial complex associated to $X$, and refer to $x'$ as a ${\delta}$-sink for ${\sigma}$ (where ${\sigma}$ and $x'$ should be clear from context). Since $R_{{\delta},X}$ is an increasing sequence of sets, it follows that ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ is an increasing sequence of simplicial complexes. In particular, for ${\delta}'\geq {\delta}$, there is a natural inclusion map ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} {\hookrightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}',X}$. We write ${{\mathfrak}{D}^{\operatorname{si}}}_X$ to denote the filtration $\{{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} {\hookrightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}',X}\}_{ {\delta}\leq {\delta}'}$ associated to $X$. We call this the Dowker sink filtration on $X$. We will denote the $k$-dimensional persistence diagram arising from this filtration by ${\operatorname{Dgm}}_k^{{\operatorname{si}}}(X)$. \(2) at (2,0)[$a$]{}; (3) at (3,2)[$b$]{}; (4) at (4,0)[$c$]{}; \(5) at (5,1)[ $ {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}= \begin{cases} {\varnothing}&: {\delta}<-1 \\ \{[a]\} &:-1\leq {\delta}<0\\ \{[a],[b],[c]\} &: 0\leq {\delta}< 1\\ \{[a],[b],[c],[ab],[bc],[ac],[abc]\} &: {\delta}\geq 1 \end{cases} $ ]{}; \(50) at (5,1)[ $ {{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}= \begin{cases} {\varnothing}&: {\delta}<-1 \\ \{[a]\} &:-1\leq {\delta}<0\\ \{[a],[b],[c]\} &: 0\leq {\delta}< 1\\ \{[a],[b],[c],[ab],[ac]\} &: 1\leq {\delta}< 2\\ \{[a],[b],[c],[ab],[bc],[ac],[abc]\} &: {\delta}\geq 2 \end{cases} $ ]{}; \(0) at (4,0); (1) at (6,0); \(2) edge \[loop left\] node\[left\][$-1$]{}(2); (3) edge \[loop right\] node\[right\][$0$]{}(3); (4) edge \[loop right\] node\[right\][$0$]{}(4); (2) edge \[bend left\] node\[above\][$1$]{} (3); (3) edge \[\] node\[below\][$1$]{} (2); (2) edge \[\] node\[above\][$2$]{} (4); (3) edge \[\] node\[below\][$2$]{} (4); (4) edge \[bend left\] node\[below\][$1$]{} (2); (4) edge \[bend right\] node\[below\][$2$]{} (3); Note that we can define a dual construction as follows: $$\label{eq:d-src} {{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}:={\left\{{\sigma}=[x_0,\ldots, x_n] : \text{ there exists } x'\in X \text{ such that } (x',x_i)\in R_{{\delta},X} \text{ for each } x_i\right\}}.$$ We call ${{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}$ the Dowker ${\delta}$-source simplicial complex associated to $X$. The filtration $\{{{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X} {\hookrightarrow}{{\mathfrak}{D}^{\operatorname{so}}}_{{\delta}',X}\}_{ {\delta}\leq {\delta}'}$ associated to $X$ is called the Dowker source filtration, denoted ${{\mathfrak}{D}^{\operatorname{so}}}_X$. We denote the $k$-dimensional persistence diagram arising from this filtration by ${\operatorname{Dgm}}_k^{{\operatorname{so}}}(X)$. Notice that any construction using ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ can also be repeated using ${{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}$, so we focus on the case of the sink complexes and restate results in terms of source complexes where necessary. In particular, we will prove in §\[sec:dowker-dual\] that $${\operatorname{Dgm}}_k^{{\operatorname{si}}}(X) = {\operatorname{Dgm}}_k^{{\operatorname{so}}}(X) \text{ for any } k\in {\mathbb{Z}}_+,$$ so it makes sense to talk about “the" Dowker diagram associated to $X$. The sink and source filtrations are not equal in general; this is illustrated in Figure \[fig:dowker-three-node\]. As in the case of the Rips filtration, both the Dowker sink and source filtrations are stable. \[prop:dowker-stab\] Let $(X,{\omega}_X), (Y,{\omega}_Y) \in {\mathcal{N}}$. Then ${d_{\operatorname{B}}}({\operatorname{Dgm}}_k^{\bullet}{(X)},{\operatorname{Dgm}}_k^{\bullet}(Y)) \leq 2{d_{\mathcal{N}}}(X,Y).$ Here ${\operatorname{Dgm}}^{\bullet}$ refers to either of ${\operatorname{Dgm}}^{{\operatorname{si}}}$ and ${\operatorname{Dgm}}^{{\operatorname{so}}}$. Both cases are similar, so we just prove the result for ${\operatorname{Dgm}}^{{\operatorname{si}}}$. Let $\eta=2{d_{\mathcal{N}}}(X,Y)$. Then by Proposition \[prop:dn-ko\], there exist maps ${\varphi}:X {\rightarrow}Y, \psi: Y {\rightarrow}X$ such that $$\max({\operatorname{dis}}({\varphi}),{\operatorname{dis}}(\psi), C_{X,Y}({\varphi},\psi),C_{Y,X}(\psi,{\varphi}))\leq \eta.$$ First we check that ${\varphi},\psi$ induce simplicial maps ${\varphi}_{\delta}:{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} {\rightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+\eta,Y}$ and $\psi_{\delta}:{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},Y} {\rightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+\eta,Y}$ for each ${\delta}\in {\mathbb{R}}$. Let ${\delta}' \geq {\delta}\in {\mathbb{R}}$. Let ${\sigma}=[x_0,\ldots, x_n] \in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$. Then there exists $x' \in X$ such that ${\omega}_X(x_i,x') \leq {\delta}$ for each $0\leq i \leq n$. Fix such an $x'$. Since ${\operatorname{dis}}({\varphi}) \leq \eta$, we have the following for each $i$: $$\vert {\omega}_X(x_i,x') - {\omega}_Y({\varphi}(x_i),{\varphi}(x'))\vert \leq \eta.$$ So ${\omega}_Y({\varphi}(x_i),{\varphi}(x')) \leq {\omega}_X(x_i,x') + \eta \leq {\delta}+ \eta$ for each $0\leq i\leq n$. Thus ${\varphi}_{{\delta}}({\sigma}):={\left\{{\varphi}(x_0),\ldots, {\varphi}(x_n)\right\}}$ is a simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+ \eta,Y}$. Thus the map on simplices ${\varphi}_{\delta}$ induced by ${\varphi}$ is simplicial for each ${\delta}\in {\mathbb{R}}$. Similarly we can check that the map $\psi_{\delta}$ on simplices induced by $\psi$ is simplicial. Now to prove the result, it will suffice to check the contiguity conditions in the statement of Lemma \[lem:stab\]. Consider the following diagram: $$\begin{tikzcd}[column sep=large] {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} \arrow{r}{s_{{\delta},{\delta}'}} \arrow[swap]{dr}{{\varphi}_{\delta}} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}',X}\arrow{dr}{{\varphi}_{{\delta}'}} & \\ &{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+\eta,Y} \arrow{r}{t_{{\delta}+\eta,{\delta}'+\eta}} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}'+\eta,Y} \end{tikzcd}$$ Here $s_{{\delta},{\delta}'}$ and $t_{{\delta}+\eta,{\delta}'+\eta}$ are the inclusion maps. We claim that $t_{{\delta}+\eta,{\delta}'+\eta} \circ {\varphi}_{\delta}$ and ${\varphi}_{{\delta}'} \circ s_{{\delta},{\delta}'}$ are contiguous simplicial maps. To see this, let ${\sigma}\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$. Since $s_{{\delta},{\delta}'}$ is just the inclusion, it follows that $t_{{\delta}+\eta,{\delta}'+\eta}({\varphi}_{\delta}({\sigma})) \cup {\varphi}_{{\delta}'}(s_{{\delta},{\delta}'}({\sigma}))={\varphi}_{\delta}({\sigma}),$ which is a simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+\eta,Y}$ because ${\varphi}_{\delta}$ is simplicial, and hence a simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}'+\eta,Y}$ because the inclusion $t_{{\delta}+\eta,{\delta}'+\eta}$ is simplicial. Thus $t_{{\delta}+\eta,{\delta}'+\eta} \circ {\varphi}_{\delta}$ and ${\varphi}_{{\delta}'} \circ s_{{\delta},{\delta}'}$ are contiguous, and their induced linear maps for homology are equal. By a similar argument, one can show that $s_{{\delta}+\eta,{\delta}'+\eta}\circ \psi_{\delta}$ and $\psi_{{\delta}'}\circ t_{{\delta},{\delta}'}$ are contiguous simplicial maps as well. Next we check that the maps $\psi_{{\delta}+\eta}\circ {\varphi}_{\delta}$ and $s_{{\delta},{\delta}+2\eta}$ in the figure below are contiguous. $$\begin{tikzcd} {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} \arrow{rr}{s_{{\delta},{\delta}+2\eta}} \arrow[swap]{dr}{{\varphi}_{\delta}} & {} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+2\eta,X}\\ {} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+\eta,Y} \arrow[swap]{ur}{\psi_{{\delta}+\eta}} & {} \end{tikzcd}$$ Let $x_i\in {\sigma}$. Note that for our fixed ${\sigma}= [x_0,\ldots, x_n]\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ and $x'$, we have: $$\begin{aligned} \|{\omega}_X(x_i,x')-{\omega}_X(\psi({\varphi}(x_i)),\psi({\varphi}(x')))\|&\leq \|{\omega}_X(x_i,x') - {\omega}_Y({\varphi}(x_i),{\varphi}(x'))\|\\ &+ \|{\omega}_Y({\varphi}(x_i),{\varphi}(x')) - {\omega}_X(\psi({\varphi}(x_i)),\psi({\varphi}(x')))\| \\ &\leq 2\eta.\\ \text{Thus we obtain }\hfill {\omega}_X(\psi({\varphi}(x_i)),\psi({\varphi}(x')))&\leq {\omega}_X(x_i,x')+ 2\eta \leq {\delta}+2\eta.\end{aligned}$$ Since this holds for any $x_i \in {\sigma}$, it follows that $\psi_{{\delta}+\eta}({\varphi}_{\delta}({\sigma})) \in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+2\eta,X}$. We further claim that $${\tau}:={\sigma}\cup \psi_{{\delta}+\eta}({\varphi}_{\delta}({\sigma})) = {\left\{x_0,x_1,\ldots, x_n, \psi({\varphi}(x_0)),\ldots, \psi({\varphi}(x_n))\right\}}$$ is a simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+2\eta,X}$. Let $0\leq i\leq n$. It suffices to show that ${\omega}_X(x_i,\psi({\varphi}(x')) \leq {\delta}+ 2\eta$. Notice that from the reformulation of ${d_{\mathcal{N}}}$ (Proposition \[prop:dn-ko\]), we have $$C_{X,Y}({\varphi},\psi) = \max_{(x,y)\in X\times Y}\|{\omega}_X(x,\psi(y)) - {\omega}_Y({\varphi}(x),y)\| \leq \eta .$$ Let $y = {\varphi}(x')$. Then $\|{\omega}_X(x_i,\psi(y)) - {\omega}_Y({\varphi}(x_i),y)\| \leq \eta$. In particular, $${\omega}_X(x_i,\psi({\varphi}(x'))) \leq {\omega}_Y({\varphi}(x_i),{\varphi}(x')) + \eta \leq {\omega}_X(x_i,x') + 2\eta \leq {\delta}+2\eta.$$ Since $0\leq i\leq n$ were arbitrary, it follows that ${\tau}\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}+2\eta,X}$. Thus $\psi_{{\delta}+\eta}\circ {\varphi}_{\delta}$ and $s_{{\delta},{\delta}+2\eta}$ are contiguous. Similarly, one can use the ${\operatorname{dis}}(\psi)$ and $C_{Y,X}(\psi,{\varphi})$ terms to show that $t_{{\delta},{\delta}+2\eta}$ and ${\varphi}_{{\delta}+\eta}\circ \psi_{\delta}$ are contiguous. The result now follows by an application of Lemma \[lem:stab\]. \[rem:dowker-benefits\] The preceding proposition shows that the Dowker persistence diagram is robust to noise in input data, and that the bottleneck distance between Dowker persistence diagrams arising from two networks can be used as a proxy for computing the actual network distance. Note the analogy with Remark \[rem:rips-benefits\]. Both the Dowker and Rips filtrations are valid methods for computing persistent homology of networks, by virtue of their stability results (Propositions \[prop:rips-stab\] and \[prop:dowker-stab\]). However, we present the Dowker filtration as an appropriate method for capturing directionality information in directed networks. In §\[sec:symmetry\] we discuss this particular feature of the Dowker filtration in full detail. In the setting of symmetric networks, the Dowker sink and source simplicial filtrations coincide, and so we automatically obtain ${\operatorname{Dgm}}_k^{{\operatorname{so}}}(X)={\operatorname{Dgm}}_k^{{\operatorname{si}}}(X)$ for any $k\in {\mathbb{Z}}_+$ and any $(X,{\omega}_X)\in {\mathcal{N}}$. When restricted to the setting of metric spaces, the Dowker complex resembles a construction called the witness complex [@de2004topological]. In particular, a version of the Dowker complex for metric spaces, constructed in terms of landmarks and witnesses, was discussed in [@chazal2014persistence], along with stability results. When restricted to the special networks that are pseudo-metric spaces, our definitions and results agree with those presented in [@chazal2014persistence]. The Functorial Dowker Theorem and equivalence of diagrams {#sec:dowker-dual} --------------------------------------------------------- Let $(X,{\omega}_X)\in {\mathcal{N}}$, and let ${\delta}\in {\mathbb{R}}$ be such that $R_{{\delta},X}$ is nonempty. By applying Dowker’s theorem (Theorem \[thm:dowker\]) to the setting $Y=X$, we have $H_k({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}) \cong H_k({{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X})$, for any $k \in {\mathbb{Z}}_+$. We still have this equality in the case where $R_{{\delta},X}$ is empty, because then ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ and ${{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}$ are both empty. Thus we obtain: \[cor:dowker\] Let $(X,{\omega}_X)\in{\mathcal{N}}$, ${\delta}\in {\mathbb{R}}$, and $k\in {\mathbb{Z}}_+$. Then, $$H_k({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}) \cong H_k({{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}).$$ In the persistent setting, Theorem \[thm:dowker\] and Corollary \[cor:dowker\] suggest the following question: > Given a network $(X,{\omega}_X)$ and a fixed dimension $k\in {\mathbb{Z}}_+$, are the persistence diagrams of the Dowker sink and source filtrations of $(X,{\omega}_X)$ necessarily equal? In what follows, we provide a positive answer to the question above. Our strategy is to use the Functorial Dowker Theorem (Theorem \[thm:dowker-functorial\]), for which we will provide a complete proof below. The Functorial Dowker Theorem implies equality between sink and source persistence diagrams. \[cor:dowker-dual\] Let $(X,{\omega}_X)\in {\mathcal{N}}$, and let $k\in {\mathbb{Z}}_+$. Then, $${\operatorname{Dgm}}_k^{{\operatorname{si}}}(X)={\operatorname{Dgm}}_k^{{\operatorname{so}}}(X).$$ Thus we may call either of the diagrams above the $k$-dimensional Dowker diagram of $X$, denoted ${\operatorname{Dgm}}_k^{{\mathfrak}{D}}(X)$. Before proving the corollary, we state an ${\mathbb{R}}$-indexed variant of the Persistence Equivalence Theorem [@edelsbrunner2010computational]. This particular version follows from the isometry theorem [@bauer-isom], and we refer the reader to [@chazal2012structure Chapter 5] for an expanded presentation of this material. \[thm:pet\] Consider two persistence vector spaces ${\mathcal{U}}=\{U^{{\delta}} {\xrightarrow}{\mu_{{\delta},{\delta}'}} U^{{\delta}'}\}_{{\delta}\leq{\delta}' \in {\mathbb{R}}}$ and ${\mathcal{V}}=\{V^{{\delta}} {\xrightarrow}{\nu_{{\delta},{\delta}'}} V^{{\delta}'}\}_{{\delta}\leq {\delta}'\in {\mathbb{R}}}$ with connecting maps $f_{{\delta}}:U^{{\delta}}{\rightarrow}V^{{\delta}'}$. \(00) at (-3,0)[$\cdots$]{}; (1) at (0,0)[$V^{{\delta}}$]{}; (2) at (3,0)[$V^{{\delta}'}$]{}; (3) at (6,0)[$V^{{\delta}''}$]{}; (01) at (9,0)[$\cdots$]{}; \(02) at (-3,2)[$\cdots$]{}; (4) at (0,2)[$U^{{\delta}}$]{}; (5) at (3,2)[$U^{{\delta}'}$]{}; (6) at (6,2)[$U^{{\delta}''}$]{}; (03) at (9,2)[$\cdots$]{}; \(00) edge\[->\] (1); (3) edge\[->\] (01); (02) edge\[->\] (4); (6) edge\[->\] (03); \(1) edge\[->\] (2); (2) edge\[->\] (3); (4) edge\[->\] (5); (5) edge\[->\] (6); (1) edge\[<-\] node\[right\][$f_{{\delta}}$]{} (4); (2) edge\[<-\] node\[right\][$f_{{\delta}'}$]{} (5); (3) edge\[<-\] node\[right\][$f_{{\delta}''}$]{} (6); If the $f_{{\delta}}$ are all isomorphisms and each square in the diagram above commutes, then: $${\operatorname{Dgm}}({\mathcal{U}}) = {\operatorname{Dgm}}({\mathcal{V}}).$$ Let ${\delta}\leq {\delta}' \in {\mathbb{R}}$, and consider the relations $R_{{\delta},X} \subseteq R_{{\delta}',X} \subseteq X\times X$. Suppose first that $R_{{\delta},X}$ and $R_{{\delta}',X}$ are both nonempty. By applying Theorem \[thm:dowker-functorial\], we obtain homotopy equivalences between the source and sink complexes that commute with the canonical inclusions up to homotopy. Passing to the $k$-th homology level, we obtain persistence vector spaces that satisfy the commutativity properties of Theorem \[thm:pet\]. The result follows from Theorem \[thm:pet\]. In the case where $R_{{\delta},X}$ and $R_{{\delta}',X}$ are both empty, there is nothing to show because all the associated complexes are empty. Suppose $R_{{\delta},X}$ is empty, and $R_{{\delta}',X}$ is nonempty. Then ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}$ and ${{\mathfrak}{D}^{\operatorname{so}}}_{{\delta},X}$ are empty, so their inclusions into ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}',X}$ and ${{\mathfrak}{D}^{\operatorname{so}}}_{{\delta}',X}$ induce zero maps upon passing to homology. Thus the commutativity of Theorem \[thm:pet\] is satisfied, and the result follows by Theorem \[thm:pet\]. #### The proof of the Functorial Dowker Theorem It remains to prove Theorem \[thm:dowker-functorial\]. Because the proof involves numerous maps, we will adopt the notational convention of adding a subscript to a function to denote its codomain—e.g. we will write $f_B$ to denote a function with codomain $B$. First we recall the construction of a combinatorial barycentric subdivision (see [@dowker1952homology §2], [@lefschetz1942algebraic §4.7], [@barmak2011algebraic Appendix A]). \[def:subdivision\] For any simplicial complex ${\Sigma}$, one may construct a new simplicial complex ${\Sigma}^{(1)}$, called the first barycentric subdivision, as follows: $${\Sigma}^{(1)}:={\left\{[{\sigma}_1,{\sigma}_2,\ldots, {\sigma}_p] : {\sigma}_1 \subseteq {\sigma}_2 \subseteq \ldots \subseteq {\sigma}_p, \text{ each } {\sigma}_i \in {\Sigma}\right\}}.$$ Note that the vertices of ${\Sigma}^{(1)}$ are the simplices of ${\Sigma}$, and the simplices of ${\Sigma}^{(1)}$ are nested sequences of simplices of ${\Sigma}$. Furthermore, note that given any two simplicial complexes ${\Sigma}, \Xi$ and a simplicial map $f:{\Sigma}{\rightarrow}\Xi$, there is a natural simplicial map $f{^{(1)}}:{\Sigma}{^{(1)}}{\rightarrow}\Xi{^{(1)}}$ defined as: $$f{^{(1)}}([{\sigma}_1,\ldots,{\sigma}_p]):=[f({\sigma}_1),\ldots,f({\sigma}_p)], \qquad {\sigma}_1\subseteq {\sigma}_2\subseteq \ldots, {\sigma}_p, \text{ each } {\sigma}_i\in {\Sigma}.$$ To see that this is simplicial, note that $f({\sigma}_i) \subseteq f({\sigma}_j)$ whenever ${\sigma}_i\subseteq {\sigma}_j$. As a special case, observe that any inclusion map $\iota:{\Sigma}{\hookrightarrow}\Xi$ induces an inclusion map $\iota{^{(1)}}:{\Sigma}{^{(1)}}{\hookrightarrow}\Xi{^{(1)}}$. Given a simplex ${\sigma}=[x_0,\ldots, x_k]$ in a simplicial complex ${\Sigma}$, one defines the barycenter to be the point ${\mathcal}{B}({\sigma}):= \sum_{i=0}^k \tfrac{1}{k+1}x_i \in \|{\Sigma}\|$. Then the spaces $\|{\Sigma}{^{(1)}}\|$ and $\|{\Sigma}\|$ can be identified via a homeomorphism ${\mathcal}{E}_{\|{\Sigma}\|}:\|{\Sigma}{^{(1)}}\| {\rightarrow}\|{\Sigma}\|$ defined on vertices by ${\mathcal}{E}_{\|{\Sigma}\|}({\sigma}):= {\mathcal}{B}({\sigma})$ and extended linearly. Details on the preceding list of definitions can be found in [@munkres-book §2.14-15, 2.19], [@spanier-book §3.3-4], and also [@barmak2011algebraic Appendix A]. The next proposition follows from the discussions in these references, and is a simple restatement of [@barmak2011algebraic Proposition A.1.5]. We provide a proof in the appendix for completeness. \[prop:subdiv-identity\] Let ${\Sigma}$ be a simplicial complex, and let $\Phi: \Sigma{^{(1)}}{\rightarrow}\Sigma$ be a simplicial map such that $\Phi({\sigma}) \in {\sigma}$ for each ${\sigma}\in \Sigma$. Then $\|\Phi\| \simeq {\mathcal}{E}_{\|\Sigma\|}$. We now introduce some auxiliary constructions dating back to [@dowker1952homology] that use the setup stated in Theorem \[thm:dowker-functorial\]. For any nonempty relation $R\subseteq X\times Y$, one may define [@dowker1952homology §2] an associated map $\Phi_{E_R} : E_R{^{(1)}}{\rightarrow}E_R$ as follows: first define $\Phi_{E_R}$ on vertices of $E_R{^{(1)}}$ by $\Phi_{E_R}({\sigma})=s_{{\sigma}}$, where $s_{{\sigma}}$ is the least vertex of ${\sigma}$ with respect to the total order. Next, for any simplex $[{\sigma}_1,\ldots,{\sigma}_k]$ of $E_R{^{(1)}}$, where ${\sigma}_1\subseteq \ldots \subseteq {\sigma}_k$, we have $\Phi_{E_R}({\sigma}_i)=s_{{\sigma}_i} \in {\sigma}_k$ for all $1\leq i\leq k$. Thus $[\Phi_{E_R}({\sigma}_1),\ldots,\Phi_{E_R}({\sigma}_k)]=[s_{{\sigma}_1},s_{{\sigma}_2},\ldots,s_{{\sigma}_k}]$ is a face of ${\sigma}_k$, hence a simplex of ${\Sigma}$. This defines $\Phi_{E_R}$ as a simplicial map $E_R{^{(1)}}{\rightarrow}E_R$. This argument also shows that $\Phi_{E_R}$ is order-reversing: if ${\sigma}\subseteq {\sigma}'$, then $\Phi_{E_R}({\sigma}) \geq \Phi_{E_{R}}({\sigma}')$. \[rem:subdiv-id\] Applying Proposition \[prop:subdiv-identity\] to the setup above, one sees that $\|\Phi_{E_R}\| \simeq {\mathcal}{E}_{\|E_R\|}$. After passing to a second barycentric subdivision $E_R^{(2)}$ (obtained by taking a barycentric subdivision of $E_R{^{(1)}}$) and obtaining a map $\Phi_{E_R{^{(1)}}}:E_R^{(2)} {\rightarrow}E_R{^{(1)}}$, one also has $\|\Phi_{E_R{^{(1)}}}\| \simeq {\mathcal}{E}_{\|E_R{^{(1)}}\|}$. One can also define [@dowker1952homology §3] a simplicial map $\Psi_{F_R}: E_R{^{(1)}}{\rightarrow}F_R$ as follows. Given a vertex ${\sigma}=[x_0,\ldots, x_k] \in E_R{^{(1)}}$, one defines $\Psi_{F_R}({\sigma})=y_{\sigma}$, for some $y_{\sigma}\in Y$ such that $(x_i,y_{\sigma}) \in R$ for each $i$. To see why this vertex map is simplicial, let ${\sigma}{^{(1)}}= [{\sigma}_0,\ldots, {\sigma}_k]$ be a simplex in $E_R{^{(1)}}$. Let $x \in {\sigma}_0$. Then, because ${\sigma}_0 \subseteq {\sigma}_1 \subseteq \ldots \subseteq {\sigma}_k$, we automatically have that $(x,\Psi_{F_R}({\sigma}_i)) \in R$, for each $i=0,\ldots, k$. Thus $\Psi_{F_R}({\sigma}{^{(1)}})$ is a simplex in $F_R$. This definition involves a choice of $y_{\sigma}$ when writing $\Psi_{F_R}({\sigma}) = y_{\sigma}$, but all the maps resulting from such choices are contiguous [@dowker1952homology §3]. The preceding map induces a simplicial map $\Psi_{F_R{^{(1)}}}:E_R^{(2)} {\rightarrow}F_R{^{(1)}}$ as follows. Given a vertex ${\tau}{^{(1)}}= [{\tau}_0,\ldots, {\tau}_k] \in E_R^{(2)}$, i.e. a simplex in $E_R{^{(1)}}$, we define $\Psi_{F_R{^{(1)}}}({\tau}{^{(1)}}) := [\Psi_{F_R}({\tau}_0),\ldots, \Psi_{F_R}({\tau}_k)]$. Since $\Psi_{F_R}$ is simplicial, this is a simplex in $F_R$, i.e. a vertex in $F_R{^{(1)}}$. Thus we have a vertex map $\Psi_{F_R{^{(1)}}}:E_R^{(2)} {\rightarrow}F_R{^{(1)}}$. To check that this map is simplicial, let ${\tau}^{(2)} = [{\tau}{^{(1)}}_0,\ldots, {\tau}{^{(1)}}_p]$ be a simplex in $E_R^{(2)}$. Then ${\tau}{^{(1)}}_0 \subseteq {\tau}{^{(1)}}_1 \subseteq \ldots \subseteq {\tau}{^{(1)}}_p$, and because $\Psi_{F_R}$ is simplicial, we automatically have $$\Psi_{F_R}({\tau}{^{(1)}}_0) \subseteq \Psi_{F_R}({\tau}{^{(1)}}_1) \subseteq \ldots \subseteq \Psi_{F_R}({\tau}{^{(1)}}_p).$$ Thus $\Psi_{F_R{^{(1)}}}({\tau}^{(2)})$ is a simplex in $F_R{^{(1)}}$. We write $F_R^{(2)}$ to denote the barycentric subdivision of $F_R{^{(1)}}$, and obtain simplicial maps $\Phi_{F_R{^{(1)}}}:F_R^{(2)} {\rightarrow}F_R{^{(1)}}$, $\Phi_{F_R}:F_R^{(1)} {\rightarrow}F_R$, $\Psi_{E_R{^{(1)}}}:F_R^{(2)} {\rightarrow}E_R{^{(1)}}$, and $\Psi_{F_R}:E_R^{(1)} {\rightarrow}F_R$ as above. Consider the following diagram: \(1) at (0,0)[$F_R^{(2)}$]{}; (2) at (3,-1)[$F_R^{(1)}$]{}; (3) at (6,-2)[$F_R$]{}; (4) at (8,0)[$F_{R'}^{(2)}$]{}; (5) at (11,-1)[$F_{R'}^{(1)}$]{}; (6) at (14,-2)[$F_{R'}$]{}; (7) at (-1,-3)[$E_R^{(2)}$]{}; (8) at (2,-4)[$E_R{^{(1)}}$]{}; (9) at (5,-5)[$E_R$]{}; (10) at (7,-3)[$E_{R'}^{(2)}$]{}; (11) at (10,-4)[$E_{R'}{^{(1)}}$]{}; (12) at (13,-5)[$E_{R'}$]{}; \(1) edge\[->,violet!80,thick\] node\[right\] (2); (2) edge\[->,violet!80,thick\] node\[above\] (3); (1) edge\[->\] node\[left\] (4); (4) edge\[->\] node\[above right\][$\Phi_{F_{R'}{^{(1)}}}$]{} (5); (5) edge\[->\] node\[above right\][$\Phi_{F_{R'}}$]{} (6); (3) edge\[->\] node\[above\] (6); \(7) edge\[->,teal!80,thick\] node\[below\] (8); (8) edge\[->,teal!80,thick\] node\[below\] (9); (7) edge\[->,dashed\] node\[left\] (10); (10) edge\[->,dashed\] node\[above right\] (11); (11) edge\[->,dashed\] node\[above\] (12); (9) edge\[->\] node\[above\] (12); \(1) edge\[->,orange,thick\] node\[left,pos=0.2\] (8); (8) edge\[->,orange,thick\] node\[below right,pos=0.2\] (3); (4) edge\[->,dashed\] node\[left\] (11); (11) edge\[->,dashed\] node\[below right\] (6); \(7) edge\[->,NavyBlue!80,thick\] node\[above left,pos=0.2\] (2); (2) edge\[->,NavyBlue!80,thick\] node\[left,pos=0.2\] (9); (10) edge\[->,dashed\] node\[left\] (5); (5) edge\[->,dashed\] node\[below right\] (12); \(1) at (0,0)[$E_R^{(2)}$]{}; (2) at (2,-1)[$E_R^{(1)}$]{}; (3) at (4,-2)[$E_R$]{}; (4) at (6,0)[$E_{R'}^{(2)}$]{}; (5) at (8,-1)[$E_{R'}^{(1)}$]{}; (6) at (10,-2)[$E_{R'}$]{}; (7) at (3,1)[$F_R{^{(1)}}$]{}; (8) at (9,1)[$F_{R'}{^{(1)}}$]{}; (1) edge\[->\] node\[below left\] (2); (2) edge\[->\] node\[below left\] (3); (1) edge\[->,dashed\] node\[left\] (4); (4) edge\[->,dashed\] node\[left\] (5); (5) edge\[->,dashed\] node\[left\] (6); (3) edge\[->\] node\[above\][$\iota_E$]{} (6); (1) edge\[->\] node\[above\] (7); (7) edge\[->\] node\[right\] (3); (4) edge\[->,dashed\] node\[left\] (8); (8) edge\[->\] node\[right\] (6); (7) edge\[->\] node\[above\][$\iota_F$]{} (8); \(1) at (0,0)[$F_R^{(2)}$]{}; (2) at (2,-1)[$F_R^{(1)}$]{}; (3) at (4,-2)[$F_R$]{}; (4) at (6,0)[$F_{R'}^{(2)}$]{}; (5) at (8,-1)[$F_{R'}^{(1)}$]{}; (6) at (10,-2)[$F_{R'}$]{}; (7) at (1,-3)[$E_R{^{(1)}}$]{}; (8) at (7,-3)[$E_{R'}{^{(1)}}$]{}; (1) edge\[->\] node\[right\] (2); (2) edge\[->\] node\[above\] (3); (1) edge\[->\] node\[left\] (4); (4) edge\[->\] node\[above right\][$\Phi_{F_{R'}{^{(1)}}}$]{} (5); (5) edge\[->\] node\[above\][$\Phi_{F_{R'}}$]{} (6); (3) edge\[->\] node\[above\] (6); (1) edge\[->\] node\[left\] (7); (7) edge\[->\] node\[above left\] (3); (4) edge\[->,dashed\] node\[left\] (8); (8) edge\[->\] node\[below right\] (6); (7) edge\[->\] node\[above\] (8); We proceed by claiming contiguity of the following. \(1) at (0,0)[$E_R^{(2)}$]{}; (2) at (2,0)[$E_R^{(1)}$]{}; (3) at (4,0)[$E_R$]{}; (7) at (2,2)[$F_R{^{(1)}}$]{}; (1) edge\[->,teal\] node\[below\] (2); (2) edge\[->,teal\] node\[below\] (3); (1) edge\[->,NavyBlue\] node\[left\] (7); (7) edge\[->,NavyBlue\] node\[right\] (3); \(1) at (0,2)[$F_R^{(2)}$]{}; (2) at (2,2)[$F_R^{(1)}$]{}; (3) at (4,2)[$F_R$]{}; (7) at (2,0)[$E_R{^{(1)}}$]{}; (1) edge\[->,violet\] node\[below\] (2); (2) edge\[->,violet\] node\[below\] (3); (1) edge\[->,orange\] node\[below left\] (7); (7) edge\[->,orange\] node\[below right\] (3); \(1) at (0,2)[$F_R^{(2)}$]{}; (2) at (3,2)[$F_R^{(1)}$]{}; (7) at (3,0)[$E_R{^{(1)}}$]{}; (3) at (6,0)[$E_R$]{}; (1) edge\[->,violet\] node\[below\] (2); (2) edge\[->,NavyBlue\] node\[left\] (3); (1) edge\[->,orange\] node\[below left\] (7); (7) edge\[->,teal\] node\[above\] (3); \(1) at (0,0)[$E_R^{(2)}$]{}; (2) at (3,0)[$E_R^{(1)}$]{}; (7) at (3,2)[$F_R{^{(1)}}$]{}; (3) at (6,2)[$F_R$]{}; (1) edge\[->,teal\] node\[above\] (2); (2) edge\[->,orange\] node\[below right\] (3); (1) edge\[->,NavyBlue\] node\[above left\] (7); (7) edge\[->,violet\] node\[below\] (3); \[cl:dowker-contigo-items\] More specifically: 1. $\Phi_{E_R}\circ \Phi_{E_R{^{(1)}}}$ and $\Psi_{E_R}\circ \Psi_{F_R{^{(1)}}}$ are contiguous. \[item:dowker-contigo-1\] 2. $\Phi_{F_R}\circ \Phi_{F_R{^{(1)}}}$ and $\Psi_{F_R}\circ \Psi_{E_R{^{(1)}}}$ are contiguous. \[item:dowker-contigo-2\] 3. $\Psi_{E_R}\circ \Phi_{F_R{^{(1)}}}$ and $\Phi_{E_R}\circ \Psi_{E_R{^{(1)}}}$ are contiguous. \[item:dowker-contigo-3\] 4. $\Psi_{F_R}\circ \Phi_{E_R{^{(1)}}}$ and $\Phi_{F_R}\circ \Psi_{F_R{^{(1)}}}$ are contiguous. \[item:dowker-contigo-4\] Items (\[item:dowker-contigo-1\]) and (\[item:dowker-contigo-3\]) appear in the proof of Dowker’s theorem [@dowker1952homology Lemmas 5, 6], and it is easy to see that a symmetric argument shows Items (\[item:dowker-contigo-2\]) and (\[item:dowker-contigo-4\]). For completeness, we will verify these items in this paper, but defer this verification to the end of the proof. By passing to the geometric realization and applying Proposition \[prop:contigo-props\] and Remark \[rem:subdiv-id\], we obtain the following from Item (\[item:dowker-contigo-3\]) of Claim \[cl:dowker-contigo-items\]: $$\begin{aligned} \|\Psi_{E_R}\|\circ \|\Phi_{F_R{^{(1)}}}\| &\simeq \|\Phi_{E_R}\|\circ\|\Psi_{E_R{^{(1)}}}\|,\\ \|\Psi_{E_R}\|\circ {\mathcal}{E}_{\|F_R{^{(1)}}\|} &\simeq {\mathcal}{E}_{\|E_R\|}\circ\|\Psi_{E_R{^{(1)}}}\|, &&\text{(Remark \ref{rem:subdiv-id})}\\ {\mathcal}{E}{^{-1}}_{\|E_R\|} \circ \|\Psi_{E_R}\| \circ {\mathcal}{E}_{\|F_R{^{(1)}}\|} &\simeq \|\Psi_{E_R{^{(1)}}}\|. &&\text{(${\mathcal}{E}$ is a homeomorphism, hence invertible)}\end{aligned}$$ Replacing this term in the expression for Item (\[item:dowker-contigo-2\]) of Claim \[cl:dowker-contigo-items\], we obtain: $$\begin{aligned} \|\Psi_{F_R}\| \circ \|\Psi_{E_R{^{(1)}}}\| &\simeq \|\Phi_{F_R}\|\circ \|\Phi_{F_R{^{(1)}}}\| \simeq {\mathcal}{E}_{\|F_R\|}\circ {\mathcal}{E}_{\|F_R{^{(1)}}\|},\\ \|\Psi_{F_R}\| \circ {\mathcal}{E}{^{-1}}_{\|E_R\|} \circ \|\Psi_{E_R}\| \circ {\mathcal}{E}_{\|F_R{^{(1)}}\|} &\simeq {\mathcal}{E}_{\|F_R\|}\circ {\mathcal}{E}_{\|F_R{^{(1)}}\|},\\ \|\Psi_{F_R}\| \circ {\mathcal}{E}{^{-1}}_{\|E_R\|} \circ \|\Psi_{E_R}\|\circ {\mathcal}{E}{^{-1}}_{\|F_R\|} &\simeq {\operatorname{id}}_{\|F_R\|}.\end{aligned}$$ Similarly, we obtain the following from Item (\[item:dowker-contigo-4\]) of Claim \[cl:dowker-contigo-items\]: $$\|\Psi_{F_R}\|\circ \|\Phi_{E_R{^{(1)}}}\| \simeq \|\Phi_{F_R}\| \circ \|\Psi_{F_R{^{(1)}}}\|, \text{ so } {\mathcal}{E}{^{-1}}_{\|F_R\|} \circ \|\Psi_{F_R}\| \circ {\mathcal}{E}_{\|E_R{^{(1)}}}\| \simeq \|\Psi_{F_R{^{(1)}}}\|.$$ Replacing this term in the expression for Item (\[item:dowker-contigo-1\]) of Claim \[cl:dowker-contigo-items\], we obtain: $$\begin{aligned} \|\Psi_{E_R}\|\circ \|\Psi_{F_R{^{(1)}}}\| & \simeq\|\Phi_{E_R}\|\circ \|\Phi_{E_R{^{(1)}}}\| \simeq {\mathcal}{E}_{\|E_R\|}\circ {\mathcal}{E}_{\|E_R{^{(1)}}\|}, \\ \|\Psi_{E_R}\|\circ {\mathcal}{E}{^{-1}}_{\|F_R\|} \circ \|\Psi_{F_R}\| \circ {\mathcal}{E}_{\|E_R{^{(1)}}}\| &\simeq {\mathcal}{E}_{\|E_R\|}\circ {\mathcal}{E}_{\|E_R{^{(1)}}\|}\\ \|\Psi_{E_R}\|\circ {\mathcal}{E}{^{-1}}_{\|F_R\|} \circ \|\Psi_{F_R}\|\circ {\mathcal}{E}{^{-1}}_{\|E_R\|} &\simeq {\operatorname{id}}_{\|E_R\|}\end{aligned}$$ Define $\Gamma_{\|E_R\|}: \|F_R\| {\rightarrow}\|E_R\|$ by $\Gamma_{\|E_R\|} := \|\Psi_{E_R}\|\circ {\mathcal}{E}{^{-1}}_{\|F_R\|}$. Then $\Gamma_{\|E_R\|}$ is a homotopy equivalence, with inverse given by $\|\Psi_{F_R}\|\circ {\mathcal}{E}{^{-1}}_{\|E_R\|}$. This shows that $\|F_R\|\simeq \|E_R\|$, for any nonempty relation $R\subseteq X\times Y$. Next we need to show that $\Gamma_{\|E_R\|}$ commutes with the canonical inclusion. Consider the following diagram, where the $\iota_\bullet$ maps denote the respective canonical inclusions (cf. Definition \[def:subdivision\]): \(1) at (0,0)[$F_R{^{(1)}}$]{}; (2) at (3,0)[$F_{R'}{^{(1)}}$]{}; (3) at (0,-1.5)[$F_R$]{}; (4) at (3,-1.5)[$F_{R'}$]{}; (5) at (-2,-3)[$E_R$]{}; (6) at (5,-3)[$E_{R'}$]{}; (1) edge\[->\] node\[above\] (2); (3) edge\[->\] node\[above\] (4); (5) edge\[->\] node\[above\] (6); (1) edge\[->\] node\[right\] (3); (2) edge\[->\] node\[left\] (4); (1) edge\[->\] node\[left\] (5); (2) edge\[->\] node\[right\] (6); \[cl:dowker-func-1\] $\iota_E\circ \Psi_{E_R}$ and $\Psi_{E_{R'}}\circ \iota_{F{^{(1)}}}$ are contiguous. \[cl:dowker-func-2\] $\iota_F\circ \Phi_{F_R}$ and $\Phi_{F_{R'}}\circ \iota_{F{^{(1)}}}$ are contiguous. Suppose Claim \[cl:dowker-func-2\] is true. Then, upon passing to geometric realizations, we have: $$\begin{aligned} \|\iota_F\| \circ {\mathcal}{E}_{\|F_R\|} \simeq \|\iota_F\|\circ \|\Phi_{F_R}\| \simeq \|\Phi_{F_{R'}}\|\circ \|\iota_{F{^{(1)}}}\| &\simeq {\mathcal}{E}_{\|F_{R'}\|}\circ \|\iota_{F{^{(1)}}}\|,\\ {\mathcal}{E}{^{-1}}_{\|F_{R'}\|}\circ \|\iota_F\| \circ {\mathcal}{E}_{\|F_R\|} &\simeq \|\iota_{F{^{(1)}}}\|.\end{aligned}$$ Suppose also that Claim \[cl:dowker-func-1\] is true. Then we have: $$\begin{aligned} \|\Psi_{E_{R'}}\| \circ \|\iota_{F{^{(1)}}}\| &\simeq \|\iota_E\|\circ \|\Psi_{E_R}\|,\\ \|\Psi_{E_{R'}}\| \circ {\mathcal}{E}{^{-1}}_{\|F_{R'}\|}\circ \|\iota_F\| \circ {\mathcal}{E}_{\|F_R\|} &\simeq \|\iota_E\|\circ \|\Psi_{E_R}\|,\\ \|\Psi_{E_{R'}}\| \circ {\mathcal}{E}{^{-1}}_{\|F_{R'}\|}\circ \|\iota_F\| &\simeq \|\iota_E\|\circ \|\Psi_{E_R}\| \circ {\mathcal}{E}{^{-1}}_{\|F_R\|}, \text{ i.e. }\\ \Gamma_{\|E_{R'}\|}\circ \|\iota_F\| &\simeq \|\iota_E\| \circ \Gamma_{\|E_R\|}.\end{aligned}$$ This proves the theorem. It only remains to prove the various claims. In proving Claim \[cl:dowker-contigo-items\], we supply the proofs of Items (\[item:dowker-contigo-2\]) and (\[item:dowker-contigo-4\]). These arguments are adapted from [@dowker1952homology Lemmas 1, 5, and 6], where the proofs of Items (\[item:dowker-contigo-1\]) and (\[item:dowker-contigo-3\]) appeared. For Item (\[item:dowker-contigo-2\]), let ${\tau}^{(2)}=[{\tau}_0{^{(1)}},\ldots,{\tau}_k{^{(1)}}]$ be a simplex in $F_R^{(2)}$, where ${\tau}_0{^{(1)}}\subseteq \ldots \subseteq {\tau}_k{^{(1)}}$ is a chain of simplices in $F_R{^{(1)}}$. By the order-reversing property of the map $\Phi_{F_R{^{(1)}}}$, we have that $\Phi_{F_R{^{(1)}}}({\tau}_0{^{(1)}}) \supseteq \Phi_{F_R{^{(1)}}}({\tau}_i{^{(1)}})$ for each $i=0,\ldots, k$. Define $x:= \Psi_{E_R}(\Phi_{F_R{^{(1)}}}({\tau}_0{^{(1)}}))$. Then $(x,y) \in R$ for each $y \in \Phi_{F_R{^{(1)}}}({\tau}_0{^{(1)}})$. But we also have $(x,\Phi_{F_R}(\Phi_{F_R{^{(1)}}}({\tau}_i{^{(1)}}))) \in R$ for each $i=0,\ldots, k$, because $\Phi_{F_R}(\Phi_{F_R{^{(1)}}}({\tau}_i{^{(1)}})) \in \Phi_{F_R{^{(1)}}}({\tau}_i{^{(1)}}) \subseteq \Phi_{F_R}({\tau}_0{^{(1)}})$ for each $i=0,\ldots, k$. Next let $0\leq i \leq k$. For each ${\tau}\in {\tau}{^{(1)}}_i$, we have $\Psi_{E_R}({\tau}) \in \Psi_{E_R{^{(1)}}}({\tau}{^{(1)}}_i)$ (by the definition of $\Psi_{E_R{^{(1)}}}$). Because $\Phi_{F_R{^{(1)}}}({\tau}_0{^{(1)}}) \in {\tau}_0{^{(1)}}\subseteq {\tau}_i{^{(1)}}$, we then have $x = \Psi_{E_R}(\Phi_{F_R{^{(1)}}}({\tau}_0{^{(1)}})) \in \Psi_{E_R{^{(1)}}}({\tau}{^{(1)}}_i)$, which is a vertex of $E_R{^{(1)}}$ or alternatively a simplex of $E_R$. But then, by definition of $\Psi_{F_R}$, we have that $(x,\Psi_{F_R}(\Psi_{E_R{^{(1)}}}({\tau}_i{^{(1)}}))) \in R$. This holds for each $0\leq i \leq k$. Since ${\tau}^{(2)}$ was arbitrary, this shows that $\Phi_{F_R}\circ \Phi_{F_R{^{(1)}}}$ and $\Psi_{F_R}\circ \Psi_{E_R{^{(1)}}}$ are contiguous. For Item (\[item:dowker-contigo-4\]), let ${\sigma}^{(2)} = [{\sigma}{^{(1)}}_0,\ldots,{\sigma}{^{(1)}}_k]$ be a simplex in $E_R^{(2)}$. Let $0\leq i \leq k$. Then ${\sigma}{^{(1)}}_0 \subseteq \ldots \subseteq {\sigma}{^{(1)}}_k$, and $\Phi_{E_R{^{(1)}}}({\sigma}{^{(1)}}_i) \in {\sigma}{^{(1)}}_i \subseteq {\sigma}{^{(1)}}_k$. So $\Psi_{F_R}(\Phi_{E_R{^{(1)}}}({\sigma}_i{^{(1)}})) \in \Psi_{F_R{^{(1)}}}({\sigma}_k{^{(1)}})$. On the other hand, we have $\Psi_{F_R{^{(1)}}}({\sigma}_i{^{(1)}}) \subseteq \Psi_{F_R{^{(1)}}}({\sigma}_k{^{(1)}})$. Then $\Phi_{F_R}(\Psi_{F_R{^{(1)}}}({\sigma}_i{^{(1)}})) \in \Psi_{F_R{^{(1)}}}({\sigma}_i{^{(1)}}) \subseteq \Psi_{F_R{^{(1)}}}({\sigma}_k{^{(1)}})$. Since $i$ was arbitrary, this shows that $\Psi_{F_R}\circ \Phi_{E_R{^{(1)}}}$ and $\Phi_{F_R}\circ \Psi_{F_R{^{(1)}}}$ both map the vertices of ${\sigma}^{(2)}$ to the simplex $\Psi_{F_R{^{(1)}}}({\sigma}_k{^{(1)}})$, hence are contiguous. This concludes the proof of the claim. Let ${\tau}{^{(1)}}=[{\tau}_0,{\tau}_1,\ldots, {\tau}_k] \in F_R{^{(1)}}$, where ${\tau}_0 \subseteq {\tau}_1\subseteq \ldots \subseteq {\tau}_k$ is a chain of simplices in $F_R$. Then $\iota_{F{^{(1)}}}({\tau}{^{(1)}}) = {\tau}{^{(1)}}$, and $\Psi_{E_{R'}}({\tau}{^{(1)}})=[x_{{\tau}_0},\ldots, x_{{\tau}_k}]$, for some choice of $x_{{\tau}_i}$ terms. Also we have $\iota_E\circ \Psi_{E_{R}}({\tau}{^{(1)}}) = [x'_{{\tau}_0},\ldots, x'_{{\tau}_k}]$ for some other choice of $x'_{{\tau}_i}$ terms. For contiguity, we need to show that $$[x_{{\tau}_0},\ldots, x_{{\tau}_k}, x'_{{\tau}_0},\ldots, x'_{{\tau}_k}] \in E_{R'}.$$ But this is easy to see: letting $y \in {\tau}_0$, we have ${\left\{(x_{{\tau}_0},y),\ldots,(x_{{\tau}_k},y),(x'_{{\tau}_0},y),\ldots,(x'_{{\tau}_k},y)\right\}} \subseteq R$. Since ${\tau}{^{(1)}}$ was arbitrary, it follows that we have contiguity. Let ${\tau}{^{(1)}}=[{\tau}_0,{\tau}_1,\ldots, {\tau}_k] \in F_R{^{(1)}}$, where ${\tau}_0 \subseteq {\tau}_1\subseteq \ldots \subseteq {\tau}_k$ is a chain of simplices in $F_R$. Then $\Phi_{F_R}({\tau}_i) \in {\tau}_k$ for each $0\leq i \leq k$. Thus $\iota_F\circ \Phi_{F_R}({\tau}{^{(1)}})$ is a face of ${\tau}_k$. Similarly, $\Phi_{F_{R'}}\circ \iota_{F{^{(1)}}}({\tau}{^{(1)}})$ is also a face of ${\tau}_k$. Since ${\tau}{^{(1)}}$ was an arbitrary simplex of $F_R{^{(1)}}$, it follows that $\iota_F\circ \Phi_{F_R}$ and $\Phi_{F_{R'}}\circ \iota_{F{^{(1)}}}$ are contiguous. The equivalence between the finite FDT and the simplicial FNTs {#sec:dowker-nerve-equiv} -------------------------------------------------------------- In this section, we present our answer to Question \[q:f-nerve-f-dowker\]. We begin with a weaker formulation of Theorem \[thm:dowker-functorial\] and some simplicial Functorial Nerve Theorems. \[thm:dowker-functorial-finite\] Let $X,Y$ be two totally ordered sets, and without loss of generality, suppose $X$ is finite. Let $R\subseteq R' \subseteq X\times Y$ be two nonempty relations, and let $E_R, F_R, E_{R'}, F_{R'}$ be their associated simplicial complexes (as in Theorem \[thm:dowker-functorial\]). Then there exist homotopy equivalences $\Gamma_{\|E_R\|}:\|F_R\| {\rightarrow}\|E_R\|$ and $\Gamma_{\|E_{R'}\|}: \|F_{R'}\| {\rightarrow}\|E_{R'}\|$ that commute up to homotopy with the canonical inclusions. The finite FDT (Theorem \[thm:dowker-functorial-finite\]) is an immediate consequence of the general FDT (Theorem \[thm:dowker-functorial\]). Let ${\mathcal}{A} = {\left\{A_i\right\}}_{i\in I}$ be a family of nonempty sets indexed by $I$. The nerve of ${\mathcal}{A}$ is the simplicial complex ${\mathcal}{N}({\mathcal}{A}):= \{{\sigma}\in {\operatorname{pow}}(I) : {\sigma}\text{ is finite, nonempty, and } \cap_{i \in {\sigma}}A_i \neq {\varnothing}\}$. Let ${\Sigma}$ be a simplicial complex. Then a collection of subcomplexes ${\mathcal}{A}_{\Sigma}= \{{\Sigma}_i\}_{i\in I}$ is said to be a cover of subcomplexes for ${\Sigma}$ if ${\Sigma}= \cup_{i\in I}{\Sigma}_i$. Furthermore, ${\mathcal}{A}_{\Sigma}$ is said to be a cover of simplices if each ${\Sigma}_i \in {\mathcal}{A}_{\Sigma}$ has the property that ${\Sigma}_i = {\operatorname{pow}}(V({\Sigma}_i))$. In this case, each ${\Sigma}_i$ has precisely one top-dimensional simplex, consisting of the vertex set $V({\Sigma}_i)$. We present two simplicial formulations of the Functorial Nerve Theorem that turn out to be equivalent; the statements differ in that one is about covers of simplices and the other is about covers of subcomplexes. \[thm:nerve-functorial-I\] Let ${\Sigma}\subseteq {\Sigma}'$ be two simplicial complexes, and let ${\mathcal}{A}_{\Sigma}=\{{\Sigma}_i\}_{i\in I}$, ${\mathcal}{A}_{{\Sigma}'}=\{{\Sigma}'_i\}_{i\in I'}$ be finite covers of simplices for ${\Sigma}$ and ${\Sigma}'$ such that $I\subseteq I'$ and ${\Sigma}_i \subseteq {\Sigma}'_i$ for each $i \in I$. In particular, ${\operatorname{card}}(I') < \infty$. Suppose that for each finite subset ${\sigma}\subseteq I'$, the intersection $\cap_{i \in {\sigma}}{\Sigma}'_i$ is either empty or contractible (and likewise for $\cap_{i \in {\sigma}}{\Sigma}_i$). Then $\|{\Sigma}\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{\Sigma})\|$ and $\|{\Sigma}'\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})\|$, via maps that commute up to homotopy with the canonical inclusions. \[thm:nerve-functorial-II\] The statement of Theorem \[thm:nerve-functorial-I\] holds even if ${\mathcal}{A}_{\Sigma}$ and ${\mathcal}{A}_{{\Sigma}'}$ are covers of subcomplexes. Explicitly, the statement is as follows. Let ${\Sigma}\subseteq {\Sigma}'$ be two simplicial complexes, and let ${\mathcal}{A}_{\Sigma}=\{{\Sigma}_i\}_{i\in I}$, ${\mathcal}{A}_{{\Sigma}'}=\{{\Sigma}'_i\}_{i\in I'}$ be finite covers of subcomplexes for ${\Sigma}$ and ${\Sigma}'$ such that $I\subseteq I'$ and ${\Sigma}_i \subseteq {\Sigma}'_i$ for each $i \in I$. In particular, ${\operatorname{card}}(I') < \infty$. Suppose that for each finite subset ${\sigma}\subseteq I'$, the intersection $\cap_{i \in {\sigma}}{\Sigma}'_i$ is either empty or contractible (and likewise for $\cap_{i \in {\sigma}}{\Sigma}_i$). Then $\|{\Sigma}\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{\Sigma})\|$ and $\|{\Sigma}'\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})\|$, via maps that commute up to homotopy with the canonical inclusions. The following result summarizes our answer to Question \[q:f-nerve-f-dowker\]. \[thm:dowker-nerve-eq\] The finite FDT, the FNT I, and the FNT II are all equivalent. Moreover, all of these results are implied by the FDT, as below: (fdt) at (-2,0)[Theorem \[thm:dowker-functorial\]]{}; (ffdt) at (1,0)[Theorem \[thm:dowker-functorial-finite\]]{}; (sfnt1) at (3,1)[Theorem \[thm:nerve-functorial-I\]]{}; (sfnt2) at (5,0)[Theorem \[thm:nerve-functorial-II\]]{}; (fdt) – (ffdt); (ffdt) – (sfnt1); (sfnt1) – (sfnt2); (sfnt2) – (ffdt); We present the proof of Theorem \[thm:dowker-nerve-eq\] over the course of the next few subsections. By virtue of Theorem \[thm:dowker-nerve-eq\], we will write simplicial FNT to mean either of the FNT I or FNT II. Theorem \[thm:dowker-functorial-finite\] implies Theorem \[thm:nerve-functorial-I\] {#theorem-thmdowker-functorial-finite-implies-theorem-thmnerve-functorial-i .unnumbered} ----------------------------------------------------------------------------------- Let $V, V'$ denote the vertex sets of ${\Sigma},{\Sigma}'$, respectively. We define the relations $R\subseteq V\times I$ and $R' \subseteq V'\times I'$ as follows: $(v,i) \in R \iff v \in {\Sigma}_i$ and $(v',i') \in R' \iff v' \in {\Sigma}_i'.$ Then $R \subseteq R'$, the set $I'$ is finite by assumption, and so we are in the setting of the finite FDT (Theorem \[thm:dowker-functorial-finite\]) (perhaps invoking the Axiom of Choice to obtain the total order on $V'$). It suffices to show that $E_R = {\Sigma}$, $E_{R'} = {\Sigma}'$, $F_R = {\mathcal}{N}({\mathcal}{A}_{\Sigma})$, and $F_{R'} = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})$, where $E_R, E_{R'}, F_R, F_{R'}$ are as defined in Theorem \[thm:dowker-functorial\]. First we claim the $E_R = {\Sigma}$. By the definitions of $R$ and $E_R$, we have $E_R = \{{\sigma}\subseteq V : \; \exists i \in I, \; (v,i) \in R \; \forall \; v\in {\sigma}\} = \{{\sigma}\subseteq V : \;\exists i \in I, \; v\in {\Sigma}_i \; \forall \; v\in {\sigma}\}.$ Let ${\sigma}\in E_R$, and let $i \in I$ be such that $v \in {\Sigma}_i$ for all $v \in {\sigma}$. Then ${\sigma}\subseteq V({\Sigma}_i)$, and since ${\Sigma}_i = {\operatorname{pow}}(V({\Sigma}_i))$ by the assumption about covers of simplices, we have ${\sigma}\in {\Sigma}_i \subseteq {\Sigma}$. Thus $E_R \subseteq {\Sigma}$. Conversely, let ${\sigma}\in {\Sigma}$. Then ${\sigma}\in {\Sigma}_i$ for some $i$. Thus for all $v \in {\sigma}$, we have $(v,i) \in R$. It follows that ${\sigma}\in E_R$. This shows $E_R = {\Sigma}$. The proof that $E_{R'} = {\Sigma}'$ is analogous. Next we claim that $F_R = {\mathcal}{N}({\mathcal}{A}_{\Sigma})$. By the definition of $F_R$, we have $F_R = \{{\tau}\subseteq I : \; \exists v \in V, \; (v,i) \in R \; \forall \; i\in {\tau}\}.$ Let ${\tau}\in F_R$, and let $v \in V$ be such that $(v,i) \in R$ for each $i \in {\tau}$. Then $\cap_{i \in {\tau}}{\Sigma}_i \neq {\varnothing}$, and so ${\tau}\in {\mathcal}{N}({\mathcal}{A}_{\Sigma})$. Conversely, let ${\tau}\in {\mathcal}{N}({\mathcal}{A}_{\Sigma})$. Then $\cap_{i \in {\tau}}{\Sigma}_i \neq {\varnothing}$, so there exists $v \in V$ such that $v \in {\Sigma}_i$ for each $i \in {\tau}$. Thus ${\sigma}\in F_R$. This shows $F_R = {\mathcal}{N}({\mathcal}{A}_{\Sigma})$. The case for $R'$ is analogous. An application of Theorem \[thm:dowker-functorial-finite\] now completes the proof. Theorem \[thm:nerve-functorial-II\] implies Theorem \[thm:dowker-functorial-finite\] {#theorem-thmnerve-functorial-ii-implies-theorem-thmdowker-functorial-finite .unnumbered} ------------------------------------------------------------------------------------ Let $X$ and $Y$ be two sets, and suppose $X$ is finite. Let $R\subseteq R' \subseteq X \times Y$ be two relations. Consider the simplicial complexes $E_R, F_R, E_{R'}, F_{R'}$ as defined in Theorem \[thm:dowker-functorial\]. Let $V_R:=V(E_R)$. For each $x \in V_R$, define $A_x:=\{{\tau}\in F_R : (x,y) \in R \text{ for all } y \in {\tau}\}$. Then $A_x$ is a subcomplex of $F_R$. Furthermore, $\cup_{x\in V_R}A_x = F_R$. To see this, let ${\tau}\in F_R$. Then there exists $x\in X$ such that $(x,y) \in R$ for all $y\in {\tau}$, and so ${\tau}\in A_x$. Let ${\mathcal}{A}:= \{A_x : x \in V_R\}$. We have seen that ${\mathcal}{A}$ is a cover of subcomplexes for $F_R$. It is finite because the indexing set $V_R$ is a subset of $X$, which is finite by assumption. Next we claim that ${\mathcal}{N}({\mathcal}{A}) = E_R$. Let ${\sigma}\in E_R$. Then there exists $y \in Y$ such that $(x,y) \in R$ for all $x\in {\sigma}$. Thus $\cap_{x\in {\sigma}}A_x \neq {\varnothing}$, and so ${\sigma}\in {\mathcal}{N}({\mathcal}{A})$. Conversely, let ${\sigma}\in {\mathcal}{N}({\mathcal}{A})$. Then $\cap_{x \in {\sigma}}A_x \neq {\varnothing}$, and so there exists $y \in Y$ such that $(x,y) \in R$ for all $x \in {\sigma}$. Thus ${\sigma}\in E_R$. Next we check that nonempty finite intersections of elements in ${\mathcal}{A}$ are contractible. Let ${\sigma}\in {\mathcal}{N}({\mathcal}{A}) = E_R$. Let $V_{\sigma}:= \cap_{x \in {\sigma}}V(A_x) \subseteq V(F_R)$. We claim that $\cap_{x \in {\sigma}}A_x = {\operatorname{pow}}(V_{\sigma})$, i.e. that the intersection is a full simplex in $F_R$, hence contractible. The inclusion $\cap_{x \in {\sigma}}A_x \subseteq {\operatorname{pow}}(V_{\sigma})$ is clear, so we show the reverse inclusion. Let ${\tau}\in {\operatorname{pow}}(V_{\sigma})$, and let $y \in {\tau}$. Then $y \in \cap_{x \in {\sigma}}A_x$, so $(x,y) \in R$ for each $x \in {\sigma}$. This holds for each $y \in {\tau}$, so it follows that ${\tau}\in \cap_{x\in {\sigma}}A_x$. Thus $\cap_{x \in {\sigma}}A_x = {\operatorname{pow}}(V_{\sigma})$. We remark that this also shows that ${\mathcal}{A}$ is a cover of simplices for $F_R$. Now for each $x \in V(E_{R'})$, define $A'_x:= \{{\tau}\in F_{R'} : (x,y) \in R' \text{ for all } y \in {\tau}\}$. Also define ${\mathcal}{A}':= \{A'_x : x \in V(E_{R'})\}$. The same argument shows that ${\mathcal}{A}'$ is a finite cover of subcomplexes (in particular, a cover of simplices) for $F_{R'}$ with all finite intersections either empty or contractible, and that $E_{R'} = {\mathcal}{N}({\mathcal}{A}')$. An application of Theorem \[thm:nerve-functorial-II\] now shows that $\|E_R\| \simeq \|F_R\|$ and $\|E_{R'}\| \simeq \|F_{R'}\|$, via maps that commute up to homotopy with the inclusions $\|E_R\| {\hookrightarrow}\|E_{R'}\|$ and $\|F_R\| {\hookrightarrow}\|F_{R'}\|$. Theorem \[thm:nerve-functorial-I\] implies Theorem \[thm:nerve-functorial-II\] {#theorem-thmnerve-functorial-i-implies-theorem-thmnerve-functorial-ii .unnumbered} ------------------------------------------------------------------------------ We lead with some remarks about the ideas involved in this proof. Theorem \[thm:nerve-functorial-II\] is a functorial statement in the sense that it is about an arbitrary inclusion ${\Sigma}\subseteq {\Sigma}'$. Restricting the statement to just ${\Sigma}$ would lead to a non-functorial statement. A proof of this non-functorial statement, via a non-functorial analogue of Theorem \[thm:nerve-functorial-I\], can be obtained using techniques presented in [@bjorner1985homotopy] (see also [@kozlov2007combinatorial Theorem 15.24]). We strengthen these techniques to our functorial setting and thus obtain a proof of Theorem \[thm:nerve-functorial-II\] via Theorem \[thm:nerve-functorial-I\]. We first present a lemma related to barycentric subdivisions and several lemmas about gluings and homotopy equivalences. These will be used in proving Theorem \[thm:nerve-functorial-II\]. Let ${\Sigma}$ be a simplicial complex, and let $\Delta$ be a subcomplex. Then $\Delta$ is an induced subcomplex if $\Delta = {\Sigma}\cap {\operatorname{pow}}(V(\Delta))$. \[lem:induced-cplx\] Let ${\Sigma}$ be a simplicial complex, and let $\Delta$ be a subcomplex. Then $\Delta{^{(1)}}$ is an induced subcomplex of ${\Sigma}{^{(1)}}$, i.e. $\Delta{^{(1)}}= {\Sigma}{^{(1)}}\cap {\operatorname{pow}}(V(\Delta{^{(1)}}))$. Let ${\sigma}$ be a simplex of $\Delta{^{(1)}}$. Then ${\sigma}$ belongs to ${\Sigma}{^{(1)}}$, and also to the full simplex ${\operatorname{pow}}(V(\Delta{^{(1)}}))$. Thus $\Delta{^{(1)}}\subseteq {\Sigma}{^{(1)}}\cap {\operatorname{pow}}(V(\Delta{^{(1)}}))$. Conversely, let ${\sigma}\in {\Sigma}{^{(1)}}\cap {\operatorname{pow}}(V(\Delta{^{(1)}}))$. Since ${\sigma}\in {\Sigma}{^{(1)}}$, we can write ${\sigma}= [{\tau}_0,\ldots, {\tau}_k]$, where ${\tau}_0 \subseteq \ldots \subseteq {\tau}_k$. Since ${\sigma}\in {\operatorname{pow}}(V(\Delta{^{(1)}}))$ and the vertices of $\Delta{^{(1)}}$ are simplices of $\Delta$, we also know that each ${\tau}_i$ is a simplex of $\Delta$. Thus ${\sigma}\in \Delta{^{(1)}}$. The equality follows. \[lem:carrier\] Let $X$ be a topological space, and let ${\Sigma}$ be a simplicial complex. Also let $f,g:X {\rightarrow}\|{\Sigma}\|$ be any two continuous maps such that $f(x), g(x)$ belong to the same simplex of $\|{\Sigma}\|$, for any $x \in X$. Then $f \simeq g$. \[lem:glue\] Let ${\Sigma}$ be a simplicial complex, and let $U \subseteq V({\Sigma})$. Suppose $\|{\Sigma}\cap {\operatorname{pow}}(U)\|$ is contractible. Then there exists a homotopy equivalence ${\varphi}: \|{\Sigma}\cup {\operatorname{pow}}(U)\| {\rightarrow}\|{\Sigma}\|$. The Gluing and Carrier Lemmas presented above are classical. We provide full details for the Gluing lemma inside the proof of the following functorial generalization of Lemma \[lem:glue\]. \[lem:glue-func\] Let ${\Sigma}\subseteq {\Sigma}'$ be two simplicial complexes. Also let $U \subseteq V({\Sigma})$ and $U' \subseteq V({\Sigma}')$ be such that $U \subseteq U'$. Suppose $\|{\Sigma}\cap {\operatorname{pow}}(U)\|$ and $\|{\Sigma}' \cap {\operatorname{pow}}(U')\|$ are contractible. Then, 1. There exists a homotopy equivalence ${\varphi}: \|{\Sigma}\cup {\operatorname{pow}}(U)\| {\rightarrow}\|{\Sigma}\|$ such that ${\varphi}(x)$ and ${\operatorname{id}}_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|}(x)$ belong to the same simplex of $\|{\Sigma}\cup {\operatorname{pow}}(U)\|$ for each $x \in \|{\Sigma}\cup {\operatorname{pow}}(U)\|$. Furthermore, the homotopy inverse is given by the inclusion $\iota: \|{\Sigma}\| {\hookrightarrow}\|{\Sigma}\cup {\operatorname{pow}}(U)\|$. 2. Given a homotopy equivalence ${\varphi}: \|{\Sigma}\cup {\operatorname{pow}}(U)\| {\rightarrow}\|{\Sigma}\|$ as above, there exists a homotopy equivalence ${\varphi}': \|{\Sigma}' \cup {\operatorname{pow}}(U')\| {\rightarrow}\|{\Sigma}'\|$ such that ${\varphi}'\|_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|} = {\varphi}$, and ${\varphi}'(x)$ and ${\operatorname{id}}_{\|{\Sigma}' \cup {\operatorname{pow}}(U')\|}(x)$ belong to the same simplex of $\|{\Sigma}'\cup {\operatorname{pow}}(U')\|$ for each $x \in \|{\Sigma}' \cup {\operatorname{pow}}(U')\|$. Furthermore, the homotopy inverse is given by the inclusion $\iota': \|{\Sigma}'\| {\hookrightarrow}\|{\Sigma}' \cup {\operatorname{pow}}(U')\|$. The proof uses this fact: any continuous map of an $n$-sphere ${\mathbb{S}}^n$ into a contractible space $Y$ can be continuously extended to a mapping of the $(n+1)$-disk ${\mathbb{D}}^{n+1}$ into $Y$, where ${\mathbb{D}}^{n+1}$ has ${\mathbb{S}}^n$ as its boundary [@spanier-book p. 27]. First we define ${\varphi}$. On $\|{\Sigma}\|$, define ${\varphi}$ to be the identity. Next let ${\sigma}$ be a minimal simplex in $\|{\operatorname{pow}}(U) \setminus {\Sigma}\|$. By minimality, the boundary of ${\sigma}$ (denoted ${\operatorname{Bd}}({\sigma})$) belongs to $\|{\Sigma}\cap {\operatorname{pow}}(U)\|$, and $\|{\Sigma}\|$ in particular. Thus ${\varphi}$ is defined on ${\operatorname{Bd}}({\sigma})$, which is an $n$-sphere for some $n \geq 0$. Furthermore, ${\varphi}$ maps ${\operatorname{Bd}}({\sigma})$ into the contractible space $\|{\Sigma}\cap {\operatorname{pow}}(U)\|$. Then we use the aforementioned fact to extend ${\varphi}$ continuously to all of ${\sigma}$ so that ${\varphi}$ maps ${\sigma}$ into $\|{\Sigma}\cap {\operatorname{pow}}(U)\|$. Furthermore, both ${\operatorname{id}}_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|}({\sigma}) = {\sigma}$ and ${\varphi}({\sigma})$ belong to the simplex $\|{\operatorname{pow}}(U)\|$. By iterating this procedure, we obtain a retraction ${\varphi}: \|{\Sigma}\cup {\operatorname{pow}}(U)\| {\rightarrow}\|{\Sigma}\|$ such that ${\varphi}(x)$ and $x$ belong to the same simplex in $\|{\Sigma}\cup {\operatorname{pow}}(U)\|$, for each $x \in \|{\Sigma}\cup {\operatorname{pow}}(U)\|$. Thus ${\varphi}$ is homotopic to ${\operatorname{id}}_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|}$ by Lemma \[lem:carrier\]. Thus we have a homotopy equivalence: $${\operatorname{id}}_{\|{\Sigma}\|}= {\varphi}\circ \iota, \; \iota \circ {\varphi}\simeq {\operatorname{id}}_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|} \tag{here $\iota:= \iota_{\|{\Sigma}\| {\hookrightarrow}\|{\Sigma}\cup {\operatorname{pow}}(U)\|}$} .$$ For the second part of the proof, suppose that a homotopy equivalence ${\varphi}: \|{\Sigma}\cup {\operatorname{pow}}(U)\| {\rightarrow}\|{\Sigma}\|$ as above is provided. We need to extend ${\varphi}$ to obtain ${\varphi}'$. Define ${\varphi}'$ to be equal to ${\varphi}$ on $\|{\Sigma}\cup {\operatorname{pow}}(U)\|$, and equal to the identity on $G:= \|{\Sigma}'\|\setminus \|{\Sigma}\cup {\operatorname{pow}}(U)\|$. Let ${\sigma}$ be a minimal simplex in $\|{\operatorname{pow}}(U')\| \setminus G$. Then by minimality, ${\operatorname{Bd}}({\sigma})$ belongs to $\|{\Sigma}' \cap {\operatorname{pow}}(U')\|$. As before, we have ${\varphi}'$ mapping ${\operatorname{Bd}}({\sigma})$ into the contractible space $\|{\Sigma}' \cap {\operatorname{pow}}(U')\|$, and we extend ${\varphi}'$ continuously to a map of ${\sigma}$ into $\|{\Sigma}' \cap {\operatorname{pow}}(U')\|$. Once again, ${\operatorname{id}}_{\|{\Sigma}' \cup {\operatorname{pow}}(U')\|}(x)$ and ${\varphi}'(x)$ belong to the same simplex $\|{\operatorname{pow}}(U')\|$, for all $x \in {\sigma}$. Iterating this procedure gives a continuous map ${\varphi}':\|{\Sigma}' \cup {\operatorname{pow}}(U')\| {\rightarrow}\|{\Sigma}'\|$. This map is not necessarily a retraction, because there may be a simplex ${\sigma}\in \|{\Sigma}\cup {\operatorname{pow}}(U)\| \cap \|{\Sigma}'\|$ on which ${\varphi}'$ is not the identity. However, it still holds that ${\varphi}'$ is continuous, and that $x, {\varphi}'(x)$ get mapped to the same simplex for each $x \in \|{\Sigma}'\cup {\operatorname{pow}}(U')\|$. Thus Lemma \[lem:carrier\] still applies to show that ${\varphi}'$ is homotopic to ${\operatorname{id}}_{\|{\Sigma}' \cup {\operatorname{pow}}(U')\|}$. We write $\iota'$ to denote the inclusion $\iota': \|{\Sigma}'\| {\hookrightarrow}\|{\Sigma}' \cup {\operatorname{pow}}(U')\|$. By the preceding work, we have $\iota'\circ {\varphi}' \simeq {\operatorname{id}}_{\|{\Sigma}' \cup {\operatorname{pow}}(U')\|}$. Next let $x \in \|{\Sigma}'\|$. Then either $x \in \|{\Sigma}'\| \cap \|{\Sigma}\cup {\operatorname{pow}}(U)\|$, or $x \in G$. In the first case, we know that ${\varphi}'(x) = {\varphi}(x)$ and ${\operatorname{id}}_{\|{\Sigma}'\|}(x) = {\operatorname{id}}_{\|{\Sigma}\cup {\operatorname{pow}}(U)\|}(x)$ belong to the same simplex of $\|{\Sigma}\cup {\operatorname{pow}}(U)\|$ by the assumption on ${\varphi}$. In the second case, we know that ${\varphi}'(x) = x = {\operatorname{id}}_{\|{\Sigma}'\|}(x)$. Thus for any $x \in \|{\Sigma}'\|$, we know that ${\varphi}'(x)$ and ${\operatorname{id}}_{\|{\Sigma}'\|}(x)$ belong to the same simplex in $\|{\Sigma}' \cup {\operatorname{pow}}(U')\|$. By Lemma \[lem:carrier\], we then have ${\varphi}'\|_{\|{\Sigma}'\|} \simeq {\operatorname{id}}_{\|{\Sigma}'\|}$. Thus ${\varphi}'\circ \iota' \simeq {\operatorname{id}}_{\|{\Sigma}'\|}$. This shows that ${\varphi}'$ is the necessary homotopy equivalence. Now we present the proof of Theorem \[thm:nerve-functorial-II\]. Notation. Let $I$ be an ordered set. For any subset $J\subseteq I$, we write $(J)$ to denote the sequence $(j_1,j_2,j_3,\ldots)$, where the ordering is inherited from the ordering on $I$. The first step is to functorially deform ${\mathcal}{A}_{\Sigma}$ and ${\mathcal}{A}_{{\Sigma}'}$ into covers of simplices while still preserving all associated homotopy types. Then we will be able to apply Theorem \[thm:nerve-functorial-I\]. We can assume by Lemma \[lem:induced-cplx\] that each subcomplex ${\Sigma}_i$ is induced, and likewise for each ${\Sigma}'_i$. We start by fixing an enumeration $I' = \{l_1,l_2,\ldots\}$. Thus $I'$ becomes an ordered set. #### Passing to covers of simplices. {#passing-to-covers-of-simplices. .unnumbered} We now define some inductive constructions. In what follows, we will define complexes denoted ${\Sigma}^\bullet, {\Sigma}'^\bullet$ obtained by “filling in" ${\Sigma}$ and ${\Sigma}'$ while preserving homotopy equivalence, as well as covers of these larger complexes denoted ${\Sigma}_{\star,\bullet}, {\Sigma}'_{\star,\bullet}$. First define: $$\begin{aligned} {\Sigma}^{(l_1)} &:= {\begin{cases} {\Sigma}\cup {\operatorname{pow}}(V({\Sigma}_{l_1})) &: \text{ if }l_1 \in I\\ {\Sigma}&: \text{ otherwise}. \end{cases}}\\ {\Sigma}'^{(l_1)} &:= {\Sigma}' \cup {\operatorname{pow}}(V({\Sigma}'_{l_1})).\end{aligned}$$ Next, for all $i\in I$, define $$\begin{aligned} {\Sigma}_{i,(l_1)} &:= {\begin{cases} {\Sigma}_i \cup {\operatorname{pow}}(V({\Sigma}_i) \cap V({\Sigma}_{l_1}))&: \text{ if }l_1 \in I\\ {\Sigma}_i &: \text{ otherwise}. \end{cases}}\end{aligned}$$ And for all $i \in I'$, define $$\begin{aligned} {\Sigma}'_{i,(l_1)} &:= {\Sigma}'_i \cup {\operatorname{pow}}(V({\Sigma}'_i) \cap V({\Sigma}'_{l_1})).\end{aligned}$$ Now by induction, suppose ${\Sigma}^{(l_1,\ldots, l_n)}$ and ${\Sigma}_{i,(l_1,\ldots, l_n)}$ are defined for all $i \in I$. Also suppose ${\Sigma}'^{(l_1,\ldots, l_n)}$ and ${\Sigma}'_{i,(l_1,\ldots, l_n)}$ are defined for all $i \in I'$. Then we define: $$\begin{aligned} {\Sigma}^{(l_1,\ldots, l_n,l_{n+1})} &:= {\begin{cases} {\Sigma}^{(l_1,\ldots,l_n)} \cup {\operatorname{pow}}(V({\Sigma}_{l_{n+1},(l_1,\ldots,l_n)})) &: \text{ if }l_{n+1} \in I\\ {\Sigma}^{(l_1,\ldots,l_n)} &: \text{ otherwise}. \end{cases}}\\ {\Sigma}'^{(l_1,\ldots, l_n,l_{n+1})} &:= {\Sigma}'^{(l_1,\ldots,l_n)} \cup {\operatorname{pow}}(V({\Sigma}'_{l_{n+1},(l_1,\ldots,l_n)})).\end{aligned}$$ For all $i \in I$, we have $$\begin{aligned} {\Sigma}_{i,(l_1,l_2,\ldots, l_{n+1})} &:= {\begin{cases} {\Sigma}_{i,(l_1,l_2,\ldots, l_n)} \cup {\operatorname{pow}}(V({\Sigma}_{i,(l_1,l_2,\ldots, l_n)}) \cap V({\Sigma}_{l_{n+1},(l_1,l_2,\ldots, l_n)}))&: \text{ if }l_{n+1} \in I\\ {\Sigma}_{i,(l_1,l_2,\ldots, l_n)} &: \text{ otherwise}. \end{cases}}\end{aligned}$$ And for all $i \in I'$, we have $$\begin{aligned} {\Sigma}'_{i,(l_1,l_2,\ldots, l_{n+1})} &:= {\Sigma}'_{i,(l_1,l_2,\ldots, l_n)} \cup {\operatorname{pow}}(V({\Sigma}'_{i,(l_1,l_2,\ldots, l_n)}) \cap V({\Sigma}'_{l_{n+1},(l_1,l_2,\ldots, l_n)})).\end{aligned}$$ Finally, for any $n \leq {\operatorname{card}}(I')$, we define ${\mathcal}{A}_{{\Sigma},(l_1,\ldots, l_n)} :=\{{\Sigma}_{i,(l_1,\ldots, l_{n})} : i \in I\}$ and ${\mathcal}{A}_{{\Sigma}',(l_1,\ldots, l_n)} :=\{{\Sigma}'_{i,(l_1,\ldots, l_{n})} : i \in I'\}$. We will show that these are covers of ${\Sigma}^{(l_1,l_2,\ldots, l_n)}$ and ${\Sigma}'^{(l_1,l_2,\ldots, l_n)}$, respectively. The next step is to prove by induction that for any $n \leq {\operatorname{card}}(I')$, we have $\|{\Sigma}\|\simeq \|{\Sigma}^{(l_1,l_2,\ldots, l_n)}\|$ and $\|{\Sigma}'\|\simeq \|{\Sigma}'^{(l_1,l_2,\ldots, l_n)}\|$, that ${\mathcal}{N}({\mathcal}{A}_{\Sigma}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma},(l_1,l_2,\ldots, l_n)})$ and ${\mathcal}{N}({\mathcal}{A}_{{\Sigma}'}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(l_1,l_2,\ldots, l_n)})$, and that nonempty finite intersections of the new covers ${\mathcal}{A}_{{\Sigma},(l_1,l_2,\ldots, l_n)}, {\mathcal}{A}_{{\Sigma}',(l_1,l_2,\ldots, l_n)}$ remain contractible. For the base case $n=0$, we have ${\Sigma}= {\Sigma}^{()}$, ${\Sigma}' = {\Sigma}'^{()}$. Thus the base case is true by assumption. We present the inductive step next. \[cl:nerve-cover-of-simplices\] For this claim, let $\bullet$ denote $l_1,\ldots, l_n$, where $0 < n < {\operatorname{card}}(I')$. Define $l:= l_{n+1}$. Suppose the following is true: 1. The collections ${\mathcal}{A}_{{\Sigma},(\bullet)}$ and ${\mathcal}{A}_{{\Sigma}',(\bullet)}$ are covers of ${\Sigma}^{(\bullet)}$ and ${\Sigma}'^{(\bullet)}$. 2. The nerves of the coverings are unchanged: ${\mathcal}{N}({\mathcal}{A}_{\Sigma}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma},(\bullet)})$ and ${\mathcal}{N}({\mathcal}{A}_{{\Sigma}'}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(\bullet)})$. 3. Each of the subcomplexes ${\Sigma}_{i,(\bullet)}$, $i\in I$, and ${\Sigma}'_{j,(\bullet)}$, $j\in I'$ is induced in ${\Sigma}^{(\bullet)}$ and ${\Sigma}'^{(\bullet)}$, respectively. 4. Let ${\sigma}\subseteq I$. If $\cap_{i\in {\sigma}}{\Sigma}_{i,(\bullet)}$ is nonempty, then it is contractible. Similarly, let ${\tau}\subseteq I'$. If $\cap_{i\in {\tau}}{\Sigma}'_{i,(\bullet)}$ is nonempty, then it is contractible. 5. We have homotopy equivalences $\|{\Sigma}\|\simeq \|{\Sigma}^{(\bullet)}\|$ and $\|{\Sigma}'\|\simeq \|{\Sigma}'^{(\bullet)}\|$ via maps that commute with the canonical inclusions. Then the preceding statements are true for ${\Sigma}^{(\bullet,l)}$, ${\Sigma}'^{(\bullet,l)}$, ${\mathcal}{A}_{{\Sigma},(\bullet,l)}$, and ${\mathcal}{A}_{{\Sigma}',(\bullet,l)}$ as well. For the first claim, we have ${\Sigma}^{(\bullet, l)} = {\Sigma}^{(\bullet)} \cup {\operatorname{pow}}(V({\Sigma}_{l,(\bullet)})) \subseteq \cup_{i \in I}{\Sigma}_{i,(\bullet, l)}$. For the inclusion, we used the inductive assumption that ${\Sigma}^{(\bullet)} = \cup_{i \in I}{\Sigma}_{i,(\bullet)}$. Similarly, ${\Sigma}'^{(\bullet, l)} \subseteq \cup_{i \in I'} {\Sigma}'_{i,(\bullet,l)}$. For the second claim, let $i \in I$. Then $V({\Sigma}_{i,(l_1)}) = V({\Sigma}_i)$, and in particular, we have $V({\Sigma}_{i,(\bullet,l)}) = V({\Sigma}_{i,(\bullet)}) = V({\Sigma}_i)$. Next observe that for any ${\sigma}\subseteq I$, the intersection $$\cap_{i \in {\sigma}}{\Sigma}_i \neq {\varnothing}\iff \cap_{i \in {\sigma}}V({\Sigma}_i) \neq {\varnothing}\iff \cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet,l)}) \neq {\varnothing}\iff \cap_{i \in {\sigma}}{\Sigma}_{i,(\bullet,l)} \neq {\varnothing}.$$ Thus ${\mathcal}{N}({\mathcal}{A}_{\Sigma}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma},(\bullet)}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma},(\bullet,l)})$, and similarly ${\mathcal}{N}({\mathcal}{A}_{{\Sigma}'}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(\bullet)}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(\bullet,l)})$. For the third claim, again let $i \in I$. If $l \not\in I$, then ${\Sigma}_{i,(\bullet,l)} = {\Sigma}_{i,(\bullet)}$, so we are done by the inductive assumption. Suppose $l \in I$. Since ${\Sigma}_{i,(\bullet)}$ is induced by the inductive assumption, we have: $$\begin{aligned} {\Sigma}_{i,(\bullet,l)} &= {\Sigma}_{i,(\bullet)} \cup ({\operatorname{pow}}(V({\Sigma}_{i,(\bullet)})\cap V({\Sigma}_{l,(\bullet)})))\\ &= ({\Sigma}^{(\bullet)} \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet)}))) \cup ({\operatorname{pow}}(V({\Sigma}_{i,(\bullet)})) \cap {\operatorname{pow}}(V({\Sigma}_{l,(\bullet)})))\\ & = ({\Sigma}^{(\bullet)} \cup {\operatorname{pow}}(V({\Sigma}_{l,(\bullet)}))) \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet)}))\\ &= {\Sigma}^{(\bullet, l)} \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet)})) = {\Sigma}^{(l)} \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet,l)})).\end{aligned}$$ Thus ${\Sigma}_{i,(\bullet,l)}$ is induced. The same argument holds for the $I'$ case. For the fourth claim, let ${\sigma}\subseteq I$, and suppose $\cap_{i \in {\sigma}} {\Sigma}_{i,(\bullet,l)}$ is nonempty. By the previous claim, each ${\Sigma}_{i,(\bullet,l)}$ is induced. Thus we write: $$\begin{aligned} \cap_{i\in {\sigma}}{\Sigma}_{i,(\bullet,l)} &= {\Sigma}^{(\bullet,l)} \cap {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet,l)})) \\ &= {\Sigma}^{(\bullet,l)} \cap {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}))\\ &= ({\Sigma}^{(\bullet)} \cup {\operatorname{pow}}(V({\Sigma}_{l,(\bullet)}))) \cap {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}))\\ &= (\cap_{i \in {\sigma}} ({\Sigma}^{(\bullet)} \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet)}))) ) \cup {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}) \cap V({\Sigma}_{l,(\bullet)}))\\ &= (\cap_{i \in {\sigma}} {\Sigma}_{i,(\bullet)} ) \cup {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}) \cap V({\Sigma}_{l,(\bullet)})).\end{aligned}$$ For convenience, define $A:=(\cap_{i \in {\sigma}} {\Sigma}_{i,(\bullet)} )$ and $B:= {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}) \cap V({\Sigma}_{l,(\bullet)}))$. Then $\|A\|$ is contractible by inductive assumption, and $\|B\|$ is a full simplex, hence contractible. Also, $A\cap B$ has the form $$\begin{aligned} &(\cap_{i \in {\sigma}} ({\Sigma}^{(\bullet)} \cap {\operatorname{pow}}(V({\Sigma}_{i,(\bullet)}))) ) \cap {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}) \cap V({\Sigma}_{l,(\bullet)})) \\ =& {\Sigma}^{(\bullet)} \cap {\operatorname{pow}}(\cap_{i \in {\sigma}}V({\Sigma}_{i,(\bullet)}) \cap V({\Sigma}_{l,(\bullet)})) \\ =& \cap_{i \in {\sigma}}{\Sigma}_{i,(\bullet)} \cap {\Sigma}_{l,(\bullet)},\end{aligned}$$ and the latter is contractible by inductive assumption. Thus by Lemma \[lem:glue\], we have $\|A\cup B\|$ contractible. This proves the claim for the case ${\sigma}\subseteq I$. The case ${\tau}\subseteq I'$ is similar. Now we proceed to the final claim. Since ${\Sigma}_{l,(\bullet)}$ is induced, we have ${\Sigma}_{l,(\bullet)} = {\Sigma}^{(\bullet)} \cap {\operatorname{pow}}(V({\Sigma}_{l,(\bullet)}))$. By the contractibility assumption, we know that $\|{\Sigma}_{l,(\bullet)}\|$ is contractible. Also we know that $\|{\Sigma}'_{l,(\bullet)}\| = \|{\Sigma}'^{(\bullet)} \cap {\operatorname{pow}}(V({\Sigma}'_{l,(\bullet)}))\|$ is contractible. By assumption we also have $V({\Sigma}_{l,(\bullet)}) \subseteq V({\Sigma}'_{l,(\bullet)})$. Thus by Lemma \[lem:glue-func\], we obtain homotopy equivalences $\Phi_l : \|{\Sigma}^{(\bullet,l)}\| {\rightarrow}\|{\Sigma}^{(\bullet)}\|$ and $\Phi'_l : \|{\Sigma}'^{(\bullet,l)}\| {\rightarrow}\|{\Sigma}'^{(\bullet)}\|$ such that $\Phi'_l$ extends $\Phi_l$. Furthermore, the homotopy inverses of $\Phi_l$ and $\Phi'_l$ are just the inclusions $\|{\Sigma}^{(\bullet)}\| {\hookrightarrow}\|{\Sigma}^{(\bullet,l)}\|$ and $\|{\Sigma}'^{(\bullet)}\| {\hookrightarrow}\|{\Sigma}'^{(\bullet,l)}\|$. Now let $\iota: \|{\Sigma}^{(\bullet)}\| {\rightarrow}\|{\Sigma}'^{(\bullet)}\|$ and $\iota_l: \|{\Sigma}^{(\bullet,l)}\| {\rightarrow}\|{\Sigma}'^{(\bullet,l)}\|$ denote the canonical inclusions. We wish to show the equality $\Phi'_l\circ \iota_l = \iota\circ \Phi_l$. Let $x \in\|{\Sigma}^{(\bullet,l)}\|$. Because $\Phi'_l$ extends $\Phi_l$ (this is why we needed the functorial gluing lemma), we have $$\Phi'_l(\iota_l(x))) = \Phi'_l(x) = \Phi_l(x) = \iota(\Phi_l(x)).$$ Since $x \in\|{\Sigma}^{(\bullet,l)}\|$ was arbitrary, the equality follows immediately. By the inductive assumption, we already have homotopy equivalences $\|{\Sigma}^{(\bullet)}\| {\rightarrow}\|{\Sigma}\|$ and $\|{\Sigma}'^{(\bullet)}\| {\rightarrow}\|{\Sigma}'\|$ that commute with the canonical inclusions. Composing these maps with $\Phi_l$ and $\Phi'_l$ completes the proof of the claim. By the preceding work, we replace the subcomplexes ${\Sigma}_l, {\Sigma}'_l$ by full simplices of the form ${\Sigma}_{l,(\bullet,l)},{\Sigma}_{l,(\bullet,l)}'$. In this process, the nerves remain unchanged and the complexes ${\Sigma}, {\Sigma}'$ are replaced by homotopy equivalent complexes ${\Sigma}^{(\bullet,l)}, {\Sigma}'^{(\bullet,l)}$. Furthermore, this process is functorial—the homotopy equivalences commute with the canonical inclusions ${\Sigma}{\hookrightarrow}{\Sigma}^{(\bullet,l)}$ and ${\Sigma}' {\hookrightarrow}{\Sigma}'^{(\bullet,l)}$. Repeating the inductive process in Claim \[cl:nerve-cover-of-simplices\] for all the finitely many $l \in I$ yields a simplicial complex ${\Sigma}^{(I)}$ along with a cover of simplices ${\mathcal}{A}_{{\Sigma},(I)}$. We also perform the same procedure for all $l \in I' \setminus I$ (this does not affect ${\Sigma}^{(I)}$) to obtain a simplicial complex ${\Sigma}'^{(I')}$ along with a cover of simplices ${\mathcal}{A}_{{\Sigma}',(I')}$. Furthermore, ${\Sigma}^{(I)}$ and ${\Sigma}'^{(I)}$ are related to ${\Sigma}$ and ${\Sigma}'$ by a finite sequence of homotopy equivalences that commute with the canonical inclusions. Also, we have ${\mathcal}{N}({\mathcal}{A}_{{\Sigma}}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma},(I)})$ and ${\mathcal}{N}({\mathcal}{A}_{{\Sigma}'}) = {\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(I')})$. Thus we obtain the following picture: (SI’) at (3,0)[$\|{\Sigma}'^{(I')}\|$]{}; (00) at (0,0)[$\cdots$]{}; (1) at (-3,0)[$\|{\Sigma}'^{(l_1)}\|$]{}; (2) at (-6,0)[$\|{\Sigma}'\|$]{}; (3) at (9,0)[$\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})\|$]{}; (01) at (6,0)[$\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}',(I')})\|$]{}; (SI) at (3,2)[$\|{\Sigma}^{(I)}\|$]{}; (02) at (0,2)[$\cdots$]{}; (4) at (-3,2)[$\|{\Sigma}^{(l_1)}\|$]{}; (5) at (-6,2)[$\|{\Sigma}\|$]{}; (6) at (9,2)[$\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}})\|$]{}; (03) at (6,2)[$\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma},(I)})\|$]{}; (SI) edge\[->\]node\[above\][$\simeq$]{} (02); (SI’) edge\[->\]node\[above\][$\simeq$]{} (00); (00) edge\[->\] (1); (3) edge\[double distance=2pt\] (01); (02) edge\[->\] (4); (6) edge\[double distance=2pt\] (03); \(1) edge\[->\]node\[above\][$\simeq$]{} (2); (00) edge\[->\]node\[above\][$\simeq$]{} (1); (4) edge\[->\]node\[above\][$\simeq$]{} (5); (02) edge\[->\]node\[above\][$\simeq$]{} (4); (1) edge\[<-\] node\[right\][$\iota_{(l_1)}$]{} (4); (2) edge\[<-\] node\[right\][$\iota$]{} (5); (3) edge\[<-\] node\[right\][$\iota_{{\mathcal}{N}}$]{} (6); (01) edge\[<-\] node\[right\][$\iota_{{\mathcal}{N},(I)}$]{} (03); (SI) edge\[->\]node\[right\][$\iota_{(I)}$]{} (SI’); (SI) edge\[->\]node\[above\][$\simeq$]{} (03); (SI’) edge\[->\]node\[above\][$\simeq$]{} (01); By applying Theorem \[thm:nerve-functorial-I\] to the block consisting of $\|{\Sigma}^{(I)}\|$, $\|{\Sigma}'^{(I')}\|$, $\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma},I})\|$ and $\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}',I'})\|$, we obtain a square that commutes up to homotopy. Then by composing the homotopy equivalences constructed above, we obtain a square consisting of $\|{\Sigma}\|$, $\|{\Sigma}'\|$, $\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}})\|$, and $\|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})\|$ that commutes up to homotopy. Thus we obtain homotopy equivalences $\|{\Sigma}\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{\Sigma})\|$ and $\|{\Sigma}'\| \simeq \|{\mathcal}{N}({\mathcal}{A}_{{\Sigma}'})\|$ via maps that commute up to homotopy with the canonical inclusions. Dowker persistence diagrams and asymmetry {#sec:symmetry} ========================================= From the very definition of the Rips complex at any given resolution, one can see that the Rips complex is blind to asymmetry in the input data (Remark \[rem:rips-symm\]). In this section, we argue that either of the Dowker source and sink complexes is sensitive to asymmetry. Thus when analyzing datasets containing asymmetric information, one may wish to use the Dowker filtration instead of the Rips filtration. In particular, this property suggests that the Dowker persistence diagram is a stronger invariant for directed networks than the Rips persistence diagram. In this section, we consider a family of examples, called cycle networks, for which the Dowker persistence diagrams capture meaningful structure, whereas the Rips persistence diagrams do not. We then probe the question “What happens to the Dowker or Rips persistence diagram of a network upon reversal of one (or more) edges?" Intuitively, if either of these persistence diagrams captures asymmetry, we would see a change in the diagram after applying this reversal operation to an edge. Cycle networks {#sec:cycle} -------------- For each $n\in {\mathbb{N}}$, let $(X_n,E_n,W_{E_n})$ denote the weighted graph with vertex set $X_n:={\left\{x_1,x_2,\ldots,x_n\right\}}$, edge set $E_n:={\left\{(x_1,x_2),(x_2,x_3),\ldots,(x_{n-1},x_n),(x_n,x_1)\right\}}$, and edge weights $W_{E_n}:E_n {\rightarrow}{\mathbb{R}}$ given by writing $W_{E_n}(e)=1$ for each $e\in E_n$. Next let ${\omega}_{G_n}:X_n\times X_n {\rightarrow}{\mathbb{R}}$ denote the shortest path distance induced on $X_n\times X_n$ by $W_{E_n}$. Then we write $G_n:=(X_n,{\omega}_{G_n})$ to denote the network with node set $X_n$ and weights given by ${\omega}_{G_n}$. Note that ${\omega}_{G_n}(x,x)=0$ for each $x\in X_n$. See Figure \[fig:cycle\] for some examples. We say that $G_n$ is the cycle network of length n. One can interpret cycle networks as being highly asymmetric, because for every consecutive pair of nodes $(x_i,x_{i+1})$ in a graph $G_n$, where $1\leq i\mod(n)\leq n$, we have ${\omega}_{G_n}(x_i,x_{i+1})=1$, whereas ${\omega}_{G_n}(x_{i+1},x_i)={\operatorname{diam}}(G_n) = n-1$, which is much larger than 1 when $n$ is large. To provide further evidence that Dowker persistence is sensitive to asymmetry, we computed both the Rips and Dowker persistence diagrams, in dimensions 0 and 1, of cycle networks $G_n$, for values of $n$ between 3 and 6. Computations were carried out using `Javaplex` in Matlab with ${\mathbb{Z}}_2$ coefficients. The results are presented in Figure \[fig:cycle\]. Based on our computations, we were able to conjecture and prove the result in Theorem \[thm:cycleH1\], which gives a precise characterization of the 1-dimensional Dowker persistence diagram of a cycle network $G_n$, for any $n$. Furthermore, the 1-dimensional Dowker persistence barcode for any $G_n$ contains only one persistent interval, which agrees with our intuition that there is only one nontrivial loop in $G_n$. On the other hand, for large $n$, the 1-dimensional Rips persistence barcodes contain more than one persistent interval. This can be seen in the Rips persistence barcode of $G_6$, presented in Figure \[fig:cycle\]. Moreover, for $n=3,4$, the 1-dimensional Rips persistence barcode does not contain any persistent interval at all. This suggests that Dowker persistence diagrams/barcodes are an appropriate method for analyzing cycle networks, and perhaps asymmetric networks in general. [0.3]{} \(1) at (0,1.5)[$x_1$]{}; (2) at (-1,0)[$x_2$]{}; (3) at (1,0)[$x_3$]{}; (4) at (0,-1); \(1) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (2); (2) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (3); (3) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (1); [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](dowker-G3.png "fig:"){width="\linewidth"} [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](rips-G3.png "fig:"){width="\linewidth"} [0.3]{} \(1) at (0,1.5)[$x_1$]{}; (2) at (-1.5,0)[$x_2$]{}; (3) at (0,-1.5,0)[$x_3$]{}; (4) at (1.5,0)[$x_4$]{}; (5) at (0,-2); \(1) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (2); (2) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (3); (3) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (4); (4) edge \[bend right\] node\[above,pos=0.5\][$1$]{} (1); [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](dowker-G4.png "fig:"){width="\linewidth"} [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](rips-G4.png "fig:"){width="\linewidth"} [0.3]{} \(1) at (0,1.5)[$x_1$]{}; (2) at (1.5,0)[$x_2$]{}; (3) at (1,-1.75)[$x_3$]{}; (4) at (-1,-1.75)[$x_4$]{}; (5) at (-1.5,0)[$x_5$]{}; (6) at (0,-2); \(1) edge node\[above,pos=0.5\][$1$]{} (2); (2) edge node\[left,pos=0.5\][$1$]{} (3); (3) edge node\[above,pos=0.5\][$1$]{} (4); (4) edge node\[right,pos=0.5\][$1$]{} (5); (5) edge node\[above,pos=0.5\][$1$]{} (1); [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](dowker-G5.png "fig:"){width="\linewidth"} [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](rips-G5.png "fig:"){width="\linewidth"} [0.3]{} \(1) at (-1,1.5)[$x_1$]{}; (2) at (1,1.5)[$x_2$]{}; (3) at (1.5,0)[$x_3$]{}; (4) at (1,-1.5)[$x_4$]{}; (5) at (-1,-1.5)[$x_5$]{}; (6) at (-1.5,0)[$x_6$]{}; (7) at (0,-2); \(1) edge node\[above,pos=0.5\][$1$]{} (2); (2) edge node\[left,pos=0.5\][$1$]{} (3); (3) edge node\[left,pos=0.5\][$1$]{} (4); (4) edge node\[above,pos=0.5\][$1$]{} (5); (5) edge node\[right,pos=0.5\][$1$]{} (6); (6) edge node\[right,pos=0.5\][$1$]{} (1); [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](dowker-G6.png "fig:"){width="\linewidth"} [0.33]{} ![The first column contains illustrations of cycle networks $G_3,G_4,G_5$ and $G_6$. The second column contains the corresponding Dowker persistence barcodes, in dimensions 0 and 1. Note that the persistent intervals in the 1-dimensional barcodes agree with the result in Theorem \[thm:cycleH1\]. The third column contains the Rips persistence barcodes of each of the cycle networks. Note that for $n=3,4$, there are no persistent intervals in dimension 1. On the other hand, for $n=6$, there are two persistent intervals in dimension 1. []{data-label="fig:cycle"}](rips-G6.png "fig:"){width="\linewidth"} Notation. In the remainder of this section, we will prove results involving Dowker sink complexes of the cycle networks $G_n$ and associated vector spaces at a range of resolutions ${\delta}$. For convenience, we will write ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}:= {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},G_n}$ (where $n$ will be fixed) and $C_k^{\delta}:= C_k({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})$, the $k$-chain vector space associated to ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ for each $k\in {\mathbb{Z}}_+$. For each $k\in {\mathbb{Z}}_+$, the boundary map from $C_k^{\delta}$ to $C_{k-1}^{\delta}$ will be denoted ${\partial}_k^{\delta}$. Whenever we write $x_i$ to denote a vertex of $G_n$, the subscript $i$ should be understood as $i{\ (\mbox{mod}\ n)}$. We write $e_i$ to denote the 1-simplex $[x_i,x_{i+1}]$ for each $1\leq i \leq n$, where $x_{n+1}$ is understood to be $x_1$. Given an element ${\gamma}\in \ker({\partial}_k^{\delta})\subseteq C_1^{\delta}$, we will write ${\langle}{\gamma}{\rangle}_{\delta}$ to denote its equivalence class in the quotient vector space $\ker({\partial}_k^{\delta})/{\operatorname{im}}({\partial}^{\delta}_k)$. We will refer to the operation of taking this quotient as passing to homology. The following theorem contains the characterization result for 1-dimensional Dowker persistence diagrams of cycle networks. \[thm:cycleH1\] Let $G_n=(X_n,{\omega}_{G_n})$ be a cycle network for some $n\in {\mathbb{N}}$, $n\geq 3$. Then we obtain: $${\operatorname{Dgm}}^{{\mathfrak}{D}}_1(G_n)={\left\{(1,{\left \lceil{n/2}\right \rceil })\in {\mathbb{R}}^2\right\}}.$$ Thus ${\operatorname{Dgm}}^{{\mathfrak}{D}}_1(G_n)$ consists of precisely the point $(1,{\left \lceil{n/2}\right \rceil })\in {\mathbb{R}}^2$ with multiplicity 1. The proof occurs in three stages: first we show that a 1-cycle appears at ${\delta}=1$, next we show that this 1-cycle does not become a boundary until ${\delta}={\left \lceil{n/2}\right \rceil }$, and finally that any other 1-cycle belongs to the same equivalence class upon passing to homology (this shows that the single point in the persistence diagram has multiplicity 1). Note that for ${\delta}<1$, there are no 1-simplices in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, and so $H_1({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})$ is trivial. Suppose $1\leq {\delta}< 2$. \[cl:cycle-2simp\] There are no 2-simplices in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ for $1\leq {\delta}< 2$. To see this, let $x_i,x_j,x_k$ be any three distinct vertices in $X_n$. Assume towards a contradiction that there exists $x\in X_n$ such that $(x_i,x),(x_j,x),(x_k,x)\in R_{{\delta},X_n},$ where $R_{{\delta},X_n}$ is as given by Equation \[eq:relation\]. Thus ${\omega}_{G_n}(x_i,x)\in \{0,1\}$, so either $x=x_i$ or $x=x_{i+1}$. Similarly we get that $x=x_j$ or $x=x_{j+1}$, and that $x=x_k$ or $x=x_{k+1}$. But this is a contradiction, since $x_i,x_j,x_k$ are all distinct. By the claim, there are no 2-simplices in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, so ${\operatorname{im}}({\partial}^{\delta}_2)$ is trivial and the only 1-chains are linear combinations of $e_i$ terms. Next, we define: $$v_n:=e_1 + e_2 + \ldots + e_n=[x_1,x_2]+[x_2,x_3]+ \ldots + [x_n,x_1].$$ Note that $v_n\in C_1^{\delta}$ for all ${\delta}\geq 1$. One can further verify that ${\partial}^{\delta}_1(v_n) = 0$, for any ${\delta}\geq 1$. In other words, $v_n$ is a 1-cycle for any ${\delta}\geq 1$. \[cl:cycle-v\] Let $1\leq {\delta}< 2$. Then $v_n$ generates $\ker({\partial}^{\delta}_1)\subseteq C_1^{\delta}$. The only 1-simplices in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ are of the form $e_i$, for $1\leq i \leq n$. So it suffices to show that any linear combination of the $e_i$ terms is a multiple of $v_n$. Let $u = \sum_{i=1}^na_ie_i \in \ker({\partial}^{\delta}_1)$, for some $a_1,\ldots, a_n \in \mathbb{K}$. Then, $$\begin{aligned} 0={\partial}^{\delta}_1(u)=\sum_{i=1}^na_i{\partial}^{\delta}_1(e_i) &= \sum_{i=1}^na_i([x_{i+1}]-[x_i])\\ &=\sum_{i=1}^n(a_{i-1}-a_i)[x_i], &&\text{where $x_0$ is understood to be $x_n$.}\end{aligned}$$ Since all the $[x_i]$ are linearly independent, it follows that $a_1=a_2=\ldots=a_n$. Thus it follows that $u$ is a constant multiple of $v_n$. This proves the claim. By the two preceding claims, it follows that ${\left\{{\langle}v_n {\rangle}_{\delta}\right\}}$ is a basis for $H_1({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})$, for ${\delta}\in [1,2)$. More specifically, ${\langle}v_n {\rangle}_{\delta}$ is a cycle that appears at ${\delta}=1$ and does not become a boundary until at least ${\delta}=2$, and any other cycle in $C_1^{\delta}$, for ${\delta}\in [1,2)$, is in the linear span of $v_n$. Next, suppose ${\delta}\geq 2$. Note that this allows the appearance of cycles that are not in the span of $v_n$. In the next claim, we show that upon passing to homology, the equivalence class of any such cycle coincides with that of $v_n$. This will show that there can be at most one nontrivial element in ${\operatorname{Dgm}}_1^{{\operatorname{si}}}(G_n)$. \[cl:cycle-decomp\] Let ${\delta}\geq 2$, and let $y = \sum_{i=1}^pa_i{\sigma}_i \in \ker({\partial}^{\delta}_1)$ for some $p\in {\mathbb{N}}$, some $a_1,\ldots,a_p \in \mathbb{K}$, and some ${\sigma}_1,\ldots, {\sigma}_p\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$. Then there exists a choice of coefficients $(b_i)_{i=1}^n \in \mathbb{K}^n$ such that $z=\sum_{i=1}^nb_ie_i \in \ker({\partial}^{\delta}_1)$ and $y-z \in {\operatorname{im}}({\partial}^{\delta}_2)$. Moreover, we obtain ${\langle}y {\rangle}_{\delta}= {\langle}z {\rangle}_{\delta}= {\langle}v_n {\rangle}_{\delta}$ upon passing to homology. ![Given two points $x_j,x_k \in X_n$, we have either ${\omega}_{G_n}(x_j,x_k)\leq n/2$, or ${\omega}_{G_n}(x_k,x_j)< n/2$. To see this, note that ${\omega}_{G_n}(x,x')+{\omega}_{G_n}(x',x)=n$ for any $x\neq x'\in X_n$.[]{data-label="fig:dowker-cycle-prop"}](dowker-cycle-prop.jpg) ![Three possible locations for a ${\delta}$-sink $x$ of a simplex $[x_j,x_k]$, assuming that ${\omega}_{G_n}(x_j,x_k)\leq n/2$. For the figure on the left, note that ${\omega}_{G_n}(x_k,x)\geq n/2 \geq {\omega}_{G_n}(x_j,x_k)$. For the figure in the middle, note that ${\omega}_{G_n}(x_j,x)\geq {\omega}_{G_n}(x_j,x_k)$. Finally, for the figure on the right, where $x=x_k$, note that ${\omega}_{G_n}(x_j,x)={\omega}_{G_n}(x_j,x_k)$ and ${\omega}_{G_n}(x_k,x)=0$.[]{data-label="fig:dowker-cycle-prop-sink"}](dowker-cycle-prop-sink.png) To see this, fix ${\sigma}_i \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, and write ${\sigma}_i= [x_j,x_k]$ for some $1\leq j,k \leq n$. If $k=j+1$ (resp. $k=j-1$), then we already have ${\sigma}_i=e_j$ (resp. ${\sigma}_i=e_k$), so there is nothing more to show. Assume $k\not\in \{j+1,j-1\}$. Since ${\omega}_{G_n}(x_j,x_k) + {\omega}_{G_n}(x_k,x_j) = n$, we have two cases: (1) ${\omega}_{G_n}(x_j,x_k)\leq n/2$, or (2) ${\omega}_{G_n}(x_k,x_j)< n/2$. In the first case, we have $k=j+l $ for some integer $l\in [2,n/2]$ (all numbers are taken modulo $n$). In the second case, $j = k+l$ for some integer $l\in [2,n/2)$ (also modulo $n$). The situation is illustrated in Figure \[fig:dowker-cycle-prop\]. Both cases are similar, so we only prove the case ${\omega}_{G_n}(x_j,x_k) \leq n/2$. Recall that any ${\delta}$-sink $x\in X_n$ for $[x_j, x_k]$ satisfies $\max({\omega}_{G_n}(x_j,x), {\omega}_{G_n}(x_k,x)) \leq {\delta}$, by the ${\delta}$-sink condition (Equation \[eq:d-sink\]). Also note that such a ${\delta}$-sink $x$ satisfies $$\max({\omega}_{G_n}(x_j,x),{\omega}_{G_n}(x_k,x)) \geq {\omega}_{G_n}(x_j,x_k),$$ as can be seen from Figure \[fig:dowker-cycle-prop-sink\]. So whenever some $x\in X_n$ is a ${\delta}$-sink for $[x_j,x_k]$, we have $x_k$ as a valid ${\delta}$-sink for $[x_j,x_k]$. Since $[x_j,x_k] \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, it must have a ${\delta}$-sink $x\in X_n$. Thus $x_k$ is a valid ${\delta}$-sink for $[x_j,x_k]$. Next let $l\in [2,n/2]$ be an integer such that $k=j+l$ (modulo $n$). Notice that: $$0={\omega}_{G_n}(x_k,x_k)={\omega}_{G_n}(x_{j+l},x_k)< {\omega}_{G_n}(x_{j+l-1},x_k)< \ldots < {\omega}_{G_n}(x_{j+1},x_k)<{\omega}_{G_n}(x_j,x_k) \leq {\delta}.$$ Then observe that: $$[x_j,x_{j+1},x_k],[x_{j+1},x_{j+2},x_k],\ldots, [x_{k-2},x_{k-1},x_k] \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta},$$ since $x_k$ is a ${\delta}$-sink for all these 2-simplices. One can then verify the following: $$\begin{aligned} &{\partial}^{\delta}_2\left([x_j,x_{j+1},x_k]+[x_{j+1},x_{j+2},x_k]+\ldots+[x_{k-2},x_{k-1},x_k]\right)\\ &={\partial}_2^{\delta}\left(\sum_{q=0}^{k-j-2}[x_{j+q},x_{j+q+1},x_k]\right)\\ &=\sum_{q=0}^{k-j-2}[x_{j+q+1},x_k]-\sum_{q=0}^{k-j-2}[x_{j+q},x_k]+\sum_{q=0}^{k-j-2}[x_{j+q},x_{j+q+1}]\\ &=\sum_{q=0}^{k-j-2}[x_{j+q+1},x_k] - [x_j,x_k] - \sum_{q=0}^{k-j-3}[x_{j+q+1},x_k] +\sum_{q=0}^{k-j-2}[x_{j+q},x_{j+q+1}]\\ &=[x_j,x_{j+1}]+[x_{j+1},x_{j+2}]+\ldots+[x_{k-1},x_k]-[x_j,x_k]\\ &=e_j+e_{j+1} + \ldots + e_{k-1} -{\sigma}_i.\end{aligned}$$ Thus $a_i(e_j+e_{j+1}+\ldots+e_{k-1})-a_i{\sigma}_i \in{\operatorname{im}}({\partial}^{\delta}_2)$. Repeating this process for all ${\sigma}_i$, $i\in \{1,\ldots,p\}$, we may obtain the coefficients $(b_i)_{i=1}^n$ such that $\sum_{i=1}^pa_i{\sigma}_i - \sum_{i=1}^nb_ie_i \in {\operatorname{im}}({\partial}^{\delta}_2).$ Let $z=\sum_{i=1}^nb_ie_i$. Then $y-z \in {\operatorname{im}}({\partial}^{\delta}_2)$. Moreover, since ${\partial}^{\delta}_1\circ{\partial}^{\delta}_2 = 0$, it follows that ${\partial}^{\delta}_1(y)-{\partial}^{\delta}_1(z)=0$, so $z\in \ker({\partial}^{\delta}_1)$. Finally, note that an argument analogous to that of Claim \[cl:cycle-v\] shows that $b_1=b_2=\ldots=b_n$. Hence it follows that $z$ is a multiple of $v_n$. Thus ${\langle}z {\rangle}_{\delta}= {\langle}v_n {\rangle}_{\delta}$. This proves the claim. By Claims \[cl:cycle-v\] and \[cl:cycle-decomp\], it follows that $H_1({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})$ is generated by ${\langle}v_n {\rangle}_{\delta}$ for all ${\delta}\geq 1$, so $\dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{\delta}))\leq 1$ for all ${\delta}\geq 1$. It remains to show that ${\langle}v_n {\rangle}_{\delta}$ does not become trivial until ${\delta}={\left \lceil{n/2}\right \rceil }$. The cases $n=3,4$ can now be completed quickly, so we focus on these simpler situations first. For either of $n=3,4$, we have ${\left \lceil{n/2}\right \rceil }=2.$ Suppose ${\delta}=2$ and $n=3$. Then we have $[x_1,x_2,x_3]\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ because ${\operatorname{diam}}(G_n)=2$ and any of $x_1,x_2,x_3$ can be a $2$-sink for $[x_1,x_2,x_3]$. Then, $${\partial}_2^{\delta}([x_1,x_2,x_3])=[x_2,x_3]-[x_1,x_3]+[x_1,x_2]=e_1+e_2+e_3 = v_3.$$ Recall that by Claim \[cl:cycle-2simp\], $v_3\not\in {\operatorname{im}}({\partial}_2^{\delta})$ for any ${\delta}<2$. Thus by Claim \[cl:cycle-v\] and the preceding equation, $v_3$ generates $\ker({\partial}_1^{\delta})$ for $1\leq {\delta}<2$, and becomes a boundary for precisely ${\delta}\geq 2$. Thus ${\operatorname{Dgm}}_1^{{\operatorname{si}}}(G_3)={\left\{(1,2)\right\}}.$ Next, suppose ${\delta}=2$ and $n=4$. Then we have $[x_1,x_2,x_3],[x_1,x_3,x_4]\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ with $x_3, x_1$ as $2$-sinks, respectively. By a direct computation, we then have: $${\partial}_2^{\delta}([x_1,x_2,x_3]+[x_1,x_3,x_4])=e_1+e_2+e_3+e_4 = v_4.$$ By following the same argument as for the case $n=3$, we see that ${\operatorname{Dgm}}_1^{{\operatorname{si}}}(G_4)={\left\{(1,2)\right\}}.$ In the sequel, we assume that $n > 4$. Recall that it remains to show that ${\langle}v_n {\rangle}_{\delta}$ does not become trivial until ${\delta}={\left \lceil{n/2}\right \rceil }$, and that ${\langle}v_n {\rangle}_{\delta}= 0$ for all ${\delta}\geq {\left \lceil{n/2}\right \rceil }$. We have already shown that ${\langle}v_n {\rangle}_{\delta}$ is not trivial for ${\delta}\in [1,2)$. We proceed by defining the following: $${\gamma}_n:=[x_1,x_2,x_3] + [x_1,x_3,x_4] + \ldots + [x_1,x_{n-1},x_n] = \sum_{i=1}^{n-2}[x_1,x_{i+1},x_{i+2}].$$ For each ${\delta}\geq {\left \lceil{n/2}\right \rceil }$, we have ${\gamma}_n \in C_2^{\delta}$ and ${\partial}_2^{\delta}({\gamma}_n)=v_n$. In particular, ${\langle}v_n {\rangle}_{\delta}= 0$ for all such ${\delta}$. Let ${\delta}\geq {\left \lceil{n/2}\right \rceil }$. Notice that $${\omega}_{G_n}(x_{{\left \lceil{n/2}\right \rceil }+1},x_1) = n - {\omega}_{G_n}(x_1,x_{{\left \lceil{n/2}\right \rceil }+1}) = n- {\left \lceil{n/2}\right \rceil } \leq n/2 \leq {\left \lceil{n/2}\right \rceil } \leq {\delta},$$ so ${\omega}_{G_n}(x_i,x_1) \leq {\delta}$ for each $i \in \{{\left \lceil{n/2}\right \rceil }+1,{\left \lceil{n/2}\right \rceil }+2,\ldots, n\}$. Then for each $i\in \{{\left \lceil{n/2}\right \rceil }+1,{\left \lceil{n/2}\right \rceil }+2,\ldots, n-1\}$, we have $[x_i,x_{i+1},x_1] \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, with $x_1$ as a ${\delta}$-sink. Also notice that for each $i\in \{1,\ldots,{\left \lceil{n/2}\right \rceil }\}$, $${\omega}_{G_n}(x_i,x_{{\left \lceil{n/2}\right \rceil }+1})\leq{\omega}_{G_n}(x_1,x_{{\left \lceil{n/2}\right \rceil }+1})={\left \lceil{n/2}\right \rceil } \leq {\delta},$$ so ${\omega}_{G_n}(x_i,x_{{\left \lceil{n/2}\right \rceil }+1}) \leq {\delta}$. Thus for any $i\in \{2,\ldots,{\left \lceil{n/2}\right \rceil }\},$ we have $[x_1,x_i,x_{i+1}] \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, with $x_{{\left \lceil{n/2}\right \rceil }+1}$ as a ${\delta}$-sink. Combining the two preceding observations, we see that for any $i\in \{2,\ldots, n-2\}$, we have $[x_1,x_{i+1},x_{i+2}]\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$. It follows that ${\gamma}_n \in C_2^{\delta}$. Next we observe the following: $$\begin{aligned} {\partial}_2^{\delta}({\gamma}_n)&={\partial}_2^{\delta}\left(\sum_{i=1}^{n-2}[x_1,x_{i+1},x_{i+2}]\right)\\ &=\sum_{i=1}^{n-2}[x_{i+1},x_{i+2}] - \sum_{i=1}^{n-2}[x_1,x_{i+2}]+ \sum_{i=1}^{n-2}[x_1,x_{i+1}]\\ &=\sum_{i=1}^{n-2}[x_{i+1},x_{i+2}] - \sum_{i=1}^{n-2}[x_1,x_{i+2}]+[x_1,x_2]+ \sum_{i=2}^{n-2}[x_1,x_{i+1}]\\ &=\sum_{i=1}^{n-2}[x_{i+1},x_{i+2}] + [x_1,x_2] - [x_1,x_n] = v_n.\end{aligned}$$ It follows that for any ${\delta}\geq {\left \lceil{n/2}\right \rceil }$, we have $v_n \in {\operatorname{im}}({\partial}_2^{\delta})$, and so ${\langle}v_n {\rangle}_{\delta}=0$ for each such ${\delta}$. ![Placement of $x_{{\left \lceil{n/2}\right \rceil }}$ and $x_{{\left \lceil{n/2}\right \rceil }-1}$, depending on whether $n$ is even or not.[]{data-label="fig:dowker-cycle-prop2"}](dowker-cycle-prop2.jpg) There does not exist ${\delta}\in [2,{\left \lceil{n/2}\right \rceil })$ such that ${\langle}v_n {\rangle}_{\delta}$ is trivial. Let $2\leq {\delta}< {\left \lceil{n/2}\right \rceil }$. As a first step, we wish to show that ${\gamma}_n \not\in C_2^{\delta}$. For this step, it suffices to show that the 2-simplex ${\sigma}:= [x_1,x_{{\left \lceil{n/2}\right \rceil }},x_{{\left \lceil{n/2}\right \rceil }+1}]$ does not belong to ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$. The placement of $x_{{\left \lceil{n/2}\right \rceil }}$ and $x_{{\left \lceil{n/2}\right \rceil }+1}$ is illustrated in Figure \[fig:dowker-cycle-prop2\]. By an argument similar to that used in Figure \[fig:dowker-cycle-prop-sink\], one can verify that there exists a ${\delta}$-sink for ${\sigma}$ if and only if at least one of $x_1,x_{{\left \lceil{n/2}\right \rceil }}, x_{{\left \lceil{n/2}\right \rceil }+1}$ is a ${\delta}$-sink for ${\sigma}$. But note the following: $$\begin{aligned} {\omega}_{G_n}(x_{{\left \lceil{n/2}\right \rceil }},x_1)= n-({\left \lceil{n/2}\right \rceil }-1) &= \begin{cases} n/2 + 1 &: n \text{ even}\\ {\left \lceil{n/2}\right \rceil } &: n\text{ odd} \end{cases}\\ &\geq {\left \lceil{n/2}\right \rceil } > {\delta},\end{aligned}$$ so $x_1$ cannot be a ${\delta}$-sink for ${\sigma}$. Similarly we note that ${\omega}_{G_n}(x_{{\left \lceil{n/2}\right \rceil }+1},x_{{\left \lceil{n/2}\right \rceil }}) = n > {\delta}$ and ${\omega}_{G_n}(x_1,x_{{\left \lceil{n/2}\right \rceil }+1})={\left \lceil{n/2}\right \rceil } > {\delta},$ so neither $x_{{\left \lceil{n/2}\right \rceil }}$ nor $x_{{\left \lceil{n/2}\right \rceil }+1}$ can be ${\delta}$-sinks for ${\sigma}$. Thus ${\sigma}\not \in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, and so ${\gamma}_n \not\in C_2^{\delta}$. Suppose there exists ${\gamma}' \in C_2^{\delta}$ such that ${\partial}_2^{\delta}({\gamma}') = v_n$. Since $[x_1,x_2]$ is a summand of $v_n$, we must have $a_j[x_1,x_2,x_j]$ as a summand of ${\gamma}'$, for some coefficient $a_j$ and some $3\leq j \leq n$. First suppose that $x_j$ is a sink for $[x_1,x_2,x_j]$. We claim that ${\gamma}'$ is homologous to a chain containing $[x_1,x_2,x_3]$ as a summand. If $j=3$, then we are done, so suppose $j > 3$. Then we also know that $[x_1,x_2,x_3,x_j]$ is a 3-simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}}$. Let ${\gamma}''$ be the chain obtained from ${\gamma}'$ by replacing $[x_1,x_2,x_j]$ with $[x_1,x_2,x_3] - [x_2,x_3,x_j] + [x_1,x_3,x_j]$. Since ${\partial}_3^{\delta}([x_1,x_2,x_3,x_j]) = [x_2,x_3,x_j] - [x_1,x_3,x_j] + [x_1,x_2,x_j] -[x_1,x_2,x_3]$ and ${\partial}_2^{\delta}\circ {\partial}_3^{\delta}= 0$, we know that ${\partial}_2^{\delta}({\gamma}'') = {\partial}_2^{\delta}({\gamma}') = v_n$. Now ${\partial}_2^{\delta}([x_1,x_2,x_3])$ contributes an $[x_1,x_3]$ summand which does not appear in $v_n$, so it must be cancelled by some other terms in ${\gamma}'$ (resp. ${\gamma}''$). Thus there must exist another 2-simplex $[x_1,x_3,x_k]$ in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$, where $k\neq 2$. But we can repeat the preceding argument to obtain a chain homologous to ${\gamma}'$ containing both $[x_1,x_2,x_3]$ and $[x_1,x_3,x_4]$ as summands. Proceeding in this way, we obtain a chain homologous to ${\gamma}'$ that contains $[x_1,x_{{\left \lceil{n/2}\right \rceil }},x_{{\left \lceil{n/2}\right \rceil }+1}]$ as a summand. But this is a contradiction to what we have shown previously, i.e. that $[x_1,x_{{\left \lceil{n/2}\right \rceil }},x_{{\left \lceil{n/2}\right \rceil }+1}]$ is not a simplex in ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$. In the case where $x_j$ is not a sink for $[x_1,x_2,x_j]$, we must have $x_2$ as a sink instead. Using similar reasoning as above, we can replace ${\gamma}'$ in this instance by a homologous chain containing $[x_n,x_1,x_2]$ as a summand. Since $[x_n,x_2]$ is not a summand of $v_n$, we can obtain another homologous chain containing $[x_{n-1},x_n,x_1]$ as a summand, then a homologous chain containing $[x_{n-1},x_n,x_1],[x_{n-2},x_{n-1},x_1]$ as summands, and so on until we again obtain a homologous chain containing $[x_1,x_{{\left \lceil{n/2}\right \rceil }},x_{{\left \lceil{n/2}\right \rceil }+1}]$ as a summand. Once again, this is a contradiction. This proves the claim. Thus we have shown that $v_n$ is a nontrivial cycle that appears at ${\delta}=1$, and becomes a boundary at exactly ${\delta}={\left \lceil{n/2}\right \rceil }$. Furthermore, we have shown that upon passing to homology, the equivalence classes of all cycles coincide with that of $v_n$. Thus there is only one off-diagonal point $(1,{\left \lceil{n/2}\right \rceil })$ on the 1-dimensional persistence diagram, which appears with multiplicity one. This concludes the proof. From our experimental results (see Figure \[fig:cycle\]), it appears that the 1-dimensional Rips persistence diagram of a cycle network does not admit a characterization as simple as that given by Theorem \[thm:cycleH1\] for the 1-dimensional Dowker persistence diagram. Moreover, the Rips complexes ${\mathfrak}{R}^{\delta}_{G_n}, {\delta}\in {\mathbb{R}}, n\in {\mathbb{N}}$ correspond to certain types of independence complexes that appear independently in the literature, and whose homotopy types remain open [@engstrom2009complexes Question 5.3]. On a related note, we point the reader to [@adamaszek2015vietoris] for a complete characterization of the homotopy types of Rips complexes of points on the circle (equipped with the restriction of the arc length metric). To elaborate on the connection to [@engstrom2009complexes], we write $H^k_n$ to denote the undirected graph with vertex set ${\left\{1,\ldots, n\right\}}$, and edges given by pairs $(i,j)$ where $1\leq i < j \leq n$ and either $j-i < k$ or $(n+i) - j < k$. Next we write ${\operatorname{Ind}}(H^k_n)$ to denote the independence complex of $H^k_n$, which is the simplicial complex consisting of subsets ${\sigma}\subseteq {\left\{1,2,\ldots, n\right\}}$ such that no two elements of ${\sigma}$ are connected by an edge in $H^k_n$. Then we have ${\operatorname{Ind}}(H^k_n) = {\mathfrak}{R}^{n-k}_{G_n}$ for each $k, n\in {\mathbb{N}}$ such that $k < n$. To gain intuition for this equality, fix a basepoint $1$, and consider the values of $j\in {\mathbb{N}}$ for which the simplex $[1,j]$ belongs to ${\operatorname{Ind}}(H^k_n)$ and to ${\mathfrak}{R}^{n-k}_{G_n}$, respectively. In either case, we have $k+1 \leq j \leq n-k+1$. Using the rotational symmetry of the points, one can then obtain the remaining 1-simplices. Rips complexes are determined by their 1-skeleton, so this suffices to construct ${\mathfrak}{R}^{n-k}_{G_n}$. Analogously, ${\operatorname{Ind}}(H^k_n)$ is determined by the edges in $H^k_n$, and hence also by its 1-skeleton. In [@engstrom2009complexes Question 5.3], the author writes that the homotopy type of ${\operatorname{Ind}}(H^k_n)$ is still unsolved. Characterizing the persistence diagrams ${\operatorname{Dgm}}_k^{{\mathfrak}{R}}(G_n)$ thus seems to be a useful future step, both in providing a computational suggestion for the homotopy type of ${\operatorname{Ind}}(H^k_n)$, and also in providing a valuable example in the study of persistence of directed networks. Theorem \[thm:cycleH1\] has the following implication for data analysis: nontrivial 1-dimensional homology in the Dowker persistence diagram of an asymmetric network suggests the presence of directed cycles in the underlying data. Of course, it is not necessarily true that nontrivial 1-dimensional persistence can occur only in the presence of a directed circle. Our motivation for studying cycle networks is that they constitute directed analogues of circles, and we were interested in seeing if the 1-dimensional Dowker persistence diagram would be able to capture this analogy. Theorem \[thm:cycleH1\] shows that this is indeed the case: we get a single nontrivial 1-dimensional persistence interval, which is what we would expect when computing the persistent homology of a circle in the metric space setting. We further studied the 2-dimensional Dowker persistence diagrams of cycle networks. Our computational examples, some of which are illustrated in Figure \[fig:cycle-2dim\], enabled us to conjecture: \[conj:dowker-dgm-2\] Let $n\in {\mathbb{N}}$, $n \geq 3$, and let $G_n$ be a cycle network. If $n$ is odd, then ${\operatorname{Dgm}}_2^{{\mathfrak}{D}}(G_n)$ is trivial. If $n$ is even, then ${\operatorname{Dgm}}_2^{{\mathfrak}{D}}(G_n) = [(\frac{n}{2},\frac{n}{2}+1)\in {\mathbb{R}}^2],$ and the multiplicity of this point is $\frac{n}{2}-1$. This computationally motivated conjecture is in fact true; moreover, we have a full characterization of the persistence diagram of a cycle network across all dimensions $k \in {\mathbb{Z}}_+$. This characterization relies on results in [@adamaszek2015vietoris] and [@adamaszek2016nerve], and is stated for even and odd dimensions below: [theorem]{}[thmdowkercyceven]{} \[thm:dowker-cyc-even\] Fix $n\in {\mathbb{N}}$, $n \geq 3$. If $l\in {\mathbb{N}}$ is such that $n$ is divisible by $(l+1)$, and $k:=\tfrac{nl}{l+1}$ is such that $0\leq k \leq n-2$, then ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l}(G_n)$ consists of precisely the point $(\tfrac{nl}{l+1},\tfrac{nl}{l+1} + 1)$ with multiplicity $\tfrac{n}{l+1} -1$. If $l$ or $k$ do not satisfy the conditions above, then ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l}(G_n)$ is trivial. As a special case, Theorem \[thm:dowker-cyc-even\] proves Conjecture \[conj:dowker-dgm-2\] by setting $l=1$. If $n$ is odd, then it is not divisible by $(l+1) = 2$, and so ${\operatorname{Dgm}}^{{\mathfrak}{D}}_2(G_n)$ is trivial. If $n$ is even, then it is divisible by $(l+1)=2$, and $\tfrac{nl}{l+1} = \tfrac{n}{2} \leq n-2$ because $n$ is at least 4. Thus ${\operatorname{Dgm}}^{{\mathfrak}{D}}_2(G_n)$ consists of the point $(\tfrac{n}{2},\tfrac{n}{2} + 1)$ with multiplicity $\tfrac{n}{2} - 1$. [theorem]{}[thmdowkercycodd]{}\[thm:dowker-cyc-odd\] Fix $n\in {\mathbb{N}}$, $n\geq 3$. Then for $l\in {\mathbb{N}}$, define $M_l:={\left\{m \in {\mathbb{N}}: \tfrac{nl}{l+1} < m < \tfrac{n(l+1)}{l+2}\right\}}$. If $M_l$ is empty, then ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l+1}(G_n)$ is trivial. Otherwise, we have: $${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l+1}(G_n) = {\left\{\left(a_l,{\left \lceil{\tfrac{n(l+1)}{l+2}}\right \rceil }\right)\right\}},$$ where $a_l:=\min{\left\{m \in M_l\right\}}.$ We use set notation (instead of multisets) to mean that the multiplicity is 1. In particular, for $l=0$, we have $\tfrac{nl}{l+1} = 0$ and $\tfrac{n(l+1)}{l+2} = \tfrac{n}{2} \geq 3/2$, so $1 \in M_l$. Thus we have ${\operatorname{Dgm}}^{{\mathfrak}{D}}_1(G_n) = {\left\{\left(1,{\left \lceil{\tfrac{n}{2}}\right \rceil }\right)\right\}},$ and so Theorem \[thm:dowker-cyc-odd\] recovers Theorem \[thm:cycleH1\] as a special case. However, whereas the proof of Theorem \[thm:cycleH1\] is elementary and pedagogical (it relies on intuitive observations about the structure of a cycle network), the proofs of Theorems \[thm:dowker-cyc-even\] and \[thm:dowker-cyc-odd\] use sophisticated machinery developed across [@adamaszek2015vietoris] and [@adamaszek2016nerve]. We provide details for Theorem \[thm:cycleH1\] in the body of the paper, and relegate full details of Theorems \[thm:dowker-cyc-even\] and \[thm:dowker-cyc-odd\] to Appendix \[sec:cycle-addendum\]. [0.48]{} [0.48]{} [0.48]{} Sensitivity to network transformations -------------------------------------- We first make the following: Let $(X,{\omega}_X) \in {\mathcal{N}}$ be a network. For any $z,z'\in X$, define the $(z,z')$-swap of $(X,{\omega}_X)$ to be the network $S_X(z,z'):=(X^{z,z'},{\omega}_X^{z,z'})$ defined as follows: $$\begin{aligned} X^{z,z'}&:= X,\\ \text{For any $x,x'\in X^{z,z'}$,}\qquad {\omega}_X^{z,z'}(x,x')&:=\begin{cases} {\omega}_X(x',x) &: x=z,x'=z'\\ {\omega}_X(x',x) &: x'=z,x=z'\\ {\omega}_X(x,x') &: \text{otherwise.} \end{cases}\end{aligned}$$ We then pose the following question: > Given a network $(X,{\omega}_X)$ and an $(x,x')$-swap $S_X(x,x')$ for some $x,x'\in X$, how do the Rips or Dowker persistence diagrams of $S_X(x,x')$ differ from those of $(X,{\omega}_X)$? This situation is illustrated in Figure \[fig:3-node-networks\]. Example \[ex:3-node\] shows an example where the Dowker persistence diagram captures the variation in a network that occurs after a pair swap, whereas the Rips persistence diagram fails to capture this difference. Furthermore, Remark \[rem:swap-rips\] shows that Rips persistence diagrams always fail to do so. We also consider the extreme situation where all the directions of the edges of a network are reversed, i.e. the network obtained by applying the pair swap operation to each pair of nodes. We would intuitively expect that the persistence diagrams would not change. The following discussion shows that the Rips and Dowker persistence diagrams are invariant under taking the transpose of a network. \[prop:si-so\] Recall the transposition map ${\mathfrak}{t}$ and the shorthand notation $X^{\top}={\mathfrak}{t}(X)$ from Definition \[defn:sym-trans\]. Let $k\in {\mathbb{Z}}_+$. Then ${\operatorname{Dgm}}_k^{{\operatorname{si}}}(X)={\operatorname{Dgm}}_k^{{\operatorname{so}}}(X^\top),$ and therefore ${\operatorname{Dgm}}_k^{{\mathfrak}{D}}(X)={\operatorname{Dgm}}_k^{{\mathfrak}{D}}(X^\top)$ by Theorem \[thm:dowker-functorial\]. \[rem:swap-rips\] Let $(X,{\omega}_X)\in {\mathcal{N}}$, let $z,z'\in X$, and let ${\sigma}\in {\operatorname{pow}}(X)$. Then we have: $$\max_{x,x'\in {\sigma}}{\omega}_X(x,x') = \max_{x,x'\in {\sigma}}{\omega}_X^{z,z'}(x,x').$$ Using this observation, one can then repeat the arguments used in the proof of Proposition \[prop:si-so\] to show that: $${\operatorname{Dgm}}_k^{{\mathfrak}{R}}(X)={\operatorname{Dgm}}_k^{{\mathfrak}{R}}(S_X(z,z')),\text{ for each } k \in {\mathbb{Z}}_+.$$ This encodes the intuitive fact that Rips persistence diagrams are blind to pair swaps. Moreover, succesively applying the pair swap operation over all pairs produces the transpose of the original network, and so it follows that ${\operatorname{Dgm}}_k^{{\mathfrak}{R}}(X)={\operatorname{Dgm}}_k^{{\mathfrak}{R}}(X^\top)$. On the other hand, $k$-dimensional Dowker persistence diagrams are not necessarily invariant to pair swaps when $k\geq 1$. Indeed, Example \[ex:3-node\] below constructs a space $X$ for which there exist points $z,z'\in X$ such that $${\operatorname{Dgm}}_1^{{\mathfrak}{D}}(X)\neq{\operatorname{Dgm}}_1^{{\mathfrak}{D}}(S_X(z,z')).$$ However, 0-dimensional Dowker persistence diagrams are still invariant to pair swaps: \[prop:dowker0pair\] Let $(X,{\omega}_X)\in {\mathcal{N}}$, let $z,z'$ be any two points in $Z$, and let ${\sigma}\in {\operatorname{pow}}(X)$. Then we have: $${\operatorname{Dgm}}_0^{{\mathfrak}{D}}(X)={\operatorname{Dgm}}_0^{{\mathfrak}{D}}(S_X(z,z')).$$ \[ex:3-node\] Consider the three node dissimilarity networks $(X,{\omega}_X)$ and $(Y,{\omega}_Y)$ in Figure \[fig:3-node-networks\]. Note that $(Y,{\omega}_Y)$ coincides with $S_X(a,c)$. We present both the Dowker and Rips persistence barcodes obtained from these networks. Note that the Dowker persistence barcode is sensitive to the difference between $(X,{\omega}_X)$ and $(Y,{\omega}_Y)$, whereas the Rips barcode is blind to this difference. We refer the reader to §\[sec:exp\] for details on how we compute these barcodes. \(1) at (0,2.5)[$a$]{}; (2) at (-1.5,0)[$b$]{}; (3) at (1.5,0)[$c$]{}; at (0,-1)[$(X,{\omega}_X)$]{}; \(4) at (6.5,2.5)[$a$]{}; (5) at (5,0)[$b$]{}; (6) at (8,0)[$c$]{}; at (6.5,-1)[$(Y,{\omega}_Y)$]{}; \(1) edge \[loop above\] node\[above,pos=0.5\][$0$]{} (1); (2) edge \[loop left\] node\[above,pos=0.5\][$0$]{} (2); (3) edge \[loop right\] node\[above,pos=0.5\][$0$]{} (3); \(4) edge \[loop above\] node\[above,pos=0.5\][$0$]{} (4); (5) edge \[loop left\] node\[above,pos=0.5\][$0$]{} (5); (6) edge \[loop right\] node\[above,pos=0.5\][$0$]{} (6); \(1) edge \[bend left,in=180,out=0\] node\[above,pos=0.5\][$6$]{} (2); (2) edge \[bend left\] node\[above,pos=0.5\][$1$]{} (1); (1) edge \[bend left\] node\[above,pos=0.5\][$4$]{} (3); (3) edge \[bend left,out=0,in=180\] node\[above,pos=0.5\][$2$]{} (1); (2) edge \[ bend left\] node\[below,pos=0.5\][$5$]{} (3); (3) edge \[bend left\] node\[above,pos=0.5\][$3$]{} (2); \(4) edge \[bend left,in=180,out=0\] node\[above,pos=0.5\][$6$]{} (5); (5) edge \[bend left\] node\[above,pos=0.5\][$1$]{} (4); (4) edge \[bend left\] node\[above,pos=0.5\][$2$]{} (6); (6) edge \[bend left,in=180,out=0\] node\[above,pos=0.5\][$4$]{} (4); (5) edge \[ bend left\] node\[below,pos=0.5\][$5$]{} (6); (6) edge \[bend left\] node\[above,pos=0.5\][$3$]{} (5); [0.49]{} [0.49]{} [0.49]{} [0.49]{} To show how the Dowker complex is constructed, we also list the Dowker sink complexes of the networks in Figure \[fig:3-node-networks\], and also the corresponding homology dimensions across a range of resolutions. Note that when we write $[a,b](a)$, we mean that $a$ is a sink corresponding to the simplex $[a,b]$. $$\begin{aligned} {{\mathfrak}{D}^{\operatorname{si}}}_{0,X} = {\left\{[a],[b],[c]\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{0,X})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{1,X} = {\left\{[a],[b],[c],[a,b](a)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{1,X})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{2,X} = {\left\{[a],[b],[c],[a,b](a),[a,c](a),[b,c](a),[a,b,c](a)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{2,X})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{3,X} = {\left\{[a],[b],[c],[a,b](a),[a,c](a),[b,c](a),[a,b,c](a)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{3,X})) = 0\end{aligned}$$ $$\begin{aligned} {{\mathfrak}{D}^{\operatorname{si}}}_{0,Y} = {\left\{[a],[b],[c]\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{0,Y})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{1,Y} = {\left\{[a],[b],[c],[a,b](a)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{1,Y})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{2,Y} = {\left\{[a],[b],[c],[a,b](a),[a,c](c)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{2,Y})) = 0\\ {{\mathfrak}{D}^{\operatorname{si}}}_{3,Y} = {\left\{[a],[b],[c],[a,b](a),[a,c](c),[b,c](b)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{3,Y})) = 1\\ {{\mathfrak}{D}^{\operatorname{si}}}_{4,Y} = {\left\{[a],[b],[c],[a,b](a),[a,c](a),[b,c](a),[a,b,c](a)\right\}} && \dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{4,Y})) = 0\\ \end{aligned}$$ Note that for ${\delta}\in [3,4)$, $\dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},Y})) = 1$, whereas $\dim(H_1({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X}))=0$ for each ${\delta}\in {\mathbb{R}}$. Based on the discussion in Remark \[rem:swap-rips\], Proposition \[prop:dowker0pair\], and Example \[ex:3-node\], we conclude the following: Moral: Unlike Rips persistence diagrams, Dowker persistence diagrams are truly sensitive to asymmetry. We summarize some of these results: \[thm:sym-trans-summary\] Recall the symmetrization and transposition maps ${\mathfrak}{s}$ and ${\mathfrak}{t}$ from Definition \[defn:sym-trans\]. Then: 1. ${\mathfrak}{R}\circ {\mathfrak}{s} = {\mathfrak}{R}$, 2. ${\mathfrak}{D}^{{\operatorname{so}}}\circ {\mathfrak}{t} = {\mathfrak}{D}^{{\operatorname{si}}}$, and 3. ${\mathfrak}{D}^{{\operatorname{si}}}\circ {\mathfrak}{t} = {\mathfrak}{D}^{{\operatorname{so}}}$. Also, there exist $(X,{\omega}_X), (Y,{\omega}_Y) \in {\mathcal{N}}$ such that $({\mathfrak}{D}^{{\operatorname{si}}}\circ {\mathfrak}{s})(X) \neq {\mathfrak}{D}^{{\operatorname{si}}}(X)$, and $({\mathfrak}{D}^{{\operatorname{so}}}\circ {\mathfrak}{s})(Y) \neq {\mathfrak}{D}^{{\operatorname{so}}}(Y)$. These follow from Example \[ex:3-node\], Remark \[rem:rips-symm\], and Proposition \[prop:si-so\]. Implementation and an experiment on network classification {#sec:exp} ========================================================== In this section, we present the results of an experiment where we applied our methods to perform a classification task on a database of networks. All persistent homology computations were carried out using the `Javaplex` package for Matlab. A full description of Javaplex can be found in [@tausz2011javaplex]. We used $\mathbb{K}={\mathbb{Z}}_2$ as the field of coefficients for all our persistence computations. The dataset and software used for our computations are available as part of the `PersNet` software package on <https://research.math.osu.edu/networks/Datasets.html>. A version of our simulated hippocampal networks experiment has appeared in [@dowker-asilo]. All networks in the following experiment were normalized to have weights in the range $[0,1]$. For each network, we computed Dowker sink complexes at resolutions ${\delta}= 0.01,0.02,0.03,\ldots,1.00$. This filtration was then passed into Javaplex, which produced 0 and 1-dimensional Dowker persistence barcodes. Simulated hippocampal networks {#sec:exp-arenas} ------------------------------ In the neuroscience literature, it has been shown that as an animal explores a given environment or arena, specific “place cells" in the hippocampus show increased activity at specific spatial regions, called “place fields" [@o1971hippocampus]. Each place cell shows a spike in activity when the animal enters the place field linked to this place cell, accompanied by a drop in activity as the animal moves far away from this place field. To understand how the brain processes this data, a natural question to ask is the following: Is the time series data of the place cell activity, referred to as “spike trains", enough to detect the structure of the arena? Approaches based on homology [@curto2008cell] and persistent homology [@dabaghian2012topological] have shown positive results in this direction. In [@dabaghian2012topological], the authors simulated the trajectory of a rat in an arena containing “holes." A simplicial complex was then built as follows: whenever $n+1$ place cells with overlapping place fields fired together, an $n$-simplex was added. This yield a filtered simplicial complexed indexed by a time parameter. By computing persistence, it was then shown that the number of persistent bars in the 1-dimensional barcode of this filtered simplicial complex would accurately represent the number of holes in the arena. We repeated this experiment with the following change in methodology: we simulated the movement of an animal, and corresponding hippocampal activity, in arenas with a variety of obstacles. We then induced a directed network from each set of hippocampal activity data, and computed the associated 1-dimensional Dowker persistence diagrams. We were interested in seeing if the bottleneck distances between diagrams arising from similar arenas would differ significantly from the bottleneck distance between diagrams arising from different arenas. To further exemplify our methods, we repeated our analysis after computing the 1-dimensional Rips persistence diagrams from the hippocampal activity networks. In our experiment, there were five arenas. The first was a square of side length $L=10$, with four circular “holes" or “forbidden zones" of radius $0.2L$ that the trajectory could not intersect. The other four arenas were those obtained by removing the forbidden zones one at a time. In what follows, we refer to the arenas of each type as 4-hole, 3-hole, 2-hole, 1-hole, and 0-hole arenas. For each arena, a random-walk trajectory of 5000 steps was generated, where the animal could move along a square grid with 20 points in each direction. The grid was obtained as a discretization of the box $[0,L]\times [0,L]$, and each step had length $0.05L$. The animal could move in each direction with equal probability. If one or more of these moves took the animal outside the arena (a disallowed move), then the probabilities were redistributed uniformly among the allowed moves. Each trajectory was tested to ensure that it covered the entire arena, excluding the forbidden zones. Formally, we write the time steps as a set $T:={\left\{1,2,\ldots, 5000\right\}}$, and denote the trajectory as a map $\operatorname{traj}:T {\rightarrow}[0,L]^2$. For each of the five arenas, 20 trials were conducted, producing a total of 100 trials. For each trial $l_k$, an integer $n_k$ was chosen uniformly at random from the interval $[150,200]$. Then $n_k$ place fields of radius $0.05L$ were scattered uniformly at random inside the corresponding arena for each $l_k$. An illustration of the place field distribution is provided in Figure \[fig:arenas-rasters\]. A spike on a place field was recorded whenever the trajectory would intersect it. So for each $1\leq i\leq n_k$, the spiking pattern of cell $x_i$, corresponding to place field PF$_i$, was recorded via a function $r_i:T{\rightarrow}{\left\{0,1\right\}}$ given by: $$r_i(t)=\begin{cases} 1 &:\text{if } \operatorname{traj}(t)\text{ intersects } \text{PF}_i,\\ 0 &: \text{otherwise}\end{cases} \qquad\qquad t\in T.$$ The matrix corresponding to $r_i$ is called the raster of cell $x_i$. A sample raster is illustrated in Figure \[fig:arenas-rasters\]. For each trial $l_k$, the corresponding network $(X_k,{\omega}_{X_k})$ was constructed as follows: $X_k$ consisted of $n_k$ nodes representing place fields, and for each $1\leq i,j\leq n_k$, the weight ${\omega}_{X_k}(x_i,x_j)$ was given by: $$\begin{aligned} {\omega}_{X_k}(x_i,x_j) &:=1-\frac{N_{i,j}(5)}{\sum_{i=1}^{n_k}N_{i,j}(5)},\\ \text{ where } N_{i,j}(5)&={\operatorname{card}}\left({\left\{(s,t)\in T^2:t\in [2,5000], t-s\in [1,5], r_j(t)=1,r_i(s)=1\right\}}\right).\end{aligned}$$ In words, $N_{i,j}(5)$ counts the pairs of times $(s,t), s < t,$ such that cell $x_j$ spikes (at a time $t$) after cell $x_i$ spikes (at a time $s$), and the delay between the two spikes is fewer than 5 time steps. The idea is that if cell $x_j$ frequently fires within a short span of time after cell $x_i$ fires, then place fields PF$_i$ and PF$_j$ are likely to be in close proximity to each other. The column sum of the matrix corresponding to ${\omega}_{X_k}$ is normalized to 1, and so ${\omega}_{X_k}^\top$ can be interpreted as the transition matrix of a Markov process. [0.3]{} [0.3]{} [0.3]{} [0.3]{} [0.3]{} [0.3]{} Next, we computed the 1-dimensional Dowker persistence diagrams of each of the 100 networks. Note that ${\operatorname{Dgm}}_1^{{\mathfrak}{D}}({\omega}_X)={\operatorname{Dgm}}_1^{{\mathfrak}{D}}({\omega}_X^\top)$ by Proposition \[prop:si-so\], so we are actually obtaining the 1-dimensional Dowker persistence diagrams of transition matrices of Markov processes. We then computed a $100\times 100$ matrix consisting of the bottleneck distances between all the 1-dimensional persistence diagrams. The single linkage dendrogram generated from this bottleneck distance matrix is shown in Figure \[fig:dendro-dowker-arenas\]. The labels are in the format `env-<nh>-<nn>`, where `nh` is the number of holes in the arena/environment, and `nn` is the number of place fields. Note that with some exceptions, networks corresponding to the same arena are clustered together. We conclude that the Dowker persistence diagram succeeded in capturing the intrinsic differences between the five classes of networks arising from the five different arenas, even when the networks had different sizes. We then computed the Rips persistence diagrams of each network, and computed the $100\times 100$ bottleneck distance matrix associated to the collection of 1-dimensional diagrams. The single linkage dendrogram generated from this matrix is given in Figure \[fig:dendro-rips-arenas\]. Notice that the Rips dendrogram does not do a satisfactory job of classifying arenas correctly. We note that an alternative method of comparing the networks obtained from our simulations would have been to compute the pairwise network distances, and plot the results in a dendrogram. But ${d_{\mathcal{N}}}$ is NP-hard to compute—this follows from the fact that computing ${d_{\mathcal{N}}}$ includes the problem of computing Gromov-Hausdorff distance between finite metric spaces, which is NP-hard [@schmiedl]. So instead, we are computing the bottleneck distances between 1-dimensional Dowker persistence diagrams, as suggested by Remark \[rem:dowker-benefits\]. Discussion ========== We provided a complete description of the Rips and Dowker persistence diagrams of general networks. The stability results we have provided give quantitative guarantees on the robustness of these persistence diagrams. As a building block, we proved a functorial generalization of Dowker’s theorem, which also yields an independent proof of a folklore strengthening of Dowker’s theorem. We have provided numerous examples suggesting that Dowker persistence diagrams are an appropriate method for analyzing general asymmetric networks. For a particular class of such examples, the family of cycle networks, we have fully characterized their Dowker persistence diagrams in all dimensions. Finally, we have implemented our methods for a classification task on a database of networks, and provided interpretations for our results. We believe that the story of “persistent homology of asymmetric networks" has more aspects to be uncovered. Of particular interest to us is the analysis of alternative methods of producing simplicial complexes from asymmetric networks, for example, the directed flag complex construction of [@dlotko2016topological]. Yet another interesting extension to the non-metric framework has appeared in [@edelsbrunner2016topological], in the context of computing generalized Čech and Rips complexes for Bregman divergences. We remark that a persistent homology framework for the directed flag complex has been proposed by [@turner], but the computational aspects of this construction have not been addressed in the current literature. Another approach for computing persistence diagrams from asymmetric networks, which bypasses the construction of any simplicial complex and operates directly at the chain level is given in [@pph]. Some other interesting questions relate to cycle networks: for example, we would like to obtain a characterization of the Rips persistence diagrams of cycle networks for any dimension $k\geq 1$. Finally, it is important to devise more efficient implementations for the Dowker complexes we present here. It is likely that ideas from the literature on efficient construction of Čech complexes [@dantchev2012efficient; @edelsbrunner2016topological] will be helpful in this regard. #### Acknowledgments. This work was supported by NSF grants IIS-1422400 and CCF-1526513. We thank Pascal Wild and Zhengchao Wan for pointing out errors on an early preprint, and also Osman Okutan and Tim Porter for useful discussions. We are especially thankful to Henry Adams for numerous helpful observations and suggestions, especially regarding the material in Appendix \[sec:cycle-addendum\], and for suggesting the proof strategy for Theorem \[thm:cyc-cech-main\]. Proofs {#app:proofs} ====== The first inequality holds by the Algebraic Stability Theorem. For the second inequality, note that the contiguous simplicial maps in the diagrams above induce chain maps between the corresponding chain complexes, and these in turn induce equal linear maps at the level of homology vector spaces. To be more precise, first consider the maps $t_{{\delta}+\eta,{\delta}'+\eta}\circ {\varphi}_{\delta}$ and ${\varphi}_{{\delta}'}\circ s_{{\delta},{\delta}'}$. These simplicial maps induce linear maps $(t_{{\delta}+\eta,{\delta}'+\eta}\circ {\varphi}_{\delta})_\#, ({\varphi}_{{\delta}'}\circ s_{{\delta},{\delta}'})_\#: H_k({\mathfrak}{F}^{\delta}) {\rightarrow}H_k({\mathfrak}{G}^{{\delta}'+\eta})$. Because the simplicial maps are assumed to be contiguous, we have: $$(t_{{\delta}+\eta,{\delta}'+\eta}\circ {\varphi}_{\delta})_\# = ({\varphi}_{{\delta}'}\circ s_{{\delta},{\delta}'})_\#.$$ By invoking functoriality of homology, we then have: $$(t_{{\delta}+\eta,{\delta}'+\eta})_\# \circ ({\varphi}_{\delta})_\# = ({\varphi}_{{\delta}'})_\#\circ (s_{{\delta},{\delta}'})_\#.$$ Analogous results hold for the other pairs of contiguous maps. Thus we obtain commutative diagrams upon passing to homology, and so ${\mathcal{H}}_k({\mathfrak}{F}), {\mathcal{H}}_k({\mathfrak}{G})$ are $\eta$-interleaved for each $k\in {\mathbb{Z}}_+$. Thus we get: $${d_{\operatorname{I}}}({\mathcal{H}}_k({\mathfrak}{F}), {\mathcal{H}}_k({\mathfrak}{G}))\leq \eta. \qedhere$$ First we show that: $${d_{\mathcal{N}}}(X,Y) \geq \tfrac{1}{2}\inf\{\max({\operatorname{dis}}({\varphi}),{\operatorname{dis}}(\psi),C_{X,Y}({\varphi},\psi), C_{Y,X}(\psi,{\varphi})) : {\varphi}:X {\rightarrow}Y, \psi:Y {\rightarrow}X \text{ any maps}\}.$$ Let $\eta = {d_{\mathcal{N}}}(X,Y)$, and let $R$ be a correspondence such that ${\operatorname{dis}}(R) = 2\eta$. We can define maps ${\varphi}:X{\rightarrow}Y$ and $\psi:Y{\rightarrow}X$ as follows: for each $x\in X$, set ${\varphi}(x)=y$ for some $y$ such that $(x,y)\in R$. Similarly, for each $y\in Y$, set $\psi(y)=x$ for some $x$ such that $(x,y)\in R$. Thus for any $x \in X, y\in Y$, we obtain $\|{\omega}_X(x,\psi(y)) - {\omega}_Y({\varphi}(x),y)\| \leq 2\eta$ and $\|{\omega}_X(\psi(y),x) - {\omega}_Y(y,{\varphi}(x))\| \leq 2\eta$. So we have both $C_{X,Y}({\varphi},\psi) \leq 2\eta$ and $C_{Y,X}(\psi,{\varphi}) \leq 2\eta$. Also for any $x,x' \in X$, we have $(x,{\varphi}(x)),(x',{\varphi}(x')) \in R$. Thus we also have $$\|{\omega}_X(x,x') - {\omega}_Y({\varphi}(x),{\varphi}(x'))\| \leq 2\eta.$$ So ${\operatorname{dis}}({\varphi}) \leq 2\eta$ and similarly ${\operatorname{dis}}(\psi) \leq 2\eta$. This proves the “$\geq$" case. Next we wish to show: $${d_{\mathcal{N}}}(X,Y) \leq \tfrac{1}{2}\inf\{\max({\operatorname{dis}}({\varphi}),{\operatorname{dis}}(\psi),C_{X,Y}({\varphi},\psi), C_{Y,X}(\psi,{\varphi})) : {\varphi}:X {\rightarrow}Y, \psi:Y {\rightarrow}X \text{ any maps}\}.$$ Suppose ${\varphi}, \psi$ are given, and $\frac{1}{2}\max({\operatorname{dis}}({\varphi}),{\operatorname{dis}}(\psi),C_{X,Y}({\varphi},\psi),C_{Y,X}(\psi,{\varphi})) = \eta$. Let $R_X = {\left\{(x,{\varphi}(x) : x\in X\right\}}$ and let $R_Y = {\left\{(\psi(y),y) : y\in Y\right\}}$. Then $R = R_X \cup R_Y$ is a correspondence. We wish to show that for any $z = (a,b), z' = (a',b') \in R$, $$\|{\omega}_X(a,a') - {\omega}_Y(b,b')\| \leq 2\eta.$$ This will show that ${\operatorname{dis}}(R) \leq 2\eta$, and so ${d_{\mathcal{N}}}(X,Y) \leq \eta$. To see this, let $z,z' \in R$. Note that there are four cases: (1) $z,z' \in R_X$, (2) $z,z' \in R_Y$, (3) $z \in R_X, z' \in R_Y$, and (4) $z\in R_Y, z'\in R_X$. In the first two cases, the desired inequality follows because ${\operatorname{dis}}({\varphi}), {\operatorname{dis}}(\psi) \leq 2\eta$. The inequality follows in cases (3) and (4) because $C_{X,Y}({\varphi},\psi) \leq 2\eta$ and $C_{Y,X}(\psi,{\varphi}) \leq 2\eta$, respectively. Thus ${d_{\mathcal{N}}}(X,Y) \leq \eta$. It suffices to show that $\Phi$ is a simplicial approximation to ${\mathcal}{E}_{\|\Sigma\|}$, i.e. whenever ${\mathcal}{E}_{\|\Sigma\|}(x) \in \|{{\sigma}}\|$ for some vertex $x \in \|{\Sigma}{^{(1)}}\|$ and some simplex ${\sigma}\in \|{\Sigma}\|$, we also have $\|\Phi\|(x) \in \|{{\sigma}}\|$ [@spanier-book §3.4]. Here $\|{\sigma}\|$ denotes the closed simplex of ${\sigma}$; for any simplex ${\sigma}=[v_0,\ldots, v_k]$, this is the collection of formal convex combinations $\sum_{i=0}^ka_iv_i$ with $a_i \geq 0$ for each $0\leq i \leq k$ and $\sum_{i=0}^ka_i =1$. Let $x = \sum_{i=0}^ka_i{\sigma}_i$ be a vertex in $\|\Sigma{^{(1)}}\|$, with each $a_i > 0$. Then we have ${\mathcal}{E}_{\|\Sigma\|}(x) = \sum_{i=0}^ka_i{\mathcal}{B}({\sigma}_i) = \sum_{i=0}^ka_i\sum_{v\in {\sigma}_i}v/{{\operatorname{card}}({\sigma}_i)},$ a vertex in $\|{\sigma}_k\|$. Also we have $\|\Phi\|(x) = \sum_{i=0}^ka_i\Phi({\sigma}_i)$, a vertex in $\|{{\sigma}_k}\|$. Thus $\Phi$ is a simplicial approximation to ${\mathcal}{E}_{\|\Sigma\|}$, and so we have $\|\Phi\|\simeq {\mathcal}{E}_{\|\Sigma\|}$. Let ${\delta}\in {\mathbb{R}}$. We first claim that ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}(X) = {{\mathfrak}{D}^{\operatorname{so}}}_{\delta}(X^\top)$. Let ${\sigma}\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}(X)$. Then there exists $x'$ such that ${\omega}_X(x,x')\leq {\delta}$ for any $x\in {\sigma}$. Thus ${\omega}_{X^\top}(x',x)\leq {\delta}$. So ${\sigma}\in {{\mathfrak}{D}^{\operatorname{so}}}_{\delta}(X^{\top})$. A similar argument shows the reverse containment. This proves our claim. Thus for ${\delta}\leq {\delta}' \leq {\delta}''$, we obtain the following diagram: $$\begin{tikzcd} {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}(X) \arrow{r}\ar[-, double equal sign distance=3pt]{d} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}'}(X) \arrow{r}\ar[-, double equal sign distance=3pt]{d} & {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}''}(X) \arrow{r}\ar[-, double equal sign distance=3pt]{d} & \ldots \\ {{\mathfrak}{D}^{\operatorname{so}}}_{\delta}(X^\top) \arrow{r} & {{\mathfrak}{D}^{\operatorname{so}}}_{{\delta}'}(X^\top)\arrow{r} & {{\mathfrak}{D}^{\operatorname{so}}}_{{\delta}''}(X^\top)\arrow{r} & \ldots \end{tikzcd}$$ Since the maps ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}{\rightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}'}$, ${{\mathfrak}{D}^{\operatorname{so}}}_{\delta}{\rightarrow}{{\mathfrak}{D}^{\operatorname{so}}}_{{\delta}'}$ for ${\delta}'\geq {\delta}$ are all inclusion maps, it follows that the diagrams commute. Thus at the homology level, we obtain, via functoriality of homology, a commutative diagram of vector spaces where the intervening vertical maps are isomorphisms. By the Persistence Equivalence Theorem (\[thm:pet\]), the diagrams ${\operatorname{Dgm}}_k^{{\operatorname{si}}}(X)$ and ${\operatorname{Dgm}}_k^{{\operatorname{so}}}{(X^\top)}$ are equal. By invoking Corollary \[cor:dowker-dual\], we obtain ${\operatorname{Dgm}}_k^{{\mathfrak}{D}}(X)={\operatorname{Dgm}}_k^{{\mathfrak}{D}}(X^\top)$. Let ${\delta}\in {\mathbb{R}}$. For notational convenience, we write, for each $k\in {\mathbb{Z}}_+$, ---------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------ ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}:={{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X} $ $C_k^{\delta}:=C_k({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},X})$ ${\partial}_k^{\delta}:={\partial}_k^{\delta}: C_k^{\delta}{\rightarrow}C_{k-1}^{\delta}$ ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}:={{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S_X(z,z')}$ $C_k^{{\delta},S}:=C_k({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S_X(z,z')})$ ${\partial}_k^{{\delta},S}:={\partial}_k^{{\delta},S}:C_k^{{\delta},S} {\rightarrow}C_{k-1}^{{\delta},S}.$ ---------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------ First note that pair swaps do not affect the entry of 0-simplices into the Dowker filtration. More precisely, for any $x\in X$, we can unpack the definition of $R_{{\delta},X}$ (Equation \[eq:relation\]) to obtain: $$[x]\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}\iff {\omega}_X(x,x)\leq {\delta}\iff {\omega}_X^{z,z'}(x,x)\leq {\delta}\iff [x]\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}.$$ Thus for any ${\delta}\in {\mathbb{R}}$, we have $C_0^{\delta}= C_0^{{\delta},S}$. Since all 0-chains are automatically 0-cycles, we have $\ker({\partial}_0^{\delta})=\ker({\partial}_0^{{\delta},S})$. Next we wish to show that ${\operatorname{im}}({\partial}_1^{{\delta}})={\operatorname{im}}({\partial}_1^{{\delta},S})$ for each ${\delta}\in {\mathbb{R}}$. Let ${\gamma}\in C_1^{\delta}$. We first need to show the forward inclusion, i.e. that ${\partial}_1^{\delta}({\gamma}) \in {\operatorname{im}}({\partial}_1^{{\delta},S})$. It suffices to show this for the case that ${\gamma}$ is a single 1-simplex $[x,x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$; the case where ${\gamma}$ is a linear combination of 1-simplices will then follow by linearity. Let ${\gamma}=[x,x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ for $x,x'\in X$. Then we have the following possibilities: 1. $x'' \in X\setminus\{z,z'\}$ is a ${\delta}$-sink for $[x,x']$. 2. $z$ (or $z'$) is the only ${\delta}$-sink for $[x,x']$, and $x,x'\not\in {\left\{z,z'\right\}}$. 3. $z$ (or $z'$) is the only ${\delta}$-sink for $[x,x']$, and either $x$ or $x'$ belongs to ${\left\{z,z'\right\}}$. 4. $z$ (or $z'$) is the only ${\delta}$-sink for $[x,x']$, and both $x,x'$ belong to ${\left\{z,z'\right\}}$. In cases (1), (2), and (4), the $(z,z')$-pair swap has no effect on $[x,x']$, in the sense that we still have $[x,x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}$. So $[x']-[x]={\partial}_1^{\delta}({\gamma})={\partial}_1^{{\delta},S}({\gamma})\in {\operatorname{im}}({\partial}_1^{{\delta},S})$. Next consider case (3), and assume for notational convenience that $[x,x']=[z,x']$ and $z'$ is the only ${\delta}$-sink for $[z,x']$. By the definition of a ${\delta}$-sink, we have $\overline{{\omega}}_X(z,z')\leq {\delta}$ and $\overline{{\omega}}_X(x',z')\leq {\delta}$. Notice that we also have: $$[z,z'],[z',x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{\delta}, \text{ with $z'$ as a ${\delta}$-sink}.$$ After the $(z,z')$-pair swap, we still have $\overline{{\omega}}_X^{z,z'}(x',z')\leq {\delta}$, but possibly $\overline{{\omega}}_X^{z,z'}(z,z')> {\delta}$. So it might be the case that $[z,x']\not\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}$. However, we now have: $$\begin{aligned} &[z',x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}, \text{ with $z'$ as a ${\delta}$-sink, and}\\ &[z,z'] \in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}, \text{ with $z$ as a ${\delta}$-sink}.\end{aligned}$$ Then we have: $$\begin{aligned} {\partial}_1^{\delta}({\gamma})={\partial}_1^{\delta}([z,x'])=x'-z &=z'-z + x'-z'\\ &={\partial}_1^{{\delta}}([z,z'])+{\partial}_1^{{\delta}}([z',x'])\\ &={\partial}_1^{{\delta},S}([z,z'])+{\partial}_1^{{\delta},S}([z',x'])\in {\operatorname{im}}({\partial}_1^{{\delta},S}),\end{aligned}$$ where the last equality is defined because we have checked that $[z,z'],[z',x']\in {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}$. Thus ${\operatorname{im}}({\partial}_1^{{\delta}})\subseteq {\operatorname{im}}({\partial}_1^{{\delta},S})$, and the reverse inclusion follows by a similar argument. Since ${\delta}\in {\mathbb{R}}$ was arbitrary, this shows that ${\operatorname{im}}({\partial}_1^{{\delta}})= {\operatorname{im}}({\partial}_1^{{\delta},S})$ for each ${\delta}\in {\mathbb{R}}$. Previously we had $\ker({\partial}_0^{\delta})=\ker({\partial}_0^{{\delta},S})$ for each ${\delta}\in {\mathbb{R}}$. It then follows that $H_0({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})=H_0({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S})$ for each ${\delta}\in {\mathbb{R}}$. Next let ${\delta}'\geq {\delta}\in {\mathbb{R}}$, and for any $k\in {\mathbb{Z}}_+$, let $f_k^{{\delta},{\delta}'}:C_k^{\delta}{\rightarrow}C_k^{{\delta}'}, g_k^{{\delta},{\delta}'}:C_k^{{\delta},S} {\rightarrow}C_k^{{\delta}',S}$ denote the chain maps induced by the inclusions ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}{\hookrightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}'}, {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S} {\hookrightarrow}{{\mathfrak}{D}^{\operatorname{si}}}_{{\delta}',S}$. Since ${{\mathfrak}{D}^{\operatorname{si}}}_{\delta}$ and ${{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S}$ have the same 0-simplices at each ${\delta}\in {\mathbb{R}}$, we know that $f_0^{{\delta},{\delta}'}\equiv g_0^{{\delta},{\delta}'}$. Let ${\gamma}\in \ker({\partial}_0^{\delta})=\ker({\partial}_0^{{\delta},S})$, and let ${\gamma}+{\operatorname{im}}({\partial}_1^{\delta}) \in H_0({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})$. Then observe that $$\begin{aligned} (f_0^{{\delta},{\delta}'})_\#({\gamma}+ {\operatorname{im}}({\partial}_1^{\delta})) &=f_0^{{\delta},{\delta}'}({\gamma}) + {\operatorname{im}}({\partial}_1^{{\delta}'}) &&\text{($f_0^{{\delta},{\delta}'}$ is a chain map)} \\ &=g_0^{{\delta},{\delta}'}({\gamma}) + {\operatorname{im}}({\partial}_1^{{\delta}'}) &&\text{($f_0^{{\delta},{\delta}'}\equiv g_0^{{\delta},{\delta}'}$)}\\ &=g_0^{{\delta},{\delta}'}({\gamma}) + {\operatorname{im}}({\partial}_1^{{\delta}',S}) &&\text{(${\operatorname{im}}({\partial}_1^{{\delta}'})={\operatorname{im}}({\partial}_1^{{\delta}',S})$)}\\ &=(g_0^{{\delta},{\delta}'})_\#({\gamma}+ {\operatorname{im}}({\partial}_1^{{\delta},S})).&&\text{($g_0^{{\delta},{\delta}'}$ is a chain map)}\end{aligned}$$ Thus $(f_0^{{\delta},{\delta}'})_\#=(g_0^{{\delta},{\delta}'})_\#$ for each ${\delta}'\geq {\delta}\in {\mathbb{R}}$. Since we also have $H_0({{\mathfrak}{D}^{\operatorname{si}}}_{\delta})=H_0({{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},S})$ for each ${\delta}\in {\mathbb{R}}$, we can then apply the Persistence Equivalence Theorem (Theorem \[thm:pet\]) to conclude the proof. Higher dimensional Dowker persistence diagrams of cycle networks {#sec:cycle-addendum} ================================================================ The contents of this section rely on results in [@adamaszek2015vietoris] and [@adamaszek2016nerve]. We introduce some minimalistic versions of definitions from the referenced papers to use in this section. The reader should refer to these papers for the original definitions. Given a metric space $(M,d_M)$ and $m\in M$, we will write $\overline{B(m,{\varepsilon})}$ to denote a closed ${\varepsilon}$-ball centered at $m$, for any ${\varepsilon}> 0$. For a subset $X\subseteq M$ and some ${\varepsilon}>0$, the Čech complex of $X$ at resolution ${\varepsilon}$ is defined to be the following simplicial complex: $${\operatorname{\mathbf{\check{C}}}}(X,{\varepsilon}):={\left\{{\sigma}\subseteq X : \cap_{x\in {\sigma}}\overline{B(x,{\varepsilon})} \neq {\varnothing}\right\}}.$$ In the setting of metric spaces, the Čech complex coincides with the Dowker source and sink complexes. We will be interested in the special case where the underlying metric space is the circle. We write $S^1$ to denote the circle with unit circumference. Next, for any $n\in {\mathbb{N}}$, we write ${\mathbb}{X}_n:={\left\{0,\tfrac{1}{n},\tfrac{2}{n},\ldots, \tfrac{n-1}{n}\right\}}$ to denote the collection of $n$ equally spaced points on $S^1$ with the restriction of the arc length metric on $S^1$. Also let $G_n$ denote the $n$-node cycle network with vertex set ${\mathbb}{X}_n$ (in contrast with ${\mathbb}{X}_n$, here $G_n$ is equipped with the asymmetric weights defined in §\[sec:cycle\]). The connection between ${\mathbb}{X}_n$ and Dowker complexes of the cycle networks $G_n$ is highlighted by the following observation: \[prop:dowker-cech-cplx\] Let $n\in {\mathbb{N}}$. Then for any ${\delta}\in [0,1]$, we have ${\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\frac{{\delta}}{2}) = {\mathfrak}{D}^{{\operatorname{si}}}_{n{\delta},G_n}.$ The scaling factor arises because $G_n$ has diameter $\sim n$, whereas ${\mathbb}{X}_n\subseteq S^1$ has diameter $\sim 1/2$. This proposition provides a pedagogical step which helps us transport results from the setting of [@adamaszek2015vietoris] and [@adamaszek2016nerve] to that of the current paper. For ${\delta}=0$, both the Čech and Dowker complexes consist of the $n$ vertices, and are equal. Similarly for ${\delta}=1$, both ${\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,1)$ and ${\mathfrak}{D}^{{\operatorname{si}}}_{n,G_n}$ are equal to the $(n-1)$-simplex. Now suppose ${\delta}\in (0,1)$. Let ${\sigma}\in {\mathfrak}{D}^{{\operatorname{si}}}_{n{\delta},G_n}$. Then ${\sigma}$ is of the form $[\tfrac{k}{n},\tfrac{k+1}{n},\ldots, \tfrac{{\left \lfloor{k+n{\delta}}\right \rfloor }}{n}]$ for some integer $0\leq k \leq n-1$, where the $n{\delta}$-sink is $\tfrac{{\left \lfloor{k+n{\delta}}\right \rfloor }}{n}$ and all the numerators are taken modulo $n$. We claim that ${\sigma}\in {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{{\delta}}{2})$. To see this, observe that $d_{S^1}(\tfrac{k}{n},\tfrac{{\left \lfloor{k+n{\delta}}\right \rfloor }}{n}) \leq {\delta}$, and so $\overline{B(\tfrac{k}{n},\tfrac{{\delta}}{2})} \cap \overline{B(\tfrac{{\left \lfloor{k+n{\delta}}\right \rfloor }}{n},\tfrac{{\delta}}{2})} \neq {\varnothing}$. Then we have ${\sigma}\in \bigcap_{i=0}^{n{\delta}}\overline{B\left(\tfrac{{\left \lfloor{k+i}\right \rfloor }}{n},\tfrac{{\delta}}{2}\right)}$, and so ${\sigma}\in {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{{\delta}}{2})$. Now let ${\sigma}\in {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{{\delta}}{2})$. Then ${\sigma}$ is of the form $[\tfrac{k}{n},\tfrac{k+1}{n},\ldots, \tfrac{k+j}{n}]$ for some integer $0\leq k\leq n-1$, where $j$ is an integer such that $\tfrac{j}{n} \leq {\delta}$. In this case, we have ${\sigma}= {\mathbb}{X}_n \cap_{i=0}^j\overline{B\left(\tfrac{k+i}{n},{\delta}\right)}$. Then in $G_n$, after applying the scaling factor $n$, we have ${\sigma}\in {\mathfrak}{D}^{{\operatorname{si}}}_{n{\delta},G_n}$, with $\tfrac{k+j}{n}$ as an $n{\delta}$-sink in $G_n$. This shows equality of the two simplicial complexes. \[thm:cech-S1\] Fix $n\in {\mathbb{N}}$, and let $0\leq k \leq n-2$ be an integer. Then, $${\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k}{2n})\simeq \begin{cases} \bigvee^{n-k-1}S^{2l} & \text{if } \tfrac{k}{n} = \tfrac{l}{l+1},\\ S^{2l+1} &\text{or if } \tfrac{l}{l+1} < \tfrac{k}{n} < \tfrac{l+1}{l+2}, \end{cases}$$ for some $l \in {\mathbb{Z}}_+$. Here $\bigvee$ denotes the wedge sum, and $\simeq$ denotes homotopy equivalence. Let $l \in {\mathbb{N}}$ be such that $(l+1)$ divides $n$ and $0\leq k\leq n-2$. Then ${\mathfrak}{D}^{{\operatorname{si}}}_{k,G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k}{2n})$ has the homotopy type of a wedge sum of $(n-k-1)$ copies of $S^{2l}$, by Theorem \[thm:cech-S1\]. Here the equality follows from Proposition \[prop:dowker-cech-cplx\]. Notice that $n-k-1 = \tfrac{n}{l+1}-1$. Furthermore, by another application of Theorem \[thm:cech-S1\], it is always possible to choose ${\varepsilon}>0$ small enough so that ${\mathfrak}{D}^{{\operatorname{si}}}_{k-{\varepsilon},G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k-{\varepsilon}}{2n})$ and ${\mathfrak}{D}^{{\operatorname{si}}}_{k+{\varepsilon},G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k+{\varepsilon}}{2n})$ have the homotopy types of odd-dimensional spheres. Thus the inclusions ${\mathfrak}{D}^{{\operatorname{si}}}_{k-{\varepsilon},G_n} \subseteq {\mathfrak}{D}^{{\operatorname{si}}}_{k,G_n} \subseteq {\mathfrak}{D}^{{\operatorname{si}}}_{k+{\varepsilon},G_n}$ induce zero maps upon passing to homology. It follows that ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l}(G_n)$ consists of the point $(\tfrac{nl}{l+1},\tfrac{nl}{l+1} + 1)$ with multiplicity $\tfrac{n}{l+1} -1$. If $l \in {\mathbb{N}}$ does not satisfy the condition described above, then there does not exist an integer $1\leq j \leq n-2$ such that $j/n = l/(l+1)$. So for each $1\leq j \leq n-2$, ${\mathfrak}{D}^{{\operatorname{si}}}_{j,G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{j}{2n})$ has the homotopy type of an odd-dimensional sphere by Theorem \[thm:cech-S1\], and thus does not contribute to ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l}(G_n)$. If $l$ satisfies the condition but $k \geq n-1$, then ${\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k}{2n})$ is just the $(n-1)$-simplex, hence contractible. Theorem \[thm:dowker-cyc-even\] gives a characterization of the even dimensional Dowker persistence diagrams of cycle networks. The most interesting case occurs when considering the 2-dimensional diagrams: we see that cycle networks of an even number of nodes have an interesting barcode, even if the bars are all short-lived. For dimensions 4, 6, 8, and beyond, there are fewer and fewer cycle networks with nontrivial barcodes (in the sense that only cycle networks with number of nodes equal to a multiple of 4, 6, 8, and so on have nontrivial barcodes). For a complete picture, it is necessary to look at odd-dimensional persistence diagrams. This is made possible by the next set of constructions. We have already recalled the definition of a Rips complex of a metric space. To facilitate the assessment of the connection to [@adamaszek2015vietoris], we temporarily adopt the notation ${\operatorname{\mathbf{VR}}}(X,{\varepsilon})$ to denote the Vietoris-Rips complex of a metric space $(X,d_X)$ at resolution ${\varepsilon}>0$, i.e. the simplicial complex ${\left\{{\sigma}\subseteq X : {\operatorname{diam}}({\sigma}) \leq {\varepsilon}\right\}}$. \[thm:vr-cech-cd\] Let $0< r < \tfrac{1}{2}$. Then there exists a map $T_r: {\operatorname{pow}}(S^1) {\rightarrow}{\operatorname{pow}}(S^1)$ and a map $\pi_r: S^1 {\rightarrow}S^1$ such that there is an induced homotopy equivalence $${\operatorname{\mathbf{VR}}}(T_r(X), \tfrac{2r}{1+2r}) {\xrightarrow}{\simeq} {\operatorname{\mathbf{\check{C}}}}(X,r).$$ Next suppose $X\subseteq S^1$ and let $0< r \leq r' < \tfrac{1}{2}$. Then there exists a map $\eta: S^1 {\rightarrow}S^1$ such that the following diagram commutes: $$\begin{tikzcd}[column sep=large] {\operatorname{\mathbf{VR}}}(T_r(X), \tfrac{2r}{1+2r}) \arrow{r}{\eta} \arrow{d}{\simeq}[swap]{\pi_r} & {\operatorname{\mathbf{VR}}}(T_{r'}(X), \tfrac{2r'}{1+2r'}) \arrow{d}{\simeq}[swap]{\pi_{r'}}\\ {\operatorname{\mathbf{\check{C}}}}(X,r) \arrow[hookrightarrow]{r}{\subseteq} & {\operatorname{\mathbf{\check{C}}}}(X,r') \end{tikzcd}$$ \[thm:cyc-cech-main\] Consider the setup of Theorem \[thm:vr-cech-cd\]. If ${\operatorname{\mathbf{\check{C}}}}(X,r)$ and ${\operatorname{\mathbf{\check{C}}}}(X,r')$ are homotopy equivalent, then the inclusion map between them is a homotopy equivalence. Before providing the proof, we show how it implies Theorem \[thm:dowker-cyc-odd\]. By Proposition \[prop:dowker-cech-cplx\] and Theorem \[thm:cech-S1\], we know that ${{\mathfrak}{D}^{\operatorname{si}}}_{k,G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k}{2n}) \simeq S^1 $ for integers $0 < k < \tfrac{n}{2}$. Let $b \in {\mathbb{N}}$ be the greatest integer less than $n/2$. Then by Theorem \[thm:cyc-cech-main\], we know that each inclusion map in the following chain is a homotopy equivalence: $${{\mathfrak}{D}^{\operatorname{si}}}_{1,G_n} \subseteq \ldots \subseteq {{\mathfrak}{D}^{\operatorname{si}}}_{b,G_n} = {{\mathfrak}{D}^{\operatorname{si}}}_{{\left \lceil{n/2}\right \rceil }^-,G_n}.$$ It follows that ${\operatorname{Dgm}}^{{\mathfrak}{D}}_1(G_n) = {\left\{\left(1,{\left \lceil{\tfrac{n}{2}}\right \rceil }\right)\right\}}$. The notation in the last equality means that ${{\mathfrak}{D}^{\operatorname{si}}}_{b,G_n} = {{\mathfrak}{D}^{\operatorname{si}}}_{{\delta},G_n}$ for all ${\delta}\in [b,b+1)$, where $b+1 = {\left \lceil{n/2}\right \rceil }$. In the more general case, let $l \in {\mathbb{N}}$ and let $M_l$ be as in the statement of the result. Suppose first that $M_l$ is empty. Then by Proposition \[prop:dowker-cech-cplx\] and Theorem \[thm:cech-S1\], we know that ${{\mathfrak}{D}^{\operatorname{si}}}_{k,G_n}$ has the homotopy type of a wedge of even-dimensional spheres or an odd-dimensional sphere of dimension strictly different from $(2l+1)$, for any choice of integer $k$. Thus ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l+1}(G_n)$ is trivial. Next suppose $M_l$ is nonempty. By another application of Proposition \[prop:dowker-cech-cplx\] and Theorem \[thm:cech-S1\], we know that ${{\mathfrak}{D}^{\operatorname{si}}}_{k,G_n} = {\operatorname{\mathbf{\check{C}}}}({\mathbb}{X}_n,\tfrac{k}{2n}) \simeq S^{2l+1} $ for integers $\tfrac{nl}{l+1} < k < \tfrac{n(l+1)}{l+2}$. Write $a_l:=\min{\left\{m\in M_l\right\}}$ and $b_l:=\max{\left\{m\in M_l\right\}}$. Then by Theorem \[thm:cyc-cech-main\], we know that each inclusion map in the following chain is a homotopy equivalence: $${{\mathfrak}{D}^{\operatorname{si}}}_{a_l,G_n} \subseteq \ldots \subseteq {{\mathfrak}{D}^{\operatorname{si}}}_{b_l,G_n} = {{\mathfrak}{D}^{\operatorname{si}}}_{{\left \lceil{n(l+1)/(l+2)}\right \rceil }^-,G_n}.$$ It follows that ${\operatorname{Dgm}}^{{\mathfrak}{D}}_{2l+1}(G_n) = {\left\{\left(a_l,{\left \lceil{\tfrac{n(l+1)}{l+2}}\right \rceil }\right)\right\}}$. It remains to provide a proof of Theorem \[thm:cyc-cech-main\]. For this, we need some additional machinery. #### Cyclic maps and winding fractions We introduce some more terms from [@adamaszek2015vietoris], but for efficiency, we try to minimize the scope of the definitions to only what is needed for our purpose. Recall that we write $S^1$ to denote the circle with unit circumference. So any $x\in S^1$ can be naturally identified with a point in $[0,1)$. We fix a choice of $0\in S^1$, and for any $x,x' \in S^1$, the length of a clockwise arc from $x$ to $x'$ is denoted by $\overrightarrow{d_{S^1}}(x,x')$. Then, for any finite subset $X\subseteq S^1$ and any $r \in (0,1/2)$, the directed Vietoris-Rips graph ${\overrightarrow{\operatorname{VR}}}(X,r)$ is defined to be the graph with vertex set $X$ and edge set $\{(x,x') : 0 < \overrightarrow{d_{S^1}}(x,x') < r\}$. Next, let $\overrightarrow{G}$ be a Vietoris-Rips graph such that the vertices are enumerated as $x_0,x_1,\ldots, x_{n-1}$, according to the clockwise order in which they appear. A cyclic map between $\overrightarrow{G}$ and a Vietoris-Rips graph $\overrightarrow{H}$ is a map of vertices $f$ such that for each edge $(x,x') \in \overrightarrow{G}$, we have either $f(x)=f(x')$, or $(f(x),f(x')) \in \overrightarrow{H}$, and $\sum_{i=0}^{n-1}\overrightarrow{d_{S^1}}(f(x_i),f(x_{i+1})) = 1$. Here $x_n:=x_0$. Next, the winding fraction of a Vietoris-Rips graph $\overrightarrow{G}$ with vertex set $V(\overrightarrow{G})$ is defined to be the infimum of numbers $\tfrac{k}{n}$ such that there is an order-preserving map $V(\overrightarrow{G}) {\rightarrow}{\mathbb{Z}}/n{\mathbb{Z}}$ such that each edge is mapped to a pair of numbers at most $k$ apart. A key property of the winding fraction, denoted ${\operatorname{wf}}$, is that if there is a cyclic map between Vietoris-Rips graphs $\overrightarrow{G} {\rightarrow}\overrightarrow{H}$, then ${\operatorname{wf}}(\overrightarrow{G}) \leq {\operatorname{wf}}(\overrightarrow{H})$. \[thm:vrcirc-main\] Let $X\subseteq S^1$ be a finite set and let $0 < r < \tfrac{1}{2}$. Then, $${\operatorname{\mathbf{VR}}}(X,r) \simeq \begin{cases} S^{2l+1} &: \tfrac{l}{2l+1} < {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(X,r)) < \tfrac{l+1}{2l+3} \text{ for some } l\in {\mathbb{Z}}_+,\\ \bigvee^j S^{2l} &: {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(X,r)) = \tfrac{l}{2l+1}, \text{ for some } j\in {\mathbb{N}}. \end{cases}$$ Next let $X' \subseteq S^1$ be another finite set, and let $r \leq r' < \tfrac{1}{2}$. Suppose $f:{\overrightarrow{\operatorname{VR}}}(X,r) {\rightarrow}{\overrightarrow{\operatorname{VR}}}(X',r')$ is a cyclic map between Vietoris-Rips graphs and $\tfrac{l}{2l+1} < {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(X,r)) \leq {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(X',r')) < \tfrac{l+1}{2l+3}$. Then $f$ induces a homotopy equivalence between ${\operatorname{\mathbf{VR}}}(X,r)$ and ${\operatorname{\mathbf{VR}}}(X',r')$. We now have the ingredients for a proof of Theorem \[thm:cyc-cech-main\]. Since the maps $\pi_r$ and $\pi_{r'}$ induce homotopy equivalences, it follows that $${\operatorname{\mathbf{VR}}}(T_r(X),\tfrac{2r}{1+2r}) \simeq {\operatorname{\mathbf{VR}}}(T_{r'}(X),\tfrac{2r'}{1+2r'}).$$ By the characterization result in Theorem \[thm:vrcirc-main\], we know that there exists $l \in {\mathbb{Z}}_+$ such that $$\tfrac{l}{2l+1} < {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(T_r(X),\tfrac{2r}{1+2r})) \leq {\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(T_{r'}(X),\tfrac{2r'}{1+2r'})) < \tfrac{l+1}{2l+3}.$$ The map $\eta$ in Theorem \[thm:vr-cech-cd\] appears in [@adamaszek2015vietoris Proposition 9.5] through an explicit construction. Moreover, it is shown that $\eta$ induces a cyclic map ${\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(T_r(X),\tfrac{2r}{1+2r})) {\rightarrow}{\operatorname{wf}}({\overrightarrow{\operatorname{VR}}}(T_{r'}(X),\tfrac{2r'}{1+2r'}))$. Thus by Theorem \[thm:vrcirc-main\], $\eta$ induces a homotopy equivalence between ${\operatorname{\mathbf{VR}}}(T_r(X),\tfrac{2r}{1+2r})$ and ${\operatorname{\mathbf{VR}}}(T_{r'}(X),\tfrac{2r'}{1+2r'})$. Finally, the commutativity of the diagram in Theorem \[thm:vr-cech-cd\] shows that the inclusion ${\operatorname{\mathbf{\check{C}}}}(X,r) \subseteq {\operatorname{\mathbf{\check{C}}}}(X,r')$ induces a homotopy equivalence. The analogue of Theorem \[thm:cyc-cech-main\] for Čech complexes appears as Proposition 4.9 of [@adamaszek2015vietoris] for Vietoris–Rips complexes. We prove Theorem \[thm:cyc-cech-main\] by connecting Čech and Vietoris-Rips complexes using Proposition 9.5 of [@adamaszek2015vietoris]. However, as remarked in §9 of [@adamaszek2015vietoris], one could prove Theorem \[thm:cyc-cech-main\] directly using a parallel theory of winding fractions for Čech complexes. [^1]: A thread with ideas towards the proof of Theorem \[thm:dowker-strong\] was discussed in [@Nlab-dowker last accessed 4.24.2017], but the proposed strategy was incomplete. We have inserted an addendum in [@Nlab-dowker] proposing a complete proof with a slightly different construction.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $X$ be a separable Banach space, $Y$ a Banach space and $f: X \to Y$ an arbitrary mapping. Then the following implication holds at each point $x \in X$ except a $\sigma$-directionally porous set: If the one-sided Hadamard directional derivative $f''_{H+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span is dense in $X$, then $f$ is Hadamard differentiable at $x$. This theorem improves and generalizes a recent result of A.D. Ioffe, in which the linear span of $S_x$ equals $X$ and $Y = {{\mathbb R}}$. An analogous theorem, in which $f$ is pointwise Lipschitz, and which deals with the usual one-sided derivatives and G\^ ateaux differentiability is also proved. It generalizes a result of D. Preiss and the author, in which $f$ is supposed to be Lipschitz.' address: 'Charles University, Faculty of Mathematics and Physics, Sokolovská 83, 186 75 Praha 8-Karlín, Czech Republic' author: - Luděk Zajíček title: 'G\^ ateaux and Hadamard differentiability via directional differentiability' --- [^1] Introduction ============ The following result ([@PZ Theorem 5]) was used in the proof of the till now strongest version of Rademacher’s theorem on G\^ ateaux differentiability of Lipschitz mappings on a separable Banach space. \[th\_PZ\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a Lipschitz mapping. Then there exists a $\sigma$-directionally porous set $A \subset G$ such that for every $x \in G \setminus A$ the set $U_x$ of those directions $u \in X$ in which the one-sided derivative $f'_+(x,u)$ exists is a closed linear subspace of $X$. Moreover, the mapping $u \mapsto f'_+(x,u)$ is linear on $U_x$. An immediate consequence of this result is the following. \[cor\_PZ\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a Lipschitz mapping. Then the following implication holds at each point $x \in G$ except a $\sigma$-directionally porous set: If the one-sided directional derivative $f'_{+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span is dense in $X$, then $f$ is G\^ ateaux differentiable at $x$. Note that each $\sigma$-directionally porous subset of a separable Banach spaces $X$ is not only a first category set, but it is also “measure null”: it is Aronszajn (=Gauss) null, and so also Haar null, (see [@BL p. 164 and Chap. 6]) and also $\Gamma$-null (see ). It is an easy well-known fact that if $f:X \to Y$ is a Lipschitz mapping between Banach spaces, then the Hadamard (one-sided) directional derivatives coincide with the usual (one-sided) directional derivatives and the Hadamard derivative coincides with the G\^ ateaux derivative. A.D. Ioffe in [@Io] recently observed that some known results dealing with the usual directional derivatives (and G\^ ateaux differentiability) of [Lipschitz]{} functions can be generalized to results dealing with the Hadamard directional derivatives (and Hadamard differentiability) of [arbitrary]{} functions. (Note that this Ioffe’s idea was followed in [@Za1].) Ioffe’s [@Io Theorem 3.7(b)] can be reformulated (see Remark \[dom\] below) in the following way. \[th\_I\] Let $X$ be a separable Banach space, $G \subset X$ an open set, and $f: G \to {{\mathbb R}}$ an arbitrary function. Then the following implication holds at each point $x \in G$ except a $\sigma$-directionally porous set: If the one-sided Hadamard directional derivative $f'_{H+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span equals to $X$, then $f$ is Hadamard differentiable at $x$. So, Theorem I is a partial generalization of Corollary PZ (since in Theorem I is a stronger assumption on $S_x$ and $Y= {{\mathbb R}}$). We will prove (see Corollary \[hadcom\] below) that the corresponding full generalization of Corollary PZ holds. Moreover, Theorem \[had\] below on Hadamard derivatives generalizes Theorem PZ. Further, we generalize Theorem PZ in another direction showing that, in this theorem, it is sufficient to suppose that $f$ is pointwise Lipschitz (see Corollary \[plux\] below). Note that the methods of proofs from [@PZ] and [@Io] cannot be easily used in the case when $f: X \to Y$ is not continuous and $Y$ is nonseparable, since then $f(X)$ need not be separable. From this reason we use an alternative method based on a small trick in the proof of Lemma \[rozd\] (which shows that $f'_+(x,w_x)$ and $f'(y_k, w_{y_k})$ are “automatically” close to one another). The main results are proved in Section 3. In Section 2 we recall basic definitions and known results we need. Preliminaries ============= In the following, by a Banach space we mean a real Banach space. The symbol $B(x,r)$ will denote (in a metric space) the open ball with center $x$ and radius $r$. If $X$ is a Banach space, we set $S_X:= \{x \in X: \\|x\\|=1\}$. Further, if $x \in X$, $v \in S_X$ and $\delta >0$, then we define the open [cone $C(x,v,\delta)$]{} as the set of all $y \neq x$ for which $\\|v - \frac{y-x}{\\|y-x\\|}\\| < \delta$. Let $X$, $Y$ be Banach spaces, $G \subset X$ an open set, and $f:G \to Y$ a mapping. We say that [$f$ is Lipschitz at $x \in G$]{} if $\limsup_{y \to x} \frac{\\|f(y)-f(x)\\|}{\\|y-x\\|} < \infty$. We say that $f$ is [pointwise Lipschitz]{} (on $G$) if $f$ is Lipschitz at all points of $G$. The directional and one-sided directional derivatives of $f$ at $x\in G$ in the direction $v\in X$ are defined respectively by $$f'(x,v) := \lim_{t \to 0} \frac{f(x+tv)-f(x)}{t}\ \ \text{and}\ \ f'_+(x,v) := \lim_{t \to 0+} \frac{f(x+tv)-f(x)}{t}.$$ The [Hadamard directional and one-sided directional derivatives]{} of $f$ at $x\in G$ in the direction $v\in X$ are defined respectively by $$f'_H(x,v) := \lim_{z \to v, t \to 0} \frac{f(x+tz)-f(x)}{t}\ \ \text{and}\ \ f'_{H+}(x,v) := \lim_{z \to v, t \to 0+} \frac{f(x+tz)-f(x)}{t}.$$ The following facts are well-known and easy to prove. \[f\] Let $X$, $Y$ be Banach spaces, $G \subset X$ an open set, $x \in G$, $v \in X$, and $f:G \to Y$ a mapping. Then the following assertions hold. 1. The derivative $f'(x,v)$ (resp. $f'_H(x,v)$) exists if and only if $f'_+(x,-v) = -f'_+(x,v)$ (resp. $f'_{H+}(x,-v) = - f'_{H+}(x,v)$). 2. If $f'_H(x,v)$ (resp. $f'_{H+}(x,v)$) exists, then $f'(x,v)=f'_H(x,v)$ (resp. $f'_+(x,v)=f'_{H+}(x,v)$). 3. If $f$ is locally Lipschitz on $G$, then $f'(x,v)=f'_H(x,v)$ (resp. $f'_+(x,v)=f'_{H+}(x,v)$) whenever one of these two derivatives exists. 4. If $f'_+(x,v)$ (resp. $f'_{H+}(x,v)$) exists and $t>0$, then $f'_+(x,tv)= t f'_+(x,v)$ (resp. $f'_{H+}(x,tv)= t f'_{H+}(x,v)$). 5. If $f$ is Lipschitz at $x$ with $\limsup_{y \to x} \frac{\\|f(y)-f(x)\\|}{\\|y-x\\|}\leq K < \infty$ and $f'_+(x,v)$ exists, then $\\|f'_+(x,v)\\| \leq K \\|v\\|$. 6. If $V_x$ is the set of all $u \in X$, for which $f'_{H+}(x,u)$ exists, then the mapping $u \mapsto f'_{H+}(x,u)$ is continuous on $V_x$. 7. If $v \in S_X$ and $f'_{H+}(x,v)$ exists, then there exists a cone $C=C(x,v,\delta)$ such that $\limsup_{y \to x, y \in C} \frac{\\|f(y)-f(x)\\|}{\\|y-x\\|} < \infty$. We will need the following easy fact (see [@Za2 Lemma 2.3]). \[hnli\] Let $X$ be a Banach space, $Y$ a Banach space, $G \subset X$ an open set, $a \in G$, and $f:G \to Y$ a mapping. Then the following are equivalent. 1. $f'_{H+}(a,0)$ exists, 2. $f'_{H}(a,0)$ exists, 3. $f'_{H}(a,0)=0$, 4. $f$ is Lipschitz at $a$. The usual modern definition of the Hadamard derivative is the following: A continuous linear operator $L: X \to Y$ is said to be the [Hadamard derivative]{} of $f$ at a point $x \in X$ if $$\lim_{t \to 0} \frac{f(x+tv)-f(x)}{t} = L(v)\ \ \ \text{for each}\ \ \ v \in X$$ and the limit is uniform with respect to $v \in C$, whenever $C\subset X$ is a compact set. In this case we set $f'_H(x) := L$. The following fact is well-known (see [@Sha]): \[eh\] Let $X$, $Y$ be Banach spaces, $\emptyset \neq G \subset X$ an open set, $x \in G$, $f: G \to Y$ a mapping and $L: X \to Y$ a continuous linear operator. Then the following conditions are equivalent: 1. $f'_H(x) = L$, 2. $f'_H(x,v)= L(v)$ for each $v \in X$, 3. if ${\varphi}: [0,1] \to X$ is such that ${\varphi}(0)=x$ and ${\varphi}_+'(0)$ exists, then $(f \circ {\varphi})_+'(0)= L({\varphi}_+'(0))$. Let $X$ be a Banach space. We say that $A \subset X$ is [directionally porous at a point $x \in X$]{}, if there exist $0 \neq v \in X$, $p>0$ and a sequence $t_n \to 0$ of positive real numbers such that $B(x+ t_n v, p t_n) \cap A = \emptyset$. (In this case we say that $A$ is [porous at $x$ in the direction $v$]{}.) We say that $A \subset X$ is [directionally porous]{} if $A$ is directionally porous at each point $x \in A$. We say that $A \subset X$ is [$\sigma$-directionally porous]{} if it is a countable union of directionally porous sets. We will need the obvious fact that $A$ is not porous at $x$ in the direction $v\neq 0$ if and only if $$\label{nepo} \text{for each $\omega>0$ there exists $\delta>0$ such that $B(x+tv, \omega t) \cap A \neq \emptyset$ for $0<t<\delta.$}$$ It is easy to see that if, for some $v \in S_X$, $\delta>0$, $r>0$, $$\label{uhpo} \text{$C(x,v,\delta) \cap B(x,r) \cap A = \emptyset$, then $A$ is porous at $x$ in direction $v$.}$$ Moreover, we will need the following results from [@Za2]. \[smerdh\] $($[@Za2 Proposition 3.2]$)$ Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a mapping. Let $M$ be the set of all $x \in G$ at which $f$ is Lipschitz and there exists $v\in X$ such that $f'_+(x,v)$ exists but $f'_{H+}(x,v)$ does not exist. Then $M$ is $\sigma$-directionally porous. \[cll\] ([@Za2 Proposition 3.1]) Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f:G \to Y$ a mapping. Let $A$ be the set of all points $x \in G$ for which there exists a cone $C= C(x,v,\delta)$ such that $\limsup_{y \to x, y \in C} \frac{\\|f(y)-f(x)\\|}{\\|y-x\\|} < \infty$ and $f$ is not Lipschitz at $x$. Then $A$ is a $\sigma$-directionally porous set. Main results ============ \[rozd\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a mapping. Let $A$ be the set of all $x \in G$ at which $f$ is Lipschitz, and for which there exist $v, w \in X$ such that $f'_+(x,v)$ and $f'_+(x,w)$ exist but either $f'_+(x, v-w)$ does not exist or $f_+'(x, v-w) \neq f'_+(x,v)- f'_+(x,w)$. Then $A$ is a $\sigma$-directionally porous set. Let $M$ be the $\sigma$-directionally porous set from Proposition \[smerdh\]. It is sufficient to prove that $A^:= A \setminus M$ is $\sigma$-directionally porous. To each $x \in A$, choose a corresponding pair $v=v_x$, $w=w_x$. Let $\{d_j: j \in {{\mathbb N}}\}$ be a dense subset of $X$. Now consider an arbitrary $x \in A^$. We can choose $n_x \in {{\mathbb N}}$ such that $$\label{enix} \\|f(y)-f(x)\\| \leq n_x \, \\|y-x\\|\ \ \text{whenever}\ \ \\|y-x\\| \leq \frac{1}{n_x}.$$ Further, we can choose $p_x \in {{\mathbb N}}$ and a sequence $1> t_i^x \searrow 0$ such that, for each $i \in {{\mathbb N}}$, $$\label{poti} \left \\| \frac{f(x+t_i^x(v_x-w_x)) - f(x)}{t_i^x} - (f'_+(x,v_x) - f'_+(x, w_x))\right\\| > \frac{6}{p_x}.$$ Since $x \notin M$, we have $f'_+(x,w_x) = f'_{H+}(x,w_x)$ and so we can choose $k_x \in {{\mathbb N}}$ such that $$\label{powx} \left \\| \frac{f(x+tz) - f(x)}{t} - f'_+(x, w_x)\right\\|< \frac{1}{p_x}\ \ \text{whenever}\ \ \\|z-w_x\\|< \frac{1}{k_x}\ \ \text{and}\ \ 0< t < \frac{1}{k_x},$$ and $$\label{povx} \left \\| \frac{f(x+tv_x) - f(x)}{t} - f'_+(x, v_x)\right\\|< \frac{1}{p_x}\ \ \text{whenever}\ \ 0< t < \frac{1}{k_x}.$$ Finally choose $j_x \in {{\mathbb N}}$ such that $\\|w_x - d_{j_x}\\| < (4 k_x)^{-1}$. Now, for natural numbers $n, p, k, j$ denote by $A^_{n,p,k,j}$ the set of all $x \in A^$ for which $n_x=n$, $p_x=p$, $k_x=k$, $j_x=j$. It is clearly sufficient to prove that, for any fixed quadruple $n,p,k,j$, the set $S:= A^_{n,p,k,j}$ is directionally porous. So fix an arbitrary $x \in S$. Denote $t_i := t_i^x$ and choose $\eta$ such that $$\label{eta} 0 < \eta < \min( (pn)^{-1}, (2k)^{-1}).$$ We will prove that $S$ is porous at $x$ in the direction $v_x-w_x$. To show this, it is sufficient to prove that there exists $i_0 \in {{\mathbb N}}$ such that $$\label{prpr} S \cap B(x+t_i(v_x-w_x),\eta t_i) = \emptyset\ \ \text{whenever }\ \ i \geq i_0.$$ To this end, consider $i \in {{\mathbb N}}$ and $y_i \in S \cap B(x+t_i(v_x-w_x),\eta t_i)$. First observe that $\\|(x + t_i(v_x-w_x))-y_i\\| < \eta t_i < \eta < (pn)^{-1}$ and so (with $x:=y_i$ and $y:= x + t_i(v_x-w_x)$) implies $$\label{yibl} \\|f(x + t_i(v_x-w_x))-f(y_i)\\| \leq n \eta t_i < \frac{t_i}{p}.$$ Now we will show (and this is the main trick of the proof) that, if $i$ is sufficiently large, then the difference of $f'_+(x, w_x)$ and $f'_+(y_i, w_{y_i})$ is “small” (see ). To this end, set $t^:= (2k)^{-1}$ and $\xi: = x+ t^* w_x$. By we immediately obtain $$\label{xxi} \left \\| \frac{f(\xi) - f(x)}{t^} - f'_+(x, w_x)\right\\|< \frac{1}{p}.$$ Further set $z^ := (t^)^{-1} (\xi-y_i)$. Then $$\\|z^- w_{y_i}\\| \leq \\|z^-w_x\\| + \\|w_x- d_j\\|+ \\|w_{y_i}-d_j\\| \leq \\|z^-w_x\\| + \frac{1}{2k}$$ and $$\\|z^-w_x\\| = \\|(t^)^{-1} (\xi-y_i) - (t^)^{-1}(\xi -x)\\|= \\|(t^)^{-1} (x-y_i)\\|\leq 2k t_i (\\|v_x - w_x\\| + \eta).$$ So there exists $i_1 \in {{\mathbb N}}$ such that $i \geq i_1$ implies $ \\|z^- w_{y_i}\\| < 1/k$, and consequently also (by with $x: =y_i \in S$) $$\label{yixi} \left \\| \frac{f(\xi) - f(y_i)}{t^} - f'_+(y_i, w_{y_i})\right\\| = \left \\| \frac{f(y_i +t^z^) - f(y_i)}{t^} - f'_+(y_i, w_{y_i}) \right\\|< \frac{1}{p}.$$ Further there exists $i_1 <i_2 \in {{\mathbb N}}$ such that $i \geq i_2$ implies $$\\|x-y_i\\|\leq t_i (\\|v_x - w_x\\| + \eta) < (2knp)^{-1} < n^{-1}$$ and consequently also (by ) $$\left\\| \frac{f(\xi) - f(x)}{t^} - \frac{f(\xi) - f(y_i)}{t^}\right\\| = 2k \\|f(y_i) - f(x)\\| \leq 2kn \\|y_i-x\\| < \frac{1}{p}.$$ Using this together with and , we obtain that $i \geq i_2$ implies $$\label{auto} \\|f'_+(x, w_x) - f'_+(y_i, w_{y_i})\\| < \frac{3}{p}.$$ Denote $a_i:= x + t_i v_x$. There exists $i_0>i_2$ such that $i \geq i_0$ implies $t_i < 1/k$, and consequently also (by ) $$\label{xzi} \left\\| \frac{f(a_i)-f(x)}{t_i}- f'_+(x,v_x)\right\\| < \frac{1}{p}.$$ Set $z_i:= (t_i)^{-1} ( a_i -y_i)= (t_i)^{-1} (x+ t_i v_x - y_i)$. Then $$\\|z_i- w_{y_i}\\| \leq \\|z_i-w_x\\| + \\|w_x- d_j\\|+ \\|w_{y_i}-d_j\\| \leq \\|z_i-w_x\\| + \frac{1}{2k}$$ and $$\\|z_i-w_{x}\\| = \left\\|\frac{x+t_i v_x -y_i}{t_i} -w_x\right\\| = \left\\|\frac{x+t_i(v_x-w_x) -y_i)}{t_i}\right\\| < \eta < \frac{1}{2k}.$$ So $ \\|z_i- w_{y_i}\\| < 1/k$ and $i \geq i_0$ implies $t_i < 1/k$, and consequently (by with $x: =y_i)$) $$\label{ziyi} \left \\| \frac{f(a_i) - f(y_i)}{t_i} - f'_+(y_i, w_{y_i})\right\\| = \left \\| \frac{f(y_i +t_iz_i) - f(y_i)}{t_i} - f'_+(y_i, w_{y_i}) \right\\|< \frac{1}{p}.$$ Thus, if $i \geq i_0$, then , , and imply $$\begin{gathered} \left \\| \frac{f(x + t_i(v_x-w_x)) - f(x)}{t_i} - (f'_+(x,v_x) - f'_+(x, w_x))\right\\|\\ \leq \left \\| \frac{f(y_i) - f(x)}{t_i} - (f'_+(x,v_x) - f'_+(x, w_x))\right\\| + \frac{1}{p} \\ \leq \left \\| \frac{f(y_i) - f(x)}{t_i} - (f'_+(x,v_x) - f'_+(y_i, w_{y_i}))\right\\| + \frac{4}{p} \\ = \left \\| \left(\frac{f(a_i) - f(x)}{t_i} - f'_+(x,v_x)\right)+ \left(f'_+(y_i, w_{y_i}) - \frac{f(a_i) - f(y_i)}{t_i}\right) \right\\| + \frac{4}{p} < \frac{6}{p}, \end{gathered}$$ which contradicts . So we have proved . \[uzav\] Let $X$, $Y$ be Banach spaces, $G \subset X$ an open set, $f: G \to Y$ a mapping. For each $x \in G$ denote by $U_x$ the set of all $v \in X$ such that $f'_+(x,v)$ exists. Then the set $B$ of all $x \in G$ such that $f$ is Lipschitz at $x$ and $U_x$ is not closed is $\sigma$-directionally porous. For each $k \in {{\mathbb N}}$, set $$B_k:= \{x \in B:\ \\|f(y)-f(x)\\| < k \\|y-x\\|\ \ \text{ whenever}\ \ \\|y-x\\| <1/k\}.$$ It is clearly sufficient to prove that each $B_k$ is directionally porous. So suppose that $k \in {{\mathbb N}}$ and $x \in B_k$ are given. Since $U_x$ is not closed, it is easy to see (using Fact \[f\](iv)) that we can find $v \in S_X \setminus U_x$ and a sequence $v_n \to v$ with $v_n \in S_X \cap U_x$. We will show that $$\label{vesv} \text{$B_k$ is porous at $x$ in the direction $v$.}$$ Suppose, to the contrary, that does not hold. We will obtain a contradiction by proving that $v \in U_x$, i.e., that $\lim_{t\to 0+} t^{-1} (f(x+tv)- f(x))$ exists. Since $Y$ is a complete space, it is sufficient to prove that for each ${\varepsilon}>0$ there exists $\delta>0$ such that $$\label{bc} \left\\|\frac{f(x+t_1v)-f(x)}{t_1} - \frac{f(x+t_2v)-f(x)}{t_2}\right\\| < {\varepsilon}\ \ \text{whenever}\ \ 0<t_1<t_2<\delta.$$ So, let $1>{\varepsilon}>0$ be given. Since does not hold, by we can find $\delta_1 >0$ such that $$\label{hust} \text{$B(x+tv, {\varepsilon}t/12k) \cap B_k \neq \emptyset$ whenever $0< t < \delta_1$.}$$ Now choose $n \in {{\mathbb N}}$ such that $\\|v - v_n\\| < {\varepsilon}/12k$. Since $v_n \in U_x$, we can find $\delta_2 >0$ such that $$\label{bc2} \left\\|\frac{f(x+t_1v_n)-f(x)}{t_1} - \frac{f(x+t_2v_n)-f(x)}{t_2}\right\\| < \frac{{\varepsilon}}{3}\ \ \text{whenever}\ \ 0<t_1<t_2<\delta_2.$$ Put $\delta: = \min(\delta_1, \delta_2)$. It is sufficient to prove . To this end, let arbitrary numbers $0<t_1<t_2<\delta$ be given. By , we have $$\label{bc3} \left\\|\frac{f(x+t_1v_n)-f(x)}{t_1} - \frac{f(x+t_2v_n)-f(x)}{t_2}\right\\| < \frac{{\varepsilon}}{3}.$$ By , we can choose points $y_1 \in B(x+t_1v, {\varepsilon}t_1/12 k) \cap B_k$ and $y_2 \in B(x+t_2v, {\varepsilon}t_2/12 k) \cap B_k$. Observe that $$\\|y_1 - (x+t_1v)\\|< t_1 {\varepsilon}/12k<1/k, \ \ \ \\|(x+t_1v)-(x+t_1v_n)\\|= t_1\\|v-v_n\\| < t_1 {\varepsilon}/12 k,$$ and therefore $\\|y_1 - (x+t_1v_n)\\|< t_1 {\varepsilon}/6 k< 1/k$. Since $y_1 \in B_k$, we obtain $$\\|f(y_1)- f(x+t_1v)\\| < k t_1 {\varepsilon}/12 k = t_1{\varepsilon}/12, \ \ \ \\|f(y_1)- f(x+t_1v_n)\\| < k t_1 {\varepsilon}/6k = t_1{\varepsilon}/6.$$ Consequently $$\label{odt1} \\|f(x+t_1v)- f(x+t_1v_n)\\| < t_1{\varepsilon}/3.$$ By the same way we obtain $$\label{odt2} \\|f(x+t_2v)- f(x+t_2v_n)\\| < t_2{\varepsilon}/3.$$ So, for $i\in \{1,2\}$ we obtain $$\label{vvn} \left\\|\frac{f(x+t_iv_n)-f(x)}{t_i} - \frac{f(x+t_iv)-f(x)}{t_i}\right\\| < \frac{{\varepsilon}}{3}.$$ Using and we obtain $$\left\\|\frac{f(x+t_1v)-f(x)}{t_1} - \frac{f(x+t_2v)-f(x)}{t_2}\right\\| < {\varepsilon}$$ which completes the proof of . \[obyc\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a mapping. Then there exists a $\sigma$-directionally porous set $C \subset G$ such that if $x \in G \setminus C$ and $f$ is Lipschitz at $x$, then the set $U_x$ of those directions $u \in X$ in which the one-sided derivative $f'_+(x,u)$ exists is a closed linear subspace of $X$. Moreover, the mapping $u \mapsto f'_+(x,u)$ is linear on $U_x$. Let $A$ be the set from Lemma \[rozd\] and $B$ be the set from Lemma \[uzav\]. We will show that it is sufficient to set $C:= A \cup B$. Obviously, $C$ is $\sigma$-directionally porous. Now consider an arbitrary $x \in G \setminus C = G \setminus (A\cup B)$ at which $f$ is Lipschitz. It is obvious that $0 \in U_x$, $f'_+(x,0)=0$ and (see Fact \[f\](iv)) $$\label{poho} \text{if $u \in U_x$ and $t \geq 0$, then $tu \in U_x$ and $f'_+(x,tu) = t f'_+(x,u)$.}$$ If $w \in U_x$, set $v:=0$ and obtain (since $x \notin A$) $$\label{opac} f'_+(x,-w) = f'_+(x,0) - f'_+(x,w) = - f'_+(x,w).$$ Obviously, and imply that $$\label{homo} \text{if $u \in U_x$ and $t \in {{\mathbb R}}$, then $tu \in U_x$ and $f'_+(x,tu) = t f'_+(x,u)$.}$$ Further consider arbitrary vectors $v, u \in U_x$. Setting $w:= -u$ and using $x \notin A$ and , we obtain $$\label{adit} f'_+(x, v+u) = f'_+(x,v-w)= f'_+(x,v) - f'_+(x,w) = f'_+(x,v) + f'_+(x,u).$$ Thus and imply that $U_x$ is a linear subspace of $X$ and the mapping $u \mapsto f'_+(x,u)$ is linear on $U_x$. Since $x \notin B$, we have that $U_x$ is closed. Theorem \[obyc\] and Fact \[f\](i),(v) immediately imply the following corollaries. \[liga\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ an arbitrary mapping. Then the following implication holds at each point $x \in G$ except a $\sigma$-directionally porous set: If $f$ is Lipschitz at $x$ and the one-sided directional derivative $f'_{+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span is dense in $X$, then $f$ is G\^ ateaux differentiable at $x$. \[plux\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a pointwise Lipschitz mapping. Then there exists a $\sigma$-directionally porous set $A \subset G$ such that for every $x \in G \setminus A$ the set $U_x$ of those directions $u \in X$ in which the one-sided derivative $f'_+(x,u)$ exists is a closed linear subspace of $X$. Moreover, the mapping $u \mapsto f'_+(x,u)$ is linear on $U_x$. \[plga\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a pointwise Lipschitz mapping. Then the following implication holds at each point $x \in G$ except a $\sigma$-directionally porous set: If the one-sided directional derivative $f'_{+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span is dense in $X$, then $f$ is G\^ ateaux differentiable at $x$. Our main result on Hadamard derivatives is the following. \[had\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ a mapping. Then there exists a $\sigma$-directionally porous set $D \subset G$ such that if $x \in G \setminus D$, then either the set $V_x$ of those directions $u \in X$ in which the one-sided Hadamard derivative $f'_{H+}(x,u)$ exists is an empty set or $V_x$ is a closed linear subspace of $X$ and the mapping $u \mapsto f'_{H+}(x,u)$ is linear on $V_x$. Let $C$ be the set from Theorem \[obyc\], $M$ the set from Proposition \[smerdh\] and $A$ the set from Proposition \[cll\]. Set $D:= C \cup M \cup A$. Then $D \subset G$ is a $\sigma$-directionally porous set. Now suppose that a point $x \in G \setminus D$ is given and $V_x \neq \emptyset$. Choose $w \in V_x$. If $w=0$, then we obtain that $f$ is Lipschitz at $x$ by Lemma \[hnli\]. If $w \neq 0$, then we set $v: = w/\\|w\\|$ and apply Fact \[f\](iv),(vii). We obtain that there exists a cone $C:= C(x,v, \delta)$ such that $\limsup_{y \to x, y \in C} \frac{\\|f(y)-f(x)\\|}{\\|y-x\\|} < \infty$. Since $x \notin A$, we obtain that $f$ is Lipschitz at $x$ also in this case. Since $x \notin M$, we obtain that $f'_+(x,u)=f'_{H+}(x,u)$ whenever one of these two derivatives exists. So, since $f$ is Lipschitz at $x$ and $x \notin C$, we obtain our assertion. \[equi\] If we have in our disposal Proposition \[smerdh\], Proposition \[cll\] and Lemma \[hnli\], then we can consider our two main theorems as “equivalent”. Indeed, we have already shown how Theorem \[obyc\] (and the above mentioned propositions) easily implies Theorem \[had\]. Further, Theorem \[had\] and Proposition \[cll\] almost immediately imply Theorem \[obyc\]. Using Theorem \[had\], Fact \[f\](i),(vi) and Lemma \[eh\], we immediately obtain the following result, which improves and generalizes [@Io Theorem 3.7(b)] (see Remark \[dom\] below). \[hadcom\] Let $X$ be a separable Banach space, $Y$ a Banach space, $G \subset X$ an open set, and $f: G \to Y$ an arbitrary mapping. Then the following implication holds at each point $x \in G$ except a $\sigma$-directionally porous set: If the one-sided Hadamard derivative $f'_{H+}(x,u)$ exists in all directions $u$ from a set $S_x \subset X$ whose linear span is dense in $X$, then $f$ is Hadamard differentiable at $x$. \[dom\] Ioffe’s [@Io Theorem 3.7(b)] works with $f: X \to [-\infty,\infty]$ and ${\operatorname{dom}}(f):= \{ x \in X:\ \|f(x)\| < \infty\}$. Denote by $B$ the set of all $x \in {\operatorname{dom}}(f) \setminus {\operatorname{int}}({\operatorname{dom}}(f))$ for which there exists $0 \neq v \in X$ such that $f'_{H+}(x,v) := \lim_{z \to v, t \to 0+} \frac{f(x+tz)-f(x)}{t}$ exists and is finite. If $x \in B$, then (cf. Fact \[f\](vii)) there exists a cone $C=C(x,v,\delta)$ such that $\limsup_{y \to x, y \in C} \frac{\|f(y)-f(x)\|}{\\|y-x\\|} < \infty$, which implies that $B \subset X \setminus {\operatorname{int}}({\operatorname{dom}}(f))$ is directionally porous at $x$ (see ). So $B$ is a directionally porous set. This simple observation shows that Corollary \[hadcom\] remains true, if we work with $f: X \to [-\infty,\infty]$ ( and $f'_{H+}(x,u)$ is, by definition, finite). In other words, the assumption of [@Io Theorem 3.7(b)] that the linear span of $S_x$ equals to $X$ can be relaxed to the assumption that the linear span of $S_x$ is dense in $X$. [WWW]{} Y. Benyamini, J. Lindenstrauss, [Geometric Nonlinear Functional Analysis, Vol. 1,]{} Colloquium Publications 48, American Mathematical Society, Providence, 2000. A.D. Ioffe, [Typical convexity (concavity) of Dini-Hadamard upper (lower) directional derivatives of functions on separable Banach spaces]{}, J. Convex Anal. 17 (2010), 1019–1032. J. Lindenstrauss, D. Preiss, J. Tišer, [Fr' echet differentiability of Lipschitz maps and porous sets in Banach spaces]{}, Princeton University Press, Princeton 2012. D. Preiss, L. Zajíček, [Directional derivatives of Lipschitz functions]{}, Israel J. Math. 125 (2001), 1–-27. A. Shapiro, [On concepts of directional differentiability]{}, J. Optim. Theory Appl. 66 (1990), 477–-487. L. Zajíček, [ Singular points of order k of Clarke regular and arbitrary functions]{}, Comment. Math. Univ. Carolin. 53 (2012), 51–-63. L. Zajíček, [Hadamard differentiability via G\^ ateaux differentiability*]{}, preprint (2012), arXiv: 1210.4715v1. [^1]: The research was supported by the grant GAČR P201/12/0436.	{ "pile_set_name": "ArXiv" }
--- abstract: \| if and only if there is an expansion $T_1$ of $T$ with $\|T_1\| = \|T\| $ such that if $M$ is a $\mu$-saturated model of $T_1$ and $\|M\| \geq \kappa$ then the reduct $ M\restriction L(T)$ is $\kappa$-saturated. We characterize theories which are superstable without f.c.p., or without f.c.p. as, respectively those where saturation is $(\aleph_0,\lambda)$-transferable or $(\kappa(T),\lambda)$-transferable for all $\lambda$. Further if for some $\mu \geq \|T\|$, $2^\mu > \mu^+$, stability is equivalent to for all $\mu \geq \|T\|$, saturation is $(\mu,2^\mu)$-transferable. author: - \| John T. Baldwin[^1]\ Department of Mathematics\ University of Illinois at Chicago\ Chicago, IL 60680 - \| Rami Grossberg\ Department of Mathematics\ Carnegie Mellon University\ Pittsburgh, PA 15213 - \| Saharon Shelah[^2]\ Institute of Mathematics\ The Hebrew University of Jerusalem\ Jerusalem, 91094 Israel\ &\ Department of Mathematics\ Rutgers University\ New Brunswick, NJ 08902 title: 'TRANSFERING SATURATION, THE FINITE COVER PROPERTY, AND STABILITY [^3]' --- \[section\] \[definition\][Lemma]{} \[definition\][Theorem]{} \[definition\][Conjecture]{} \[definition\][Proposition]{} \[definition\][Corollary]{} \[definition\][[Example]{}]{} \[definition\][Problem]{} \[definition\][Observation]{} \[definition\][Claim]{} \[definition\][Fact]{} \[definition\][Remark]{} Introduction ============ The finite cover property (f.c.p.) is in a peculiar position with respect to the stability hierarchy. Theories without the f.c.p. are stable; but f.c.p. is independent from $\omega$-stability or superstability. We introduce a notion of transferability of saturation which rationalizes this situation somewhat by placing f.c.p. in a natural hierarchy of properties. For countable theories the hierarchy is $\omega$-stable without f.c.p., superstable without f.c.p., not f.c.p., and stable. For appropriate $(\mu,\kappa)$ each of these classes of theories is characterized by $(\mu,\kappa)$-transferability of saturation in following sense. if and only if there is an expansion $T_1$ of $T$ with $\|T_1\|=\|T\|$ such that if $M$ is a $\mu$-saturated model of $T_1$ and $\|M\| \geq \kappa$, then the reduct $ M\restriction L(T)$ is $\kappa$-saturated. The finite cover property was introduced by Keisler in [@Ke] to produce unsaturated ultrapowers. One of his results and a slightly later set theoretic advance by Kunen yield immediately that if for $\lambda>2^{\|T\|}$, saturation is $(\|T\|^+,\lambda)$-transferable then $T$ does not have the finite cover property. The finite cover property was also studied extensively by Shelah in [@Sh:10] and chapters II, VI and VII of [@bible]; those techniques are used here. Our notation generally follows [@bible] with a few minor exceptions: $\|T\|$ is the number of symbols in $\|L(T)\|$ plus $\aleph_0$. We do not distinguish between finite sequences and elements, i.e. we write $a \in A$ to represent that the elements of the finite sequence $a$ are from the set $A$. References of the form IV x.y are to [@bible]. There are several equivalent formulations of the finite cover property. The following, which looks like a strengthening of the compactness theorem, is most relevant here. \[fcp\] The first order theory $T$ does not have the if and only if for every formula $\phi(x;y)$ there exists an integer $n$ depending on $\phi$ such that for every $A$ contained in a model of $T$ and every subset $p$ of $\{\phi(x,a), \neg \phi(x,a); a\in A \}$ the following implication holds: if every $q \subseteq p$ with cardinality less than $n$ is consistent then $p$ is consistent. The two main tools used in this paper are the following consequence of not f.c.p. and a sufficient condition for $\lambda$-saturation. \[easyfact\] Let $T$ be a complete first order theory without the f.c.p. and $\Delta$ a finite set of $L(T)$-formulas. There is an integer $k_{\Delta}$ such that if $M \models T$ is a saturated model, $A {\subseteq}M$ with $\|A\| < \|M\|$ and $ {\bf I}$ is a set of $\Delta$-indiscernibles over $A$ with cardinality at least $k_{\Delta}$ then there exists ${\bf J} {\subseteq}M$ a set of $\Delta$-indiscernibles (over $A$) extending $ {\bf I}$ of cardinality $\|M\|$. The principal tool for establishing the transfer of saturation is \[basictool\] If a model $M$ of a stable theory if $M$ is either $F^a_{\kappa(T)}$-saturated or $\kappa(T) +\aleph_1$-saturated and for each set of indiscernibles ${\bf I}$ contained in $M$ there is an equivalent set of indiscernibles ${\bf J}$ contained in $M$ with $\|{\bf J}\| = \lambda$ then $M$ is $\lambda$-saturated. We thank Anand Pillay for raising the issue of the superstable case and the referee for the final formulation of Theorem \[nfcp\] which generalizes our earlier version and for correcting an oversight in another argument. The transferability hierarchy ============================== In this section we characterize certain combinations of stability and the finite cover property in terms of transferability of saturation. Extending the notation we write [saturation is $(0,\kappa)$-transferable in $T$]{} if and only if there is an expansion $T_1$ of $T$ with $\|T_1\|=\|T\|$ such that if $M \models T_1$ and $\|M\| \geq \kappa$, $M\|L(T)$ is $\kappa$-saturated. In particular, taking $\|M\| = \kappa$, $PC(T_1,T)$ is categorical in $\kappa$. Using this language we can reformulate an old result of Shelah ( Theorems VI.5.4 and VIII.4.1) to provide the first stage of our hierarchy. \[fact1\] For a countable theory $T$, the following are equivalent. 1. $T$ does not have the finite cover property and is $\omega$-stable. 2. For all $ \lambda > \aleph_0 $, saturation is $(0,\lambda)$-transferable in $T$. 3. For some $ \lambda > 2^{\aleph_0}$ saturation is $(0,\lambda)$-transferable in $T$. Since the proof of 1) implies 2) is not given in [@bible] and follows the line of our other arguments we sketch the proof in our discussion after Theorem \[ss\]. This result holds only for countable languages; the remainder apply to theories of arbitrary cardinality. We introduce the following special notation to uniformize the statement of the next result. $$\kappa'(T) = \left\{ \begin{array}{ll} \kappa(T) & \mbox{ if $T$ is stable}\\ \|T\|^+ & \mbox{ if $T$ is unstable} \end{array} \right.$$ \[thmA\] The following are equivalent for a complete theory $T$. 1. $T$ does not have the finite cover property. 2. For all $ \lambda \geq \|T\|^+ $, saturation is $(\kappa'(T),\lambda)$-transferable in $T$. 3. For some $ \lambda > 2^{\|T\|}$, saturation is $(\kappa'(T),\lambda)$-transferable in $T$. It is obvious that $(2)$ implies $(3)$. Now we show $(3)$ implies $(1)$ by showing that if $T$ has the f.c.p. then saturation is not even $(\|T\|^+,\lambda)$-transferable (and so certainly not $(\kappa(T) +\aleph_1,\lambda)$-transferable). Let $T_1$ be any extension of $T$ and $N_0$ an arbitrary model of $T_1$ with cardinality at least $\lambda$. By Kunen’s theorem (see [@Ku], or Theorem 6.1.4 in [@CK]) there exists an $\aleph_1$-incomplete $\|T\|^+$-good ultrafilter $D$ on $\|T\|$. Denote by $N_1$ the ultrapower $N_0^{\|T\|}/D$. By [@Ke] 1.4 and 4.1 or VI.5.3, $N_1$ is $\|T\|^+$-saturated but not $(2^{\|T\|})^+$-saturated. We now show $(1)$ implies $(2)$. Let $T$ be a theory without the f.c.p.. By II.4.1, $T$ is stable. The proof now splits into two cases depending on whether $T$ is superstable. We begin with the case that $T$ is stable but not superstable. Then $\kappa'(T) \geq \kappa(T) + \aleph_1$ and this inequality will be essential shortly. Let $L_1:=L(T) \bigcup \{F\}$ where $F$ is a binary function symbol. The theory $T_1$ consists of $T$ and the following axioms. 1. For each $x$, the function $F(x,\cdot)$ is injective. 2. For every finite $\Delta {\subseteq}L(T)$, let $k_{\Delta}$ be the integer from Fact \[easyfact\]. If $I$ is a finite set of $\Delta$-indiscernibles of cardinality at least $k_{\Delta}$ then there exists an $x_I$ such that 1. the range of $F(x_I,\cdot)$ contains $I$ and 2. the range of $F(x_I,\cdot)$ is a set of $\Delta$-indiscernibles. It should be clear that the above axioms can be formulated in first order logic in the language $L_1$. To see that $T_1$ is consistent, we expand a saturated model $N$ of $T$ to a model of $T_1$. Fix a $1$-$1$ correspondence between finite sets of $\Delta$-indiscernibles [I]{} with $\| {\bf I}\| \geq k_{\Delta}$ and elements $x_{\bf I}$ of $N$. By Fact \[easyfact\], each sufficiently large finite sequence of $\Delta$-indiscernibles [I]{} in $N$ extends to one with $\|N\|$ elements. Fix a $1$-$1$ correspondence between the universe of $N$ and this sequence. Interpret $F(x_I,x)$ as this correspondence. Suppose $N^$ is a $\kappa'(T)$-saturated model of $T_1$ of cardinality at least $\lambda$ and denote $N^\restriction L(T)$ by $N$. Since $\kappa'(T) \geq \kappa(T) +\aleph_1$, by Fact \[basictool\] we need only establish the following claim. \[claim2\] Any infinite sequence of indiscernibles ${\bf I}$ in $N$ extends to a sequence [J]{} of indiscernibles (over the empty set) with cardinality $\|N\|$. Let $ q(x)$ be the set of formulas which expresses that for each finite $\Delta$ the range of $F(x,\cdot)$ is a set of $\Delta$-indiscernibles and ${\bf I}$ is contained in the range of $F(x,\cdot)$. If $ a \in N$ realizes the type $q(x)$ then, since $F(a,\cdot)$ is $1$-$1$, ${\bf J}:= \{ F(a,b) : b \in N \}$ is as required. We now show $q(x)$ is consistent. Fix a finite $q^* {\subseteq}q(x)$ and let $\Delta $ be a finite subset of $ L(T)$ such that all the $L(T)$-formulas from $q^$ appear in $\Delta$. Let $m< \omega$ be sufficiently large so that all the elements of $I$ appearing in $q^$ are among $\{b_0, \ldots ,b_{m-1} \}$ and $m \geq k_{\Delta}$. It suffices to show that for some $a\in N$, each $b_i$ for $i<m$ is in the range of $F(a,\cdot)$ and the range of $F(a,\cdot)$ is a set of $\Delta$-indiscernibles. This follows immediately from $T_1$, by the assumption that $m \geq k_{\Delta}$. Since $q$ is over a countable set there exists an element $a \in N^$ satisfying $q^$ and we finish. $_{\ref{claim2}}$ We now prove Case 2) of $1$ implies $2$: superstable $T$. The general outline of the proof is the same but we replace $\kappa(T) + \aleph_1$-saturation with $F^a_{\kappa(T)}$-saturation and we must use a different trick to find an equivalent set of indiscernibles. The idea for guaranteeing $F^a_{\kappa(T)}$-saturation is taken from Proposition 1.6 of [@Ca]; the referee suggested moving it from a less useful place in the argument to here. \[kT\] If $T$ does not have the f.c.p. then there is an expansion $T_1$ of $T$ such that if $M$ is a $\kappa(T)$-saturated model of $ T_1$ then $M\restriction L(T)$ is $F^a_{\kappa(T)}$-saturated. Let $T$ be a theory without the f.c.p.. Form $L_1$ by adding to $L$ new $k$-ary function symbols $f^{\theta,E}_i$, for $i<m= m(\theta,E)$, for each pair of formulas $\theta(z), E(x,y,z)$ with $\lg(z) =k$ such that for any $M \models T$ and $a \in M$, if $M\models \theta(a)$ then $E(x,y,a)$ is an equivalence relation with $m$ classes. The theory $T_1$ consists of $T$ and the following axioms: For each $k$-ary sequence $z$ such that $\theta(z)$, the elements $f^{\theta,E}_i(z)$, $i<m$ provide a complete set of representatives for $E(x,y,z)$. In any model of $T$, one can choose Skolem functions $f^{\theta,E}_i(z)$ to give sets of representatives for the finite equivalence relations so $T_1$ is consistent. Now suppose that $N^* \models T_1\;$ is $\kappa(T)$-saturated. For any $q ={{\rm stp}}(d/C)$ with $\|C\|<\kappa(T)$, note that $q$ is equivalent to the $L_1$-type over $C$ consisting of the formulas $E(x,f^{\theta,E}_i(c))$ for $E$ a finite equivalence relation defined over a finite sequence $c \in C$ such that $E(d,f^{\theta,E}_i(c))$. Since this type is realized, $N=N^\restriction L(T)$ is $F^a_{\kappa(T)}$-saturated. $_{\ref{kT}}$ Now we show finish showing $(1)$ implies $(2)$ in the superstable case. Let $\lambda \geq \|T\|^+$ be given. We must find a $T_2$ to witness $(\aleph_0,\lambda)$-transferability. First expand $T$ to $T_1$ as in Lemma \[kT\] so that if $M$ is an $\aleph_0$-saturated model of $ T_1$ then $M\restriction L(T)$ is $F^a_{\aleph_0}$-saturated. Form $L_2$ by adding to $L_1$ an $n+2$-ary function symbol $F_n$ for each $n$. The theory $T_2$ consists of $T_1$ and the following axioms: 1. For each $x$ and $n$-ary sequence $z$, the function $F_n(x,z,\cdot)$ is injective. 2. For every finite $\Delta {\subseteq}L(T)$ and $n$-ary sequence $z$, let $k_{\Delta}$ be the integer from Fact \[easyfact\]. If $I$ is a finite set of $\Delta$-indiscernibles over $z$ of cardinality at least $k_{\Delta}$ then there exists an $x_I$ such that 1. the range of $F_n(x_I,z,\cdot)$ contains $I$, 2. the range of $F_n(x_I,z,\cdot)$ is a set of $\Delta$-indiscernibles over $z$. The consistency of $T_2$ is entirely analogous to the similar step in the proof of Theorem \[thmA\]. We just have to interpret each $F_n(x,z,y)$ instead of a single function of two variables. Now suppose that $N^ \models T_2\;$ is an $\aleph_0$-saturated model of cardinality at least $\lambda$. The reduct $N$ of $N^$ to $L(T)$ is $F^a_{\aleph_0}$-saturated and it suffices by Fact \[basictool\] to show for each set of indiscernibles $I$ contained in $N$ there is an equivalent set of indiscernibles $J$ with $\|J\| = \lambda$. Let ${\bf I} =\{b_n : n< \omega \}$ be such an infinite set of indiscernibles in $N$. Let $p^ = {{\rm Av}}(I,N)$ and, since $N$ is $F^a_{\aleph_0}$-saturated, choose $m < \omega$ such that for $B = \{b_0 \ldots b_{m-1}\}$, $p^\|B$ is stationary and $p^$ does not fork over $B$. We show there is a sequence [J]{} of indiscernibles based on $p^\restriction B$ with $\|{\bf J}\| = \|N\|$. Let $q_1(x)$ be a type over $B$ that contains $(\forall y)\theta(F_m(x,b_0, \ldots b_{m-1},y))$ for all $\theta(x) \in p^\restriction B$, for each $\phi(x_0,\ldots x_{n-1}) \in L(T)$ such that $N \models \phi(b_0, \ldots b_{n-1})$, the formula $ (\forall y_1) \ldots (\forall y_n) \phi( F_m(x,b_0, \ldots b_{m-1},y_1),\ldots F_m(x,b_0, \ldots b_{m-1},y_n))$ and the assertion that $F_m(x,c,\cdot)$ is injective. The definition of $T_2$ implies the consistency of $q_1$. Since $q_1$ is a type over a finite set, $q_1$ is realized by an element $a \in N^$; this guarantees the existence of a set of $\|N\|$ indiscernibles equivalent to $I$ as required. $_{\ref{thmA}}$ In the superstable case we can get one slightly stronger result which allows to characterize superstable without f.c.p. by $(\aleph_0,\lambda)$-transferability. \[ss\] If for some $\lambda > 2^{\|T\|}$, saturation is $(\aleph_0,\lambda)$-transferable in $T$ then $T$ is superstable without the f.c.p. By Theorem \[thmA\] we deduce that $T$ does not have the f.c.p. (using $\lambda > 2^{\|T\|}$) and so $T$ is stable. Suppose for contradiction there are a stable but not superstable $T$ and a $T_1$ which witnesses $(\aleph_0,\lambda)$-transferability in $T$. Apply VIII.3.5 to ${\rm PC}(T_1,T)$ taking $\kappa = \aleph_0$, $\mu = (2^{\|T\|})^+$ and $\lambda \geq\mu$. There are $2^{\mu}$ models of $T_1 $ with cardinality $\lambda$, which are $\aleph_0$-saturated, whose reducts to $L(T)$ are nonisomorphic. So some are not $\lambda$-saturated. $_{\ref{ss}}$. We were unable to find a uniform argument for $1$ implies $2$ of Theorem \[thmA\]; there seem to be two quite different ideas for making the large set of indiscernibles equivalent to the given set. The proof of 1) implies 2) of Fact \[fact1\] proceeds along similar lines with the following variation. Since $T$ is $\omega$-stable every $\omega$-saturated model is $F^a_{\omega}$-saturated. Again using the $\omega$ stability, it easy to Skolemize with countably many functions so that each type over a finite set is realized. Then the same trick as in Theorem \[ss\] guarantees the existence of large equivalent indiscernible sets. The proof of Theorem \[thmA\] yields somewhat more than is necessary. The theory $T_1$ which is found in the implication (1) implies (2) does not depend on $\lambda$ and contains only a single additional function symbol. We could obtain a stronger result than $(3)$ implies $(1)$ with the same proof by demanding in a modified $(3)$ that the model witnessing $(\|T\|^+,\lambda)$-transferability have cardinality $\lambda = \lambda^{\|T\|} > 2^{\|T\|}$. As pointed out by the referee, we can combine the arguments for Theorem \[thmA\] and Theorem \[ss\] to characterize $ \kappa(T)$ for theories without the finite cover property if $\kappa(T)$ satisfies the set-theoretic conditions of Theorem VIII.3.5. For example, under the GCH if $\kappa(T)$ is not the successor of a singular cardinal and $T$ does not have the f.c.p. $\kappa(T)$ is the least $\kappa$ such that there is $\lambda > 2^{\|T\|}$ for which saturation is $(\kappa,\lambda)$-transferable. \[stable\] Suppose that there exists a cardinal $\mu \geq \|T\|$ such that $\;2^{\mu}>\mu^+$. For a complete theory $T$, the following are equivalent: 1. $T$ is stable. 2. For all $ \mu \geq \|T\|$, saturation is $(\mu^+,2^{\mu})$-transferable in $T$. 3. For some $ \mu \geq \|T\|$, saturation is $(\mu^+,\mu^{++})$-transferable in $T$. The condition $\mu^+ < 2^{\mu}$ is used only for $(2)$ implies $(3)$ (which is obvious with that hypothesis). In the next two lemmas we prove in ZFC that $(1)$ implies $(2)$, and that $(3)$ implies $(1)$. This shows in ZFC that stability is bracketed between two transferability conditions. \[B1->2\] If $T$ is stable and $ \mu \geq \|T\|$, saturation is $(\mu^+,2^{\mu})$-transferable in $T$. We must find an expansion $T_1$ of $T$ such that if $M$ is a $\mu^+$-saturated model of $T_1$ and $\|M\| \geq 2^{\mu}$, $M\restriction L$ is $2^{\mu}$-saturated. Form $L_1$ by adding one additional binary predicate $E(x,y)$ and add axioms asserting that $E$ codes all finite sets. (I.e., for every set of $k$ elements $x_i$ there is a unique $y$ such that $E(z,y)$ if and only if $z$ is one of the $x_i$.) For any model $M_1$ of $T_1$ and any element $b$ of $M_1$, let $[b]:=\{a\in M_1 : M_1\models E[a,b]\}$. Now let $M_1$ be a $\mu^+$-saturated model of $T_1$ and $M$ the reduct of $M_1$ to $L$. Suppose $A \subseteq M$ has cardinality less than $2^{\mu}$ and $p \in S^1(A)$. We must show $p$ is realized in $M$. By the definition of $\kappa (T)$ there exists $B\subseteq A$ of cardinality less than $\kappa (T)$ such that $p $ does not fork over $B$. Since $M_1$ is $\|T\|^+$-saturated, we may take $p \restriction B$ to be stationary. Let ${\hat p}\in S(M)$ be an extension of $p$ that does not fork over $B$. Since $\mu^+>\|T\|\geq\kappa (T)$, by $\mu^+$-saturation of $M$ there exists $I:=\{a_n : n<\omega\}\subseteq M$ such that $a_n\models {\hat p}\restriction (B \cup \{a_k : k<n\})$. Since the sequence is chosen over a stationary type, $I$ is a set of indiscernibles. Define an $L_1$-type $q$ over $I$ so that if $b$ realizes $q$, $[b] \bigcup I$ is a set of indiscernibles over the empty set. Since the relation $E$ codes finite sets, and $I$ is a set of indiscernibles $q$ is consistent. By the $\aleph_1$-saturation of $M_1$ there exists $b\in M$ realizing the type $q$. If $[b]$ has $2^{\mu}$ elements we are finished since for each formula $\phi({ x},{\overline y})$ and each ${\overline a} \in A$ with $\phi({ x},{\overline a})\in p$, only finitely many elements of $[b]$ satisfy $\neg\phi({ x},{\overline a})$. To show $[b]$ is big enough, using the $\mu^+$-saturation of $M$, we define inductively for $\eta \in 2^{\leq\mu}$ elements $c_{\eta}\in M$ such that 1. $c_{\emptyset}=b$ 2. For any $\eta$, $[c_{\eta\frown 0}]$ and $ [c_{\eta\frown 1}]$ are disjoint subsets of $[c_{\eta}]$. 3. If $\lg(\eta)$ is a limit ordinal $\alpha$, $[c_{\eta}] {\subseteq}\cap_{i<\alpha}[c_{\eta\restriction i}]$ Now for $s \in 2^{\mu}$, the $c_s$ witness that $\|[b]\| = 2^{\mu}$. $_{\ref{B1->2}}$. \[B3->1\] If $ \mu \geq \|T\|$ and saturation is $(\mu^+,\mu^{++})$-transferable in $T$ then $T$ is stable. For the sake of contradiction suppose $T$ is an unstable theory and that there is a $T_1 \supseteq T$ such that if $M$ is a $\mu^+$-saturated model of $T_1$ with cardinality at least $\mu^{++}$, $M\restriction L(T)$ is $\mu^{++}$-saturated. Fix $M_0\models T_1$ with cardinality at least $\mu^{++}$. Let $D$ be a $\mu$-regular ultrafilter on $I=\mu$. Construct an ultralimit sequence $\langle M_{\alpha}:\alpha < \mu^+\rangle$ as in VI.6 with $M_{\alpha+1} = M_{\alpha}^I/D$ and taking unions at limits. By VI.6.1 $M_{\mu^+}$ is $\mu^+$-saturated. But by VI.6.2, since $T$ is unstable, $M_{\mu^+}$ is not $\mu^{++}$-saturated. $_{\ref{B3->1}}$ The methods and concerns of this paper are similar to those in the recent paper of E. Casanovas [@Ca]. He defines a model to be expandable if every consistent expansion of ${\rm Th}(M)$ with at most $\|M\|$ additional symbols can be realized as an expansion of $M$. His results are orthogonal to those here. He shows for countable stable $T$ that $T$ has an expandable model which is not saturated of cardinality greater than the continuum if and only if $T$ is not superstable or $T$ has the finite cover property. By varying the parameters in $(\mu,\kappa)$-transferability of saturation we have characterized four classes of countable theories: $\omega$-stable without f.c.p., superstable without f.c.p., not f.c.p., and stable. For uncountable $\lambda$, they correspond respectively to: $(0,\lambda)$-transferability, $(\aleph_0,\lambda)$-transferability, $(\aleph_1,\lambda)$-transferability, $(\aleph_1,2^{\aleph_0})$-transferability. Although the analogous results for uncountable languages are more cumbersome to summarise, countability of the language is only essential for the $\omega$-stable characterization. [99]{} Enrique Casanovas, Compactly expandable models and stability, [Journal of Symbolic Logic]{} [60]{} , 1995, pages 673–683. C.C. Chang and H. Jerome Keisler, [Model Theory]{}, North-Holland Pub.l Co. 1990. H. Jerome Keisler, Ultraproducts which are not saturated, [Journal of Symbolic Logic]{} [32]{} , 1967, pages 23–46. Kenneth Kunen, Ultrafilters and independent sets, [Trans. Amer. Math. Soc.]{}, [172]{}, 1972, pages 199–206. Saharon Shelah, , , [3]{}, pages 271–362, 1971. — [MR:]{} 47:6475, (02H05) Saharon Shelah, [Classification Theory and the Number of Nonisomorphic Models*]{}, Rev. Ed., North-Holland, 1990, Amsterdam. [^1]: Partially supported by NSF grant 9308768 [^2]: This is item $\#$ 570 on Shelah’s list of publications. [^3]: The authors thank the United States - Israel Binational Science foundation for supporting this research as well as the Mathematics department of Rutgers University where part of this research was carried out.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Progress is reviewed in the understanding of color confinement.' author: - Adriano Di Giacomo title: 'CONFINEMENT IN QCD: RESULTS AND OPEN PROBLEMS' --- [^1] Introduction ============ History ------- The existence of quarks was first hypothesized by M. Gell-mann in the sixties[@Gell1]. The history of the idea is instructive.\ By analogy to the electromagnetic interaction it had been realized that the vector current of ${\beta}$ decay was the Noether current of isospin conservation (Conserved Vector Current Hypothesis[@FG]).\ However the electric charge $Q$ is not a generator of the isospin symmetry group, since it contains an isoscalar part proportional to the hypercharge $Y=N + S$, the sum of the strangeness $S$ and of the baryon number $N$. $$Q= T_3 + {Y\over2}$$ To put weak and electromagnetic interactions on the same footing, the symmetry group had to be enlarged, to include $Y$ among the generators, to a group of rank $2$ containing the $SU(2) $ group of isospin as a subgroup. Among the two possible candidates $G_2$ and $SU(3)$ the latter was found to be the correct choice, with hadrons assigned to the representations $1$,$8$,$10$,$\overline{10}$, the so called eightfold way[@Gell2].\ All the representations of $SU(3)$ are sum of products of the fundamental representations $3$ and $\overline3$, so that an obvious question was about the existence of particles in these representations, the quarks and the antiquarks, as fundamental constituents of hadronic matter. Their charges as predicted by eq(1) are fractional $\pm 1/3$, $\pm 2/3$, a clear experimental signature.\ An intense search for quarks was immediately started, but after 40 years no quark has ever been found, and only upper limits have been established for their production cross sections and abundance.\ It was also realized that there was a problem with Pauli principle. If the ${\Delta}_{3/2}$ is made of three quarks, the state with charge $2$ and spin component ${3\over 2}$ is symmetric under exchange of the quarks since for any reasonable potential the three $u$ quarks are in $S$ state. A possible way out was to assign an extra quantum number to the quarks[@Gell3], which was named color, so that each quark could exist in three different color states.\ After the quantization of the gauge theories it was suggested that the color symmetry could be an $SU(3)$ gauge symmetry with quarks in the fundamental representation, and eight gauge bosons, the gluons mediating their interaction. The theory was named Quantum Chromodynamics (QCD)[@FGL]. Experiments provide evidence for the existence of quarks and gluons at short distances, but quarks never appear at large distances as free particles.\ This phenomenon is known as Confinement of Color. Experiments ----------- The ratio $R\equiv{{n_q}\over {n_p}}$ of the abundances of quarks and antiquarks to the abundance of nucleons has been investigated typically by Millikan-like experiments. No particle with fractional charge has ever been found, with an upper limit [@PDG] $$R\le10^{-27}$$ The expectation for $R$ in the absence of Confinement can be evaluated in the Standard Cosmological Model[@Okun] as follows.\ At $\approx 10^{-9}$seconds after Big-Bang when the temperature was $T\cong 10Gev$ and the effective quark mass $m_q$ of the same order of magnitude, quarks would burn to produce hadrons by the esothermic reactions\ $q+ \overline q \to hadrons$\ $q + q \to \overline q + hadrons$\ Putting ${\sigma_0} = \lim_{v\to0}v{\sigma}$, the burning rate is given by $n_q{\sigma_0}$. The expansion rate in the model is equal to $G_N^{1/2}T^2$ with $G_N$ Newton gravitational constant ant $T$ the temperature. The decoupling of relic quarks will occur when due to the burning processes the quark density will decrease to a value such that the burning rate is smaller than the expansion rate, or when $$n_q {\sigma}_0 = G_N^{1/2}T^2$$ Since the abundance of photons is $n_{\gamma}\simeq T^3$, dividing both sides of eq(3) by $T^3$ gives $${n_q \over n_{\gamma}} = {G_N^{1/2}\over {T{\sigma}_0}}$$ By use of the experimental values ${n_{\gamma}\over n_p}\simeq10^9$, $G_N = 10^{-19}/m_p$, and assuming ${\sigma}_0\simeq m_{\pi}^{-2} $, $T\simeq 10Gev$ we get $$R_{expected }\equiv {n_q\over n_p }= {n_q\over n_{\gamma} }{n_{\gamma}\over n_p } \simeq 10^{-12}$$ Quarks have been also searched as products of particle reactions [@PDG], again with no result. As an example for the inclusive cross section ${\sigma}_q\equiv {\sigma}( p + p\to q(\overline q) + X)$ the experimental upper limit is $${\sigma}_q \le 10^{-40} cm^2$$ The expected value in the absence of confinement is ${\sigma_q}_{expected}\cong {\sigma}_{Total}\cong 10^{-25} cm^2$ The ratios of the upper limits to the expectations are then $${R\over {R_{expected}}}\le 10^{-15} , \\ {{\sigma}_q\over {\sigma_q}_{expected}}\le 10^{-15}$$ $10^{-15}$ is a small number. The only natural possibility is that the ratios are zero, i.e. that confinement is an absolute property, due to some symmetry of the system.\ This is similar to what happens in superconductivity, where the explanation for the upper limits on the resistivity is that it is exactly zero, due to the Higgs breaking of the conservation of electric charge, or in electrodynamics where the natural explanation for the upper limit to the photon mass is that it is exactly zero, the symmetry being gauge invariance.\ No experimental evidence exists for the confinement of the gluons.\ We shall, anyhow, define confinement as absence of colored particles in asymptotic states. Only color singlet particles can propagate as free particles.\ As a working hypothesis we shall assume that some symmetry of the ground state is responsible for confinement. The deconfinement transition. ============================= A limiting temperature exists in hadron physics, known as Hagedorn temperature [@Hag] $T_H$, due to the property of strong interactions to convert excess of energy into creation of particles. It was first conjectured in 1975[@CP] that its existence could be the indication of a deconfining phase transition from hadron to a plasma of quarks and gluons.\ This transition has not yet been detected experimentally, but extensive experimental programs and dedicated machines are being devoted to it at CERN SPS,at Brookhaven (RHIC),and at CERN LHC. The transition has been observed in Lattice QCD.\ Both in experiments and in lattice simulations the main problem is to define and to detect the transition, i.e. to give an operational definition of confined and deconfined. In a way this problem will be the main object of my lectures.\ Finite temperature QCD ---------------------- To deal with a system of fields at non zero temperature T one has to compute the partition function $$Z= Tr [ exp(-{H\over T})]$$ with $H$ the Hamiltonian.\ It can easily be proved that $Z$ is equal to the Feynman Euclidean path integral with the time axis compactified to the interval $(0, {1\over T})$, with periodic boundary conditions for boson fields, antiperiodic for fermions. $$Z = \int [d{\phi}] e^{-\int d^3x \int ^{1\over T} _0 L[{\phi}(\vec x, t)]dt}$$ A system at $T=0$ is simulated on a lattice which is in all directions bigger than the physical correlation length. To have a finite temperature the size in the time direction $L_t$ must be such that $$T = {1\over {aL_t}}$$ where $a=a(\beta,m)$ is the lattice spacing in physical units, which depends on $\beta=\equiv {2N_c\over g^2}$ and on the quark masses $m$. The size $L_s$ in the space directions, instead, must be larger than all physical scales. An asymmetric lattice is therefore needed $L_t\times L_s^3$ with $L_s\gg L_t$.\ The dependence of the lattice spacing $a(\beta, m)$ on $\beta$ is dictated by renormalization group equations. At large enough $\beta$’s $$a\cong {1\over{\Lambda}} e^{\beta\over {2b_0}}$$ with $b_0$ the coefficient of the lowest order term of the beta function, which is negative because of asymptotic freedom. For the temperature T of eq(2.3) we obtain $$T\cong {\Lambda\over {L_t}} e^{\beta\over {\|2b_0\|}}$$ $T$ is an increasing exponential function of $\beta$, i.e. a decreasing function of the coupling constant $g^2$. This is a peculiar behavior : when the coupling constant is big, and the fluctuations are large, i.e. in the disordered phase the temperature is small. In the ordered phase, instead, where the coupling constant and the fluctuations are small the temperature is large. In ordinary thermal systems T plays the role of the coupling constant, low temperature corresponds to order, high temperature to disorder.\ The key word to understand what happens is Duality. Duality ------- Duality is a deep concept in statistical mechanics which has been exported into field theory and string theory.\ It was first introduced in [@KW] and then developed in [@KC] in the frame of the 2d Ising model which, being solvable, is a prototype system for it.\ The Ising model in 2d is defined on a simple square lattice by associating to each site a dichotomic field variable $\sigma = {\underline +}1$. The partition function is $$Z[\beta, {\sigma}]= \sum exp(-\sum_{ij} {\beta} {\sigma_i}{\sigma_j})$$ The sum in the action runs on nearest neighbors and ${\beta=1/T}$ is the inverse temperature in units of the interaction constant. The model is exactly solvable.\ A second order Curie phase transition takes place at $T_c={2\over {\ln(1+\sqrt2)}}$ from an ordered ferromagnetic low temperature phase in which $<\sigma>\neq 0$ to a disordered phase in which the magnetization vanishes.\ The model can be considered as a discretized field theory in (1+1) dimensions, and the lagrangean can be written, apart from an irrelevant constant, as\ $$L={\beta} \sum_{\mu=1,2} {\Delta}_{\mu} {\sigma} {\Delta}_{\mu} {\sigma}$$\ with ${\Delta}_{\mu} {\sigma} \equiv {\sigma} (n+\hat {\mu})- {\sigma} (n)$. The equation of motion is ${\Delta}^2{\sigma}=0$ and a topological conserved current exists $j_{\mu} = {\epsilon}_{\mu\nu}{\Delta}_{\nu}{\sigma}$. $${\Delta}_{\mu}j_{\mu} =0$$ because of the antisymmetry of the tensor ${\epsilon}_{\mu\nu}$. The corresponding conserved charge is $$Q= \sum_{n_1} j_0( n_0,n_1)=\sum_{n_1}{\epsilon_{01}}{\Delta_1}{\sigma} (n_0,n_1)\\= {\sigma} (n_0,+ \infty) - {\sigma}(n_0,- \infty)$$ In a continuum version of the model, when the correlation length goes large compared to the lattice spacing, the value at spatial infinity being a discrete variable becomes a topological quantum number. Typical spacial configurations with non trivial topology are the kinks for which $\sigma$ is negative below some point $\overline n_1$ and positive above it. An anti-kink has opposite signs.\ It can be shown that the operator which creates a kink $\mu(n_0,n_1)$ is a dichotomic variable like $\sigma$ and that the partition function obeys the duality equation $$Z[\beta, \sigma] = Z[\beta^,\mu]$$ with $$sinh(2\beta^)= {1\over sinh(2\beta)}$$ and the same functional form of $Z$ on both sides of eq(2.7).\ The system admits two equivalent descriptions :\ 1) a ’direct’ description in terms of the fields $\sigma$ whose vacuum expectation values are the order parameters, which is convenient in the ordered phase, i.e. in the weak coupling regime. In this description kinks are non local objects with non trivial topology.\ 2) a ’dual’ description in which the topological excitations become local and the original fields non local excitations. The duality mapping eq(2.7) maps the weak coupling regime of the direct description into the strong coupling regime of the dual excitations and viceversa. The dual description is convenient in the strong coupling regime of the direct description.\ The 2-d ising model is self-dual, being the form of the dual partition function the same as that of the direct description, but this is not a general fact.\ Other examples of duality are : the duality angles-vortices in the 3-d X-Y model[@DDPT], the duality magnetization - Weiss domains in the 3d Heisenberg model [@DMP], the duality $A_{\mu}$-monopoles in compact U(1) gauge theory [@FM][@DP][@PC], the duality fields-monopoles in N=1 SUSY SU(2) gauge theory, and many examples in string theory[@ST].\ The idea is then to look for dual, topologically non trivial excitations in QCD, which we shall generically denote by $\mu$, which are ordered in the confining phase $<\mu>\neq0$, thus defining the dual symmetry. The deconfinement transition on the Lattice ------------------------------------------- The same problem as in experiments exists for Lattice simulations : how to define and detect the confined and the deconfined phase.\ In pure gauge theory (no quarks, quenched ) the Polyakov criterion is used, which consists in measuring the $q\bar q$ potential at large distances. If it grows linearly with distance $$V(R) _{R\to\infty} \sim \sigma R$$ there is confinement. If it goes to a constant $$V(R)_{R\to\infty} \sim C + {C'\over R}$$ the phase is deconfined. The potential is measured through the correlator of Polyakov lines. A Polyakov line is defined as the parallel transport along the time axis across the Lattice. $$L(\vec x)\equiv P exp(\int^{1\over T}_0 igA_0(\vec x,t)dt)$$ In terms of the correlator of two Polyakov lines\ $$G(\vec x- \vec y)= <\bar L(\vec x) L(\vec y)>$$\ the static potential $V(\vec x- \vec y)$ acting between a quark and an antiquark is given by $$V(\vec x- \vec y) = -T ln(G(\vec x- \vec y))$$ At large distances, by cluster property, $$<\bar L(\vec x) L(\vec y)> _{\|\vec x -\vec y\|\to \infty \|}\approx \|<L>\|^2 + K exp(-{ {\sigma \|\vec x -\vec y\|}\over T})$$ If $\|<L>\|\neq 0$ then $V(R)\to constant$ as $R\to\infty$ and there is no confinement. If, instead, $\|<L>\| = 0$ then, at large $R$, $V(R)\approx \sigma R $ and there is confinement.\ $\|<L>\|$ is an order parameter for confinement, $Z_N$, the centre of the gauge group, being the relevant symmetry.\ Indeed it can be shown that $$\|<L>\| = exp(- {F_q\over T})$$ with $F_q$ the chemical potential of a quark. In the confined phase $F_q$ diverges and $\|<L>\|\to 0$.\ There is a problem in the continuum limit since $F_q$ diverges also in the deconfined phase due to the self-energy of the quark, and a renormalization is needed[@ren].\ A transition is observed on the Lattice at a temperature $T_c$ from a low temperature phase where $\|<L>=0\|$ (confinement) to a high temperature phase where $\|<L>\neq 0\|$ (deconfinement).\ For gauge group SU(2) ${T_c\over \sqrt \sigma}=.50$ and the transition is second order in the universality class of the $3d$ ising model.\ For gauge group SU(3) ${T_c\over \sqrt \sigma}=.630(5)$ and the transition is weak first order. With the usual convention $2\pi\sigma=1 Gev$ this gives $T_c\approx 270 Mev $. The order of the transition is determined by use of finite size scaling techniques, which are nothing but renormalization group equations \[See e.g. [@cardy]\].\ The density of free energy $F$ by dimensional arguments depends on the spacial size $L_s$ of the system in the form $$f = F(L_s) L_s^4 = \Phi ({a \over \xi}, {\xi \over L_s})$$ where $a$ is the lattice spacing and $\xi$ is the correlation length.\ In the vicinity of $T_c$ $\xi$ goes large with respect to $a$, so that ${a\over \xi}\approx 0$.\ Since $\xi$ diverges as $\tau\equiv (1 -{T\over T_c})\to 0$ as $$\xi \propto \tau^{-\nu}$$ the variable ${\xi\over L_s}$ can be traded with the variable $\tau L_s^{1\over \nu}$, and $$f = \phi (\tau L_s^{1\over \nu})$$\ For the specific heat $C_V = -{1\over V}{\partial^2 \over {\partial T^2}}$ and for the susceptibility $\chi_{<L>} \equiv \int d^3x<\bar L(\vec x) L(\vec 0)>$ the resulting scaling laws are $$\begin{aligned} C_V-C_0 = L_s^{\alpha\over \nu} \phi_C(\tau L_s^{1\over \nu})\\ \chi_{<L>} = L_s^{\gamma\over \nu} \phi_{<L>}(\tau L_s^{1\over \nu}) \end{aligned}$$ From the measured behavior with $L_s$ of these quantities the critical indexes $\alpha, \gamma, \nu$ can be determined, which identify the universality class of the transition.\ For 3d ising $\alpha =.11, \gamma =1.43, \nu =.63$.\ For a weak first order $\alpha =1, \gamma =1, \nu =1/3$.\ In the presence of quarks $Z_3$ is not a symmetry any more and $<L>$ is not an order parameter. The string breaks also in the confined phase and its energy is converted into pions. How to define confined and deconfined?\ The phase diagram for the case $N_f=2$ with two quarks of equal mass $m$ is shown in Fig.\[1\]. ![The Phase Diagram of $N_f = 2$ QCD []{data-label="1"}](PhaseDia.eps){width="4in"} A line exists across which $<L>, <\bar\psi \psi>, <E> $ all experience a rapid change, so that their susceptibilities have a peak. All these peaks happen to coincide within errors. Conventionally the phase below the line is called confined, the one above it deconfined. An order parameter is needed, which must exist if, as we have argued, the transition is order-disorder. The dual excitations have to be identified. The dual excitations of QCD =========================== The general idea is that the low temperature phase of QCD (strong coupling) can be described, in a dual language, in terms of topologically non trivial excitations which are non local in terms of gluons and quarks, but are local fields in a dual language, and weakly coupled[@'tHooft78][@'tHooft81][@SW].\ There exist two main proposals for these excitations, both due to ’tHooft.\ 1) Vortices[@'tHooft78]\ 2) Monopoles[@'tHooft81],[@'tHooft75],[@mandl].\ In the vicinity of the deconfining transition $(T\leq T_c)$ the free energy density should depend on the dual fields in a form dictated by symmetry and scale invariance. The deconfining transition is a change of symmetry : the disorder parameter $<\Phi_{dual}>\neq 0$ for $T <T_c$, $<\Phi_{dual}>=0$ for $T \ge T_c$.\ Two main approaches have been developed in the literature:\ a) Expose in the lattice configurations the dual excitations and show that by removing them confinement gets lost. (Vortex dominance, abelian dominance, monopole dominance)\ b) Study the symmetry and the change of symmetry across $T_c$.\ Vortices -------- Vortices are one dimensional defects associated to closed lines $C$, $V(C)$. If $W(C')$ is the Wilson loop, i.e. the parallel transport along the line $C'$ then $$V(C)W(C') = W(C') V(C)e^{i{{n_{CC'}\pi }\over N_c}}$$ $n_{CC'}$ is the linking number of the two curves $C,C'$, which is well defined in $3d$.\ In $(2+1)d$, $<V(C)>\neq 0$ means spontaneous breaking of a symmetry, the conservation of the number of vortices minus the number of anti-vortices, $<V(C)>=0$ means super-selection of that number, and $V(C)$ can be an order parameter for confinement. In $(3+1)d$ this statements have no special meaning.\ In any case, as a consequence of Eq(3.1) whenever $V(C)$ obeys the area law $W(C')$ obeys the perimeter law, and viceversa whenever $W(C')$ obeys the area law $V(C)$ obeys the perimeter law.\ The ’tHooft loop, defined as the expectation value of a vortex going straight across the lattice, or the dual of the Polyakov line, is non zero in the confined phase, zero in the deconfined phase[@ddl]. The corresponding symmetry is $Z_3$, which, however, does not survive the introduction of dynamical quarks.\ Monopoles --------- Monopoles exist as solitons in Higgs gauge theories with the Higgs in the adjoint representation [@t][@p]. They are stable for topological reasons.\ If the gauge group is $SU(2)$ they are hedgehog-like configurations for the Higgs field $\phi$, with $\phi^i (\vec r)\propto r^i$, and are characterized by a zero of $\phi$ corresponding to the position of the monopole. These configurations are called monopoles because of the non trivial topology of the mapping of the sphere at spacial infinity $S_2$ on the sphere of the possible values of $<\vec \phi>$.\ Physically this can be understood in terms of the ’tHooft tensor, $F_{\mu \nu}$. $$F_{\mu \nu}\equiv \hat{\phi}\vec G_{\mu \nu} - {1\over g}\hat{\phi}(D_{\mu}\hat{\phi}\wedge D_{\nu}\hat{\phi})$$ where $g$ is the gauge coupling constant, $D_{\mu}\hat\phi$ the covariant derivative of $\phi$,\ $$D_{\mu}\hat \phi= [\partial_{\mu}- g\vec A_{\mu}\wedge ]\hat \phi$$ and $\hat\phi\equiv {{\vec \phi}\over {\|\vec \phi\|}}.$\ $F_{\mu\nu} $ is gauge invariant by construction. Moreover the bilinear terms in $A_{\mu}A_{\nu}$ cancel between the two terms of eq(3,2) and\ $$F_{\mu \nu}= \partial_{\mu} (\hat \phi \vec A_{\nu} ) - \partial_{\nu} (\hat \phi \vec A_{\mu})-{1\over g} \hat \phi (\partial_{\mu} \hat \phi \wedge \partial_{\nu} \hat \phi)$$\ In the unitary gauge $\vec \phi = (0,0,1)$, the last term vanishes and\ $$F_{\mu\nu} = \partial_{\mu} A^3_{\nu} - \partial_{\nu} A^3_{\mu}$$\ an abelian field.\ $F_{\mu\nu}$ obeys Bianchi identities $$\partial_{\mu} F_{\mu\nu}^=0$$ with $F_{\mu\nu}^\equiv {1\over 2} \epsilon _{\mu\nu\rho\sigma} F_{\rho\sigma}$ the dual tensor. The identity can be violated at the location of singularities, where a non zero magnetic current exists\ $$\partial_{\mu} F_{\mu\nu}^\equiv j_{\nu}$$\ In any case due to the antisymmetry of $F_{\mu\nu}^$\ $$\partial_{\nu} J_{\nu}=0$$ For the monopole solution [@t]\ $$F_{\mu 0}= 0 , (\vec E=0)$$\ $${1\over 2}\epsilon_{ijk} F_{jk}= {1\over {2g}} {r_i\over {r^3}} + Dirac- string$$\ A Dirac monopole. The string is produced by the singularity of the transformation to the unitary gauge at the zero of $\phi$. The transformation to the unitary gauge is called Abelian Projection.\ For $SU(N)$ gauge group one can inquire about the existence of monopole solitons and what the analog of $\hat \phi$ is[@ddlp]. If we denote by $\Phi= \Sigma _a \Phi^a T^a $ the Higgs field, by $A_{\mu}= \Sigma _a A^a_{\mu} T^a$ the gauge field and by $G_{\mu\nu}=\partial_{\mu} A_{\nu}-\partial _{\nu} A_{\mu} +ig [A_{\mu},A_{\nu}]$ the field strength tensor, with $T^a$ the generators of the gauge group in the fundamental representation, normalized as $Tr(T^a T^b)= \delta ^{ab}$,we can define the generalized ’tHooft tensor as $$F_{\mu \nu}= Tr ( \Phi G_{\mu\nu} - {i\over g} \Phi [D_{\mu} \Phi, D_{\nu} \Phi])$$ The necessary and sufficient condition to have abelian projection, i.e. cancellation of bilinear terms in $A_{\mu}A_{\nu}$ is that\ $$\Phi = \Phi^a , \Phi^a = U(x)^{\dagger }\Phi^a_{diag} U(x)$$\ with U(x) an arbitrary gauge transformation and $$\Phi^a_{diag} = diag({a\over N},{a\over N},...{a\over N}^{(N-a) times}, -{{(N-a)}\over N},-{{(N-a)}\over N}....-{{(N-a)}\over N}) a= 1, 2,....(N-1)$$ The residual symmetry is $SU(a)\otimes SU(N-a)\otimes U(1)$.\ For each $\Phi^a $ one has $$F_{\mu \nu}^a= Tr ( \Phi^a G_{\mu\nu} - {i\over g} \Phi^a [D_{\mu} \Phi^a, D_{\nu} \Phi^a])\\= \partial_{\mu} Tr(\Phi^a A_{\nu}) - \partial_{\nu} Tr(\Phi^a A_{\mu})-{i\over g} Tr(\Phi^a [\partial _{\mu} \Phi^a, \partial_{\nu} \Phi^a])$$ Transforming to the unitary gauge where $\Phi^a=\Phi^a_{diag}$ gives $$F_{\mu \nu}^a= \partial_{\mu} Tr(\Phi^a_{diag} A_{\nu}) - \partial_{\nu} Tr(\Phi^a_{diag} A_{\mu})$$ Expanding the diagonal part of $A_{\mu}$ as a sum of simple roots of the algebra of the group, $\alpha ^a$, which obey the orthogonality relations $Tr(\alpha ^a \Phi ^b_{diag})= \delta^ {ab}$, $A^{\mu}_{diag} = \Sigma _a A^a_{\mu} \alpha^a$,one gets $$F^a_{\mu\nu} = \partial_{\mu} A^a_{\nu} - \partial_{\nu} A^a_{\mu}$$ which is an abelian field. The simple roots have the form\ $$\alpha ^a = diag( 0, 0,....0,1,-1,0,... 0)$$\ with the 1 at the a-th entry. A monopole soliton solution exists for each value of $a$ in the $SU(2)$ subspace spanned by the elements $+1$ and $-1$. For the Higgs field one has\ $$\Phi (x) = U(x)^{\dagger } \Phi (x)_{diag} U(x)$$\ where $ \Phi (x)_{diag}$ is defined with eigenvalues in decreasing order. Expanding $\Phi (x)_{diag}$ in the complete basis $\Phi^a_{diag}$,\ $$\Phi (x)_{diag} = \Sigma_a c^a (x) \Phi^a_{diag}$$\ one gets $$\Phi(x) = \Sigma_a c^a(x) \Phi^a(x)$$ The transformation $U(x)$ is singular at the sites where some $c^a(x)$ vanishes, i.e. wherever two subsequent eigenvalues of $\Phi$ coincide: these points are the locations of the monopoles. The field strength $F^a_{\mu\nu}$ can be defined also in the absence of a Higgs field in the lagrangean simply as $$\begin{aligned} F_{\mu \nu}^a= Tr ( \Phi^a G_{\mu\nu} - {i\over g} \Phi^a [D_{\mu} \Phi^a, D_{\nu} \Phi^a])\\ {\Phi}^a(x) =U^{\dagger} (x) \Phi^a_{diag} U(x) \end{aligned}$$ U(x) an arbitrary gauge transformation which can have non trivial topology or singular points. $F^a_{\mu\nu}$ depends on the choice of $U(x)$. It obeys the Bianchi identities Eq.(3.3), apart from singularities where the magnetic current can be non zero.\ In any case the magnetic current will obey the conservation law eq(3.4).\ The theory has (N-1) topological symmetries built in, corresponding to the conservation of magnetic charges.\ If these symmetries are realized a la Wigner the Hilbert space will be superselected. If they are Higgs broken the system will be a dual superconductor.\ Our working hypothesis will be that the dual symmetry of QCD is the conservation of (N-1) magnetic charges.The change of symmetry at $T_c$ is a transition from Higgs-broken to superselected. Dual excitations carry magnetic charge.\ Our program will then be to construct magnetically charged operators $\mu^a$ and study their vacuum expectation values $<\mu^a>$.\ $<\mu^a>\neq 0$ means dual superconductivity.\ $<\mu^a>=0$ means normal vacuum.\ This should hold both in quenched theory and with dynamical quarks, in agreement with the ideas of $N_c\to \infty$ limit of QCD. Construction of $<\mu^a>$ ------------------------- The basic idea is simply that $$e^{ipa}\|x> = \|x + a>$$ if $x$ is a position variable and $p$ its conjugate momentum.\ Specifically $$\mu^a(\vec x, t) = exp ( i\int d^3y Tr[\Phi^a(\vec y, t) \vec E (\vec y, t)]\vec b_{\perp} (\vec x -\vec y))$$ where $\vec E$ is the electric field operator, and $\vec b(\vec x-\vec y)_{\perp}$ is the vector potential produced by a static monopole sitting at $\vec x$ in $\vec y$.\ $$\vec \nabla \vec b_{\perp}=0,$$ $$\vec \nabla\wedge \vec b_{\perp }={2\pi\over g}{\vec r\over r^3} + Dirac- string$$\ $$\Phi^a(x)= U^{\dagger}(x)\Phi^a_{diag} U(x)$$ with U(x) a generic gauge transformation.\ $\mu^a$ is gauge invariant. In the gauge $$\Phi^a=\Phi^a_{diag}$$ $$\mu^a(\vec x,t) =exp(i\int Tr[\Phi^a_{diag}\vec E(\vec y, t)]\vec b_{\perp} (\vec x-\vec y)d^3y\\ =exp(i \int \vec E^a_{\perp}(\vec y,t)\vec b_{\perp}(\vec x-\vec y)d^3y$$ $\vec E^a_{\perp}(\vec y,t)$ is the conjugate momentum to $\vec A^a_{\perp}(\vec y, t)$ so that $$\mu^a(\vec x,t)\|\vec A^a_{\perp} (\vec y, t)> = \|\vec A^a_{\perp} (\vec y, t) + \vec b_{\perp} (\vec x -\vec y)>$$ A Dirac monopole has been added to the abelian projected configuration. There are (N-1) species of monopoles, corresponding to a=1,..... N-1.\ $\mu^a$ creates a singularity (monopole) in a selected gauge and in all the gauges obtained from it by a transformation which is continuous in a neighborhood of the singularity. The number of monopoles per $fm^3$ is finite as illustrated in Fig.2 where an histogram is displayed of the distribution of the difference between two eigenvalues of a plaquette operator for different values of the lattice spacing[@ddlpp]. Therefore creating a monopole in an abelian projection implies that a monopole is also created in any other abelian projection, apart from a set of zero measure. The statements $<\mu^a>=0$ and $<\mu^a\neq0>$ are independent of the abelian projection, so that the statement that QCD vacuum is or is not a dual superconductor are absolute, projection independent statements.\ Measuring $<\mu^a>$ ------------------- By construction ![Distribution of the differences of the phases of the eigenvalues of the Polyakov line, for three lattices with the same physical volume and different lattice spacing. A monopole on any site would correspond to a non zero value at zero angle. []{data-label="2"}](poly_distr.eps){width="4in"} $$<\mu^a> ={Z^a\over Z}$$ where $Z$ is the partition function of the theory, and $Z^a$ the one modified by the insertion of the monopoles. Eq(3.17) implies that at $\beta=0$ $<\mu^a>=1$.\ Taking advantage of that it is convenient, instead of measuring $<\mu^a>$ directly, to measure its susceptibility $\rho^a= {\partial\over \partial \beta}ln<\mu^a>$ which is much less noisy and will prove more suitable for our purposes. From Eq(3.17) one immediately gets $$\rho^a = <S>_S - <S^a>_{S^a}$$ One also has $$<\mu^a> = exp[\int d\beta'\rho^a(\beta')]$$ It follows from Eq(3.19) that, in the infinite volume limit :\ (i) $<\mu^a>\neq 0$ for $T,T_c$ iff $\rho^a$ tends to a finite limit.\ (ii) $<\mu^a>=0$ for $T>T_c$ iff $\rho^a\to \infty$ The property (ii) is much easier to check on $\rho$ than by a direct measurement of $<\mu>$ which can only give limits, due to statistical errors.\ In the critical region $T\approx T_c$ a strong negative peak is expected due to a rapid decrease of $<\mu^a>$, and scaling laws corresponding to the fact that the correlation length goes large with respect to the lattice spacing. The renormalization group equations read[@ddp] $$<\mu^a> = L_s^\kappa \Phi^a(\tau L_s^{1\over \nu},mL_s^{y_h})$$ Sending $L_s\to\infty$ keeping $\tau L_s^{1\over \nu}$ fixed gives [@ddp] $$<\mu^a> \approx m^{{-\kappa}\over {y_h}}\phi^a (\tau L_s^{1\over \nu})$$ or $$\rho^a /{L_s^{1\over \nu}}= f(\tau L_s^{1\over \nu})$$ a scaling law from which the critical index $\nu$ can be determined.\ The prototype theory is compact $U(1)$ in $4d$, where everything is understood analytically at the level of theorems [@FM] [@DP][@DGP]. There is a phase transition at $\beta_c \approx 1.01$ which is first order, from a confined phase to a deconfined phase, and $<\mu>$ is non zero below $\beta_c$ and zero above $\beta_c$.\ Moreover $\mu$ is proved to be a gauge invariant charged operator of the Dirac type. A numerical determination provides a check of the approach. The result is shown in Fig.s \[3\],\[4\],\[5\] ![$\rho$ vs $\beta$. The peak signals the transition []{data-label="3"}](fig21.ps){width="4in"} A strong negative peak signals the transition. At low $\beta 's$ $\rho$ is size independent, at large $\beta 's$ it is proportional to $L_s$ with a negative coefficient, implying that $<\mu>$ is strictly zero in the thermodynamic limit. ![Size dependence of $\rho$ below $\beta_c$ []{data-label="4"}](fig3.ps){width="4in"} Finite size scaling agrees with a first order transition. ![Finite size scaling. []{data-label="5"}](fig41.ps){width="4in"} For quenched $SU(2)$ theory the deconfining transition is detected in a similar way at the right value of $\beta$ and the critical index $\nu$ is that of the $3d$ ising model[@dlmp1]. Quenched $SU(3)$ also shows a first order transition at the right temperature [@dlmp2]. The numerical check of the independence on the choice of the abelian projection is contained in [@cddlp]. The case of $N_f=2$ $QCD$ can be approached in the same way. The results are displayed in the Fig.’s\[6\], \[7\],\[8\], \[9\]. The finite size scaling is that of a first order transition, and definitely excludes a second order transition in the universality class of O(4), O(2) model[@ddlpp].\ ![$N_f=2$.Size dependence of $\rho$ below $\beta_c$ []{data-label="6"}](full2.eps){width="4in"} ![$\rho$ peaks at different spatial sizes. []{data-label="7"}](full3.eps){width="4in"} ![$\rho$ tends to $-\infty$ in the thermodynamical limit, or $<\mu>\to 0$ []{data-label="8"}](full4.eps){width="4in"} The implications of this fact, together with a finite size scaling analysis of other quantities, like the specific heat, the chiral condensate and its susceptibility will be the object of the next section. ![Scaling of $\rho$ consistent with a first order transition. []{data-label="9"}](eqs.eps){width="4in"} $N_f=2$ QCD =========== QCD with two flavors of light quarks is a good approximation to nature, and also a specially instructive system from the theoretical point of view. For the sake of simplicity we shall consider two quarks of equal mass $m$.\ The phase diagram is shown in Fig.\[1\]. For $m\ge 2Gev$ the system is quenched to all effects, the phase transition is first order and $<L>$ is a good order parameter, $Z_3$ the relevant symmetry. At $m\approx 0$ a phase transition takes place from the spontaneously broken phase to a symmetric phase, and $<\bar\psi\psi>$ is the order parameter. In the intermediate region of $m$’s chiral symmetry is broken by the mass, $Z_3$ is broken by the coupling to quarks and apparently there is no order parameter. Also The $U_A(1)$ symmetry which is broken by the anomaly is expected to be restored about at the same temperature as the chiral symmetry. Three transitions, deconfinement, chiral, $U_A(1)$: are they independent? Of course a definition of deconfinement is needed to answer this question.\ The Chiral Transition --------------------- If one assumes, following reference [@PW] that low mass scalars and pseudoscalars are the relevant degrees of freedom, the order parameters are $$\underline \Phi : \Phi_{ij}=<\bar \Psi_i(1+\gamma_5)\Psi_j>\, , i,j=1...N_f$$ Under the symmetry group $SU(N_f)\otimes SU(N_f)\otimes U_A(1)$\ $$\Phi \to e^{i\alpha}U_L\Phi U_R$$. The most general effective Lagrangean (density of free energy) invariant under the symmetry group is $$L_{eff} = {1\over 2} Tr(\partial_{ \mu}\Phi^{\dagger}\partial_{\mu}\Phi) - {m^2\over 2}Tr(\Phi^{\dagger}\Phi) -{ {\pi}^2\over 3}g_1[Tr(\Phi^{\dagger}\Phi)] ^2-{ {\pi}^2\over 3}g_2[Tr(\Phi^{\dagger}\Phi)^2] +\\ c[det{\Phi} + det{\Phi}^{\dagger}]$$ Terms with higher dimension have been neglected since they become irrelevant at the critical point.\ Infrared stable fixed points indicate second order phase transitions. A $(4-{\epsilon})$ extrapolation to $3d$ is intended. The last term in Eq(4.2) is the Wess-Zumino term describing the anomaly: it is invariant under $SU(N_f)\otimes SU(N_f)$ but not under $U_A(1)$, and has dimension $N_f$. For $N_f\geq3$ it is irrelevant and no IR stable fixed point exists, so that the transition is weak first order.\ For $N_f=2$ instead the Wess-Zumino term has dimension $2$ so that its square and its product with the mass term are also relevant. If $c=0$ at the fixed point the symmetry is $O(4)\otimes O(2)$ and no IR fixed point exists, so that the transition is 1st order. Physically this happens if the mass of the $\eta \prime $, $m_{\eta\prime }$, vanishes at $T_c$.\ If $c\neq 0$, or if $m_{\eta\prime }$ is non zero at $T_c$, the symmetry is $O(4)$ and the transition can be second order. If this is the case the transition is a crossover around the critical point \[see Fig. \[1\]\], a tricritical point is expected at non zero chemical potential [@SS] which could be observed in heavy ion collisions. No evidence of it has emerged from experiments to date.\ If, instead, the transition is first order it will also be such in the vicinity of the chiral point and possibly all along the transition line, and no tricritical point exists.\ This issue is fundamental to understand confinement: a first order phase transition is a real transition and can correspond to a change of symmetry and to the existence of an order parameter.\ A crossover means that one can go continuously from the confined region to the deconfined one and that confinement is not an absolute property of the QCD vacuum. Thermodynamics -------------- The order of the transition can be determined by a finite size scaling analysis [@fischer][@brezin] of lattice simulations.\ Let $\tau = (1-{T \over T_c})$ be the reduced temperature. As $\tau \to 0$ the correlation length of the order parameter, $\xi$, diverges as\ $$\xi \approx \tau^{-\nu}$$\ so that the ratio of the lattice spacing $a$ to $\xi$ is negligible and there is scaling. If $L_s$ is the spacial extension of the lattice, the scaling laws read $$C_V-C_0 \approx L_s^{\alpha\over \nu} \Phi_C(\tau L_s^{1\over\nu}, mL_s^{y_h})$$ and $$\chi \approx L_s^{\gamma \over \nu}\Phi_{\chi}(\tau L_s^{1\over\nu}, mL_s^{y_h})$$ Here $C_V={{\partial E}\over {\partial T}}_V$, and $$\chi \equiv\int d^3x<\Phi(\vec x)\Phi(\vec 0)>_{conn}$$is the susceptibility of the order parameter $\Phi$.\ The critical indexes $\alpha, \beta, \gamma, \nu$ are the anomalous dimensions of the operators and identify the order and the universality class of the transition. Eq’s(4.3) and (4.4) are nothing but the renormalization group equations. The subtraction needed for $C_V$ corresponds to an additive renormalization[@brezin].\ Notice that the scaling law of the specific heat is unambiguous whilst that for $\chi$ only holds if $\Phi$ is the order paramenter : the equality of the index ${\nu}$ for the two scalings can a legitimation of the order parameter.\ The scaling laws Eq.’s(4.3) and (4.4) involve two scales, a fact which makes the analysis complicated with respect to the simpler case of quenched QCD. To simplify the problem one can study the dependence on one scale by keeping the other one fixed [@ddp]. One possibility is to vary $m$ and $L_s$ keeping the quantity $mL_s^{1\over \nu}$ which appears in the scaling laws fixed. The scaling equation (4.3) becomes then $$C_V-C_0 \approx L_s^{\alpha\over \nu} \Phi_C(\tau L_s^{1\over\nu}, mL_s^{y_h}=M)$$ so that the peak scales as $$(C_V-C_0)_{peak}\propto L_s^{\alpha\over \nu}$$ This allows a determination of ${\alpha\over \nu}$. The critical index $y_h$ is the same within errors for $O(4)$ and $O(2)$ universality classes, so that the same simulations can be used to check both universality classes; moreover the index $\alpha$ is negative for both, implying that the peak should decrease with increasing volume. (see Table 1) $y_h$ $\nu$ $\alpha$ $\gamma$ $\delta$ ---------------- ---------- ----------- ----------- ----------- ----------- $O(4)$ 2.487(3) 0.748(14) -0.24(6) 1.479(94) 4.852(24) $O(2)$ 2.485(3) 0.668(9) -0.005(7) 1.317(38) 4.826(12) $MF$ $9/4$ $2/3$ 0 1 3 $1^{st} Order$ 3 $1/3$ 1 1 $\infty$ : Critical exponents.[]{data-label="CRITEXP"} Fig.\[10\] shows a test of Eq(4.5),Fig.\[12\] a test of Eq(4.6). $O(4)$ and $O(2)$ universality classes are excluded with a high confidence level ($\chi^2 /dof \simeq 20$) : the peaks increase rapidly with the volume instead of decreasing. ![Test of scaling Eq.(4.5), with $M= 74.7$ (left) and $M= 149.4 $(right) : the curves should coincide.[]{data-label="10"}](f10-1.eps "fig:"){width="1.9in"}![Test of scaling Eq.(4.5), with $M= 74.7$ (left) and $M= 149.4 $(right) : the curves should coincide.[]{data-label="10"}](f10-2.eps "fig:"){width="1.9in"} For the details of the determination of the subtraction $C_0$ see [@ddp]. ![Scaling Eq.(4.7) of the chiral susceptibility, with $M= 74.7$ (left) and $M= 149.4$ (right) : the curves should coincide.[]{data-label="11"}](f11-1.eps "fig:"){width="1.9in"}![Scaling Eq.(4.7) of the chiral susceptibility, with $M= 74.7$ (left) and $M= 149.4$ (right) : the curves should coincide.[]{data-label="11"}](f11-2.eps "fig:"){width="1.9in"} A similar result is obtained for the susceptibility of $<\bar\psi \psi>$ which is believed to be a good order parameter near $T_c$ (Fig.\[11\] and Fig. \[12\]) $$\chi \approx \Phi_{\chi}(\tau L_s^{1\over\nu}, mL_s^{y_h}=M)$$ and $$\chi_{peak}\propto L_s^{\gamma\over \nu}$$ For this test $\chi ^2/dof \approx 10$. ![Test of the scaling Eq.’s (4.6) and (4.8) []{data-label="12"}](f12-1.eps "fig:"){width="1.9in"}![Test of the scaling Eq.’s (4.6) and (4.8) []{data-label="12"}](f12-2.eps "fig:"){width="1.9in"}\ ![Test of the scaling Eq.’s (4.6) and (4.8) []{data-label="12"}](f12-3.eps "fig:"){width="1.9in"}![Test of the scaling Eq.’s (4.6) and (4.8) []{data-label="12"}](f12-3.eps "fig:"){width="1.9in"} As a result we can state that the transition is neither in the universality class of $O(4)$ nor in that of $O(2)$. Another possibility is to look at the large volume limit keeping the first variable fixed: $\tau L_s^{1\over \nu}$ is related to the ratio ${\xi \over L_s}$ of the correlation length to the spacial size of the lattice, while the other variable $m L_s^{y_h}$ is related to the ratio of the pion Compton wave length ${1\over {m_{\pi}}} $ to $L_s$. As $L_s$ goes much larger than ${1\over {m_{\pi}} }$ a finite limit is reached and[@ddp] $$C_V-C_0 \approx m^{{-\alpha}\over {\nu y_h}} f_C(\tau L_s^{1\over\nu})$$ and $$\chi \approx m^{\gamma \over {\nu y_h}}\Phi_{\chi}(\tau L_s^{1\over\nu})$$ The result of this analysis is shown in Fig.\[13\]and Fig\[14\]. No scaling is observed assuming second order transition with $O(4)$ or $O(2)$, but a good scaling for first order. ![Scaling Eq.(4.9) for O(4).[]{data-label="13"}](f13.eps){width="4in"} ![Scaling Eq.(4.9) for first order. []{data-label="14"}](f14.eps){width="4in"} A similar result is obtained for the scaling Eq.(4.10) Fig.\[15\] ![Testing Eq.(4.10) for O(4) and first order. []{data-label="15"}](f15-1.eps "fig:"){width="1.9in"}![Testing Eq.(4.10) for O(4) and first order. []{data-label="15"}](f15-2.eps "fig:"){width="1.9in"} Finally one can investigate the so called magnetic equation of state: $$<\bar \psi \psi> = m^{1\over \delta} f(\tau m^{-1\over {\nu y_h}})$$ For $O(4)$ $\delta= 4.85$, for first order $\delta = \infty$ Again no scaling is observed assuming $O(4)$ or $O(2)$ second order transition Fig.\[16\], and good scaling for first order, Fig\[17\]. The issue is fundamental and deserves further attention. ![Scaling of the magnetic equation of state (4.11) assuming O(4) []{data-label="16"}](f16.eps){width="4in"} ![Scaling of the magnetic equation of state (4.11) assuming first order []{data-label="17"}](f17.eps){width="4in"} Concluding remarks =================== We have discussed the experimental evidence for confinement and how it naturally implies that there exists a dual symmetry in QCD whose breaking is responsible for confinement. We have presented the two most accredited candidates for dual topological excitations, vortices and monopoles.\ We have then shown how the working hypothesis that monopoles confine via dual superconductivity of the vacuum can be tested by numerical simulations on the lattice, through an order parameter which is the $vev $ of operators carrying magnetic charge. The numerical tests strongly support the validity of this idea, which can be put in a consistent form and made independent on the choice of the abelian projection. This holds both for quenched QCD and in the presence of dynamical quarks.\ A prerequisite is that deconfinement is a true order-disorder phase transition, and not a crossover, which would allow a continuous path from confined to deconfined phase. We have thus discussed a test with $N_f=2$ QCD where an unsolved dilemma exists between the existence of a crossover and a first order phase transition. We definitely exclude a second order chiral transition which would imply a crossover at non zero quark mass, whilst we find evidence for a first order transition. The issue is fundamental and deserves further studies.\ From what we have seen we can conclude that the dual excitations of $QCD$ are magnetically charged, or that dual superconductivity of the vacuum can be the mechanism of confinement.\ However we are not yet able to identify them. [99]{} M. Gell-mann:[Phys.Lett.]{}[8]{} (1964) 214 R.P.Feynman, M. Gell-Mann :[Phys.Rev.]{}[109]{} (1958) 193 M. Gell-Mann:[Phys. Rev.]{}[125]{} (1962) 1067 M. Gell-Mann:[Acta Phys. Austriaca Suppl.]{}[9]{} (1972) 733 H. Fritzsch, M. Gell-Mann, H. Leutwyler:[Phys.Lett.]{}[47B]{} (1973) 365 Review of Particle Physics:[EPJ]{}[15]{} (2000) L. Okun: Leptons and Quarks, [North Holland]{} (1982) R. Hagedorn: [Nuovo Cim. Suppl. ]{}[3]{} (1965) 147 N. Cabibbo, G. Parisi: [Phys. Lett.]{}[59B]{} (1975) 67 H.A. Kramers, G.H. Wannier:[Phys. Rev.]{}[60]{} (1941) 252 L.P.Kadanoff, H. Ceva:[Phys. Rev. ]{}[B3]{} (1971) 3918 G. Di Cecio, A. Di Giacomo, G. Paffuti, A. Trigiante:[Nucl. Phys. ]{}[B489]{} (1997) 739 A. Di Giacomo, D. Martelli, G. Paffuti:[Phys. Rev. ]{}[D60]{}(1999) 094511 J. Froelich, P.A.Marchetti:[Commun. Math. Phys.]{} [112]{} (1987) 343 A. Di Giacomo, G. Paffuti: [Phys. Rev. ]{}[D56]{} (1997) 6816 V. Cirigliano, G. Paffuti: [Commun. Math. Phys.]{}[200]{} (1999) 381 N. Seiberg, E. Witten :[Nucl. Phys. ]{}[B 341]{} (1994) 484 T. Banks, W. Fischler, S. H. Shenker, L. Susskind:[Phys. Rev. ]{}[D55]{} (1997) 5112 O. Kaczmarek, F. Karsch, P. Petreczky, F. Zantow:[Nucl. Phys. Proc Suppl. ]{}[B129]{} (2004) 560 Finite size scaling, J.L.Cardy ed. North Holland (1988) G. ’tHooft:[Nucl. Phys.]{}[B138]{} (1978) 1 G. ’tHooft:[Nucl. Phys.]{}[B190]{} (1981) 455 G. ’tHooft :[High Energy Physics EPS International Conference]{} Palermo 1975,A. Zichichi ed. S. Mandelstam:[Phys. Repts]{}[23C]{} (1976) 245 L. Del Debbio, A. Di Giacomo, B. Lucini:[Nucl. Phys.]{}[bf B594]{} (2001) 287 G. ’tHooft:[Nucl. Phys.]{}[B79]{} (1974) 276 A. M. Polyakov:[Pis’ma JETP]{}[20]{} (1974) 430 L.Del Debbio, A. Di Giacomo, B. Lucini, G. Paffuti. [Abelian projection in SU(N) gauge theories]{} hep-lat/0203023 M. D’Elia, A. Di Giacomo, B. Lucini, G. Paffuti, C. Pica:[Phys. Rev.]{}[D71]{} (2005) 114502 M. D’Elia, A. Di Giacomo, C. Pica:[hep-lat/0503030]{} Submitted for publication. A. Di Giacomo, G. Paffuti:[Nucl. Phys. Proc. Suppl.]{}[106]{}( 2002) 664 A. Di Giacomo, B. Lucini, L. Montesi, G. Paffuti :[it Phys. Rev. ]{}[D61]{}(2000)034503 A. Di Giacomo, B. Lucini, L. Montesi, G. Paffuti :[it Phys. Rev. ]{}[D61]{}(2000)034504 J.Carmona,M. D’Elia, A. Di Giacomo, B. Lucini, G. Paffuti:[Phys. Rev. ]{}[D64]{}(2001)114507 R.D.Pisarski, F.Wilczek:[Phys.Rev. ]{}[D29]{}, (1984) 338 M.A. Stephanov, K.Rajagopal, E.V. Shuryak:[Phys. Rev. Lett. ]{}[81]{} (1998) 4816 M.E.Fischer, M.N. Barber:[Phys. Rev. Lett.]{}[28]{} (1972) 1516 E. Brezin :[Journal de Physique]{} [43]{} (1982) 15 K.G.Wilson: [Phys. Rev.]{} [D10]{} (1974) 2445 [^1]: Presented at The Cracow School of Theoretical Physics, XLV Course.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a method to edit a target portrait footage by taking a sequence of audio as input to synthesize a photo-realistic video. This method is unique because it is highly dynamic. It does not assume a person-specific rendering network yet capable of translating arbitrary source audio into arbitrary video output. Instead of learning a highly heterogeneous and nonlinear mapping from audio to the video directly, we first factorize each target video frame into orthogonal parameter spaces, i.e., expression, geometry, and pose, via monocular 3D face reconstruction. Next, a recurrent network is introduced to translate source audio into expression parameters that are primarily related to the audio content. The audio-translated expression parameters are then used to synthesize a photo-realistic human subject in each video frame, with the movement of the mouth regions precisely mapped to the source audio. The geometry and pose parameters of the target human portrait are retained, therefore preserving the context of the original video footage. Finally, we introduce a novel video rendering network and a dynamic programming method to construct a temporally coherent and photo-realistic video. Extensive experiments demonstrate the superiority of our method over existing approaches. Our method is end-to-end learnable and robust to voice variations in the source audio. Some results are shown in Fig. \[fig:feature-graphic\]. Video results are shown on our project page.' author: - \| Linsen Song$^{1}\thanks{This work was done during an internship at SenseTime Research.}$Wayne Wu$^{2,3}$Chen Qian$^{2}$Ran He$^{1}$Chen Change Loy$^{3}$\ $^1$NLPR, CASIA $^2$SenseTime Research $^3$Nanyang Technological University\ [[email protected]]{}\ [[email protected]]{} bibliography: - 'egbib.bib' title: 'Everybody’s Talkin’: Let Me Talk as You Want' --- ![image](many_demo.png){width="\textwidth"} \[fig:feature-graphic\] Introduction ============ Video portrait editing is a highly sought-after technique in view of its wide applications, such as filmmaking, video production, and telepresence. Commercial video editing applications, such as Adobe Premiere and Apple iMovie, are resource-intensive tools. Indeed, editing audio-visual content would require one or more footages to be reshot. Moreover, the quality of the edited video is highly dependent on the prowess of editors. Audio-based approach is an attractive technique for automatic video portrait editing. Several methods [@Chen2018LipMG; @Zhu2018HighResolutionTF] are proposed to animate the mouth region of a still image to follow an audio speech. The result is an animated static image rather than a video, hence sacrificing realism. Audio-driven 3D head animation [@Taylor2016AudiotoVisualSC] is an audio-based approach but aiming at a different goal, namely to drive stylized 3D computer graphic avatars, rather than to generate a photo-realistic video. Suwajanakorn [@Suwajanakorn2017SynthesizingOL] attempted to synthesize photo-realistic videos driven by audio. While impressive performance was achieved, the method assumes the source audio and target video to come from the same identity. The method is only demonstrated on the audio tracks and videos of Barack Obama. Besides, it requires long hours of single-identity data (up to 17 hours [@Suwajanakorn2017SynthesizingOL]) for training using relatively controlled and high-quality shots. In this paper, we investigate a learning-based framework that can perform many-to-many audio-to-video translation, , without assuming a single identity of source audio and the target video. We further assume a scarce number of target video available for training, , at most a 15-minute footage of a person is needed. Such assumptions make our problem non-trivial: 1) Without sufficient data, especially in the absence of source video, it is challenging to learn direct mapping from audio to video. 2) To apply the framework on arbitrary source audios and target videos, our method needs to cope with large audio-video variations between different subjects. 3) Without explicitly specifying scene geometry, materials, lighting, and dynamics, as in the case of a standard graphics rendering engine, it is hard for a learning-based framework to generate photo-realistic yet temporally coherent videos. To overcome the aforementioned challenges, we propose to use the expression parameter space, rather than the full pixels, as the target space for audio-to-video mapping. This facilitates the learning of more effective mapping, since the expression is semantically more relevant to the audio source, compared to other orthogonal spaces, such as geometry and pose. In particular, we manipulate the expression of a target face by generating a new set of parameters through a novel LSTM-based Audio-to-Expression Translation Network. The newly generated expression parameters, combined with geometry and pose parameters of the target human portrait, allow us to reconstruct a 3D face mesh with the same identity and head pose of the target but with new expression (, lip movements) that matches the phonemes of the source audio. We further propose an Audio ID-Removing Network that keeps audio-to-expression translation agnostic to the identity of the source audio. Thus, the translation is robust to variations in the voices of different people in different source audio. Finally, we solve the difficult face generation problem as a face completion problem conditioned on facial landmarks. Specifically, after reconstructing a 3D face mesh with new expression parameters, we extract the associated 2D landmarks from the mouth region and represent them as heatmaps. These heatmaps are combined with target frames where the mouth region is masked. Taking the landmark heatmaps and the masked target frames as inputs, a video rendering network is then used to complete the mouth region of each frame guided by dynamics of the landmarks. We summarize our contributions as follows: 1) We make the first attempt at formulating an end-to-end learnable framework that supports audio-based video portrait editing. We demonstrate coherent and photo-realistic results by focusing specifically on expression parameter space as the target space, from which source audios can be effectively translated into target videos. 2) We present an Audio ID-Removing Network that encourages an identity-agnostic audio-to-expression translation. This network allows our framework to cope with large variations in voices that are present in arbitrary audio sources. 3) We propose a Neural Video Rendering Network based on the notion of face completion with a masked face as input and mesh landmarks as conditions. This approach facilitates the generation of photo-realistic video for arbitrary people within one single network. Ethical Considerations. Our method could contribute greatly towards advancement in video editing. We envisage relevant industries, such as filmmaking, video production, and telepresence to benefit immensely from this technique. We do acknowledge the potential of such forward-looking technology being misused or abused for various malevolent purposes, , aspersion, media manipulation, or dissemination of malicious propaganda. Therefore, we strongly advocate and support all safeguarding measures against such exploitative practices. We welcome enactment and enforcement of legislation to obligate all edited videos to be distinctly labeled as such, to mandate informed consent be collected from all subjects involved in the edited video, as well as to impose hefty levy on all law defaulters. Being at the forefront of developing creative and innovative technologies, we strive to develop methodologies to detect edited video as a countermeasure. We also encourage the public to serve as sentinels in reporting any suspicious-looking videos to the authority. Working in concert, we shall be able to promote cutting-edge and innovative technologies without compromising the personal interest of the general public. Related Work ============ Audio-based Facial Animation. Driving a facial animation of a target 3D head model by input source audio learns to associate phonemes or speech features of source audio with visemes. Taylor [@Taylor2017ADL] propose to directly map phonemes of source audio to face rig. Afterward, many speech-driven methods have been shown superior to phoneme-driven methods under different 3D models, face rig [@Zhou2018Visemenet; @Edwards2016JALIAA], face mesh [@Karras2017AudiodrivenFA], and expression blendshapes [@Pham2017SpeechDriven3F]. Compared to driving a 3D head model, driving a photo-realistic portrait video is much harder since speaker-specific appearance and head pose are crucial for the quality of the generated portrait video. Taylor [@Taylor2016AudiotoVisualSC] present a sliding window neural network that maps speech feature window to visual feature window encoded by active appearance model (AAM) [@Cootes1998ActiveAM] parameters. To improve the visual quality, in several methods [@Jamaludin2019YouST; @Vougioukas2018EndtoEndSF; @Zhou2018TalkingFG], a still face image is taken as a reference for video generation. However, the result is an animation of a still image rather than a natural video. Recently, Suwajanakornet [@Suwajanakorn2017SynthesizingOL] obtain the state-of-the-art result in synthesizing the Obama video portrait. However, it assumes the source and target to have the same identity and requires long hour of training data (up to 17 hours). Thus, it is not applicable in audio-based video editing that need to cope with different sources of voice and target actors. In addition, the target video data is relatively scarce. Fried [@Fried2019TextbasedEO] proposed a method to edit a talking-head video based on its transcript to produce a realistic video. While it produce compelling results, a person-specific face rendering network need to be trained for each target person. Besides, it takes a long time for viseme search (up to 2 hours for 1-hour recording) and relies on phoneme, thus it cannot be scaled to different languages. Video-based Facial Reenactment. It is inherently difficult to synthesize mouth movements based solely on speech audio. Therefore, many methods turn to learning mouth movements from videos comprising the same/intended speech content. [@Geng2018WarpguidedGF; @Wiles2018X2FaceAN; @Thies2016Face2FaceRF; @Kim2018DeepVP; @Nagano2018paGANRA]. From source portrait video, facial landmarks [@Geng2018WarpguidedGF; @shengju2019makeaface] or expression parameters [@Wiles2018X2FaceAN] are estimated to drive the target face image. In all of these methods [@Geng2018WarpguidedGF; @shengju2019makeaface; @Wiles2018X2FaceAN], the generated portrait videos are frame-wise realistic but they suffer from poor temporal continuity. ReenactGAN [@wayne2018reenactgan] is the first end-to-end learnable video-based facial reenactment method. It introduces a notion of “boundary latent space” to perform many-to-one face reenactment. However, ReenactGAN needs person-specific transformers and decoders, which makes the model size increase linearly with persion identities raising. Many model-based methods [@Thies2016Face2FaceRF; @Kim2018DeepVP; @Nagano2018paGANRA] leverage a 3D head model to disentangle facial geometry, expression, and pose. Face2Face [@Thies2016Face2FaceRF] transfers expressions in parameter space from the source to the target actor. To synthesize a realistic target mouth region, the best mouth image of the target actor is retrieved and warped [@Thies2016Face2FaceRF]. Kim [@Kim2018DeepVP] present a method that transfers expression, pose parameters, and eye movement from the source to the target actor. These methods are person-specific [@Kim2018DeepVP] therefore rigid in practice and suffer audio-visual misalignment [@Thies2016Face2FaceRF; @Kim2018DeepVP; @Nagano2018paGANRA] therefore creating artifacts leading to unrealistic results. Deep Generative Models. Inspired by the successful application of GAN [@Goodfellow2014GenerativeAN] in image generation [@Radford2015UnsupervisedRL; @Mao2016LeastSG; @Zhu2017UnpairedIT; @Huang2017BeyondFR; @wayne2019transgaga], many methods [@Chen2018LipMG; @Zhou2018TalkingFG; @Zhu2018HighResolutionTF; @Pham2018GenerativeAT; @Kim2018DeepVP] leverage GAN to generate photo-realistic talking face images conditioned on coarse rendering image [@Kim2018DeepVP], fused audio, and image features [@Chen2018LipMG; @Zhu2018HighResolutionTF; @Zhou2018TalkingFG]. Generative inpainting networks [@Pathak2016ContextEF; @Iizuka2017GloballyAL; @Liu2018ImageIF; @Yu2018FreeFormII] are capable of modifying image content by imposing guided object edges or semantic maps [@Yu2018FreeFormII; @Song2018GeometryAwareFC]. We convert talking face generation into an inpainting problem of mouth region, since mouth movement is primarily induced by input speeches. Monocular 3D Face Reconstruction. Reconstructing 3D face shape and texture from a single face image has extensive applications in face image manipulation and animation [@Fyffe2014DrivingHF; @Garrido2015VDubMF; @Thies2016Face2FaceRF; @Roth2016Adaptive3D; @Suwajanakorn2017SynthesizingOL]. In general, monocular 3D face reconstruction produces facial shape, expression, texture, and pose parameters by solving a non-linear optimization problem constrained by a statistical linear model of facial shape and texture, such as Basel face model [@Paysan2009A3F], FaceWarehouse model [@Cao2014FaceWarehouseA3], and Flame [@Li2017LearningAM]. Recently, different 3D head models have been increasingly applied in talking portrait video synthesis [@Thies2016Face2FaceRF; @Suwajanakorn2017SynthesizingOL; @Kim2018DeepVP; @Nagano2018paGANRA; @Fried2019TextbasedEO; @Kim2019NeuralSV]. ![image](architecture.png){width="1.0\linewidth"} \[fig:fig\_architecture\] Methodology =========== The architecture of the proposed method is shown in Fig. \[fig:fig\_architecture\]. First, we register a parametric 3D face model [@Cao2014FaceWarehouseA3] in the target video, for every portrait video frame to extract face geometry, pose, and expression parameters. Then, the Audio-to-Expression Translation Network learns the mapping from the source audio feature to face expression parameters. We design an Audio ID-Removing Network to alleviate the issues on large variations caused by multiple speakers. Lastly, we formulate the talking face generation problem as a face completion problem guided by mouth region landmarks, in which the landmarks are projected from the restructured 3D facial mesh. We propose a Neural Video Rendering network to complete the mouth region of each frame, guided by the dynamics of the landmarks to generate a photo-realistic portrait video. 3D Face Modeling ---------------- We leverage a parametric 3D face model [@Cao2014FaceWarehouseA3] on portrait video frame to recover low dimensional geometry, expression, and pose parameters. To reduce parameter dimension, geometry and expression bases are computed based on high-quality head scans [@Blanz1999AMM] and facial blendshapes [@Cao2014FaceWarehouseA3; @Alexander2010TheDE] via principal component analysis (PCA). The geometry parameters $s\in \mathbb{R}^{199}$ and the expression parameters $e\in \mathbb{R}^{29}$ are the coefficients of geometry and expression principle components in the PCA, respectively. The pose of the head $p\in \mathbb{R}^{6}$ which contains 3 head rotation coefficients, 2 translation coefficients ($x$ and $y$ directions on the screen surface), and 1 scaling coefficient. All the parameters are computed by solving a non-linear optimization problem, constrained by the statistical linear 3D face model [@Blanz1999AMM]. By optimizing the geometry, expression, and pose parameters of a given monocular face image based on its detected facial landmarks, protrait video frames will be automatically annotated with low dimensional vectors [@Blanz1999AMM]. The recovered expression parameters are used as the learning target in the Audio-to-Expression Translation Network. Then, the recovered geometry and pose parameters, together with the expression parameters inferred by the Audio-to-Expression Translation Network, are employed for reconstructing the 3D facial mesh. Audio-to-Expression Translation ------------------------------- ### Audio ID-Removing Network We empirically find that identity information embedded in the speech feature degrades the performance of mapping speech to mouth movement. Inspired by recent advances of the speaker adaptation method in the literature of speech recognition [@Visweswariah2002ST; @Povey2012AB], we transfer the speech feature lies in different speaker domains onto a “global speaker” domain by applying a linear transformation, in the form [@Visweswariah2002ST]: $$\begin{array}{c} x^{'}=W_i x + b_i = \bar{W}_i \bar{x}, \\ {\rm where}\ \bar{W_i}=(W_i,\ b_i),\ \bar{x}=(x;\ 1),\\ \bar{W_i}=I+\sum_{j=1}^{k} \lambda_j \bar{W}^j \end{array} \label{eq_audio}$$ Here, $x$ and $x^{'}$ represent the raw and transferred speech feature, respectively, while $\bar{W_i}=I+\sum_{j=1}^{k} \lambda_j \bar{W}^j$ represents the speaker-specific adaptation parameter that is factorized into an identity matrix $I$ plus weighted sum of $k$ components $\bar{W}^j$ [@Povey2012AB]. In speech recognition, these parameters are iteratively optimized by fMLLR [@Digalakis1995SA; @Gales1998ML] and EM algorithms. We formulate the above method into a neural network to be integrated with our end-to-end deep learning network. From Eq.(\[eq\_audio\]), the parameters $\lambda_j$ need to be learned from the input speech feature, while the matrix components $\bar{W}^j\ $ is general speech features of different speakers. Thus, we design an LSTM+FC network to infer $\lambda_j$ from the input and set the matrix components $\bar{W}^j$ as the optimizing parameter of the Audio ID-Removing Network. The matrix components $\bar{W}^j$ of the Audio ID-Removing Network are updated by the gradient descent-based algorithm. The details of the network is depicted in Fig. \[fig\_id\_remove\]. The output of the Audio ID-Removing Network is a new MFCC (Mel-frequency cepstral coefficients) spectrum. We apply a pre-trained speaker identity network VGGVox [@Nagrani2017VA; @Nagrani2017VD] on the new MFCC spectrum and constrain the Audio ID-Removing Network by the following cross-entropy loss function: $$\mathcal{L}_{norm} = -\sum_{c=1}^{N} \frac{1}{N}\log p(c\|x^{'}), \label{eq_vggvox}$$ where $N$ is the number of speakers, $c$ is the speaker class label. The $p(c\|x^{'})$ is the probability of assigning MFCC $x^{'}$ to speaker $c$, which is inferred from the pre-trained VGGVox. Eq. enforces the Audio ID-Removing Network to produce an MFCC spectrum that is not distinguishable by the pre-trained VGGVox. ![Audio ID-Removing Network. We formulate the speaker adaptation method from speech recognition [@Visweswariah2002ST; @Povey2012AB] as a neural network. The network removes identity in speech MFCC spectrum by transferring it to the “global speaker” domain.](id_remove.png){width="1.0\linewidth"} \[fig\_id\_remove\] ### Audio-to-Expression Translation Network We formulate a simple but effective Audio-to-Expression Translation Network that learns the mapping from the ID-removed MFCC feature to the corresponding facial expression parameters. To infer the expression parameters at time $t$, the translation network observes a sliding window speech clip of 1 second, which contains 0.8 seconds before time $t$ and 0.2 seconds after time $t$. We empirically find it challenging to train a network to solely regress the expression parameters. The underlying reason could be that the expression parameters are defined and related to the 3DMM model that is hard to model by the network. To facilitate the learning, we introduce a shape constraint. In particular, with the predicted expression parameters from audio and the ground truth geometry/pose parameters of the video portrait, we can obtain a predicted reconstructed 3D facial mesh. Then, we project 3D points of mouth area to the 2D space to obtain the predicted 2D mouth landmarks. Using a similar method, we can obtain a set of ground-truth 2D mouth landmarks from the ground-truth expression parameters. The shape constraint can be introduced between the predicted 2D mouth landmarks and ground-truth 2D mouth landmarks. The whole process of generating mouth landmarks from expression parameters only involves linear operations and thus is differential. The loss function is written as follows: $$\mathcal{L}_{trans}=\mathcal{L}_{exp}+\mathcal{L}_{shape}=\|\|\hat{e}-e\|\|_2+\|\|\hat{l}-l\|\|_2, \label{eq_audio2expldmk}$$ where $e$ and $l$ are the ground truth expression and landmark, respectively, and $\hat{e}$ and $\hat{l}$ are the output expression and landmark of the translation network, respectively. The Audio ID-Removing and Audio-to-Expression Translation Networks are trained jointly, whose objective function is weighted sum of $L_{norm}$ (Eq. ) and $L_{trans}$ (Eq. ). Neural Video Rendering Network ------------------------------ ### Network Architecture Our final step is to generate photo-realistic talking face video that is conditioned on dynamic background portrait video and is guided by the mouth region landmark heatmap sequence. We design a completion-based generation network that completes the mouth region guided by mouth landmarks. First, to obtain the masked face images, a tailored dynamic programming based on retiming algorithm inspired by [@Suwajanakorn2017SynthesizingOL] is introduced to select frame sequence whose head shaking and blink of eyes look compatible with the source speech. Then, the mouth area that contains lip, jaw, and nasolabial folds are manually occluded by a square mask filled with random noise. To make the conversion from the landmark coordinates to heatmap differentiable, we follow [@Jakab2018conditional; @wayne2019disentangling] to generate heatmaps with Gaussian-like functions centered at landmark locations. We modify a Unet [@Ronneberger2015UNetCN; @shengju2019makeaface]-based network as our generation network. The employed skip-connection enables our network to transfer fine-scale structure information. In this way, the landmark heatmap at the input can directly guide the mouth region generation at the output, and the structure of the generated mouth obeys the heatmaps [@shengju2019makeaface; @Wang2019ExampleGuidedSC]. We composite the generated mouth region over the target face frame according to the input mouth region mask. To obtain the mouth region mask, we connect the outermost mouth landmarks as a polygon and fill it with white color, then we erode the binary mask and smooth its boundaries with a Gaussian filter [@Kim2019NeuralSV]. With the soft mask, we leverage Poisson blending [@Prez2003PoissonIE] to achieve seamless blending. To improve the temporal continuity of generated video, we apply a sliding window on the input masked video frames and heatmaps [@Kim2018DeepVP; @Kim2019NeuralSV]. The input of the Neural Video Rendering Network is a tensor stacked by 7 RGB frames and 7 heatmap gray images [@Kim2019NeuralSV]. It works well in most cases while a little lip motion jitters and appearance flicker might emerge in the final video. Then, a video temporal flicker removal algorithm improved from [@Bonneel2015BlindVT] is applied to eliminate these artifacts. Please refer to appendix for more details of the flicker removal algorithm. ### Loss Functions The loss function for training the Neural Video Rendering Network is written as follows: $$\mathcal{L}_{render} = \mathcal{L}_{recon} + \mathcal{L}_{adv} + \mathcal{L}_{vgg} + \mathcal{L}_{tv} + \mathcal{L}_{gp}. \label{eq_render}$$ The reconstruction loss $L_{recon}$ is the pixel-wise L1 loss between the ground truth and generated images. To improve the realism of the generated video, we apply the LSGAN [@Mao2017LS] adversarial loss $L_{adv}$ and add the gradient penalty term $L_{gp}$ [@Gulrajani2017IT] for faster and more stable training. We also apply the perception loss $L_{vgg}$ [@Johnson2016PL] to improve the quality of generated images by constraining the image features at different scales. The total variation regularization term $L_{tv}$ is used to reduce spike artifact that usually occurs when $L_{vgg}$ is applied [@Johnson2016PL]. The network is trained end-to-end with $L_{total} = L_{norm} + L_{trans} + L_{render}$ (Equations ,, and ) with different coefficients. Due to the limited space, we report the details of the loss function, network architecture, and experimental settings in our appendix. ![image](many_to_many.png){width="0.9\linewidth"} \[fig\_many\_to\_many\] Results ======= We show qualitative and quantitative results on a variety of videos to demonstrate the superiority of our method over existing techniques and the effectiveness of proposed components. Datasets. We evaluate our method on a talking face benchmark dataset GRID [@Cooke2006AA] and a speech video dataset we newly collected. The former contains 1,000 sentences spoken by 18 males and 16 females. We follow Chen [@Chen2018LipMG] to split training and testing sets on the GRID dataset. Since GRID only provides frontal face videos of minor head movement, we record a video dataset that contains multiple head poses and time-varying head motion. The collected dataset contains speech videos of 4 speakers. Each speaker contributes 15 minutes video for training and 2 minutes video for testing, all videos are captured from 7 viewpoints to provide 7 head poses. Resolution of each video is $1920\times1080$. We also take several videos downloaded from YouTube with the same percentage of the training and testing split of recorded data to evaluate our approach. Evaluation Metrics. To evaluate the accuracy of the expression parameters and the projected landmarks under various head poses and motions, we apply the following distance metric: $$\begin{array}{c} E_{exp} = \frac{1}{N_{exp}}\sum_{i=1}^{N_{exp}}\|\|\widehat{e}(i)-e(i)\|\|_2,\\ E_{ldmk} = \frac{1}{N_{ldmk}}\sum_{i=1}^{N_{ldmk}}\|\|\widehat{l}(i)-{l}(i)\|\|_2, \end{array} \label{eq_exp_ldmk_err}$$ where $N_{ldmk}$ and $N_{exp}$ are the number of landmarks and expression parameters respectively and $i$ is the index of landmarks or expression parameters. To quantitatively evaluate the generated quality of portrait videos, we apply common image quality metrics like PSNR [@Wang2004IQ] and SSIM [@Wang2004IQ]. To qualitatively evaluate the generated quality of portrait videos, we conduct a user study in Section \[section\_user\_study\] and demonstrate some generated video results on our project page. Audio-to-Video Translation -------------------------- Many-to-Many Results. To prove that the audio-to-expression network is capable of handling various speakers and the face completion network is generalized on multiple speakers, we present one-to-many results and many-to-one results in Fig. \[fig\_many\_to\_many\] and on our project page. In the one-to-many results, we use the speech audio of one speaker to drive different speakers. Note that different speakers share a single generator instead of multiple person-specific generators. In the many-to-one results, we use the speech audio of different speakers to drive the same speaker. This is in contrast to recent methods, where the whole pipeline [@Suwajanakorn2017SynthesizingOL] or part of components [@Kim2019NeuralSV; @Fried2019TextbasedEO] is designed for a specific person, which disables these methods in handling different voice timbres and facial appearances. Large Pose Results. The main purpose of leveraging 3D face model is to handle head pose variations in generating talking face videos. As far as we know, majority of the recent audio-driving methods focus on generating frontal face video no matter whether a 3D head model is applied [@Suwajanakorn2017SynthesizingOL; @Kim2019NeuralSV] or not [@Vondrick2016GeneratingV; @Zhou2018Visemenet; @DBLP:conf/cvpr/ChenMDX19]. Our method, however, can generate portrait videos under various large poses driven by audio input. Thanks to the decomposition of audio-driving facial animation problem in our framework, which makes the audio only relate to expression parameters of face rather than shape or pose parameters. Results are shown in Fig. \[fig\_large\_pose\] and the video on our project page. Note that in previous methods [@Suwajanakorn2017SynthesizingOL; @Zhou2018Visemenet; @DBLP:conf/cvpr/ChenMDX19], they directly learn a mapping from audio to landmarks, which involves the shape and pose information that is actually independent to the input audio. ![Large pose results. We demonstrate $4$ head poses including up, down, right and left. Video results of all of $7$ head poses can be viewed the video on our project page.](large_pose.png){width="1\linewidth"} \[fig\_large\_pose\] Audio Editing & Singing Results. Our method can also be used to edit the speech contents of a pre-recorded video by splitting and recombining the words or sentences taken from any source audio. We show our audio editing results in Fig. \[fig\_audio\_edit\] and video on our project page. In addition, we also ask a person to record singing and the audio is fed into our network. The driving result can be viewed in Fig. \[fig\_sing\] and video on our project page. This demonstrates the generalization capability of our method and its potential in more complex audio-to-video tasks. ![Audio editing. We select “knowledge is” and “virtue” from “Knowledge is one thing, virtue is another” in the source audio, then recombine them as “Knowledge is virtue” as input.](audio_edit.png){width="1\linewidth"} \[fig\_audio\_edit\] ![Singing. We evaluate our network on the singing audio clips. Video result can be viewed in the video on our project page.](sing.png){width="1\linewidth"} \[fig\_sing\] Comparison with State-of-the-Art {#sec_comparison_sota} -------------------------------- We compare our method with the recent state-of-the-art portrait video generation methods, , Audio2Obama [@Suwajanakorn2017SynthesizingOL], Face2Face [@Thies2016Face2FaceRF], DVP (Deep Video Portrait) [@Kim2018DeepVP] and Text-based Editing (TBE) [@Fried2019TextbasedEO]. The comparative results are demonstrated in Fig. \[fig\_qualitative\_compare\] and video on our project page. First, the Audio2Obama [@Suwajanakorn2017SynthesizingOL] combines a weighted median texture for synthesizing lower face texture and a teeth proxy for capturing teeth sharp details. Our GAN-based rendering network generates better texture details compare to the weighted median texture synthesis [@Suwajanakorn2017SynthesizingOL], , nasolabial folds (Fig. \[fig\_qualitative\_compare\] (a)). Then, we compare our method to Face2Face [@Thies2016Face2FaceRF] that supports talking face generation driving by source video in Fig. \[fig\_qualitative\_compare\] (b). Face2Face [@Thies2016Face2FaceRF] directly transfers facial expression of source video in the parameter space while our method infers facial expression from source audio. The similar lip movement of Face2Face and our method in Fig. \[fig\_qualitative\_compare\] (b) suggests the effectiveness of our Audio-to-Expression Translation Network in learning accurate lip movement from speech audio. Moreover, our GAN-based rendering network generates better texture details, such as mouth corners and nasolabial folds. We also compare to another video-driving method DVP [@Kim2018DeepVP] that supports talking face generation (Fig. \[fig\_qualitative\_compare\] (c)). In DVP, a rendering-to-video translation network is designed to synthesize the whole frame other than the face region. It avoids the blending of face region and background that might be easily detectable. The DVP might fail in a complex and dynamic background as shown in Fig. \[fig\_qualitative\_compare\] (c). In contrast, our method uses the original background and achieves seamless blending that is hard to distinguish. Finally, we compare our method with the contemporary text-based talking face editing method TBE [@Fried2019TextbasedEO] in Fig. \[fig\_qualitative\_compare\] (d). In TBE, the mouth region is searched by phoneme and a semi-parametric inpainting network is proposed to inpaint the seam between the retrieved mouth and the original face background. This method requires training of a person-specific network per input video while our method can generalize on multiple speakers and head poses. Besides, our generation network produces competitive mouth details as shown in Fig. \[fig\_qualitative\_compare\] (d). ![Comparison to state-of-the-art methods. Comparison between our method with Audio2Obama [@Suwajanakorn2017SynthesizingOL], Face2Face [@Thies2016Face2FaceRF], DVP [@Kim2018DeepVP], and TBE [@Fried2019TextbasedEO].](qualitative_compare.png){width="1\linewidth"} \[fig\_qualitative\_compare\] Ablation Study -------------- Evaluation of Parameter Regression. To prove the superiority of incorporating the 3D face model, we compare our network with the one that replaces Audio-to-Expression Translation Network with an Audio-to-Landmark Translation Network as performed in [@Suwajanakorn2017SynthesizingOL]. The Audio-to-Landmark Translation Network modifies the last fully connected layer of the Audio-to-Expression Translation Network so that its output dimension is the coordinate number of mouth region landmarks. The visualized comparison can be viewed in the video on our project page and Fig. \[fig:abl\_2d\_3d\_id\_remove\] (a). We also compare the quantitative metric on GRID and collected dataset as shown in Tab. \[tab\_2d\_3d\]. In the collected dataset that contains more head motion and poses, our method achieves better lip synchronization results as the mouth generated by the one that applies Audio-to-Landmark Translation Network does not even open. ![(a) 2D vs 3D qualitative comparison. 3D parameter regression outperforms 2D parameter regression for head motion and poses. (b) ID-removing qualitative comparison. Improvement can be observed on a failure case caused by not applying Audio ID-Removing Network.](abl_2d_3d_id_remove.png){width="0.9\linewidth"} \[fig:abl\_2d\_3d\_id\_remove\] Evaluation of ID-removing. Our Audio ID-Removing Network transfers the speech feature of different speakers to a “global speaker”, which can be directly proven by tSNE [@Maaten2008VisualizingDU] maps in supplementary materials. We also demonstrate the lip synchronization improvement in Fig. \[fig:abl\_2d\_3d\_id\_remove\] (b) and the video on our project page. Quantitative metrics on GRID and collected dataset also validate its effectiveness as shown in Tab. \[tab\_abl\]. Evaluation of Completion-Based Generation. We evaluate the effects of the proposed completion-based generation that benefits from jointly training on data of different people. As shown in Tab. \[tab\_gen\_abl\], jointly training completion-based generators outperform separately training person-specific generators with much fewer network parameters when the number of speakers increases, regardless of the time length of the training data. User Study {#section_user_study} ---------- To quantitatively evaluate the visual quality of generated portrait videos, following [@Fried2019TextbasedEO], we conduct a web-based user study involving 100 participants on the collected dataset. The study includes 3 generated video clips for each of the 7 cameras and for each of the 4 speakers, hence a total of 84 video clips. Similarly, we also collect $84$ ground truth video clips and mix them up with the generated video clips to perform the user study. We separately calculate the study results of the generated and ground truth video clips. In the user study, all the $84\times 2=168$ video clips are randomly shown to the participants and they are asked to evaluate its realism by evaluating if the clips are real on a likert scale of 1-5 (5-absolutely real, 4-real, 3-hard to judge, 2-fake, 1-absolutely fake) [@Fried2019TextbasedEO]. As shown in Tab. \[tab\_user\_study\], the generated and the ground truth video clips are rated as “real”(score 4 and 5) in 55.0% and 70.1% cases, respectively. Since humans are highly tuned to the slight audio-video misalignment and generation flaws, the user study results demonstrate that our method can generate deceptive audio-video content for large poses in most cases. Conclusion ========== In this work, we present the first end-to-end learnable audio-based video editing method. At the core of our approach is the learning from audio to expression space bypassing the highly nonlinearity of directly mapping audio source to target video. Audio ID-Removing Network and Neural Video Rendering Network are introduced to enable generation of photo-realistic videos given arbitrary targets and audio sources. Extensive experiments demonstrate the robustness of our method and the effectiveness of each pivotal component. We believe our approach is a step forward towards solving the important problem of audio-based video editing and we hope it will inspire more researches in this direction. Appendix {#appendix .unnumbered} ======== Details of Audio-to-Expression Translation Network {#sec:audio_to_exp} ================================================== The network architecture of our Audio-to-Expression Translation Network can be viewed in Figure \[fig:audio\_to\_exp\]. In the training phase, we use paired audio and video frames from the training footage as network input. Ground truth facial shape, expression and pose parameters are calculated from video frame by monocular reconstruction. From the input audio, our Audio-to-Expression Translation Network infers predicted expression parameters that are supervised by the ground truth expression parameters. The loss function is $L_{exp}=\|\|\hat{e}-e\|\|_2$ in Eq. \[eq:l\_trans\]. We reconstruct facial 3D mesh by predicted expression parameters and ground truth shape parameters, then we use the ground truth pose parameters to project 2D mouth landmarks that are supervised by the ground truth 2D mouth landmarks. The loss function is $L_{shape}=\|\|\hat{l}-l\|\|_2$ in Eq. \[eq:l\_trans\]. The ground truth 2D mouth landmarks are projected in a similar way where ground truth expression parameters are used. In the testing phase, the predicted expression parameters from the source audio together with ground truth shape and pose parameters from target video are used to estimate 2D mouth landmarks. The embedded mouth identity and head pose of estimated 2D mouth landmarks are the same as those of the target video while the mouth movement in accord with the source audio. $$L_{trans}=L_{exp}+L_{shape} = \|\|\hat{e}-e\|\|_2+\|\|\hat{l}-l\|\|_2 \label{eq:l_trans}$$ ![image](audio_to_expression.png){width="1\linewidth"} Temporal Flicker Removal Algorithm {#sec:deflicker} ================================== In our approach, talking face video is generated frame by frame and temporal information in video frames is only concerned in landmark estimation. During testing, we find the generated talking face videos demonstrate acceptable frame continuity even in the circumstance that the video temporal flicker removal algorithm is not applied. It is due to that audio series clips used to generate time-adjacent frames contain vast overlap in time. The remaining temporal flicker in the video can be attributed to two reasons: 1) The inferred mouth and jawline landmarks contain slight jitter. 2) Appearance flicker, especially color flicker exists in the video frames generated by the inpainting network. Based on the above analysis, our flicker removal algorithm contains two parts: mouth landmark motion smoothing and face appearance deflicker. Algorithm \[ldmk\_smooth\] demonstrates the mouth landmark motion smooth algorithm. $l_{t-1}$: mouth and jawline landmarks at time $t-1$;\ $l_{t}$: mouth and jawline landmarks at time $t$;\ $d_{th}$: mouth movement distance threshold, $s$: mouth movement smooth strength; $l_{t}^{'}$: smoothed mouth and jawline landmarks at time $t$; get mouth center position $c_t$ at time $t$ from $l_t$ get mouth center position $c_{t-1}$ at time $t-1$ from $l_{t-1}$ $l_{t}^{'}=l_{t}$ $\alpha = \exp(-s\|\|c_{t}-c_{t-1}\|\|_2)$ $l_{t}^{'}=\alpha l_{t-1}+(1-\alpha)l_{t}$ $l_{t}^{'}$ The appearance deflicker algorithm is modified from [@Bonneel2015blind]. We take mouth movement into consideration. If the mouth does not move, then the color flicker is more obvious, and then we increase the deflicker strength. We denote the generated frame and processed frame at time $t$ as $P_t$ and $O_t$, respectively. The mouth center moving distance between time $t-1$ and $t$ is denoted as $d_t$. The processed frame at $t$ is written as: $$\mathcal{F}(P_t) = \frac{4\pi^2f^2\mathcal{F}(P_t)+\lambda_t\mathcal{F}({\rm warp}(O_{t-1}))}{4\pi^2 f^2+\lambda_t}$$ where $\lambda_t=\exp(-d_t)$. Here, $\mathcal{F}$ is the Fourier transform and $f$ means frequency. Function ${\rm warp}(O_{t-1})$ uses optical flow from $P_{t-1}$ to $P_{t}$ to warp input frame $O_{t-1}$. Compared with [@Bonneel2015blind], the weight of previous frame $\lambda_t$ is measured by the strength of mouth motion instead of global frame consistency. Other Experiments {#sec:exp} ================= Quantitative Comparison on GRID dataset --------------------------------------- Our method mainly focuses on talking face video editing, which is different from the recent methods that generate full face from input audio and reference still face image [@Vondrick2016GeneratingV; @Jamaludin2019YouST; @Chen2018LipMG; @Zhu2018HighResolutionTF]. Here we quantitatively compare our method with these methods [@Vondrick2016GeneratingV; @Jamaludin2019YouST; @Chen2018LipMG; @Zhu2018HighResolutionTF] on image generation metrics. For a fair comparison, in our method, we do not apply any post-process and we also modify the input of the inpainting network to generate the full face other than the mouth region. Specifically, the original network input tensor is stacked by 7 RGB frames and 7 heatmap gray images (from time $t-6$ to time $t$), we remove the RGB frame and heatmap gray image at time $t$ and require the network to generate the complete frame image at time $t$. Table \[tab\_ssim\_psnr\] demonstrates that our approach outperforms these methods in PSNR and achieves comparable performance in SSIM. Method PSNR SSIM IS FID -------------------------------------- ----------- ---------- ----------- ---------- Vondrick [@Vondrick2016GeneratingV] 28.45 0.60 - - Jamaludin [@Jamaludin2019YouST] 29.36 0.74 - - Chen [@Chen2018LipMG] 29.89 0.73 - - Ours 30.01 0.94 23.53 9.01 : SSIM, PSNR, IS and FID score comparison of our method and recent methods on GRID dataset. for fair comparison, we generate the full face and do not apply any post process. \[tab\_ssim\_psnr\] Ablation Study on Temporal Flicker Removal Algorithm ---------------------------------------------------- The temporal flicker removal algorithm tries to smooth the output landmark coordinates and eliminate the appearance flicker. The quantitative improvement is slight as shown in Table \[tab\_abl\] but the temporal continuity improvement is obvious as shown in the video on our project page, especially when the mouth does not open. We demonstrate the quantitative results of our 3D parameter regression baseline, Audio ID-Removing Network and Temporal Flicker Removal Algorithm in Table \[tab\_abl\]. Audio ID-removing Effects in tSNE map ------------------------------------- The tSNE [@Maaten2008VisualizingDU] maps in Figure \[fig:id\_remove\_tsne\] demonstrate the 2D visualized the distribution of the input MFCC spectrum and the identity removed MFCC spectrum produced by our Audio ID-Removing Network. We can see that the speaker identity can not be distinguished after removing identity in the MFCC spectrum. ![tSNE before and after id-removing. 2D visualized distributions of input MFCC and normalized MFCC. Different color represents different speaker. Our audio normalization erases the identity information embedded in the MFCC spectrum.[]{data-label="fig:id_remove_tsne"}](tsne_id_remove.png){width="1\linewidth"} Runtime Performance {#sec:run_time} =================== We conduct the inference phrase on a commodity desktop computer with an NVIDIA GTX 1060 and an Intel Core i7-8700. The audio to expression network takes 17 ms per frame and the inpainting network takes 77 ms per frame. The post processes including deflicker and teeth proxy take 1.3s and 300 ms per frame respectively. The deflicker algorithm involves the calculation of optical flow that dominates the inference time. Thus, it takes about 1.7s/100ms to generate one video frame with/without the post-process on average. Limitations {#sec:limit} =========== Emotion: Our method does not explicitly model facial emotion or estimate the sentiment from the input speech audio. Thus, the generated video looks unnatural if the emotion of the driving audio is different from that of the source video. This problem also appears in [@Suwajanakorn2017SynthesizingOL] and we leave this to future improvement. Tongue: In our method, our Neural Video Rendering Network produces lip fiducials and the teeth proxy adds the teeth high-frequency details. Our method ignores the tongue movement when some phonemes (“Z” in the word “result”) are pronounced. The tongue texture can not be well generated according to lip fiducials and teeth proxy as shown in Figure \[fig:limitation\] (a). Accent: Our method performs poorly when the driving speech audio contains an accent. For example, the generated results driven by a speech with a strong Russian accent do not achieve visually satisfactory lip-sync accuracy as shown in Figure \[fig:limitation\] (b). We owe it to the fact that English speech with a strong accent is an outlier to our Audio-to-Expression Translation Network and we leave it to future research. ![Failure Cases. (a) Poor tongue generation result on phoneme “Z” that require the use of tongue. (b) Poor lip-sync accuracy when we use normal speech audio to drive a speaker with strong Russian accent.[]{data-label="fig:limitation"}](limitation.png){width="1\linewidth"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Achieving strong interactions between individual photons enables a wide variety of exciting possibilities in quantum information science and many-body physics. Cold atoms interfaced with nanophotonic structures have emerged as a platform to realize novel forms of nonlinear interactions. In particular, when atoms are coupled to a photonic crystal waveguide (PCW), long-range atomic interactions can arise that are mediated by localized atom-photon bound states. We theoretically show that in such a system, the absorption of a single photon can change the band structure for a subsequent photon. This occurs because the first photon affects the atoms in the chain in an alternating fashion, thus leading to an effective period doubling of the system and a new optical band structure for the composite atom-nanophotonic system. We demonstrate how this mechanism can be engineered to realize a single-photon switch, where the first incoming photon switches the system from being highly transmissive to highly reflective, and analyze how signatures can be observed via non-classical correlations of the outgoing photon field.' address: 'ICFO - Institut de Ciències Fotòniques, The Barcelona Institute of Science and Technology, 08860 Castelldefels (Barcelona), Spain.' author: - 'Andreas Albrecht, Tommaso Caneva, and Darrick E Chang' bibliography: - 'mirror\_ref.bib' title: Changing optical band structure with single photons --- Introduction ============ Creating strong, controllable interactions between individual photons enables many opportunities[@chang14rev] ranging from quantum information processing to the creation of strongly correlated quantum states of light[@hartmann06; @greentree06; @chang08; @angelakis07; @noh16]. Atoms form ideal nonlinear photonic elements to accomplish these tasks, and therefore have been used in diverse settings as cavity QED[@turchette95; @kimble08; @reiserer15] or Rydberg atomic gases[@pritchard10; @peyronel12; @pritchard13]. A promising new approach is the field of waveguide QED[@chang07; @loo13], made possible by experiments to interface atoms with propagating modes of nanophotonic systems including nanofibers[@vetsch10] and photonic crystal waveguides (PCWs)[@goban14; @goban15; @javadi15; @lodahl15]. While these platforms were originally developed to improve upon figures of merit compared to their macroscopic counterparts, more interestingly they enable unprecedented possibilities for engineering interactions with no obvious prior analogue[@petersen14; @douglas15]. Here, we identify and analyze a remarkable example, where a single propagating photon effectively changes the photonic band structure for a second photon. This effect relies on the fact that a near-resonant photon is efficiently converted into an atomic excitation, and that atomic excitations can be made to interact strongly and coherently by aligning the atomic transition frequency with band gaps of an underlying PCW[@douglas15; @john90; @kurizki90]. Specifically, an excited atom hybridizes with a localized photon to form an atom-photon bound state, the photonic component of which can couple to proximal atoms[@douglas15; @hood16; @liu16]. This mechanism has been validated experimentally only recently for atoms along PCWs[@hood16] and transmon qubits[@liu16]. Here, we exploit that for an atomic lattice trapped with the PCW periodicity, atom-atom interactions naturally alternate in sign due to the Bloch structure of the underlying optical modes. Thus, the presence of one atomic excitation effectively creates a doubling of the atomic periodicity from the standpoint of a second propagating photon, resulting in a dramatic change of its dispersion relation. This should be contrasted with the qualitatively very different physics that arises for the case of spatially smooth atomic interactions, which has been separately analyzed before for PCWs[@douglas15mol; @shahmoon14] and also lies at the heart of optical nonlinearities involving Rydberg gases[@pritchard10; @peyronel12; @pritchard13]. We focus specifically on a regime in which a periodicity doubling leads to a so-called atomic-mirror configuration [@chang12; @corzo16; @sorensen16]. This creates a situation in which the absorption of a single photon makes the medium highly reflecting for a subsequent one. In Section\[sec\_config\] we present the detailed setup of our proposal and outline how a single photon influences the propagation of subsequent photons. In Section\[sec\_lshift\], we discuss a particular scheme, in which the atom-atom interactions arising from a PCW bandgap can be utilized to create a dispersive interaction, where an excited atom shifts the resonance frequencies of proximal atoms in an alternating fashion. Subsequently, Section\[sec\_correl\] discusses the effect of these atomic interactions on light propagation. Specifically, we show how the high conditional reflectance produced by the absorption of a single photon manifests itself in the form of an anti-bunched transmitted field, given a weak classical input state. An application of the mechanism as a single-photon switch is discussed in Section\[sec\_condphoton\], in which a single-photon wavepacket is first mapped into the system to regulate the propagation of subsequent photons. This is followed by a discussion on how decoherence impacts the photon switch in Section\[sec\_dephasing\]. Dynamic atomic mirror configuration {#sec_config} =================================== We consider an array of $N$ atoms trapped along a PCW with the photonic crystal periodicity $a$ (Fig.\[b\_setup\](a)). We assume that the ${\|g\rangle}$ to ${\|e\rangle}$ transition of an atomic three-level system, with frequency $\omega_{\rm eg}$, couples to a guided mode of given polarization (e.g., transverse magnetic (TM), Fig.\[b\_setup\](b),(c)). In line with experiments, we assume a single atom can emit into the waveguide at a rate $\Gamma_{\rm 1D}$ significant in magnitude compared to its free-space rate $\Gamma'$[@goban14; @goban15]. We probe the $g$-$e$ transition by a (weak coherent) input field of the form $\mathcal{E}_{\rm in}(z,t)=\tilde{\mathcal{E}}_{\rm in}(t)e^{i(kz-\delta t)}$ through the waveguide, where $k$ denotes the wavevector, $\delta$ the detuning from atomic resonance and $\tilde{\mathcal{E}}_{\rm in}(t)$ accounts for the temporal pulse shape. Its propagation can be described in a spin-model description[@chang12; @caneva15; @lekien05; @dzsotjan10; @tudela11]. In that framework the atomic dynamics are governed by the Hamiltonian $\mathcal{H}_{\rm TM}=\mathcal{H}_{\rm 0}+\mathcal{H}_{\rm wg}$. Here, $\mathcal{H}_{0}$ describes the atom coupling to the input field and the free-space decay and takes the form $$\mathcal{H}_{0}=-i\frac{\Gamma'}{2}\sum_m \sigma_{\rm ee}^m-\sum_m \left[\tilde{\mathcal{E}}_{\rm in}(t)\,e^{ikz_m}\,e^{-i\delta\,t}\sigma_{\rm eg}^m+\mathrm{h.c.}\right] ,$$ with $z_m=m\,a$ the position of atom $m$ and the atomic operators defined as $\sigma_{ij}={\|i\rangle}{\langle j\|}$. The spin Hamiltonian $\mathcal{H}_{\rm wg}=-i\frac{\Gamma_{\rm 1D}}{2}\sum_{m,n} e^{i\,k \|z_m-z_n\|}\,\sigma_{\rm eg}^m\,\sigma_{\rm ge}^n$ accounts for the coupling of the atoms via the emission and re-absorption of photons in the TM mode. Such an effective spin Hamiltonian description for the atom-photon interaction is valid in the Markovian limit, where time retardation over the typical atomic bandwidth frequency range and associated with the free field waveguide propagation is negligible, an assumption readily fulfilled for typical atomic chain sizes[@caneva15; @shi15]. More general non-Markovian approaches and the validity range of the Markovian limit for a variety of atom-waveguide configurations have been studied in the literature[@shi15; @guimond16; @zheng13; @ballestero13; @pletyukhov12]. Once the atomic dynamics are determined from evolution under $\mathcal{H}_{\rm TM}$, the right (+)- and left (-)-going field can be recovered using the input-output relations[@chang12; @caneva15; @lalumiere13; @xu15] $ \mathcal{E}_{\pm}(z,t) =\mathcal{E}_{\rm in}^{\pm}(z,t)+i\,\frac{\Gamma_{\rm 1D}}{2}\,\sum_{z_m}\sigma_{\rm ge}^m e^{\pm i\,k(z-z_m)} $, which relate the outgoing field in terms of the atomic coherences and input fields $\mathcal{E}_{\rm in}^{\pm}$ (here $\mathcal{E}_{\rm in}^+(z,t)\equiv\mathcal{E}_{\rm in}(z,t)$, while $\mathcal{E}_{\rm in}^{-}(z,t)=0$). ![\[b\_setup\] (a) Array of atoms along a PCW, trapped with the structure period $a$. Lower: An excitation (single photon absorption, yellow atom) induces strong, long-range and alternating atom-atom interactions $\mathcal{J}$ mediated by a photonic bound state within the PCW band gap (yellow). In certain regimes, this interaction effectively produces an alternating shift of the atomic transition frequency $\omega$ from its bare value $\omega_0$ as indicated by the vertical scale. (b) Schematic band structure of the TE and TM mode. The horizontal dashed line marks the atomic transition frequencies, here without loss of generality drawn equal. (c) Atomic three-level structure and couplings (see main text for definitions). (d) Reflectance (red) and transmittance (blue) for an input field detuning $\delta$ from a two-level transition and $\Gamma_{\rm 1D}=\Gamma'$. Solid lines correspond to $N=100$, $k\,a=\pi/2$, dashed lines to a sublattice of every second atom $N=50$, $k\,a=\pi$. ](01_setup.pdf) A second transition involving metastable state ${\|s\rangle}$ and ${\|e\rangle}$ is assumed to have a different polarization, and couples to the transverse electric (TE) mode of the waveguide. The transition frequency $\omega_{\rm es}$ is assumed to lie within a TE band gap (see Fig.\[b\_setup\](b)). Thus, ${\|e\rangle}$ cannot spontaneously emit into the band due to the absence of resonant modes. Instead, the photon that the atom attempts to emit is exponentially confined to a length $L$ around the atomic position[@douglas15; @john90]. This photon can be coherently exchanged with proximal atoms, leading to a long-range interaction of the form[@douglas15; @kurizki90; @hood16] $$\label{hambg} \mathcal{H}_{\rm bg}=\mathcal{J}\,\sum_{m\neq n}\cos(q\,z_m)\cos(q\,z_n)\,e^{-\|z_m-z_n\|/L}\,\sigma_{\rm es}^m\sigma_{\rm se}^n\,.$$ Here, the cosine terms account for the standing-wave nature of the field at frequencies within a band gap[@goban15], in direct analogy to a Fabry-Perot cavity. Within the TE band gap the wavevector $q=\pi/a$, such that $\cos(qz_m)\cos(qz_n)=(-1)^{(m+n)}$ and the interaction is intrinsically of alternating sign. Whereas previous proposals neglected that sign alternation[@douglas15mol; @shahmoon14; @caneva15], it forms the crucial mechanism for creating non-linear photon interactions in our setup. The alternating character also makes our approach clearly distinct from setups based on Rydberg atoms[@pritchard10; @peyronel12; @pritchard13], where the photon nonlinearities arise from spatially smooth atom-atom interactions. We moreover assume that the longitudinal spatial extent is tuned to be nearly constant[@douglas15; @hood16], such that $\exp(-\|z_m-z_n \|/L)\simeq 1$. On one hand, this limit is readily achievable experimentally[@hood16], but conceptually it also serves as a useful baseline, given that alternating interactions have never been explored from the standpoint of nonlinear optics. In addition, we introduce a control field $\Omega$ from free-space on the band gap ($s$-$e$) transition that allows for modifying the atom-atom interaction character. It can be described by $\mathcal{H}_c=\delta_c \sum_m\sigma_{\rm ss}^m-\Omega\sum_m\left[\sigma_{\rm es}^m+\mathrm{h.c.}\right]$ with $\delta_c$ the control field detuning from the $s$-$e$ resonance. Thus the total Hamiltonian is $\mathcal{H}=\mathcal{H}_{\rm TM}+\mathcal{H}_c+\mathcal{H}_{\rm bg}$, which combined with the input-output operators allows for solving the probe field propagation dynamics. Note that, although an alternating interaction like (\[hambg\]) can also be achieved in a cavity (where $L\rightarrow\infty$)[@goldstein97], there such a coupling would naturally be probed by the output field associated with the same mode that generated the interactions. In contrast, the PCW offers the opportunity to effectively combine cavity QED and waveguide QED. In particular, the bandgap interaction produced by the TE mode can be probed by propagating fields in a co-linear TM mode. As we will see, while the atoms would respond to propagating TM photons alone like a typical atomic ensemble, the strong TE-mediated interactions will endow the propagation with novel quantum nonlinearities. We assume that the atoms are initialized in state ${\|g\rangle}$, and we restrict the analysis to two excitations in the $\{{\|e\rangle},{\|s\rangle}\}$ manifold. This corresponds to weak input states in the TM mode where the probability of more than two photons being absorbed is negligible. Spin-flip interactions (\[hambg\]) require at least one excitation in each of the states ${\|e\rangle}$ and ${\|s\rangle}$. Since all atoms are initialized in ${\|g\rangle}$, a single photon propagating in the TM mode remains unaffected by the band gap interaction. To understand two-photon propagation, we examine the atomic two-excitation spectrum. In the regime $\|\delta_c\|\gtrsim \|\Omega\|, \|\mathcal{J}\|$, the states ${\|e\rangle}$ and ${\|s\rangle}$ are not effectively mixed together. The states ${\|e_m e_n\rangle}$ and ${\|s_m s_n\rangle}$ therefore remain eigenstates within lowest-order perturbation theory, but experience an energy shift as the far off-resonant control field weakly couples these states to states $({\|s_m e_n\rangle}+{\|e_m s_n\rangle})/\sqrt{2}$. Due to $\mathcal{H}_{\rm bg}$ the energy difference between these two states is not just $\delta_c$, but shifted additionally by $\mathcal{J}_{mn}=(-1)^{m+n}\,\mathcal{J}$. This results in ac-Stark energy shifts, which in section\[sec\_lshift\] are shown to be $\Delta\omega_{\rm ee}=-2\Omega^2/(\delta_c+\mathcal{J}_{mn})$ and $\Delta\omega_{\rm ss}=2\Omega^2/(\delta_c-\mathcal{J}_{\rm mn})$ for ${\|e_m e_n\rangle}$ and ${\|s_m s_n\rangle}$, respectively. Consequently, two-excitation energy levels pick up an alternation due to the underlying Bloch modes. The transition energies can thus be divided into two sublattices of double periodicity(Fig.\[b\_setup\](a)), separated in energy by $\Delta\nu\simeq4\Omega^2 \mathcal{J}/(\delta_c^2-\mathcal{J}^2)$. We now discuss how this manifests itself in photon propagation. Suppose a first photon is initially sent into the medium. The photon has a high chance of creating an atomic excitation in ${\|e\rangle}$ (${\|s\rangle}$), if its frequency is close to single-photon (two-photon) resonance $\delta\sim 0$ ($\Delta\equiv(\delta-\delta_c)\sim 0$). The propagation of a second photon through the medium is then governed by the frequencies of the doubly excited states ${\|e_m e_n\rangle}$ (‘e-branch’) or ${\|s_m s_n\rangle}$ (‘s-branch’). The second photon thus sees a new band structure associated with this atomic “refractive index” pattern of doubled periodicity. In particular, if the splitting is resolved $\Delta\nu >\Gamma'$, the second photon would effectively either see a highly transparent medium if its frequency is similar to the bare atomic transition (as all resonances have been shifted), or see only every other atom if its frequency is aligned with one of the sublattice transitions. Even though that concept is quite general, we here choose $k\,a=\pi/2$, in which case the period-doubled configuration $k\,(2a)=\pi$ is known to result in a dramatically different photon propagation. Whereas $k\,a=\pi/2$ represents a configuration that minimizes reflection[@douglas15mol; @caneva15], $k\,a=\pi$ corresponds to the atomic-mirror configuration of constructive reflection[@chang12; @corzo16; @sorensen16]. In particular, for an array of $N/2$ two-level atoms (i.e. states $g$ and $e$ evolving under $\mathcal{H}_0+\mathcal{H}_{\rm wg}$) with this spacing, the reflectance spectrum is characterized by a Lorentzian[@chang12] $$\label{refl1} \mathcal{R}(\Gamma_{\rm 1D}, \Gamma', \delta)=\frac{\left[(N/2)\,\Gamma_{\rm 1D}\right]^2}{\left[(N/2)\,\Gamma_{\rm 1D}+\Gamma'\right]^2+(2\delta)^2}\,.$$ Here the reflectance $\mathcal{R}$ is defined as the ratio of reflected to input intensity. For $\Gamma'\ll(N/2)\,\Gamma_{\rm 1D}$, the resonance width scales like $\Delta\omega=(N/2)\,\Gamma_{\rm 1D}$. This property arises from the superradiant collective response of the atoms[@chang12], and is in fact independent of the filling fraction of trapped atoms. Thus, our choice of lattice constant $k\,a=\pi/2$ enables the resulting physics to be observed without the need for unity filling of sites. The reflectance spectrum is shown in Fig.\[b\_setup\](d) for $N/2$ atoms with spacing $k\,a=\pi$, and differs significantly from the case of $N$ atoms spaced at $k\,a=\pi/2$. Later, we will show that for the two-photon transition involving ${\|s\rangle}$, Eq.(\[refl1\]) holds but with re-scaled decay rates $\Gamma_{\rm 1D}$, $\Gamma'$. Energy shifts and effective decay rates {#sec_lshift} ======================================= As seen in the preceding section, the combination of band gap interactions ($\mathcal{H}_{\rm bg}$) and off-resonant driving ($\mathcal{H}_c$) leads to an effective energy shift of two-excitation states ${\|e_m e_n\rangle}$ and ${\|s_m s_n\rangle}$. The transition energies to these states determine the nonlinear response as seen by a probe field $\mathcal{E}_{\rm in}(z,t)$. We derive and explain these results in greater detail here, by deriving an effective Hamiltonian for the contribution $\mathcal{H}_{\rm es}=\mathcal{H}_{\rm bg}+\mathcal{H}_c$ in the framework of quasi-degenerate perturbation theory[@tannoudji04]. Such a description is valid in the dispersive limit of large control field detunings, here restricted to second order in the ratio of control field coupling strength and detuning $\sim \|\Omega\|/\|\delta_c\|$. In the single excitation manifold, the Hamiltonian $\mathcal{H}_{\rm es}$ is independent of the band gap coupling and takes on the simple form $ \mathcal{H}_{\rm es}^{(1)}=\delta_c\sum_m\,{\|s_m\rangle}{\langle s_m\|} -\Omega\sum_m \bigl({\|e_m\rangle} {\langle s_m\|}+\mathrm{h.c.}\bigr)$, where ${\|j_m\rangle}=\sigma_{jg}^m{\|g\rangle}^{\otimes N}$ denotes the state with atom $m$ excited to state $j$ and all other atoms in the ground state ${\|g\rangle}$. In the limit $\|\Omega \|\ll\|\delta_c\|$ and to second order in $\|\Omega\|/\|\delta_c\|$ one obtains the effective form $$\label{heffst} \mathcal{H}_{\rm es}^{(1)}\simeq \left(-\frac{\Omega^2}{\delta_c} \right) \sum_m{\|\tilde{e}_m\rangle} {\langle \tilde{e}_m\|}+\left(\delta_c+\frac{\Omega^2}{\delta_c} \right)\,\sum_{m}{\|\tilde{s}_m\rangle}{\langle \tilde{s}_m\|}$$ with the dressed states defined as ${\|\tilde{e}\rangle}={\|e\rangle}-\epsilon\,{\|s\rangle}$ and ${\|\tilde{s}\rangle}={\|s\rangle}+\epsilon\,{\|e\rangle}$, where $\epsilon=\Omega/\delta_c$. Eq. (\[heffst\]) thus describes the conventional (single-atom) ac-Stark shift, where a far detuned laser shifts the energies of the states involved in the transition by an amount $\pm \Omega^2/\delta_c$. In the two-excitation manifold, the Hamiltonian $\mathcal{H}_{\rm es}$ can be written as $$\begin{aligned} \label{h2c}\fl\quad\eqalign{\mathcal{H}_{\rm es}^{(2)}&=2\delta_c\sum_{(mn)} {\|s_ms_n\rangle}{\langle s_ms_n\|} +\sum_{(mn)} \left(\delta_c+\mathcal{J}_{mn}\right)\,{\|+_{mn}\rangle}{\langle +_{mn}\|} +\left( \delta_c-\mathcal{J}_{mn}\right)\,{\|-_{mn}\rangle}{\langle -_{mn}\|}\\ &-\sqrt{2}\Omega\,\sum_{(mn)}\left({\|e_m e_n\rangle} {\langle +_{mn}\|}+{\|s_ms_n\rangle} {\langle +_{mn}\|}+\mathrm{h.c.}\right)}\end{aligned}$$ where ${\|\pm_{mn}\rangle} =( {\|e_ms_n\rangle}\pm{\|s_me_n\rangle})/\sqrt{2}$ and the sum accounts for all possible two-excitation pair combinations. Thus, the control field effectively couples the states ${\|s_ms_n\rangle}$ and ${\|e_me_n\rangle}$ to ${\|+_{mn}\rangle}$, with detunings $\delta_{\rm eff}^\pm=\delta_c\pm \mathcal{J}_{mn}$ (see Fig.\[b\_hameff\](a)) that depend on the alternating sign of $\mathcal{J}_{\rm mn}$. In the limit of $\|\delta_c\pm\mathcal{J}\|\gg \|\Omega\|$, the Hamiltonian to second order in $\|\Omega\|/\|\delta_c\pm\mathcal{J}_{mn}\|$ can be written as $\mathcal{H}_{\rm es}^{(2)}=\sum_{(mn)}\mathcal{H}_{mn}$ with $$\begin{aligned} \label{eff_ham2}\fl \eqalign{\mathcal{H}_{ mn}\simeq & \frac{-2\,\Omega^2}{\delta_c+\mathcal{J}_{mn}}{\|\widetilde{e_me_n}\rangle} {\langle \widetilde{e_me_n}\|}+\left(2\delta_c+\frac{2\,\Omega^2}{\delta_c-\mathcal{J}_{mn}}\right){\|\widetilde{s_ms_n}\rangle}{\langle \widetilde{s_ms_n}\|}+\left( \delta_c-\mathcal{J}_{mn}\right){\|-_{mn}\rangle}{\langle -_{mn}\|}\\ &+\left(\delta_c+\mathcal{J}_{mn}+\frac{4\,\mathcal{J}_{mn}\,\Omega^2}{\mathcal{J}_{mn}^2-\delta_c^2}\right)\,{\|\widetilde{+}_{mn}\rangle} {\langle \widetilde{+}_{mn}\|} -\left[\frac{2\,\Omega^2 \mathcal{J}_{mn}}{\mathcal{J}_{mn}^2-\delta_c^2}\,{\|\widetilde{e_me_n}\rangle}{\langle \widetilde{s_ms_n}\|}+\mathrm{h.c.}\right]\,,} \end{aligned}$$ with the dressed states ${\|\widetilde{ee}\rangle}={\|ee\rangle}+\sqrt{2}\,\epsilon_+\,{\|+\rangle}$, ${\|\tilde{+}\rangle} = {\|+\rangle}-\sqrt{2}\epsilon_+\,{\|ee\rangle}-\sqrt{2}\,\epsilon_-\,{\|ss\rangle}$, ${\|\widetilde{ss}\rangle}={\|ss\rangle}+\sqrt{2}\,\epsilon_-\,{\|+\rangle} $ to linear order in $\epsilon_{\pm}$ where $\epsilon_{\pm}=\Omega/(\mathcal{J}_{mn}\pm\delta_c)$. In the limit $\|\delta_c\|\gg\|\Omega\|$ and $\|\mathcal{J}\|\lesssim \|\delta_c\|$ under consideration, actual transitions between dressed states, represented by the ${\|\widetilde{ee}\rangle}{\langle \widetilde{ss}\|}$ term in (\[eff\_ham2\]), are negligible, and therefore the control field coupling leads to pure energy shifts. ![\[b\_hameff\] (a) Energy levels and control field couplings and detunings in the two-excitation manifold. The band gap coupling $\mathcal{J}_{mn}=\mathcal{J}\,(-1)^{m+n}$ leads to a splitting of its eigenstates ${\|\pm_{mn}\rangle}=1/\sqrt{2}\,({\|e_ms_n\rangle}\pm{\|s_me_n\rangle})$, resulting in an effective control field detuning $\delta_{\rm eff}^\pm=\delta_c\pm \mathcal{J}_{\rm mn}$ between the transitions from states ${\|e_m e_n\rangle}$ and ${\|s_m s_n\rangle}$. (b) Frequency separation for the $e$- (left) and $s$-branch (right) in the rotating frame and for $\mathcal{J}_{mn}=\pm \mathcal{J}$, indicating the frequency with respect to the probe field detuning $\delta$. ](02_effHamiltonian.pdf) We now analyze the transition energies for the $e$- and $s$-branch. Thereby the $e$-branch is defined as the transitions involving an excitation to ${\|e\rangle}$ via a probe beam tuned near the atomic $g$ to $e$ transition ($\delta\simeq 0$). Analogously, the $s$-branch transitions are defined as the two-photon excitations to ${\|s\rangle}$ and are addressed for detunings near a zero two-photon detuning ($\Delta\equiv \delta-\delta_c\simeq 0$). For the two branches, the corresponding transition frequencies needed to create one or two excitations – effective detunings in the rotating Hamiltonian frame – are depicted in Fig.\[b\_hameff\](b). The resonances for the probe field in that picture follow by choosing $\delta$ equal to the indicated frequencies. For the $e$-branch, and an atom-atom coupling $\pm\mathcal{J}$, this results in $\delta=-\Omega^2/\delta_c$ for the absorption of a single photon from the ground state, which simply represents the level shift arising from the ac-Stark shift induced by the control field. To absorb a second photon in the presence of an atomic excitation in ${\|e\rangle}$ a detuning $\delta=-2\Omega^2/(\delta_c\pm \mathcal{J})+\Omega^2/\delta_c=-\Omega^2/\delta_c\pm(\Omega/\delta_c)^2(2\mathcal{J})+\mathcal{O}([\mathcal{J}/\delta_c]^2)$ is required, with the sign dependent on the corresponding atomic sublattice of double periodicity. Similarly for the $s$-branch, the single photon resonance follows as $\delta=\delta_c+\Omega^2/\delta_c$, and the resonance for a second excitation as $\delta=\delta_c+2\Omega^2/(\delta_c\mp\mathcal{J})-\Omega^2/\delta_c=\delta_c+\Omega^2/\delta_c\pm(\Omega/\delta_c)^2\,(2\mathcal{J})+\mathcal{O}([\mathcal{J}/\delta_c]^2)$. The linewidths of the transitions are determined by the effective coupling rates into the waveguide $\Gamma_{\rm 1D}^{\rm eff}$ and to free space $\Gamma'_{\rm eff}$, which follow from multiplying the amount of e-population in the corresponding dressed states with the bare rates $\Gamma_{\rm 1D}$ and $\Gamma'$. For the $e$-branch, and in the dispersive limit considered here, they are in good approximation given by the original rates, i.e. $\Gamma_{\rm 1D}^{\rm eff}\simeq \Gamma_{\rm 1D}$ and $\Gamma_{\rm eff}'=\Gamma'$. For the $s$-branch they follow as $\Gamma_{\rm 1D}^{\rm eff}\simeq (\Omega/\delta_c)^2\,\Gamma_{\rm 1D}$, $\Gamma_{\rm eff}'\simeq (\Omega/\delta_c)^2\,\Gamma'$ for the single excitation transition, and $\Gamma_{\rm 1D}^{\rm eff}\simeq \epsilon_-^2\,\Gamma_{\rm 1D}$, $\Gamma_{\rm eff}'\simeq \epsilon_-^2\,\Gamma'$ with $\epsilon_-= \Omega/(\mathcal{J}_{mn}-\delta_c)$ for the two-excitation transition. Despite providing an intuitive qualitative picture, the model presented above is not sufficient to fully describe the transition frequency between the states ${\|gs\rangle}$ and ${\|ss\rangle}$ (see Fig.\[b\_hameff\](b)). For the s-branch the linewidth $\Delta\omega_{\rm res}\simeq (N/2)\, \Gamma_{\rm 1D}^{\rm eff}$ decreases at a faster rate $\sim (\Omega/\delta_c)^2$ than the increase of precision for the energy shifts and thus the resonance positions in the perturbative treatment. Thus we cannot expect our estimated resonance frequency to overlap (within less than the linewidth $\Delta\omega_{\rm res}$) with the full numerical solution, even though the essential properties are still well-approximated. A more accurate prediction of the resonance frequencies, which also provides excellent quantitative agreement for the $s$-branch, is given by a transfer-matrix model introduced in \[sec\_tramodel\]. Photon and excitation correlations {#sec_correl} ================================== We now consider the effects of the change in band structure on a weak coherent state input. Here and in the following we resonantly excite atoms to the ‘s-branch’ by choosing a two-photon detuning $\Delta\sim 0$, and define $\delta=\delta_{\rm res}^{(1)}$ as the probe detuning where linear transmission is maximally attenuated. Signatures of nonlinearities are observable in two-photon correlations of the cw-probe field $\tilde{\mathcal{E}}_{\rm in}(t)=\mathcal{E}_0$; in particular we illustrate in Fig.\[b\_correl\](a) $g^{(2)}(0)={\langle \mathcal{E}_+^\dagger\mathcal{E}_+^\dagger\mathcal{E}_+\mathcal{E}_+ \rangle}/{\langle \mathcal{E}_+^\dagger\mathcal{E}_+ \rangle}^2$ of the transmitted probe field (see\[sec\_numimp\] for details of the calculation). We find two branches of photon antibunching $g^{(2)}(0)<1$, which are increasingly shifted away from the point of maximum linear attenuation $\delta=\delta_{\rm res}^{(1)}$ with increasing atomic coupling $\mathcal{J}$. The positions of the branches, which indicate a suppression of two-photon propagation, qualitatively follow the conditions for high reflectance of a second photon when a single atomic excitation is first prepared deterministically (white lines, calculated in detail later), however with a reduced splitting. In the present case of a weak coherent state input, the spatio-temporal dynamics of multi-photon propagation appear quite complicated, and we cannot find a simple predictive model for the anti-bunching dips. However, we believe that the positions of the anti-bunching dips represent a balance between efficient absorption of the first photon (maximal near $\delta_{\rm res}^{(1)}$), and decreased reflectance away from the resonances of a single sublattice (white lines). While it might be difficult to observe experimentally, numerically we can find additional confirmation of the physics by examining the atomic populations. Specifically, near the anti-bunching dips for $\delta > \delta_{\rm res}^{(1)}$, a checkerboard pattern, as the one illustrated for a selected point (black arrow of Fig.\[b\_correl\](a)) in Fig.\[b\_correl\](b), is observed in the population of states ${\|s_m s_n\rangle}$. Such a pattern reveals that indeed only atoms of a single sublattice are excited. For the opposite anti-bunching branch ($\delta < \delta_{\rm res}^{(1)}$) a complementary (inverted) pattern holds. ![\[b\_correl\] (a) Two-photon correlation $g^{(2)}(0)$ for the transmitted field vs relative field detuning $\delta-\delta_{\rm res}^{(1)}$ ($\delta_{\rm res}^{(1)}=97.62\,\Gamma'$) and atomic coupling rate $\mathcal{J}$. The overset illustrates the linear transmittance spectrum $\mathcal{T}^{(1)}={\langle \mathcal{E}_+^\dagger \mathcal{E}_+ \rangle}/\mathcal{E}_0^2$. White lines border the region of high reflectance ($\mathcal{R} >0.8$) for the case where an $s$-excitation is prepared deterministically. The numerical parameters used are $N=100$, $\Gamma_{\rm 1D}=0.5\,\Gamma'$, $\delta_c=94\,\Gamma'$, $\Omega=18.8\,\Gamma'$, $\mathcal{E}_0=10^{-4}\,\Gamma'$. (b) Wave-function population (arbitrary units) of the doubly excited states ${\|s_ms_n\rangle}$ at the position marked by the black arrow ($\delta-\delta_{\rm res}^{(1)}=1.38\,\Gamma'$, $\mathcal{J}=25\,\Gamma'$) in (a). ](03_g2correl.pdf) Conditional single photon reflection {#sec_condphoton} ==================================== We now consider the case where a “gate" photon is initially and deterministically mapped onto a spin excitation, whose presence or absence controls the propagation of a subsequent “signal” photon. Such a case is interesting because it removes the spatio-temporal complexity associated with a coherent state input, and also constitutes a logical quantum gate. The signal photon will be centered at a frequency corresponding to one of the sublattice resonances, which leads to strong reflection in the presence of the gate. We will also be primarily interested in the regime where the sublattice resonances are well-separated $\Delta\nu>(N/2)\,\Gamma_{\rm 1D}$. This implies that the signal photon is far off resonance from the atomic medium absent the gate and is thus nearly perfectly transmitted in that case. We address the regime of smaller resonance separations $\Delta\nu$ in \[sec\_Jdependence\], where it is shown that suitable reflectance resonances can also be obtained in regimes of overlapping resonances, which drastically reduces the atom-atom coupling $\mathcal{J}$ required by around an order of magnitude. ![\[b\_decspec\] (a) Left: Atomic population decay in time for an initial subradiant excitation in ${\|s\rangle}$, an atom number $N=100$, and for different ratios of waveguide $\Gamma_{\rm 1D}$ to free-space $\Gamma'$ decay. The $e$- and $s$-populations are defined as the total populations $p_e^{\rm tot}=\sum_{n=1}^{N} {\langle \sigma_{\rm ee}^{(n)} \rangle}$ and $p_s^{\rm tot}=\sum_{n=1}^{N} {\langle \sigma_{\rm ss}^{(n)} \rangle}$, respectively. Right: Degenerate (blue and red) single excitation subradiant states ${\|\psi\rangle}=\sum_n c_n\,\exp(ik a\,n)\,{\|s_n\rangle}$ for $N=100$. The upper plot illustrates the amplitude $\|c_n\|$ in the atom numbering $n$, the lower plot the amplitude in momentum space $\|c_k\|$ obtained by a discrete Fourier transform of ${\|\psi\rangle}$. (b) Conditional reflectance $\mathcal{R}$ (red, squares) and transmittance $\mathcal{T}$ (blue, circles) spectra for the s-branch in the presence of a non-decaying gate excitation and $N\,\Gamma_{\rm 1D}/\Gamma'=100$. Green and light green lines (triangles) correspond to the transmittance and reflectance in the absence of an initial gate excitation, respectively. Data points represent numerical results, solid lines the transfer-matrix model predictions. Parameters: $\delta_c=470\,\Gamma'$, $\Omega=94\,\Gamma'$, $\mathcal{J}=235\,\Gamma'$. ](04_spect.pdf) The gate photon can be mapped onto a spin excitation ${\|s\rangle}$ using, e.g., electromagnetically induced transparency (EIT). It results in a spin wave excitation ${\|\psi\rangle}\sim \sum_n e^{i\,k_{\rm EIT}\,z_n} c_n\,{\|s_n\rangle}$ with an error$\sim \Gamma'/(N\Gamma_{\rm 1D})$[@douglas15mol; @fleischhauer05; @gorshkov07]. Here $k_{\rm EIT}=\pi/(2a)$ is the wavevector corresponding to the EIT transparency window, $c_n$ is a slowly-varying envelope that depends on the details of the initial pulse shape, and ${\|s_n\rangle}\equiv \sigma_{\rm sg}^n{\|g\rangle}^{\otimes N}$. Once the gate photon is mapped in, directly applying the control field $\Omega$ would cause the spin-wave to propagate and be mapped out into a photon. To “trap” this excitation inside the atomic gas, and allow an efficient interaction with the signal photon, one can first apply a spatially varying phase (e.g. by a magnetic field gradient) to shift the central wavevector, $k_{\rm EIT}\to 2\,k_{\rm EIT}$. With the appropriate amplitude shape, the resulting state can be made subradiant to the waveguide, such that it represents an eigenstate of $\mathcal{H}_{\rm wg}$ with minimal decay rate. The form of such a subradiant state ${\|\psi_d\rangle}=\sum_n c_n\,e^{ika\,n}{\|s_n\rangle}$ is illustrated in Fig.\[b\_decspec\](a) (right plots), where both the spatial amplitude $\|c_n\|$ and the amplitude in momentum space $\|c_k\|$ are plotted. To define such a state, one first identifies the eigenstate ${\|j\rangle}$ of $\mathcal{H}_{\rm wg}$ in the single excitation manifold $\{{\|e_n\rangle} \}$, with minimal collective waveguide decay rate $\Gamma_{\rm wg}^{\rm (1D)}=-2\,{\rm Im}({\langle j\|\mathcal{H}_{\rm wg}\|j \rangle})$. Out of that subradiant state the analogue one in states ${\|s\rangle}$ is obtained by substituting ${\|e_n\rangle}\to {\|s_n\rangle}$. Two degenerate such states exist (blue and red line in Fig.\[b\_decspec\](a), respectively), which are centered in momenta around $k=0$ and $k=2k_{\rm EIT}=\pi/a$. For these subradiant states the waveguide decay scales as $\Gamma_{\rm wg}^{\rm (1D)}\propto \rho_e \Gamma_{\rm 1D} N^{-3}$ with the atom number $N$. Here, $\rho_e\,\simeq (\Omega/\delta_c)^2$ is the amount of population in ${\|e\rangle}$ that mixes with ${\|s\rangle}$ due to the control field. The decay of an initial subradiant excitation is shown in Fig.\[b\_decspec\](a) (left plot) for different ratios of $\Gamma_{\rm 1D}/\Gamma'$. One generally observes that the decay of an excitation within the chain, apart from the initial time period, is essentially independent of the waveguide decay rate and given by $\Gamma_{\rm dec}\simeq \rho_e\,\Gamma'$. More generally, the collective nature of $\mathcal{H}_{\rm wg}$ splits the single excitation manifold into a few states that radiate efficiently into the waveguide and a vast number of states with subradiant character. Thus, even though the subradiant state defined above represents the optimal choice, any state with $\Gamma_{\rm wg}^{\rm (1D)}<\Gamma'$, e.g. a product state with only a single atom $m$ excited to ${\|s_m\rangle}$, reveals a comparable decay behavior dominated by the decay to free space $\Gamma'$. Having established the fidelity for the gate photon to be mapped into a spin excitation and its lifetime, we now consider its effect on a signal photon. In order to separate the finite decay rate of the spin excitation arising from the gate photon from the propagation dynamics of the signal, we first consider a system where one atom is initialized in ${\|s\rangle}$, and whose decay rates $\Gamma_{\rm 1D}=\Gamma'=0$ are set to zero. Thus it only sees the rest of the atomic chain through the band gap coupling. The response to the signal field is then calculated by considering the system response to a weak cw-input field $\tilde{\mathcal{E}}_{\rm in}(t)=\mathcal{E}_0$. This results in a reflectance spectrum as shown by the data points in Fig.\[b\_decspec\](b). The results also show good agreement with a simpler effectively linear optical model (solid lines). Here the reflection and transmission coefficients of individual atoms are alternated by hand, and the reflectance and transmittance of the entire array is obtained by a transfer matrix calculation, whose details are given in \[sec\_tramodel\].\ In the limit $\Delta\nu>(N/2)\,\Gamma_{\rm 1D}$, the s-branch spectrum consists of two resonances ($\pm$) of Lorentzian shape. The response in reflection around these resonances can be approximated by $\mathcal{R}(\Gamma_{\rm 1D}^{\rm eff}, \Gamma'_{\rm eff}, \delta_{\pm})$ (\[refl1\]). Here $\Gamma_{\rm 1D}^{\rm eff}=\epsilon_{-}^2\, \Gamma_{\rm 1D}$ and $\Gamma'_{\rm eff}=\epsilon_{-}^2\,\Gamma'$ are the effective waveguide and free space decay rates of the transition, respectively, and $\delta_{\pm}$ the resonance detuning. The parameter $\epsilon_{-}$ characterizes the amount of $e$-amplitude mixed into $s$ in the two-excitation manifold, and has been derived in section\[sec\_lshift\] to $\epsilon_{-}\simeq \Omega/(\pm\mathcal{J}-\delta_c)$. A detrimental effect, associated with the off-resonant interactions between the optical field and the far-detuned atomic sublattice, and discussed in more detail in \[sec\_crossinterf\], appears in the peak reflectance as interference dips (inset of Fig.\[b\_decspec\](b)). These dips become more prominent for larger optical depth $\mathcal{D}=2N\Gamma_{\rm 1D}/\Gamma'$ and decreasing frequency difference between the sublattice resonances. Aside from increasing the latter, they can also be suppressed by modifying the propagation phase of the bare waveguide $k\,a=\pi\,(1+\kappa)/2$, here characterized by a small adjustment parameter $\kappa$ which is specified in detail in \[sec\_crossinterf\]. Having analyzed the gate excitation lifetime and the excitation-dependent reflectance properties independently, we now consider the full dynamics of a signal pulse interacting with a (decaying) initial gate excitation. In order to exploit the reflectance properties, the pulse time $t_0$ must be small compared to the gate excitation lifetime, and its bandwidth $\propto 1/t_0$ small compared to the reflectance resonance linewidth: $\Gamma_{\rm dec}<t_0^{-1}<\Delta\omega_{\rm res}$. We illustrate the optimal signal pulse reflectance $\mathcal{R}_{\rm pulse}$ and its scaling with $N\Gamma_{\rm 1D}/\Gamma'$ in Fig.\[b\_scaling\](a), obtained for the numerically optimized pulse time $t_0^{\rm opt}$ depicted in Fig.\[b\_scaling\](b). Here we assume the $s$-branch configuration as in Fig.\[b\_decspec\](b) with an initial (and decaying) gate excitation in $s$ prepared at $t=0$, followed by the launch of a signal pulse of shape $\tilde{\mathcal{E}}_{\rm in}(t)=\mathcal{E}_0\,\sin^2(\pi\,t/(2\,t_0))$ ($0\leq t\leq 2\,t_0$, $\tilde{\mathcal{E}}_{\rm in}(t)=0$ otherwise). The reflectance $\mathcal{R}_{\rm pulse }$ is defined as the ratio of time-integrated reflected output to input intensity, where the background field leaked by the decaying excitation has been subtracted. ![\[b\_scaling\] (a) Signal pulse reflectance $\mathcal{R}_{\rm pulse}$ vs $N\Gamma_{\rm 1D}/\Gamma'$ for an initial gate excitation prepared at $t$=0 and a signal pulse $\tilde{\mathcal{E}}_{\rm in}(t)=\mathcal{E}_0\,\sin^2(\pi\,t/(2\,t_0))$ for $0\leq t\leq 2t_0$. The temporal pulse width $t_0$ has been optimized, with the optimized values shown in (b). The red triangles and red circles correspond to the signal pulse resonant with the right (broader) s-branch resonance of Fig.\[b\_decspec\](b) with phase adjustment parameter $\kappa=9\mbox{e-4}$ and $\kappa=0$, respectively. Green triangles account for the signal pulse resonant with the left (narrower) s-branch resonance ($\kappa=0$). (c) Reflected $I_r$ (red) and transmitted $I_t$ (blue) signal pulse intensity (normalized by $\|\mathcal{E}_0\|^2$) for $N\Gamma_{\rm 1D}/\Gamma'=100$ ($N$=200, $\Gamma_{\rm 1D}$=$\Gamma'/2$, $\mathcal{E}_0=\mbox{2e-4}\,\Gamma'$) on the right (broader) s-branch resonance with optimal pulse width $t_0=2.95\,\Gamma'^{-1}$. Solid and dashed lines correspond to the presence and absence of a gate excitation, respectively. ](05_scaling.pdf) As expected a larger ratio $\Delta\omega_{\rm res}/\Gamma_{\rm dec}$ improves the reflectance. Thus, centering the gate pulse around the right (broader) resonance of Fig.\[b\_decspec\](b), yields a higher reflectance as compared to the left (narrower) resonance. In Fig.\[b\_scaling\](a), one also sees the effect of the interference dips. In particular, without compensating for them, the reflectance (red circles) saturates around $77\%$. In contrast, no saturation emerges in the compensated (propagation phase adjusted) version, reaching $83\%$ reflection for $N\,\Gamma_{\rm 1D}/\Gamma'=200$ ($\mathcal{D}=400$). An example for the probe pulse propagation is illustrated in Fig.\[b\_scaling\](c), comparing the propagation in the presence and absence of an initial excitation. A clear inversion from dominant reflection in the former to transmission in the latter case is observed. Influence of dephasing {#sec_dephasing} ====================== Here we address the impact of dephasing on the conditional reflectance of a photon. The dephasing rates of atoms near photonic crystal waveguide structures are not known yet, but have been measured for a similar system consisting of atoms trapped in the close vicinity ($\sim$200 nm) of nanofibers. For these setups decoherence (dephasing) rates in the range of $\gamma\simeq 2\pi\,\left[200\,{\rm Hz} - 50\,{\rm kHz} \right]$ have been reported[@reitz13; @sayrin15] for cesium atoms, where the lowest rates correspond to the magnetically insensitive microwave clock transition[@reitz13]. Their origin has been attributed to temperature dependent ac-Stark shifts in the optical trap geometry. In order to relate these quantities to the free space $\Gamma'$ and waveguide emission rate $\Gamma_{\rm 1D}$, we assume that $\Gamma'=2\pi\cdot 4.56\,{\rm MHz}$ as reported for Cs atoms trapped along photonic crystal waveguides in[@hood16] and $\Gamma_{\rm 1D}\simeq \Gamma'$ as reported in[@goban15]. With the dephasing parameter range above, this results in $\gamma/\Gamma'\simeq\gamma/\Gamma_{\rm 1D} \simeq [4\cdot 10^{-5} - 0.01]$. ![\[b\_dephasing\] (a) $s$-branch reflectance spectrum for different dephasing rates $\gamma$, with the parameters chosen as in Fig.\[b\_decspec\] (and $\Gamma_{\rm 1D}=\Gamma'$). Solid lines correspond to the transfer-matrix prediction, data points to a numerical simulation. (b) Maximal reflectance amplitude $\mathcal{R}_{\rm max}$ for the setup as in (a) vs dephasing rate $\gamma$ for the left (red) and right (blue) reflectance resonance. The dashed line indicates the reflectance amplitude in the absence of dephasing. (c) Waveguide decay rate $\Gamma_{\rm wg}^{\rm(1D)}$ for an initial subradiant state superposition vs atom number $N$ and for different dephasing rates $\gamma$ (double logarithmic scale). Solid lines are fits to a scaling $\Gamma_{\rm wg}^{\rm(1D)}\sim N^{-\alpha}$, except for the case of $\gamma=10^{-4}\Gamma_{\rm 1D}^{\rm eff}$, which clearly does not scale polynomially with $N$. ](06_dephasing.pdf) In the following we model dephasing as uncorrelated noise, which in the framework of a Markovian (Lindblad form) master equation can be described as ($\sigma_{mm}={\|m\rangle}{\langle m\|}$) $$\label{mdeph1} \mathcal{L}^{\rm deph}[\rho]=\sum_m\frac{\gamma_m}{2}\,\left(2\sigma_{\rm mm}\rho\,\sigma_{\rm mm}-\rho\,\sigma_{\rm mm}-\sigma_{\rm mm}\,\rho \right)$$ resulting in a decoherence rate $\gamma_{mn}=(\gamma_m+\gamma_n)/2$ on the $m$ to $n$ transition. As $\gamma\ll\Gamma'$ it is sufficient to only explicitly consider dephasing on the $g$ to $s$ transition whereas additional dephasing affecting the excited state can be ignored, and thus the rates in (\[mdeph1\]) are chosen such that only $\gamma_{gs}=\gamma\neq 0$. Dephasing affects the conditional photon reflection at two stages: It modifies the reflectance spectrum as seen by a second ‘signal’ photon (i.e. Eq.(\[refl1\]) and Fig.\[b\_decspec\](b)) and it reduces the lifetime of a first ‘gate’ photon (Fig.\[b\_decspec\](a)). The impact on the reflectance spectrum has been analyzed by re-deriving the transfer-matrix model as introduced in\[sec\_tramodel\] including dephasing of the form(\[mdeph1\]). The validity of that approach is verified in Fig.\[b\_dephasing\](a) by comparing the predicted reflectance to the one obtained by a fully numerical Hamiltonian evolution, in complete analogy to the procedure in the absence of dephasing (Fig.\[b\_decspec\](b)). It illustrates that dephasing essentially reveals itself as a reduction of the reflectance amplitude. Based on the transfer-matrix model, the maximal reflectance amplitude $\mathcal{R}_{\rm max}$ dependence on the dephasing rate $\gamma$ is shown in Fig.\[b\_dephasing\](b) for the two s-branch resonances of Fig.\[b\_decspec\](b). This analysis suggests that the impact of dephasing becomes important for $\gamma/\Gamma_{\rm 1D}>10^{-2}$, whereas for the dephasing parameters stated above we do not expect significant amplitude reductions. More generally, the influence of a dephasing $\gamma$ on the reflectance resonances can be incorporated into the reflectance formula(\[refl1\]). Around a resonance $\delta=\lambda$, and assuming that $\gamma\ll\|\lambda-\delta_c\|$, the reflectance follows as ($\delta'=\delta-\lambda$) $$\label{rfdeph} \mathcal{R}(\delta')=\frac{p_\lambda^2\,[(N/2)\Gamma_{\rm 1D}]^2}{p_\lambda^2\,[(N/2)\Gamma_{\rm 1D}+\Gamma'+\gamma\,\kappa_\lambda]^2+(2\delta')^2}\,.$$ Here $p_\lambda$ denotes the $e$-population of the underlying dressed state on resonance, which determines the effective decay rates $\Gamma_{\rm 1D}^{\rm eff}=p_\lambda\Gamma_{\rm 1D}$ and $\Gamma'_{\rm eff}=p_\lambda\,\Gamma'$, i.e. $p_\lambda=\epsilon_-^2$ for the s-branch and with the definitions of Section\[sec\_lshift\]. The coefficient $\kappa_\lambda$ increases with increasing $s$-type character of the resonance, and, in the limit when the atom-atom coupling is much smaller than the control field detuning $\|\mathcal{J}\|\ll\|\delta_c\|$, is given by $$\label{rfcoeff} \kappa_\lambda=1+\frac{2\Omega^2}{(\lambda-\delta_c)^2}\,.$$ Formulas (\[rfdeph\]) and (\[rfcoeff\]) can be analytically derived out of the transfer matrix model for a chain of $N/2$ atoms with periodicity $k\,a=\pi$ and for $\mathcal{J}=0$. Moreover, (\[rfdeph\]) has been numerically verified beyond that limiting case; i.e. it also holds true for $\|\mathcal{J}\|\lesssim \|\delta_c\|$ but with a coefficient $\kappa_\lambda$ that must be numerically inferred. Thus, in regard to the reflectance spectrum, the effect of dephasing can be folded into a modification of $\Gamma'$, i.e. substituting $\Gamma'\to\Gamma'+\gamma\,\kappa_\lambda$ with a correction $\gamma\,\kappa_\lambda$, the latter expected to be small for realistic dephasing parameters. We now discuss the influence of dephasing on an initial ‘gate’ excitation. Preparing such an excitation in a subradiant state configuration – a coherent superposition – serves to increase its lifetime, i.e. makes its collective decay into the waveguide $\Gamma_{\rm wg}^{\rm(1D)}\ll\Gamma_{\rm 1D}^{\rm eff}$ subradiant and the lifetime is then dictated by the free space decay rate $\Gamma'_{\rm eff}$ alone as seen in Section\[sec\_condphoton\]. In contrast, in the absence of coherence – in the presence of strong dephasing – $\Gamma_{\rm wg}^{\rm (1D)}=\Gamma_{\rm 1D}^{\rm eff}$ and the total decay rate is given by the combination of $\Gamma'_{\rm eff}+\Gamma_{\rm 1D}^{\rm eff}$. Importantly, conditional photon reflection in the proposed setup only depends on the presence, i.e. the lifetime, of an excitation, but not on the specific state or the coherence of that excitation. Note that the fact that the subradiant state decays into free space at a rate $\Gamma_{\rm eff}'$ even without dephasing is much different than the typical case of photon storage, where a pulse would be stored indefinitely. In particular, here, the stored excitation (in state ${\|s\rangle}$) must be used to virtually populate state ${\|e\rangle}$ via the control field $\Omega$, which enables an interaction with a second incoming photon. A weak dephasing rate, such as expected from the previously quoted experimental values, is thus expected to have negligible effect on top of the already important free-space emission $\Gamma'_{\rm eff}$. The waveguide decay rate $\Gamma_{\rm wg}^{\rm (1D)}$ and its scaling with the atom number $N$ for an initially prepared subradiant state is analyzed in Fig.\[b\_dephasing\](c) for different magnitudes of the dephasing rate $\gamma$. It has been defined as the inverse 1/e-decay time. The latter is obtained out of a numerical simulation, performed based on a two-level system evolution under the full master equation with waveguide decay rate $\Gamma_{\rm 1D}^{\rm eff}$ and $\Gamma'=0$. Dephasing leads to two effects: It increases $\Gamma_{\rm wg}^{\rm (1D)}$ and it changes the particle number scaling, which in the absence of dephasing is given by $\Gamma_{\rm wg}^{\rm (1D)}\sim 1/N^3$. For an $s$-excitation as considered here, $\Gamma_{\rm 1D}^{\rm eff}\simeq (\Omega/\delta_c)^2\,\Gamma_{\rm 1D}\simeq 0.04\Gamma_{\rm 1D}$ and analogously $\Gamma'_{\rm eff}\simeq 0.04\,\Gamma'$. As seen from Fig.\[b\_dephasing\](c), for realistic dephasing rates $\Gamma_{\rm wg}^{\rm (1D)}$ remains subradiant, thus $\Gamma_{\rm wg}^{\rm (1D)}<\Gamma'_{\rm eff}$ and the excitation lifetime essentially remains unaltered by dephasing. In conclusion, only minor modifications due to dephasing are expected for both the reflectance spectrum and the ‘gate’ excitation lifetime and thus also for the conditional reflectance of photons. The main limiting mechanism is given by the emission into free space; for the reflectance dephasing can be ascribed to a (small) correction to that rate. Conclusions {#sec_conclusion} =========== In summary, we have proposed a system, in which a single excitation (‘gate photon’) controls the propagation of a subsequent photon by effectively modifying the topology of atoms along a PCW. In particular, its presence manifests itself as an alternation of atomic energy levels, such that subsequent photons can be made to ‘see’ only every other atom. This property has a tremendous impact on the propagation of photons, e.g., switches the system from being transmissive to highly reflective in the absence and presence of a gate excitation, respectively. We have shown how this can be observed in two-photon correlations, where conditional reflectance reveals itself as anti-bunching. Our proposed mechanism involves a straightforward way to both create and probe a photon nonlinearity of alternating type within a single platform, a combination that makes the PCW setup unique and enables the access to novel types of nonlinearities. The consequences of such an alternating nonlinearity for many photons are yet unknown, and might inspire the active search for the creation and understanding of many-body states of light. Moreover, while we have only considered the limit of infinite-range interactions ($L\rightarrow\infty$) here, it would be interesting in the future to explore finite-range interactions, where a photon only changes the topology of its immediate surroundings. It would also be interesting to consider analogous effects in higher dimension[@tudela15]. The authors thank H.J. Kimble and J.S. Douglas for useful discussions. The work was supported by the ERC Starting Grant FOQAL, the MINECO Plan Nacional Grant CANS, the Spanish Ministry of Economy and Competitiveness through the “Severo Ochoa” Programme for Centres of Excellence in R&D (SEV-2015-0522), the Fundacio Cellex, and the CERCA Programme/Generalitat de Catalunya. Transfer matrix model for the reflectance and transmittance spectra {#sec_tramodel} =================================================================== In Fig.\[b\_decspec\](b) we have illustrated the reflectance and transmittance spectra, both in the presence and absence of a ‘gate’ excitation in the system. It has been shown that the spectra obtained by a fully numerical simulation can be well-approximated by a linear optical transfer-matrix model. Here we outline in detail how that transfer-matrix model calculation is performed. ![\[b\_tmodel1\] (a) Transfer matrix description for the field calculation. The input and output fields $\mathcal{E}$ are connected by a series of matrices $\mathcal{M}$ representing the scattering at an atomic site $\mathcal{M}_{\rm scatt}$ and free propagation sections $\mathcal{M}_{\rm free}$ in-between. (b) In the presence of a ‘gate’ excitation, the scattering at a single atomic site is modeled by an atom coupled to a non-dissipative ancilla via the atom-atom interaction $\pm\mathcal{J}$. The atomic steady state coherence ${\langle \sigma_{\rm ge} \rangle}={\langle \sigma_{\rm ge}^{\rm atom} \otimes \mathbbm{1}_{\rm ancilla} \rangle}$ determines the reflection coefficient $r$ at the atomic site. (c) Reflectance (red squares) in the presence of a gate excitation, and transmittance (green triangles) in its absence for the field tuned close to two-photon ($\delta-\delta_c\sim 0$) resonance ($s$-branch, upper panel) and single-photon ($\delta\sim 0$) resonance ($e$-branch, lower panel). In the former case, conditional reflectance arises from transitions to states ${\|ss\rangle}$, in the latter case from transitions to states ${\|ee\rangle}$. The transfer matrix model predictions are given by the solid orange and green line, respectively. Parameters: $N=100$, $\delta_c=470\,\Gamma'$, $\Omega=94\,\Gamma'$, $\mathcal{J}=235\,\Gamma'$. ](07_tmatrix.pdf) For the propagation of photons in one dimension, a very elegant way to obtain linear reflectance and transmittance spectra is given by the transfer matrix approach[@chang12; @deutsch95]. In that model, the propagation of photons is modeled by free propagation sections intersected by a series of scattering events at the (assumed point-like) atomic positions (see Fig.\[b\_tmodel1\](a)). The input and output fields are then related by two-dimensional matrices $\mathcal{M}$ $$\fl\left(\begin{array}{{20}{c}}\mathcal{E}_+^{r} \\ \mathcal{E}_-^{r} \end{array}\right) = \mathcal{M}_{\rm tot}\left(\begin{array}{{20}{c}} \mathcal{E}_+^l\\\mathcal{E}_-^l \end{array}\right) \quad \mathrm{with}\quad \mathcal{M}_{\rm tot}=\mathcal{M}_{\rm scatt}^{(N)}\,\mathcal{M}_{\rm free}\,\mathcal{M}_{\rm scatt}^{(N-1)}\dots\mathcal{M}_{\rm scatt}^{(1)}\mathcal{M}_{\rm free}$$ where $+$ and $-$ denote the right (‘transmitted’) and left (‘reflected’) propagating field and $l$ and $r$ denote the left input (beginning of atom chain) and right output (chain end) position, respectively. The scattering matrices for atoms $1$ to $N$ are of the form $$\label{mscatt} \mathcal{M}_{\rm scatt}=\frac{1}{t}\left(\begin{array}{{20}{c}} t^2-r^2 & r \\ -r & 1 \end{array}\right)$$ with $t$ and $r$ being the frequency-dependent reflection and transmission coefficients associated with a single atom. The free propagation of distance $a$ imprints a pure phase and is described by $$\label{mfree} \mathcal{M}_{\rm free}=\left(\begin{array}{{20}{c}}e^{ika} & 0\\ 0 & e^{-ika} \end{array}\right)\,.$$ In principle $M_{\rm free}$ is also frequency-dependent, through the frequency dependence of the wavevector $k$. Neglecting that dependence, i.e., evaluating $k$ at the atomic transition frequency, corresponds to a Markov approximation, which holds true for realistic atom numbers[@chang12; @guimond16]. For our system under consideration, we thus have $ka=\pi/2$ or $k\,a=(\pi/2)\,(1+\kappa)$ with a small adjustment $\kappa\ll 1$ (see \[sec\_crossinterf\]). The total reflected and transmitted field coefficients then follow as $r_{\rm tot}=\mathcal{M}_{\rm tot}^{12}/\mathcal{M}_{\rm tot}^{22}$ and $t_{\rm tot}=1/\mathcal{M}_{\rm tot}^{22}$, respectively, with the numerical indices referring to the matrix elements of $\mathcal{M}_{\rm tot}$. The reflectance (transmittance), defined as the ratio of reflected (transmitted) to input field intensity is then given by $\mathcal{R}=\|r_{\rm tot}\|^2$ ($\mathcal{T}=\|t_{\rm tot}\|^2$). Thus it remains to calculate the transmission and reflection coefficients $t$ and $r$. They can be obtained by solving for the steady state of the spin-model Hamiltonian $\mathcal{H}$ for a single atom. Based on the input-output relations[@chang12; @caneva15] (see also Section\[sec\_config\]) $ \mathcal{E}_{+} =\mathcal{E}_0+i\,\frac{\Gamma_{\rm 1D}}{2}\,\sigma_{\rm ge}$ and $\mathcal{E}_-=i\,\frac{\Gamma_{\rm 1D}}{2}\,\sigma_{\rm ge}$ and noting that $\mathcal{E}_+=t\,\mathcal{E}_0$, $\mathcal{E}_-=r\,\mathcal{E}_0$ with the input field $\mathcal{E}_0$, the coefficients follow as $$\label{coeff1} r=i\,\frac{\Gamma_{\rm 1D}}{2}\frac{1}{\mathcal{E}_0}\,{\langle \sigma_{\rm ge} \rangle}^{[\rm 1st]}, \quad t=1+r \,.$$ Here ${\langle \sigma_{\rm ge} \rangle}^{[\rm 1st]}$ denotes the steady state expectation value to first (linear) order in $\mathcal{E}_0$. In the absence of an initial ‘gate’ excitation the atomic scattering matrices are equal, i.e. $\mathcal{M}_{\rm scatt}^{(k)}\equiv\mathcal{M}_{\rm scatt}$. The transmission and reflection coefficients follow out of $$\label{hamlinear}\fl \mathcal{H}_{\rm atom}=-\left(\delta+i\frac{\Gamma'+\Gamma_{\rm 1D}}{2}\right)\,\sigma_{\rm ee}-(\delta-\delta_c)\,\sigma_{\rm ss}-\Omega\,(\sigma_{\rm es}+\mathrm{h.c.})-\mathcal{E}_0\,(\sigma_{\rm eg}+\mathrm{h.c.})$$ and (\[coeff1\]), and for the three-level atom are given by the known coefficients[@chang11] $$\label{refl3} r=-\frac{\Gamma_{\rm 1D}\,(\delta-\delta_c)}{(\Gamma_{\rm 1D}+\Gamma'-2 i\,\delta)\,(\delta-\delta_c)+2i \,\Omega^2}\quad \mathrm{and }\,\, t=1+r\,.$$ A priori, the case where a ‘gate’ excitation is present seems beyond what can be captured by the transfer matrix approach for linear optics, since formally the response to a subsequent signal photon involves the atomic two-excitation manifold. However, we can construct a simpler artificial model (see Fig.\[b\_tmodel1\](b)), in which the ‘gate’ excitation is held in an ancilla atom that is decoupled from the waveguide – thus it does not decay nor do photons in the waveguide directly excite this ancilla. In this model, the ancilla atom interacts via a photonic band gap, i.e. via the alternating atom-atom coupling $\pm\mathcal{J}$, to an ‘actual’ atom that does couple to the waveguide. We can therefore calculate the linear optical properties of the actual atom, which is modified by the ancilla, and we find that our artificial model captures well the dynamics of the full system of interest. In that case, $\mathcal{M}_{\rm scatt}^{(2k)}=\mathcal{M}_{\rm scatt}^{[-\mathcal{J}]}$ and $\mathcal{M}_{\rm scatt}^{(2k+1)}=\mathcal{M}_{\rm scatt}^{[+\mathcal{J}]}$ for $k\in \mathds{N}_0$ with $\pm \mathcal{J}$ referring to the sign of the alternating atom-atom coupling constant. The Hamiltonian is described by $\mathcal{H}=\mathcal{H}_{\rm atom}+\mathcal{H}_{\rm ancilla}+\mathcal{H}_{\rm int}$, denoting the system, two-level ancilla and interaction Hamiltonian, respectively. The atom (system) part $\mathcal{H}_{\rm atom}$ is given by(\[hamlinear\]), the ancilla Hamiltonian takes the form $$\mathcal{H}_{\rm ancilla} =-\delta\,\sigma_{\rm ee}^{(2)}-(\delta-\delta_c)\,\sigma_{\rm ss}^{(2)}-\Omega\,(\sigma_{es}^{(2)}+\mathrm{h.c.})$$ and the atom-atom interaction between both reads $\mathcal{H}_{\rm int}=\pm\mathcal{J}\,\sigma_{es}^{(1)}\otimes\sigma_{se}^{(2)}+\mathrm{h.c.}$. Here the indices $(1)$ and $(2)$ denote the atom and ancilla, respectively. The reflection and transmission coefficients are obtained out of (\[coeff1\]), by solving for the steady state expectation value ${\langle \sigma_{\rm ge} \rangle}^{[\rm 1st]}={\langle \sigma_{\rm ge}^{(1)}\otimes \mathbbm{1}^{(2)} \rangle}^{[\rm1st]}$ to first (linear) order in $\mathcal{E}_0$. Thereby the initial excitation (in $e$ or $s$) is set by setting the populations in the coupled optical Bloch equations to the values $$\begin{aligned} \eqalign{ {\langle {\|g\rangle}_1{\langle g\|}\otimes {\|e\rangle}_2{\langle e\|} \rangle} &=1 \quad \mbox{(e-branch)}\\ {\langle {\|g\rangle}_1{\langle g\|}\otimes {\|s\rangle}_2{\langle s\|} \rangle} &=1 \quad \mbox{(s-branch)} }\end{aligned}$$ and all remaining populations to zero. Setting the populations to numerical values in the equations explicitly assumes them to be constant to a good approximation. This limits the validity of the formula to $(\Omega/\delta_c)^2\ll1$, i.e. to a regime in which the bare states ${\|e\rangle}$ and ${\|s\rangle}$ form a good basis of the coupled system. Moreover the population exchange between system and ancilla must be negligible, that is the approach is not valid for a signal field excitation to state $s$ with an initial (ancilla) ‘gate’ $e$-excitation and vice versa. In Fig.\[b\_tmodel1\](c) the transfer matrix model predictions for the reflectance and transmittance spectra (solid lines) are compared to the results of a full numerical simulation (data points) of the spin-model Hamiltonian. The latter is obtained by numerically solving for the steady state reflectance and transmittance for a cw-input field (see also the upcoming\[sec\_numimp\]), where the initial state is set to all atoms being in the ground state (no initial ‘gate’ excitation) or to an excitation of a single atom to ${\|e\rangle}$ or ${\|s\rangle}$ whose decay rates $\Gamma', \,\Gamma_{\rm 1D}$ are set to zero (initial ‘gate’ excitation). Good agreement of the approaches is obtained, even when the parameters are not chosen to reside deep in the dispersive regime, e.g. $\Omega/\delta_c\simeq 0.2$, $\mathcal{J}/\delta_c\simeq 0.5$. Reflectance dependence on the band gap interaction parameter $\mathcal{J}$ {#sec_Jdependence} ========================================================================== In Fig.\[b\_decspec\](b) of Section\[sec\_condphoton\], a large atom-atom coupling parameter $\mathcal{J}$ (as defined in (\[hambg\])) has been chosen ($\mathcal{J}=235\,\Gamma'$) in order to achieve a large relative splitting between the two reflectance resonances. Only in such a regime are the two resonances of the Lorentzian form(\[refl1\]). Moreover, in that case the reflectance resonance frequencies are shifted away significantly from the resonances in the absence of a gate excitation, such that perfect transmission at these frequencies is obtained in the latter case. By partly relaxing at least one of these two requirements, we here show that it is possible to obtain reflectances of comparable linewidth even for signicantly lower coupling parameters $\mathcal{J}$. In Fig.\[b\_Jspect\], reflectance spectra (red lines, in the presence of a gate excitation) are shown for different magnitudes of $\mathcal{J}$. The ratios of $\Omega/\delta_c=0.2$ and $\mathcal{J}/\delta_c=0.5$ are kept fixed and correspond to the ones in Fig.\[b\_decspec\]. This implies that both the excitation lifetime Fig.\[b\_decspec\](a) and the reflectance linewidths – up to disturbances and overlaps – remain unaltered. For decreasing $\mathcal{J}$ the two resonances overlap, and finally form a common resonance around the resonance absent a gate excitation (dip position of the blue lines, which represent the transmission without any prior excitations). Moreover, disturbances and reflectance dips due to interference with the off-resonant sub-lattices become more prominent for smaller $\mathcal{J}$ and thus smaller resonance separations. However, this latter effect can be compensated for by adjusting the phase factor $k\,a =\pi/2\,(1+\kappa)$ with $\kappa$ a suitable correction factor, as further discussed in\[sec\_crossinterf\]. Such an adjustment works very well to reconstruct one of the resonances (black lines, here performed for the right resonance), even in regimes of large overlaps and disturbances (e.g., $\mathcal{J}=25\,\Gamma'$). Thus, suitable reflectance resonances can be constructed even for coupling parameters $\mathcal{J}$ that are an order of magnitude lower than in the well-separated Lorentzian configuration of Fig.\[b\_decspec\](b). ![\[b\_Jspect\] Reflectance for different band gap coupling strengths $\mathcal{J}$. Here, the ratios $\Omega/\delta_c=0.2$ and $\mathcal{J}/\delta_c=0.5$ are kept constant and correspond to the ones in Fig.\[b\_decspec\](b). Red lines and circles depict the reflectance $\mathcal{R}$ in the presence of a gate excitation, blue lines the transmittance $\mathcal{T}$ in the absence of the latter. Black lines and triangles correspond to an optimization of the right resonance by the phase adjustment parameters (a) $\kappa=2\cdot 10^{-3}$, (b) $\kappa = 4\cdot 10^{-3}$ and (d) $\kappa =8.5\cdot 10^{-3}$. Solid lines are obtained out of the transfer matrix model, data points (circles, triangles) correspond to numerical data points obtained as described in Section\[sec\_condphoton\]. ](08_spectJ.pdf) Cross-interference reflectance dips and their compensation {#sec_crossinterf} ========================================================== An atomic lattice with lattice constant $ka=\pi$ ideally produces a Lorentzian reflectance spectrum as described by(\[refl1\]). Then in the case of band-gap interactions, we argued that an array of atoms with lattice constant $ka=\pi/2$ effectively appears as two sub-lattices each of lattice constant $ka=\pi$ for a signal field, when a ‘gate’ excitation is first stored. For low atom number, the reflectance near the resonance of one sub-lattice fits well with a Lorentzian. For larger atom number, however, the off-resonant response of the other sublattice cannot be ignored, and results in a significant interference dip in the reflectance (see inset of Fig.\[b\_decspec\](b)). Here we analyze the origin of the reflectance dip in the transfer matrix formalism introduced in \[sec\_tramodel\], and discuss how it can be eliminated by a slight change in the inter-atomic propagation phase, $ka=(\pi/2)(1+\kappa)$. A repetitive unit in the transfer matrix description consists of a scattering event at the (quasi-) resonant atom $\mathcal{M}_{\rm scatt}$(\[mscatt\]), followed by a free propagation $\mathcal{M}_{\rm free}$(\[mfree\]) of distance $a$, an off-resonant scattering $\mathcal{M}_{\rm scatt}^{\rm off}$ (at the off-resonant sublattice), and another free propagation section $\mathcal{M}_{\rm free}$. For weak probe fields and large detunings, the individual off-resonant $e$- and $s$-transitions contribute independently and the scattering matrix can be written as $\mathcal{M}_{\rm scatt}^{\rm off}\simeq\mathcal{M}_{\rm scatt}^e\,\mathcal{M}_{\rm scatt}^{s}$ with $\mathcal{M}_{\rm scatt}^\mu$ the two-level scattering matrix of transition $\mu$. The two-level reflection and transmission coefficients are given by Eq.(\[refl3\]) with $\Omega=\delta_c=0$, $\Gamma_{\rm 1D}\to\Gamma_{\rm 1D}^\mu$, $\Gamma'\to\Gamma'_\mu$ and $\delta\simeq\Delta_\mu$. Here $\Gamma_{\rm 1D}^\mu$ and $\Gamma'_\mu$ correspond to the effective decay rates of transition $\mu$. The detuning $\Delta_\mu$ denotes the field detuning to the off-resonant sub-lattice resonance $\mu$. As we assume the probe field to be tuned to a particular reflectance resonance, whose reflectance dip behavior we seek to analyze, $\Delta_\mu$ simply corresponds to the reflectance peak separation. Out of (\[mscatt\]) and for large detunings $\Delta_\mu$, that is, to first order in $(\Gamma'+\Gamma_{\rm 1D})/\Delta_\mu$, this results in $$\label{mscattoff} \mathcal{M}_{\rm scatt}^{\mu}\simeq \left(\begin{array}{{20}{c}} 1-i\,\kappa_\mu\,\frac{\pi}{2} & -i\kappa_\mu\frac{\pi}{2} \\ i\kappa_\mu\,\frac{\pi}{2} & 1+i\kappa_\mu\,\frac{\pi}{2} \end{array}\right)$$ where we defined $\Gamma_{\rm 1D}^\mu/(2\,\Delta_\mu)=\kappa_\mu\,\pi/2$. The idea is to compensate for the disturbance introduced by $\mathcal{M}_{\rm scatt}^{\mu}$ through adjusting the free space propagation phase as $k\,a=\pi\,(1+\kappa)/2$. Introducing that phase into the free propagation matrix $\mathcal{M}_{\rm free}$ (\[mfree\]) results in $$\mathcal{M}_{\rm free}\simeq\left(\begin{array}{{20}{c}} e^{i\frac{\pi}{2}}\,(1+i\kappa\,\frac{\pi}{2}) & 0 \\ 0 & e^{-i\frac{\pi}{2}}\,(1-i\,\kappa\,\frac{\pi}{2}) \end{array}\right)$$ where for consistency the adjustment parameter has been expanded to first order. A single repetitive unit thus follows as $\mathcal{M}_{\rm unit}=\mathcal{M}_{\rm free}\,\mathcal{M}_{\rm scatt}^{\rm off}\,\mathcal{M}_{\rm free}\,\mathcal{M}_{\rm scatt}$. Note that if one removes the contribution of the off-resonant sub-lattice (which can be achieved by setting $\kappa_\mu=0$ in $M_{\rm scatt}^{\rm off}$), and if one also sets $\kappa=0$, then $ M_{\rm unit}$ describes a unit cell of the ideal atomic mirror configuration. This yields the Lorentzian reflection(\[refl1\]). To recover the ideal situation, our goal is then to choose $\kappa$ such that it cancels the presence of $\kappa_\mu$ to lowest order. It can readily be shown that $$\label{form_kappa} \kappa=\kappa_e+\kappa_s\quad \mathrm{with} \quad \kappa_\mu=\frac{\Gamma_{\rm 1D}^\mu}{\pi\,\Delta_\mu}\,.$$ That is, both the off-resonant sublattice $e$- and $s$-transition can contribute to the formation of reflectance dips. Besides the adjustment of the inter-atomic propagation phase by $\kappa$, reducing its magnitude e.g. by increasing the detunings $\Delta_\mu$ can be utilized to suppress the dip formation. As shown in Fig.\[b\_scaling\](a), an adjustment of the inter-atomic phase by $\kappa$, which eliminates the dips in the reflectance resonances, prevents the reflectance of a probe pulse from saturating with an increasing optical depth. Numerical reflectance and transmittance calculation {#sec_numimp} =================================================== We here illustrate the procedure used for numerically calculating the reflectance and transmittance of a probe field $\mathcal{E}_{\rm in}(z,t)$. It is based on the spin-model description[@caneva15], i.e. a time evolution of the atomic (spin) state ${\|\psi\rangle}$ under an effective non-Hermitian Hamiltonian $\mathcal{H}$, and the subsequent reconstruction of the field based on input-output relations (see Section\[sec\_config\] for the effective spin Hamiltonian and the general form of the input-output operators). In order to reduce the numerical complexity, we truncate the atomic wavefunction ${\|\psi\rangle}$ to a maximum of two excitations in $e$ or $s$, i.e. we keep only the ground state ${\|g_0\rangle}={\|g\rangle}^{\otimes N}$ with all atoms in $g$, the single-excitation states {${\|e_m\rangle}$, ${\|s_m\rangle}$} with one atom (e.g. atom $m$) excited, and the two-excitation states {${\|e_m e_n\rangle}$, ${\|s_m s_n\rangle}$, ${\|e_m s_n\rangle}$} with two excited atoms (e.g. $m$ and $n$). For that approximation to be valid, i.e. for higher order excitations to be negligible, we restrict the probe field $\mathcal{E}_{\rm in}(z,t)$ to weak coherent fields of photon number $\bar{n}\ll1$. The wavefunction is initialized in the ground state ${\|\psi(0)\rangle}={\|g_0\rangle}$ or in a single ‘gate’ excitation state ${\|\psi(0)\rangle}={\|s'\rangle}$ ($s$-branch) or ${\|\psi(0)\rangle}={\|e'\rangle}$ ($e$-branch). These latter states can be superpositions of single excitation states, e.g. the subradiant state configuration introduced in section\[sec\_condphoton\], or non-decaying ancilla states as used for the reflectance spectra calculations. Subsequently the time evolved atomic state ${\|\psi(t)\rangle}$ is numerically calculated based on the Schrödinger equation $(\mathrm{d}/\mathrm{d}t){\|\psi(t)\rangle}=-i\,\mathcal{H}\,{\|\psi(t)\rangle}$. The reflected $I_r$ and transmitted $I_t$ field intensities at a specific time $t$ and position $z$ then follow from the input-output relations. For example, the transmitted field $\mathcal{E}_+(z_e,t)=\mathcal{E}_{\rm in}(z_e,t)\,\mathbbm{1}+i\,(\Gamma_{\rm 1D}/2)\,\sum_{m=1}^N\sigma_{ge}^m\,e^{i\,k\,(z_e-z_m)}$, where $z_e=N\,a$ denotes the end position of the atomic chain. Thus, the transmitted intensity follows as $$\label{fint} I_t(z_e,t)={\langle \psi(t)\|\,\mathcal{E}_+(z_e,t)^\dagger\,\mathcal{E}_+(z_e,t)\,\|\psi(t) \rangle}$$ and involves the calculation of two-body atomic correlations of the form ${\langle \psi(t)\|\,\sigma_{eg}^m\sigma_{ge}^n\,\|\psi(t) \rangle}$. The reflected field intensity follows analogously by simply replacing the field operators with the ones of inverse propagation $\mathcal{E}_-(z=0,t)=i\,(\Gamma_{\rm 1D}/2)\sum_{m=1}^N\sigma_{ge}^m e^{ikz_m}$. The transmittance $\mathcal{T}$ then follows by normalizing the transmitted intensity by the input intensity. For a cw input field $\mathcal{E}_{\rm in}(z,t)=\mathcal{E}_0\,e^{i\,(kz-\delta\,t)}$, e.g. for the calculation of the transmittance spectra, this is performed by evaluating (\[fint\]) in the steady state regime (i.e. at a time when $I_t$ reaches a constant value $I_t^{\rm ss}$ in time), and the transmittance then follows as $\mathcal{T}=I_t^{\rm ss}/\|\mathcal{E}_0\|^2$. For a signal pulse, e.g. $\mathcal{E}_{\rm in}(z,t)=\mathcal{E}_0\,\sin^2(\pi\,t/(2 t_0))\,e^{i\,(k z-\delta t)}$, it is defined by $\mathcal{T}=(\int I_t(z_e,t')\,\mathrm{d}t')/(\int \|\mathcal{E}_{\rm in}(t')\|^2\,\mathrm{d}t')$, where in addition the background field of the dissipative ‘gate excitation’ is subtracted as discussed below. The reflectances $\mathcal{R}$ and higher order correlations as $g^{(2)}(0)$ are calculated analogously. As the time evolution follows from the Schrödinger equation in the wavefunction picture, the dissipative nature, i.e. the emission into free space ($\Gamma'$) and into the waveguide ($\Gamma_{\rm 1D}$), is only accounted for by the norm decay of the wavefunction evolving under the non-Hermitian Hamiltonian $\mathcal{H}$. In particular, quantum jumps, e.g. a jump from ${\|e\rangle}$ to ${\|g\rangle}$ by the emission of a photon, are not correctly accounted for. Thus, either the occurrence of these jumps must be infrequent, which implies that the wavefunction norm decay on the timescale of interest is small, or the state after a jump event (e.g. ${\|g\rangle}$) must have a negligible impact on the observables of interest $\mathcal{R}$ and $\mathcal{T}$. ![\[b\_bill\] Reflected $I_r$ (red) and transmitted $I_t$ (blue) signal pulse intensity in time for a signal pulse $\tilde{\mathcal{E}}_{\rm in}(t)=\mathcal{E}_0\,\sin^2(\pi t/(2 t_0))$ \[$0\leq t\leq 2\,t_0$\] centered around the broader $s$-resonance of Fig.\[b\_decspec\](b) and for an initial ‘gate’ excitation in $s$ prepared at $t=0$. The intensities are normalized by $\|\mathcal{E}_0\|^2$. (a) Including the background field leaked by the initial excitation. (b) With subtracted background field as explained in the main text. Gray lines represent the background field intensity alone. Parameters: $N=200$, $\Gamma_{\rm 1D}=0.5\,\Gamma'$, $\delta_c=470\,\Gamma'$, $\Omega=94\,\Gamma'$, $\mathcal{J}=235\,\Gamma'$, $\delta=514.9\,\Gamma'$, $t_0=2.95\,\Gamma'^{-1}$, $\mathcal{E}_0=2\,$[e-4]{}$\Gamma'$ and $\kappa=8\,$[e-4]{}. ](09_bgfield.pdf) For a weak coherent field input and in the absence of a ‘gate’ excitation, the total excited state population $\wp_{\rm ex}$ remains small and the occurrence of quantum jumps (e.g. with rate $\sim\wp_{\rm ex}\,\Gamma'$) can be neglected on the considered timescale (e.g. the time to reach a steady state that forms on a timescale $\sim \Gamma'^{-1}$). The case of an initial ‘gate’ excitation is more subtle as it involves a full excitation, whose decay is generally non-negligible. In our specific case, where our observable of interest is reflection of a signal field, including jumps would yield a negligible change. In particular, a jump to ${\|g_0\rangle}$ causes the system to be almost perfectly transmitting. This originates from the fact that the resulting state following the jump corresponds to a configuration far-detuned from the probe field, which is assumed to be tuned to the two-excitation reflectance resonance. Therefore the field then does not interact with the atomic lattice any more and thus does not contribute to the reflectance signal. The decay of reflectance associated with the gate excitation decay is correctly accounted for by the wavefunction norm decay and does not require one to consider jump events. Despite the ‘gate’ excitation being of subradiant nature with respect to the waveguide emission, the field associated with its decay, named ‘background field’ in the following, can still be significant compared to the field of a weak-coherent ‘signal’ pulse. In contrast, for a true single photon pulse it would be negligible in magnitude. As all of our calculations are performed with weak coherent input states for the signal, we must subtract out the part of the intensity that arises from leakage of the gate into the waveguide. The background subtracted pulse reflectance can be obtained in the following way: Calculating the output field intensity in the absence of a ‘signal’ pulse, i.e. for $\mathcal{E}_{\rm in}\equiv0$ and an initial gate excitation e.g. ${\|\psi(0)\rangle}={\|s'\rangle}$, and subtracting that intensity from the combined intensity in its presence ($\mathcal{E}_{\rm in}\neq0$, ${\|\psi(0)\rangle}={\|s'\rangle}$). The method has been found successful in recovering the expected pulse shapes and scalings even for background fields that are orders of magnitude larger than the reflected ‘signal’ field. As an example, we illustrate in Fig.\[b\_bill\] the reflected and transmitted intensity of a ‘signal’ pulse in the presence of a decaying ‘gate’ excitation in $s$. In Fig.\[b\_bill\](a) the total combined intensity of signal pulse field and the field leaked by the gate excitation is shown. The signal pulse shape can be identified on top of a fast-oscillating background, where the oscillations are due to the non-eigenstate nature of the initial excitation. The background subtracted intensities are depicted in Fig.\[b\_bill\](b) along with the background field alone. A clear separation of these individual components is obtained; both the oscillating nature is removed for the background subtracted pulse intensities and a proper normalization is observed, i.e. the intensity is (approaches) zero at the beginning and end of the signal pulse. References {#references .unnumbered} ==========	{ "pile_set_name": "ArXiv" }
--- abstract: 'We introduce new observables for the study of the inclusive same sign dileptons production at LHC which are built out of ratios of the observed number of same-sign dileptons, both with same $N(\ell,\ell)$ and different flavor $N(\ell,\ell'')$. As a case study we apply them to the stau coannihilation region of the constrained minimal supersymmetric standard model. We show that the new variables depend rather mildly on the center of mass energy and how these can be used to constraint the parameter space in the $(m_{1/2},\tan\beta)$ plane.' author: - 'M. Cannoni' - 'O. Panella' - 'M. Pioppi' - 'M. Santoni' title: 'Probing dark matter and CMSSM with same-sign dilepton searches at the LHC' --- #### Introduction. {#introduction. .unnumbered} The starting of the LHC era will allow us to finally shed light into the last missing piece of the standard model (SM), the Higgs boson, and hopefully to probe the supersymmetric (SUSY) extension of the SM that represents the most popular solution to the hierarchy problem, gauge coupling unification and the nature of dark matter. In the minimal supersymmetric standard model (MSSM) with R-parity conservation and in the mSUGRA inspired constrained version (CMSSM), the lightest neutralino is the lightest SUSY particle (LSP), neutral and stable and one of the favored dark matter (DM) candidates [@bertonerev]. In the framework of the SM, events in proton-proton ($pp$) collisions with two isolated same-sign leptons in the final state, or same-sign dileptons (SSD), are very rare. They may come from double gauge boson production $WZ$, $WW$ and decays, double parton scattering or $t\bar{t}W$, the last processes yet to be observed in proton-proton collisions. This makes this signature very natural to look for new physics. Isolated SSD represent also a standard search channel for SUSY models [@SSD]. The main processes that lead to inclusive final states SSD in proton-proton collisions are gluino pair production $pp\to\glui\glui $, gluino-squark associate production $pp \to \glui \squa$ and squark pair production $p p \to \squa \squa'$. Cascade decays from these pairs produce easily $\chi^{\pm}_1\chi^{\pm}_1$, $\chi^{\pm}_1\chi^{0}_2$, $\chi^{0}_2\chi^{0}_2$ that decay and eventually lead to SSD that can be of the same flavor (SFSSD) $e^\pm$$e^\pm$, $\mu^\pm\mu^\pm$, $\tau^\pm\tau^\pm$ or of different flavor (DFSSD), $e^\pm$$\mu^\pm$, $\mu^\pm\tau^\pm$, $e^\pm\tau^\pm$. The signal was searched for by CDF at Tevatron [@CDFlsd] and both the CMS [@CMSlsd] and ATLAS [@ATLASlsd; @ATLASgg+ssd] collaborations have already performed searches for these particular class of events in the data sample of the final 2011 run with the LHC working at the center of mass energy $\sqrt{s}=7$ TeV and a total integrated luminosity of about 5 fb$^{-1}$. No evidence for new physics was found and upper bounds on the number of SSD were used to set constraints on the parameter space of the CMSSM. In this brief report we show that by simply counting the number of SSD pairs $N(ee)$, $N(\mu\mu)$, $N(\tau\tau)$, $N(e\mu)$, $N(e,\tau)$, $N(\mu\tau)$ it is possible to obtain direct information on fundamental CMSSM parameters like, for example, $m_{1/2}$ and $\tan\beta$. The proposed variables, that are build from the number of SSD, may give preliminary informations on the fundamental parameters of the underlying theory. Once these are known, one has a guide to which decay chain are likely to show up in the data yields and thus measure the masses of the SUSY particles. As a practical example we study the SSD signature in connection with the so-called stau co-annihilation region () of the CMSSM parameter space that is of interest for dark matter searches. #### Observables for the SSD signal. {#observables-for-the-ssd-signal. .unnumbered} Let us consider gluino pair production. The gluino is a Majorana particle and decays with equal branching fractions ($\cal{B}$) to particles and antiparticles; this property allows to have two same sign charginos from the decay chains of the pair of sparticles produced in the $pp$ collisions. The two same-sign charginos lead to final states with SSD, neutrinos and LSP. Under this assumption, we introduce the new observables as follows. The cross section can be approximated by $\sigma(pp\to 2\chi_{1}^{+} +X)\simeq \sigma(pp\to \tilde{g}\tilde{g}) \times {\cal B}(\tilde{g}\to q\tilde{q})^2 \times {\cal B}(\tilde{q}\to {q'}\chi_{1}^{+})^2$. We synthetically call this cross section $\sigma_{\chi\chi}^{\tilde{g}\tilde{g}}$ and the branching ratios corresponding to the various chargino’s decay chains leading to a lepton plus undetected particles (neutrinos and LSP) plus hadronic jets, $\chi^{\pm}_1 \to \ell +X$ with ${\cal B}_{i,\ell}$, $\ell =e,\mu,\tau$. For a given integrated luminosity, the number of SFSSD is estimated as $N(\ell \ell) \propto \sigma_{\chi\chi}^{\tilde{g}\tilde{g}} \times ( \sum_{i} {\cal B}_{i,\ell})^2$, while the number of DFSSD is instead $N(\ell \ell') \propto 2 \sigma_{\chi\chi}^{\tilde{g}\tilde{g}} (\sum_{i} {\cal B}_{i,\ell})\times (\sum_{i} {\cal B}_{i,\ell'})$. The factor 2 takes into account the fact that the leptons come from two identical charginos. In reason of the expected similar behavior of the first two lepton generations and the peculiar role held by the leptons and sleptons of the third family, we consider the ratios: $$\begin{aligned} \frac{N(l \tau)}{N(\tau \tau)}=2R, \qquad \frac{N(l\, l)}{N(\tau \tau)}=R^2 \label{NSF} \label{eq:1}\end{aligned}$$ with $l=e,\;\mu$ and $$R=\frac{{\sum_{i} {\cal B}_{i,l}}}{{\sum_{i} {\cal B}_{i,\tau}}}. \label{eq:2}$$ We remark that when all contributing mechanisms are considered the total cross section ($\sigma_{\chi\chi}= \sigma_{\chi\chi}^{\tilde{g}\tilde{g}}+ \sigma_{\chi\chi}^{\tilde{q}\tilde{g}} + \sigma_{\chi\chi}^{\tilde{q}\tilde{q}}\dots $) drops out exactly. Considering the relations , we finally define the variables: $$\begin{aligned} \label{eqN1} N_1 &=& \frac{1}{2}\,\frac{N(e \tau)+N(\mu \tau)}{N(\tau \tau)}\, \approx 2 R\label{N1}, \label{eq:3}\\ N_2 &=& \frac{1}{4}\,\frac{N(ee)+N(\mu\mu)+N(e\mu)}{N(\tau \tau)}\, \approx R^2\label{N2}. \label{eq:4}\end{aligned}$$ We choose the normalization factors $1/2$ and $1/4$ in such a way that $N_1$ and $N_2$ are average quantities that should take values similar to the basic ratios in Eq. . By definition, these variables should depend very mildly on the production cross section and on the center of mass energy of the machine. If we relax the assumption that the SSD are produced through a pair of charginos $\chi_1^\pm\chi_1^\pm $ the above Eq. and Eq. are still valid to a very good degree of approximation although the cancellation of the cross section does not apply any longer. #### Dark matter and the CMSSM. {#dark-matter-and-the-cmssm. .unnumbered} The CMSSM is specified by assigning the value of the common gauginos mass $m_{1/2}$, the common scalars mass $m_0$, the common trilinear scalar coupling $A_0$ at a certain unification energy scale. The other fundamental parameter is $\tan\beta$, the ratio of the two vacuum expectation values of the Higgs doublet. Further we assume the Higgs mixing parameter $\mu$ to be positive. The is one of the regions [@ellisWMAP] of the parameter space in which the relic density of the neutralino as main dark matter component is compatible with the WMAP [@wmap] measurement of the cosmic microwave background anisotropies, $0.096<\Omega h^2 < 0.128$ at 3$\sigma$. In the [@stauco], the neutralino relic density is controlled by co-annihilation before the freeze-out between the LSP and the lightest stau, $\tilde{\tau}_1$, which is the next LSP with a mass splitting $\Delta m=m_{\tilde{\tau}_1} -m_{\chi_0}$ of few GeVs. In literature there exist theoretical studies discussing signatures of the at LHC. In Refs. [@Arnowitt] the authors show that with the reconstruction of the decays $\chi^0_2 \to \tilde{\tau}_1 \tau \to \tau\tau \chi^0_1 $ at the end of the gluinos and squarks cascade, it is in principle possible to measure the masses of the relevant particles and infer the values of the CMSSM parameters. In Ref. [@Godbole] it is argued that the measurement of the tau polarization in the $\tilde{\tau}_1 \to \tau \chi^0_1 $ can be useful to measure the mass difference $\Delta m$. The smallness of $\Delta m$ is an important quantity for the phenomenology at LHC [@Arnowitt], as it sets the scale of the transverse momentum of the leptons at the end of the cascade decays, but also for indirect detection signals in astrophysical searches with gamma ray, as it is responsible for spectral features at the endpoint of the gamma spectrum [@ib] and for the large cross section for the annihilation into a $\tilde{\tau}_1$ pair at the galactic center [@BHsusy]. We consider the parameter space with fixed trilinear scalar coupling $A_0 =0$, which is used as benchmark for supersymmetric studies at the LHC. The stau-coannihilation strips shown in Figure \[strips\] are obtained with $\textsf{microOMEGAs}$ [@micromegas] using [@Softsusy] as supersymmetric mass spectrum calculator, imposing WMAP constraints on the relic density accelerator constraints on the lightest Higgs, $m_h >114.4$ GeV, chargino mass $m_{\chi^{+}_1}>103.5$ GeV and the flavor physics constraint from bottom mesons decays $B_s \to X\gamma$ and $B_s \to \mu^+ \mu^-$. In Figure \[strips\] we also report the 95% CL exclusion curves obtained by CMS [@CMSlsd] and ATLAS [@ATLASlsd; @ATLASgg+ssd] with the SSD search and the CMS SUSY search with hadronic final states [@CMSjets]. As can be seen the first two years of operation of the LHC could only marginally exclude the parameter space of interest in this work. ![Regions of the CMSSM parameter space in the ($m_0,m_{1/2}$) plane which satisfy the WMAP constraint on the relic density $0.096<\Omega h^2 < 0.128$ for different values of $\tan\beta$ that is given by the numbers above the strips. The dashed, dot-double-dashed and dot-dashed lines are respectively the the 95% CL exclusion curves obtained by CMS [@CMSlsd] and ATLAS [@ATLASlsd; @ATLASgg+ssd] with the SSD search and the CMS SUSY search with hadronic final states [@CMSjets].[]{data-label="strips"}](strips.eps) ![image](N1_pt15_eff_8TeV.eps) ![image](N2_pt15_eff_8TeV.eps) ![image](N1_pt15_eff_14TeV.eps) ![image](N2_pt15_eff_14TeV.eps) #### Analysis. {#analysis. .unnumbered} To check the validity of the new observables we make a preliminary simulation selecting points along the strips with $\tan\beta =10$ and $\tan\beta =40$. The theoretical ratio $R$ defined in Eq. is calculated in each point using [@susyhit] to compute the branching fractions. All informations of the selected models are then passed to [@pythia] to generate events in p-p collisions at $\sqrt{s}= 14$ TeV. We have generated $4\times 10^6$ events for each CMSSM point requiring that in the final state there are same-sign leptons, jets and missing energy. We find that Eqs. and are well satisfied, the number of SSD coming from two same-sign $\chapm$ is correctly predicted in terms of the theoretical ratio of the chargino’s branching ratios $R$. When all production mechanisms are allowed, and especially at large $\tan\beta$, the number of taus SSD is contaminated by the decay chain involving the second neutralino, $\chi^{\pm}_1 \chi^0_2$ and $\chi^0_2 \chi^0_2$ thus, depending on the point of the parameter space, deviations up to $50\%$ are observed. We now want to relate $N_{1,2}$ to $m_{1/2}$ and $\tan\beta$, the two CMSSM parameters that are the more interesting from the dark matter phenomenology point of view. In fact in Ref. [@MCdirect] it was found that the neutralino mass along the strips of Fig. \[strips\] is roughly given by $m_{\chi}=0.44\times m_{1/2} -16\;\text{GeV}$ for all the values of $\tan\beta$. Furthermore the spin-independent neutralino-nucleon cross section, and hence direct detection rates, strongly depend on $\tan\beta$. ![image](Luminosity_N1.eps) ![image](Luminosity_N2.eps) In order to cover the full parameter space we took 20 equally spaced points (in $m_{1/2}$) along each strip of Fig. \[strips\] and the corresponding value of $m_0$ is chosen to be the center value of the strip width along the $m_0$ direction. We carry out the simulation with 4$\times 10^6$ events for each model point. For realistic results we employ the efficiency model for the SSD signal developed by the CMS collaboration that allows to obtain realistic results at the generator level bypassing the full detector simulation [@CMSlsd2; @CMSlsd]. Tau leptons are identified by their hadronic decays, $\tau_h$. We imposed all the observed leptons $e$, $\mu$, $\tau_h$ to have a transverse momentum larger than 15 GeV, and to be within the acceptance of the ATLAS and CMS detectors ($\|\eta\|<2.4$). The search region to detect experimentally SSD events is defined by two additional variables: the transverse hadronic energy $H_T=\sum_{jet} {p_T}_{jet}$, determined by all the quarks and gluons with $p_T>40$ GeV within the detector acceptance and $E^{\text{miss}}_T$, the missing transverse energy determined by all the undetectable particles (neutrinos and LSP) in the event and the visible particles outside the detector acceptance. On top of the lepton selection we require $ H_T >450$ GeV and $E^{\text{miss}}_T$$>120$ GeV. The SSD detection efficiency is simulated using the formulas given in [@CMSlsd2; @CMSlsd], where the probability for an event to pass the selection cuts is given as a function of the $H_T$, $E^{\text{miss}}_T$ and the $p_T$ of the two same sign leptons. Events with three or more selected leptons are discarded. Some of the SSD events may even originate from a multi-lepton final state, typically with 3 or 4 leptons, where only two of them satisfy the selection criteria. The introduction of the kinematic cuts (especially those on $p_T$) lessen the accuracy of the approximations in Eqs. , which have however been verified to hold when no cuts are applied. We emphasize that the results of the simulation presented here are obtained considering all the production mechanisms and without imposing any selection on the decay chain that lead to SSD. We thus build the contour maps of $N_1$ and $N_2$ in the ($m_{1/2}, \tan\beta$) plane that are shown in left panels of Fig. \[fig:2\], $\sqrt{s}=8$ TeV, and right panels, $\sqrt{s}=14$ TeV. Fig. \[Fig:lum14\] shows the luminosity necessary to achieve an accuracy of 50% in the measurement of the observables $N_1$ (left panel) of $N_2$(right panel) at $\sqrt{s}=14$ TeV. We consider here only the statistical error. Systematic errors can be neglected because $N_{1,2}$ are ratios of observed yields measured using the same selection criteria and obtained in the same experimental conditions. With the $\sqrt{s}=14 $ TeV run that will follow after the LHC upgrade, with the possibility of accumulating up to 100 - 1000 fb$^{-1}$ of luminosity, will offer the possibility to probe large portions of the parameter space in the plane $(m_{1/2},\tan\beta)$ up to values of $m_{1/2}\approx 900 $ GeV. We do not show the corresponding of Fig. \[Fig:lum14\] for $\sqrt{s}=8$ TeV because the production cross sections are smaller and the planned 20 fb$^{-1}$ are not enough to get $N_{1,2}$ with the same accuracy. The rather small differences that can be seen in Fig. \[fig:2\] in the values that $N_{1,2}$ take at the two center of mass energies and in Fig. \[Fig:lum14\] in luminosity curves, confirm our expectation that to a first approximation the observables $N_{1,2}$ are independent of the center of mass energy of the collider. #### Comments and summary. {#comments-and-summary. .unnumbered} Both the CMS and ATLAS collaborations have released results that hint for a possible evidence of a Higgs with mass around 124-126 GeV in the first $\sim 5$ fb$^{-1}$ of data obtained with $\sqrt{s}=7$ TeV [@CMShiggs; @ATLAShiggs]. The light Higgs mass in the MSSM receives a large contribution from radiative corrections thus represents a crucial quantity to test any SUSY model. Updated analysis of the CMSSM parameter space including the new Higgs data [@ellisolive] show that $A_0 \neq 0$, large $\tan\beta$ and heavy SUSY spectrum are now generally preferred. In the case of the confirmation of the discovery it would be interesting to make detailed study of the SSD yield expected from the smaller parameter space, compared to the one explored here, compatible with that Higgs mass. In our analysis we have excluded those model points found in the right edge of the strips at high $m_{1/2}$ in which the mass difference between the $\tilde{\tau}_1$ and the LSP ($\tilde{\chi}^0_1$) is less then the mass of the tau lepton ($m_\tau \approx 1.7$ GeV). In this case the two-body decay $\tilde{\tau}_1 \to \tilde{\chi}^0_1 \tau$ is forbidden. The $\tilde{\tau}_1$ decays only into suppressed three body final states. and is a long lived charged particle that have been proposed as a solution to the lithium problem [@LLS2]. Once produced they can decay outside the detector [@LLS1]. The sudden reduction of the number of tau SSD $N(\tau\tau)$ is reflected in much higher numerical values of $N_1$ and $N_2$ (in particular those of $N_2$). For a recent study connecting long-lived staus and SSD in models with the gravitino as the dark matter candidate see Ref. [@masip]. In summary, we have shown that if an excess of SSD relative to the SM yield is observed then a measurement of the proposed observables $N_{1,2}$ within a given accuracy will allow to pin down a given portion of the stau co-annihilation region () of the CMSSM parameter space in the $(m_{1/2},\tan\beta)$ plane. This will in turn also give access to other informations. For example if we restrict the possible values of $m_{1/2}$ to a certain range then this will of course also restrict the possible values of $m_0$ since these two parameters are related according to the WMAP strips of Fig. \[strips\]. The proposed method can evidently be extended to all SUSY models predicting events with final states containing SSD or to different extensions of the standard model that contain Majorana particles such as models with weak scale heavy Majorana neutrinos [@HMN]. #### Acknowledgements. {#acknowledgements. .unnumbered} The work of M. C. is supported by a MultiDark under Grant No. CSD2009-00064 of the Spanish MICINN Consolider-Ingenio 2010 Program. Further support is provided by the MICINN project FPA2011-23781. M. C., O. P. and M. S. acknowledge partial support from the Grant MICINN-INFN(PG21)AIC-D-2011-0724. Part of this work was done as partial fulfillment for the requirements of the master thesis (Laurea Magistrale) of M. S. (University of Perugia, December 2011). #### Note added. {#note-added. .unnumbered} After submitting this paper for publication both the ATLAS and CMS experiments have reported new evidence at 5$\sigma$ level for a scalar particle of mass around $125$ GeV compatible with the Higgs boson [@5s]. As already discussed above, and in view of the newly reported experimental evidence on the Higgs mass, the choice of the trilinear scalar parameter $A_0=0$ is not allowed any longer since the radiative corrections needed to achieve $m_h=125$ GeV are driven by the couplings of the third generation sfermions and especially so the Higgs-stop-stop trilinear coupling $A_t$ must be large, implying large and negative values for $A_0$. The allowed parameter space of the CMSSM in light of the Higgs discovery is being still more severely constrained and according to some authors this model is already disfavored by the data, while others authors do not agree with this conclusion. We refer the reader to Refs. [@ellisolive; @CMSSMscan] for a list of recent studies showing these different points of view. [100]{} G. Bertone, D. Hooper and J. Silk, Phys. Rept. [405]{}, 279 (2005). H. Baer, X. Tata and J. Woodside, Phys. Rev. D [45]{}, 142 (1992); R. M. Barnett, J. F. Gunion and H. E. Haber, Phys. Lett. B [315]{}, 349 (1993). M. Guchait and D. P. Roy, Phys. Rev. D [52]{}, 133 (1995). H. Baer, C. -h. Chen, F. Paige and X. Tata, Phys. Rev. D [53]{}, 6241 (1996). J. Alwall, D. Rainwater and T. Plehn, Phys. Rev. D [76]{}, 055006 (2007). H. Baer, A. Lessa and H. Summy, Phys. Lett. B [674]{}, 49 (2009). K. T. Matchev, F. Moortgat, L. Pape and M. Park, Phys. Rev. D [82]{}, 077701 (2010). A. Abulencia [et al.]{} \[CDF Collaboration\], Phys. Rev. Lett. [98]{}, 221803 (2007). S. Chatrchyan [et al.]{} \[CMS Collaboration\], arXiv:1205.6615. G. Aad [et al.]{} \[ATLAS Collaboration\], Phys. Lett. B [709]{}, 137 (2012). G. Aad [et al.]{} \[ATLAS Collaboration\], arXiv:1203.5763. J. R. Ellis, K. A. Olive, Y. Santoso and V. C. Spanos, Phys. Lett. B [565]{}, 176 (2003). D. Larson [et al.]{}, Astrophys. J. Suppl. [192]{}, 16 (2011). J. R. Ellis, T. Falk, K. A. Olive and M. Srednicki, Astropart. Phys. [13]{}, 181 (2000) \[Erratum-ibid. [15]{}, 413 (2001)\]; M. E. Gomez, G. Lazarides and C. Pallis, Phys. Rev. D [61]{}, 123512 (2000); T. Nihei, L. Roszkowski and R. Ruiz de Austri, J. High Energy Phys. 07 (2002) 024; J. Edsjo, M. Schelke, P. Ullio and P. Gondolo, J. Cosmol. Astropart. Phys. 04 (2003) 001. R. L. Arnowitt, B. Dutta, A. Gurrola, T. Kamon, A. Krislock and D. Toback, Phys. Rev. Lett. [100]{}, 231802 (2008); R. L. Arnowitt, A. Aurisano, B. Dutta, T. Kamon, N. Kolev, P. Simeon, D. A. Toback and P. Wagner, Phys. Lett. B [649]{}, 73 (2007); R. L. Arnowitt, B. Dutta, T. Kamon, N. Kolev and D. A. Toback, Phys. Lett. B [639]{}, 46 (2006). R. M. Godbole, M. Guchait and D. P. Roy, Phys. Rev. D [79]{}, 095015 (2009). T. Bringmann, L. Bergstrom and J. Edsjo, J. High Energy Phys. 01 (2008) 049; M. Cannoni, M. E. Gomez, M. A. Sanchez-Conde, F. Prada and O. Panella, Phys. Rev. D [81]{}, 107303 (2010). M. Cannoni, M. E. Gomez, M. A. Perez-Garcia and J. D. Vergados, Phys. Rev. D [85]{}, 115015 (2012). G. Belanger, F. Boudjema, P. Brun, A. Pukhov, S. Rosier-Lees, P. Salati and A. Semenov, Comput. Phys. Commun. [182]{}, 842 (2011). B. C. Allanach, Comput. Phys. Commun. [143]{}, 305 (2002). CMS-PAS-SUS-12-002, https://twiki.cern.ch/twiki/bin/view/CMSPublic/PhysicsResultsSUS12002 A. Djouadi, M. M. Muhlleitner and M. Spira, Acta Phys. Polon. B [38]{}, 635 (2007) M. Muhlleitner, A. Djouadi and Y. Mambrini, Comput. Phys. Commun. [168]{}, 46 (2005) A. Djouadi, J. Kalinowski, M. Spira, Comput. Phys. Commun. [108]{}, 56 (2005). http://home.thep.lu.se/ torbjorn/Pythia.html S. Chatrchyan [et al.]{} \[CMS Collaboration\], J. High Energy Phys. 06 (2011) 077; CMS Collaboration, CMS-PAS-SUS-11-010. M. Cannoni, Phys. Rev. D [84]{}, 095017 (2011). S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Lett. B [710]{}, 26 (2012). G. Aad [et al.]{} \[ATLAS Collaboration\], Phys. Lett. B [710]{}, 49 (2012). J. Ellis and K. A. Olive, Eur. Phys. J. C [72]{}, 2005 (2012). T. Jittoh, K. Kohri, M. Koike, J. Sato, T. Shimomura and M. Yamanaka, Phys. Rev. D [76]{}, 125023 (2007). T. Ito, K. Nakaji and S. Shirai, Phys. Lett. B [706]{}, 314 (2012). R. Barcelo, J. I. Illana, M. Masip, A. Prado and P. Sanchez-Puertas, arXiv:1206.5108. O. Panella, M. Cannoni, C. Carimalo and Y. N. Srivastava, Phys. Rev. D [65]{}, 035005 (2002). CMS-PAS-HIG-12-020,\ http://cdsweb.cern.ch/record/1460438?ln=en;\ ATLAS-CONF-2012-093,\ http://cdsweb.cern.ch/record/1460439. O. Buchmueller [et al.]{}, arXiv:1112.3564; D. Ghosh, M. Guchait, S. Raychaudhuri and D. Sengupta, arXiv:1205.2283; C. Balazs, A. Buckley, D. Carter, B. Farmer and M. White, arXiv:1205.1568; P. Bechtle [et al.]{}, arXiv:1204.4199; H. Baer, V. Barger and A. Mustafayev, arXiv:1202.4038; S. Akula, P. Nath and G. Peim, arXiv:1207.1839; H. Baer, V. Barger, P. Huang, A. Mustafayev and X. Tata, arXiv:1207.3343; J. Cao, Z. Heng, J. M. Yang and J. Zhu, arXiv:1207.3698.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We derive the optimal proposal density for Approximate Bayesian Computation (ABC) using Sequential Monte Carlo (SMC) (or Population Monte Carlo, PMC). The criterion for optimality is that the SMC/PMC-ABC sampler maximise the effective number of samples per parameter proposal. The optimal proposal density represents the optimal trade-off between favoring high acceptance rate and reducing the variance of the importance weights of accepted samples. We discuss two convenient approximations of this proposal and show that the optimal proposal density gives a significant boost in the expected sampling efficiency compared to standard kernels that are in common use in the ABC literature, especially as the number of parameters increases.' address: - 'Center for Computational Astrophysics, Flatiron Institute, 162 5th Ave, New York City, NY 10010, USA' - 'Sorbonne Universités, CNRS, UMR 7095, Institut d’Astrophysique de Paris, 98 bis boulevard Arago, 75014 Paris, France' - 'Sorbonne Universités, Institut Lagrange de Paris (ILP), 98 bis boulevard Arago, 75014 Paris, France' - 'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08540, USA' author: - - - bibliography: - 'optimalABC.bib' title: Optimal proposals for Approximate Bayesian Computation --- Introduction ============ The naïve approach to Approximate Bayesian Computation (ABC) generates (compressed) data simulations for parameters $\theta$ that are drawn from the prior $\pi(\theta)$. If the resulting simulated data $d'$ is within $\epsilon$ of the true data $d$, i.e. under a distance metric $\rho(d,d')\leq\epsilon$, then $\theta$ is accepted as a sample from the approximate posterior density, $p(\theta \| \rho(d,d')\leq\epsilon)$. If simulation is costly, it is advantageous to attempt to increase the fraction of accepted $\theta$ by proposing new candidate $\theta$s from a proposal density $q(\theta)$ that is large in parameter ranges that are preferred by the data, and small in less interesting regions of the prior volume. It is natural to base the choice of $q(\theta)$ on the current accepted samples from the approximate posterior $p(\theta \| \rho(d,d')\leq\epsilon)$. This is the approach taken in SMC-ABC algorithms (see eg., [@ABCreview2012] for a review); a proposal density $q(\theta)$ is constructed at each population iteration, which is typically a kernel density estimator (KDE) based on the accepted points from the previous population (ie, a kernel that adapts as the algorithm steps through successive population iterations [@Beaumont2009; @Mckinley2009; @Toni2009a; @Barnes2011; @Didelot2011; @Jasra2012; @Filippi2013; @Bonassi2015]). The price to pay for the increased fraction of accepted ABC samples of $\theta$ is the necessity to importance weight the accepted samples by $\pi/q$. The variance in these importance weights will reduce the effective number of samples. The more concentrated $q$ is relative to $\pi$, on parameters $\theta$ that have a high probability of being accepted, the larger the variance in the importance weights $\pi/q$. In practice, a poor choice of proposal density $q$ can lead to a proposed sample whose proposal density $q$ was low, but is subsequently accepted, leading to a large importance weight that can overwhelm the rest of the (weighted) samples leading to a very small effective sample size. The natural question, then, is how to choose the proposal density that represents the optimal trade-off between a high acceptance rate and a low importance-weight variance. Most SMC-ABC implementations propose new parameters for forward simulation via an importance weighted KDE based on the previous population’s accepted samples, with uniform [@Toni2009; @Toni2009a; @Mckinley2009], student-t [@Didelot2011] and Gaussian kernels [@Sisson2007; @Beaumont2009; @Filippi2013; @Bonassi2015] in common use. A choice has to be made for the kernel bandwidth and there are various choices in the literature: for example, for Gaussian kernels, [@Sisson2007] use the importance-weighted variance of the previous population samples, @Beaumont2009 use twice the importance-weighted variance of the previous population samples, whilst [@Bonassi2015] use standard recommendations from @West1993 and @Scott2005[^1]. Previous studies have sought kernels that are optimal in the following sense: they minimize the sum of the Kullback-Leibler divergence between the proposal KDE and the target density, and the negative log acceptance ratio, providing some trade-off between closeness of the target and proposal and the acceptance ratio [@Beaumont2009; @Filippi2013]. @Beaumont2009 showed that for a global Gaussian kernel, this optimality criterion leads to a bandwidth equal to twice the importance-weighted variance of the previous population’s accepted samples (with some further refinement and generalization by [@Filippi2013]). @Filippi2013 also considered the same optimality criterion applied to local rather than global kernels, deriving a locally-optimal kernel-covariance. Whilst this optimality criterion has proved powerful, the relative importance and utility of the KL divergence and acceptance ratio terms in this approach is ambiguous. In this paper we take a slightly different approach and derive the proposal density $q(\theta)$ for SMC-ABC sampling that maximizes the effective number of samples per parameter proposal (and hence forward simulation). This provides the optimal trade-off between high acceptance rate and low variance in the importance weights under a straightforward and pragmatic definition of optimality. Rather than restricting to a given class of perturbation kernels (eg., Gaussian kernels), we derive the optimal proposal density (in the asymptotic $\epsilon\rightarrow 0$ limit) assuming only that some density estimator for the ABC posterior is available at each population iteration. The result provides a well-motivated guide for adaptive proposal density choice for SMC-ABC sampling. Optimal SMC-ABC proposal densities {#sec:optimal} ================================== We define the sampling efficiency as the functional $\omega[q]$ that measures the effective number of samples per parameter proposal. This is composed of two components: 1) $f_{a}$, the fraction of proposed points that will be accepted, and 2) $N_w$, the effective number of points after application of the importance weights. The expected fraction of accepted points is given by $$\begin{aligned} f_{a}=P(\rho(d,d')<\epsilon)&=\int I_{\rho(d,d')<\epsilon}\, p_q(d') dd'\nonumber \\ &\underset{\epsilon \rightarrow 0}{\approx} V_\epsilon p_q(d)=V_\epsilon \int p(d\|\theta)q(\theta) d\theta =V_\epsilon \, p(d) \int \frac{q(\theta)}{\pi(\theta)}p(\theta\|d)d\theta \end{aligned}$$ where $ p_q(d)$ is the probability density of simulated data $d$ when the parameters are proposed from $q$, $V_\epsilon$ is the volume of the space for accepted samples, and “$\approx$” becomes accurate in the limit of small $\epsilon$, assuming $p(d)$ is continuous. The expected effective number of points after application of the importance weights to $N_s$ accepted samples is given by $$\begin{aligned} N_w=\frac{\left(\sum w_i\right)^2}{\sum w_i^2}= %% \frac{\left(\sum_i^{N_s} \pi(\theta_i)/q(\theta_i)\right)^2}{\sum_i^{N_s} \left(\pi(\theta_i)/q(\theta_i)\right)^2}&\\ %% \underset{N_s\rightarrow\infty}{\approx}\frac{N_s\left(\int \frac{\pi}{q} p_\mathrm{accepted}(\theta)d\theta\right)^2}{\int \frac{\pi^2}{q^2}p_\mathrm{accepted}(\theta)d\theta}&\\ %% \underset{\epsilon \rightarrow 0}{\approx}\frac{N_s V_\epsilon\,\left(\int \frac{\pi}{q} p(d\|\theta)q(\theta)d\theta\right)^2}{\int \frac{\pi^2}{q^2} p(d\|\theta)q(\theta)d\theta} &\\ %% =\frac{N_s V_\epsilon\,p(d)}{\int\frac{\pi(\theta)}{q(\theta)} p(\theta \| d) d\theta}. \end{aligned}$$ We therefore find for the sampling efficiency $\omega[q]$ (dropping $q$-independent constant pre-factors), $$\label{omega} \omega[q]\equiv \frac{A[q]}{B[q]} = \frac{\int \frac{q(\theta)}{\pi(\theta)}p(\theta \| d)d\theta}{\int \frac{\pi(\theta)}{q(\theta)}p(\theta \| d)d\theta}\propto f_a\frac{N_w}{N_s},$$ where the second equality defines the functionals $A[q]$ and $B[q]$, and the last proportionality defines the sampling efficiency as the number of effective (accepted) samples per parameter proposal (and hence forward simulation). For naïve ABC, proposing parameters from the prior, $\omega[\pi]=1$. Maximizing $\omega[q]$ with respect to $q$, under the constraint of $q$ being normalized, gives the optimal proposal $$\label{eq:optimalq} q^\ast(\theta)=\sqrt[]{\omega[q^\ast]\frac{p(\theta \| d)\pi(\theta)}{2A[q^\ast]-\frac{p(\theta\|d)}{\pi(\theta)}}}.$$ This is the main result of the paper. This implicit equation for $q^\ast$ can be solved iteratively; alternatively one can just consider $\omega^\ast\equiv\omega[q^\ast]$ and $A^\ast\equiv A[q^\ast]$ as parameters and $$q^\ast(\theta)=\max_{\omega^\ast,A^\ast} \omega[q(\theta,\omega^\ast,A^\ast)].$$ Further, there are two simple and fast approximations for $q^\ast$ that can be obtained without iteration or that can inform good starting points for the iteration; these are discussed in §\[sec:approximations\]. Note that we only need to know $A$ to be able to sample from the optimal proposal , since $\omega$ just sets the normalization. Assuming a density estimator for $p(\theta\|d)$ is available, and $\pi$ is easy to evaluate, then Markov Chain Monte Carlo (MCMC) methods can efficiently generate draws from $q^\ast$ to serve as parameter proposals for the next SMC iteration. The optimal proposal Eq. can be seen as the geometric mean of posterior and prior (ie., the numerator), with a relative boost where the posterior is larger than the prior. The denominator boosts $q^\ast$ in the region where the posterior peaks, but the peaks in $q^\ast$ are narrower than the corresponding ones in $p(\theta\|d)$ in their immediate neighborhoods to compensate for the heavier tails away from the peaks. Illustrative examples are shown in Figures \[fig:gaussian\_qplots\]–\[fig:chi2\_qplots\] and discussed in §\[sec:examples\]. Fast approximations to the optimal proposal density {#sec:approximations} --------------------------------------------------- ### Geometric mean approximation Expanding $q^\ast$ gives the convergent series $$q^\ast=\sqrt[]{\frac{p(\theta\|d)\pi(\theta)}{2B[q^\ast]}}\sum_{i=0}^{\infty} \begin{pmatrix} i-\frac12\\i \end{pmatrix}\left(\frac{p(\theta\|d)}{2A[q^\ast]\pi(\theta)}\right)^i.$$ Convergence follows from the properties of the binomial coefficient and the lower bound on $A[q^\ast]$ in Eq. (\[holders\]) below. In numerical experiments we find the $i>0$ terms give sub-dominant contributions to $\omega$ for a wide variety of choices for $q$, $p$ and $\pi$. Taking the normalized leading term of the series gives the geometric mean approximation to the optimal proposal[^2], $$\label{geometric_mean} q_0 \propto\sqrt{p(\theta \| d)\pi(\theta)}.$$ Note that for simple cases where the posterior is Gaussian under a uniform prior, the geometric mean approximation leads to a Gaussian proposal density centered on the posterior mean and with twice the posterior covariance. This is similar to taking a Gaussian KDE proposal with bandwidth equal to the estimated posterior variance; this kernel scheme is sometimes adopted in the ABC literature (eg., [@Sisson2007; @Ishida2015]). ### Bounded approximation Given the value of $A[q^]$, we could sample the optimal proposal density Eq. . Whilst $A[q^]$ is not available a priori, we can bound $A[q^]$ from above and below: using Hölder’s inequality to obtain the upper bound, and the non-negativity of probability densities for the lower bound, we find that $$\label{holders} \frac{1}{2}\sup_\theta\frac{p(\theta\|d)}{\pi(\theta)}< A[q^] \leq\sup_\theta \frac{p(\theta\|d)}{\pi(\theta)}.$$ Choosing a value of $A[q^]$ between these bounds avoids the need for iteration for the cost of a mild reduction in optimality. Fixing $A$ to be the average of the upper and the lower bound leads to the bounded approximation* for the optimal proposal, $$\label{bounded_approx} q_{\bar{A}} \propto \sqrt{ \frac{p(\theta \| d)\pi(\theta)}{2\bar{A} - \frac{p(\theta \| d)}{\pi(\theta)} }},$$ with $$\bar{A} = \frac{3}{4}\sup_\theta\frac{p(\theta\|d)}{\pi(\theta)}.$$ In numerical experiments, we find that the bounded approximation gives close to optimal sampling efficiencies (see §\[sec:examples\] and Table \[tab:sampling\_efficiencies\]). If further optimality is desired, this can be used as a starting point for iterating Eq. towards the optimal $q^\ast$. All that is required to propose samples from the optimal proposal or an approximation of it is a density estimator for the posterior. In practice, this could be a KDE or mixture model fit to the accepted samples in the previous SMC population. With a posterior density estimator in hand, the optimal proposal can be found iteratively using Eq. , or via one of the convenient approximations Eq. or , and then sampled using MCMC or otherwise to generate parameter proposals for the next SMC iteration. ![ABC proposals for Gaussian posterior $p(\theta \| d) = \mathcal{N}(0, 1)$ (grey), with a Gaussian prior $\pi(\theta)=\mathcal{N}(0,5)$ (not shown). From bottom to top at the peak: commonly used KDE proposal with bandwidth of twice the (estimated) posterior variance (blue), geometric mean approximation of the optimal proposal (red-dotted), bounded approximation of the optimal proposal with $A = 3/4\,\sup_\theta p(\theta\|d)/\pi(\theta)$ (red-dashed), optimal proposal density (red).[]{data-label="fig:gaussian_qplots"}](gaussian.pdf){width=".79\textwidth"} ![ABC proposals for a bimodal posterior $p(\theta \| d) = 1/2\,\mathcal{N}(-2, 1) + 1/2\,\mathcal{N}(2, 1)$ (grey), with a Gaussian prior $\pi(\theta)=\mathcal{N}(0,10)$ (not shown). From bottom to top at the peak: commonly used KDE proposal with bandwidth of twice the (estimated) posterior variance (blue), geometric mean approximation of the optimal proposal (red-dotted), bounded approximation of the optimal proposal with $A = 3/4\,\sup_\theta p(\theta\|d)/\pi(\theta)$ (red-dashed), optimal proposal density (red).[]{data-label="fig:bimodal_qplots"}](bimodal.pdf){width=".79\textwidth"} ![ABC proposals for $\chi^2$ posterior $p(\theta \| d) = \chi^2(\nu=3)$ (grey), with a uniform prior $\pi(\theta)=\mathcal{U}(0,30)$ (not shown). From bottom to top at the peak: commonly used KDE proposal with bandwidth of twice the (estimated) posterior variance (blue), geometric mean approximation of the optimal proposal (red-dotted), bounded approximation of the optimal proposal with $A = 3/4\,\sup_\theta p(\theta\|d)/\pi(\theta)$ (red-dashed), optimal proposal density (red).[]{data-label="fig:chi2_qplots"}](chi2.pdf){width=".79\textwidth"} [ccccc]{} Proposal, $q(\theta)$ & $A[q]$ & $B[q]$ & $\omega[q]\equiv A[q]/B[q]$\ \ $p(\theta \| d)$ & 3.57 & 1.0 & 3.57\ $q^\mathrm{KDE}_{h=\sqrt{2}\sigma_p}(\theta)$ & 2.54 & 0.41 & 6.16\ $q_0(\theta)$ & 2.96 & 0.38 & 7.71\ $q_{\bar{A}}(\theta)$ & 3.26 & 0.40 & 8.20\ $q^\ast(\theta)$ & 3.34 & 0.41 & [8.22]{}\ \ \ $p(\theta \| d)$ & 3.68 & 1.0 & 3.68\ $q^\mathrm{KDE}_{h=\sqrt{2}\sigma_p}(\theta)$ & 2.55 & 0.40 & 6.36\ $q_0(\theta)$ & 3.22 & 0.34 & 9.47\ $q_{\bar{A}}(\theta)$ & 3.48 & 0.35 & 9.93\ $q^\ast(\theta)$ & 3.52 & 0.35 & [9.94]{}\ \ \ $p(\theta \| d)$ & 4.77 & 1.0 & 4.77\ $q^\mathrm{KDE}_{h=\sqrt{2}\sigma_p}(\theta)$ & 3.80 & 0.57 & 6.68\ $q_0(\theta)$ & 3.56 & 0.35 & 10.23\ $q_{\bar{A}}(\theta)$ & 4.05 & 0.36 & 11.23\ $q^\ast(\theta)$ & 4.34 & 0.38 & [11.40]{}\ **** \[tab:sampling\_efficiencies\] Examples {#sec:examples} -------- ### Numerical examples of optimal ABC proposals and their approximations Figures \[fig:gaussian\_qplots\]–\[fig:chi2\_qplots\] and Table \[tab:sampling\_efficiencies\] illustrate the optimal ABC proposal density and its approximations in three scenarios: a Gaussian posterior (Figure \[fig:gaussian\_qplots\]), a bimodal double-Gaussian posterior (Figure \[fig:bimodal\_qplots\]), and a $\chi^2$ posterior (Figure \[fig:chi2\_qplots\]). We compare the optimal proposals to the commonly used ABC proposal scheme recommended in @Beaumont2009 for reference: a Gaussian KDE with bandwidth set to double the (estimated) posterior variance. For illustration, the proposals are compared in the converged $\epsilon\rightarrow0$ limit where the approximate posterior (in practice, a density estimator for the accepted samples) is close to the true posterior. The KDE proposal [@Beaumont2009] is shown as the convolution of the true posterior with a Gaussian with twice the posterior variance. The three examples shown in Figures \[fig:gaussian\_qplots\]–\[fig:chi2\_qplots\] and Table \[tab:sampling\_efficiencies\] display the same essential characteristics. The optimal proposals are boosted in regions of high posterior density (around the peak) to give a high acceptance rate, whilst having slightly broader tails compared to the posterior to ensure the importance-weight variance is kept under control (hence giving an improved effective sample size). In contrast, the KDE proposals are typically much broader than the optimal proposals around the posterior peak, which leads to lower expected acceptance rates. Meanwhile, using a posterior density estimate for the proposal gives a poor expected importance-weight variance, owing to the narrower tails compared to the other proposal schemes. The optimal proposal represents the trade-off between high proposal density in regions of high posterior density, and fatter tails to maintain a lower importance-weight variance. In all three examples shown, the bounded approximation for the optimal proposal Eq. performs nearly as well as the optimal proposal. This is especially clear from Table \[tab:sampling\_efficiencies\], where the sampling efficiencies for the optimal proposal versus the bounded approximation are very similar. The geometric mean approximation also provides a reasonable first approximation to the optimal proposal for the examples shown in Figures \[fig:gaussian\_qplots\]–\[fig:chi2\_qplots\], with improvements in sampling efficiencies compared to the KDE or posterior-approximation proposals (see Table \[tab:sampling\_efficiencies\]). These three one-dimensional examples have the virtue of being easy to visualise and to show the features of the optimal proposal for Gaussian, skewed and multi-modal posteriors. In the following section we demonstrate that a lower bound on the expected relative improvement of the optimal proposal improves exponentially on the performance of other kernels. ### Expected improvement as a function of data informativeness and parameter dimensionality We can obtain a lower bound on the improvement in sampling efficiency enabled by the optimal proposal by applying the geometric mean approximation to a simple toy problem. In this model, we take the prior and posterior to be $n_\theta$-dimensional multivariate Gaussians, with means in each dimension of $\mu_\pi$ and $\mu_p=0$, and diagonal covariances with variances $\sigma^2_\pi$ and $\sigma^2_p=1$, respectively. We then exploit the fact that $\omega[q_0]$ can be obtained analytically in this setting to rapidly evaluate the improvement in sampling efficiency over other choices of the proposal as a function of the dimensionality of the parameter space and the informativeness of the data (as measured by both the reduction in volume and the shift in the mean in going from prior to posterior). As the optimal proposal $q^\ast$ outperforms the geometric mean approximation $q_0$, the results in this section can be considered lower bounds on the improvement in performance realized by using the optimal proposal. Figure \[fig:approximate\_improvement\_posterior\] compares the geometric mean approximation proposal to using an estimator for the posterior density as the proposal, in the three-dimensional setting. It is clear that proposing samples from the posterior is highly suboptimal, whilst using even the geometric mean approximation to the optimal proposal instead gives improvements that are large, especially when the data are highly informative or surprising compared to prior expectations. Figure \[fig:approximate\_improvement\_literature\_3d\] compares the geometric mean approximation proposal to the Gaussian KDE proposal scheme recommended in [@Beaumont2009] (with bandwidth set to double the estimated posterior variance), again for the case where the posterior and prior are three-dimensional Gaussians. Again, gains in sampling efficiency are expected using the geometric mean approximation as opposed to the KDE scheme. Figure \[fig:approximate\_improvement\_literature\_10d\] shows the same case but for a 10-parameter rather than three-parameter set-up, showing that the relative improvement from using the optimal kernel quickly becomes larger in higher dimensions. Inspecting the analytical result for the Gaussian case confirms that the relative improvement in sampling efficiency of the geometric mean approximation scales exponentially with the number of parameters when compared to the other cases we study here. This also implies exponentially scaling improvement with number of parameters for the optimal proposal. ![Improvement $a$ in the sampling efficiency when sampling from the geometric mean approximation to the optimal proposal rather than the posterior for the case where both prior and posterior are Gaussian. This plot shows the case where $n_\theta=3$ (3-parameters). Even for this modest number of parameters the improvement is large when the data are informative. While the exact result guarantees $a>1$, the geometric mean approximation gives $a<1$ (shown in blue) for the atypical case when the posterior and prior have nearly equal width but are located far from each other. []{data-label="fig:approximate_improvement_posterior"}](neff_opt_abc_3d.pdf){width="\textwidth"} ![Improvement $a$ in the sampling efficiency when drawing proposals from the geometric mean approximation to the optimal proposal rather than the KDE scheme recommended in [@Beaumont2009] (with bandwidth equal to twice the posterior variance) for the case where both prior and posterior are Gaussian. This plot shows the case where $n_\theta=3$. While the exact result guarantees $a>1$, the approximation gives $a<1$ (shown in blue) for the atypical case when the posterior and prior have nearly equal width but are located far from each other. []{data-label="fig:approximate_improvement_literature_3d"}](neff_opt_vs_lit_abc_3d.pdf){width="\textwidth"} ![Improvement $a$ in the sampling efficiency when drawing proposals from the geometric mean approximation to the optimal proposal rather than the KDE scheme recommended in [@Beaumont2009] (with bandwidth equal to twice the posterior variance) for the case where both prior and posterior are Gaussian. This plot shows the case where $n_\theta=10$. While the exact result guarantees $a>1$, the approximation gives $a<1$ (shown in blue) for the atypical case when the posterior and prior have nearly equal width but are located far from each other. []{data-label="fig:approximate_improvement_literature_10d"}](neff_opt_vs_lit_abc_10d.pdf){width="\textwidth"} Conclusions =========== We have derived an optimal proposal scheme for SMC-ABC algorithms by maximizing the sampling efficiency, defined as the effective number of samples per parameter proposal. This represents the optimal trade-off between obtaining a high acceptance rate, whilst reducing the variance of the importance-weights of the accepted samples, hence increasing the effective sample size for the accepted samples. We derived an implicit form for the optimal proposal that can be solved iteratively, and also two convenient and simple approximations that can be evaluated quickly, provided a density estimator for the posterior is available (based on the accepted samples from the previous SMC population). We have shown that the optimal proposal scheme gives a substantial boost in the expected sampling efficiency for a range of statistical models, and the expected gain in sampling efficiency increases with the dimensionality of the problem. The derived results hence provide a guide for choosing optimal proposal densities for SMC-ABC applications, for the small cost of constructing a posterior density estimator at each SMC population iteration. We wish to thank Ethan Anderes and Tom Charnock for discussions. This work is supported by the Simons Foundation. Benjamin Wandelt acknowledges support from the Labex Institut Lagrange de Paris (ILP) (reference ANR-10- LABX-63) part of the Idex SUPER, and received financial state aid managed by the Agence Nationale de la Recherche, as part of the programme Investissements d’avenir under the reference ANR-11-IDEX-0004-02. [^1]: ie., taking the (component-wise) importance-weighted variance divided by $N^{1/3}$, for $N$ samples. [^2]: From Jensen’s inequality we have $$\int \left(\frac{p}{\pi}\right)^{2}\pi d\theta\geq \int \left(\frac{p}{\pi}\right)^{\frac32}\pi d\theta\geq 1\geq \int \left(\frac{p}{\pi}\right)^{\frac12}\pi d\theta.$$ It follows that $\omega[q_0]\geq \omega[\pi]$ and $\omega[p]\geq \omega[\pi]$ for any pdf $p$ and $\pi$. While we cannot conclude in general that $\omega[q_0]\geq \omega[p]$, Figure \[fig:approximate\_improvement\_posterior\] shows that in the Gaussian example $q_0$ outperforms the posterior in all practically relevant cases.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We point out that the measurement of target spin depolarization $D_{nn}$ in the $\bar{p}p\to\bar{\Lambda}\Lambda$ reaction may test dynamical mechanisms invoked to explain the proton spin puzzle revealed by polarized deep–inelastic scattering experiments. In particular, models with [negatively]{} polarized $\bar{s}s$ pairs in the proton wave function predict $D_{nn}<0$, whereas models with [positively]{} polarized gluons would predict $D_{nn}>0$.' --- =0.25in =0.25in =0.25in =8.50in =6.00in =0.00in =0.00in [CERN–TH/95–47\ DOE/ER/40427–04–N95]{} 0.5cm Mary Alberg$^{a,b,}$[^1], John Ellis$^{c,}$[^2] and Dmitri Kharzeev$^{c,d,}$[^3]\ 0.8cm [$^a)$ Department of Physics\ Seattle University\ Seattle, WA 98122, USA]{} [$^b)$ Department of Physics\ University of Washington\ Seattle, WA 98195, USA]{} [$ ^c)$ Theory Division\ CERN\ Geneva, Switzerland]{} [$^d)$ Physics Department\ University of Bielefeld\ 33615 Bielefeld, Germany]{} CERN–TH/95–47 DOE/ER/40427–04–N95 February 1995 The reaction $\bar{p}p\to\bar{\Lambda}\Lambda$ is a testing-ground for different approaches to non-perturba- tive QCD, in particular the quark model and meson-exchange models. In addition to total cross section and angular distribution measurements at different energies, the spin correlation of the $\Lambda$ and $\bar{\Lambda}$ have been measured [@PS185]. These were found to be predominantly in a spin-triplet state, with the spin-singlet component very small and consistent with zero within errors. This feature could easily be understood within quark models, if the $\bar{s}s$ pair that carry the $\bar{\Lambda}$ and $\Lambda$ spins in the naïve constituent quark model were produced by effective vector $(^3S_1)$ or scalar $(^3P_0)$ field exchange [@AHW]. Spin-triplet dominance could also be accomodated in a meson-exchange model, if the relative phase of the $K-$ and $K^-$ exchange amplitudes was suitably adjusted [@MEX], but spin-singlet suppression could not be regarded as a natural prediction of this class of models. 0.3cm There is a plentiful evidence from other experiments at LEAR and elsewhere that baryon wave functions may be more complicated than in the naïve constituent quark model. In particular, the experimental value of the $\pi-$nucleon $\sigma-$term [@sigma] and deep-inelastic experiments [@DIS] provide evidence for hidden $\bar{s}s$ pairs in the nucleon. Most strikingly, several recent LEAR experiments [@LEAR] find clear evidence for apparent violations of the OZI quark-line rule in $\bar{p}N\to\phi X$ annihilations, where $X=\gamma,\pi,\pi\pi$. A natural interpretation of these data is in terms of the shake-out or rearrangement of $\bar{s}s$ pairs present in the $\bar{p}N$ initial state [@EGK]. 0.3cm It recently was pointed out [@EKKS] that many features of these apparently OZI–violating hadronic processes can be understood if one assumes that the proton wave function contains an admixture of [polarized]{} $\bar{s}s$ pairs. This assumption is motivated by the experimental results on deep-inelastic scattering [@DIS] which indicate that strange quarks and antiquarks in the proton indeed have a net polarization opposite to the proton spin [@EK]. An alternative interpretation of these deep-inelastic results ascribes them to to polarized gluons in the proton [@gluon], a suggestion whose implications for low–energy $\bar{p}p$ annihilation have not yet been explored. 0.3cm The PS185 Collaboration is now proposing [@prop] an extension of its studies using a polarized target and measuring the depolarization $D_{nn}$ ($p\to\Lambda$ polarization transfer). Quark models generally predict positive values for this quantity [@pos1], whereas meson exchange models generally predict negative values [@neg]. We argue in this note that these measurements may discriminate between the polarized $\bar{s}s$ and gluon interpretations of the experimental results on polarized deep-inelastic scattering. Specifically, we find that the polarized $\bar{s}s$ model predicts [negative]{} depolarization $D_{nn}<0$, whereas the polarized gluon model predicts [positive]{} depolarization $D_{nn}>0$. Thus the proposed extension of the PS185 experiment could provide valuable insight into the proton spin puzzle. 0.3cm The mechanism which is responsible for the negative polarization of the strange sea is most probably of nonperturbative nature. Its origin can be linked to chiral dynamics [@BEK], and we shall discuss now a particular model based on this idea. We base our discussion on two starting points. First, the fact that the masses of pions and kaons are small at the typical hadronic scale can be attributed to the existence of strong attraction between quarks and antiquarks in the pseudoscalar $J^{PC}=0^{-+}$ channel. Second, from phenomenological analyses of the strange quark condensate in the framework of the QCD sum rules [@SVZ] it is known that the density of strange quark-antiquark pairs in QCD vacuum is quite high [@str]: $ <0\|\bar{s}s\|0>\simeq (0.8\pm 0.1) <0\|\bar{q}q\|0>$. Using the standard value of the light quark condensate [@SVZ], $<0\|\bar{q}q\|0>\simeq(250 MeV)^3$, we come to the conclusion that the density of strange quark-antiquark pairs in the vacuum is about $1\ fm^{-3}$. 0.3cm Let us now consider the basic $\|uud>$ proton state immersed in the QCD vacuum. The strong attraction in the spin-singlet pseudoscalar channel discussed above will induce correlations between light valence quarks from the proton wave function and vacuum strange antiquarks with opposite spins (see Fig.1). As a consequence of this, the spin of the strange antiquarks will be aligned [opposite]{} to the proton spin. Moreover, we note that in order to preserve the vacuum quantum numbers ($J^{PC}=0^{++}$), strange quark-antiquark pairs must be in a relative spin-triplet, $L=1$ $^3P_0$ state (see Fig.1). Therefore the spin of strange quarks must also be aligned [opposite]{} to the proton spin. The resulting wave function of the $\bar{s}s$ containing component, consistent with parity and spin constraints, corresponds to a spin-triplet, polarized $S_z=-1$ $\bar{s}s$ pair with angular momentum $L_z=+1$ coupled to the “usual" $S_z=1/2$ $\|uud>$ state. This wave function is similar to the one used in [@EKKS], which makes identical predictions for triplet–dominance and depolarization in $\bar{p}p\to\bar{\Lambda}\Lambda$. 0.3cm The picture advocated above should be contrasted with a similar but inequivalent one based on effective chiral theories with direct quark–Goldstone couplings [@Gold]: $$\L_{int} \sim \bar{\Psi}\gamma^{\mu}\gamma_5\Psi \partial_{\mu}\varphi, \label{L}$$ where $\Psi$ is a quark field, and $\varphi$ is the field of a (pseudoscalar) Goldstone boson. In these theories, a light quark can emit a spin-zero Goldstone boson, and this induces spin-flip of the quark. If the emitted boson is a K-meson, the emission turns the light quark into a strange quark with the opposite spin orientation (see Fig.2a). As before, this leads to the polarization of strange quarks opposite to the spin of the proton. The K-meson can in turn dissociate into a strange antiquark and light quark, which leads to formation of the $\|uud\bar{s}s>$ component considered above (see Fig.2b). However the Goldstone fields now are to be treated as elementary, spin-zero fields, and as such they dissociate into an [unpolarized]{} $(q\bar{s})$ pair. Though the net polarization carried by the $\bar{s}s$ pair is again opposite to the proton spin, the $\bar{s}s$ pair itself can be in either a spin-triplet or spin-singlet state with statistical weights. 0.3cm We observe that the mechanism responsible for the contribution of the strange sea to the proton spin can be tested in the process of proton–antiproton annihilation into the hyperon– antihyperon pair, $\bar{p}p\to\bar{\Lambda}\Lambda$. In the model we advocate above, this process can be viewed [@EKKS] as the dissociation of a spin-triplet $\bar{s}s$ pair from the initial proton or antiproton into a $\bar{\Lambda}\Lambda$ state (see Fig.3). Since the spin of the $\Lambda$ is carried by the spin of the strange quark, this (spin correlation conserving) dissociation leads to a spin–triplet final state for the two hyperons[^4]. This is indeed consistent with the experimental observation [@PS185] that the spin–singlet fraction in the $\bar{\Lambda}\Lambda$ final state is equal to zero within statistical errors. On the other hand, the effective chiral theory (\[L\]) would not lead [a priori]{} to this conclusion. 0.3cm However, dominance by the spin–triplet state of $\bar{\Lambda}\Lambda$ is not unique to the “intrinsic strangeness" model. It is also a feature of the naïve quark model approach to this process [@AHW], since in this approach the $\bar{s}s$ pair is produced through the $\bar{q}q\to\bar{s}s$ subprocess mediated either by a gluon exchange or by an effective scalar field; in both cases the structure of the $\bar{s}s$ producing vertex ($^3S_1$ and $^3P_0$ respectively) allows only a spin–triplet $\bar{\Lambda}\Lambda$ final state [@AHW]. 0.3cm There exists, however, a way to test the polarized intrinsic strangeness model in the $\bar{p}p\to\bar{\Lambda}\Lambda$ process. This model predicts more than just the dominance of the spin-triplet state in $\bar{\Lambda}\Lambda$. Since the initial $\bar{s}s$ pair carries a polarization opposite to the (anti)proton spin, it predicts that the spin of the final $S=1$ $\bar{\Lambda}\Lambda$ pair is polarized in the direction opposite to the spin of initial spin-triplet $\bar{p}p$ state. 0.3cm An experimental observable [@prop] which measures the amount of spin transferred from the initial-state proton to the final-state hyperon is the depolarization $D_{nn}$. Assuming a fully polarized proton target, the depolarization $D_{nn}$ ($\vec{n}$ is normal to the production plane) is $+1$ if the spin of the final-state $\Lambda$ hyperon is always parallel to the spin of the target, and $-1$ if the spin of the $\Lambda$ is always opposite to the spin direction of the target. The polarized intrinsic strangeness model in the idealized version described above therefore predicts $D_{nn}=-1$. 0.3cm We contrast this prediction with what we would expect within the polarized gluon interpretation of the proton spin puzzle. According to one favoured formulation of this interpretation [@gluon], the negative experimental value of the axial current matrix element $$<p\|\bar{s}\gamma_{\mu}\gamma_{5}s\|p> = \Delta s\ s_{\mu}, \label{spin}$$ where $s_{\mu}$ is the proton spin vector, is due to the $U(1)$ axial anomaly, which induces a correction: $$\Delta s = \Delta \hat{s} - {\alpha_{s} \over {2\pi}}\ \Delta G,$$ where $\Delta \hat{s}$ is the polarized $\bar{s}s$ contribution prior to quantum corrections, and $\Delta G$ is the net gluon polarization in the proton. In the most absolute version of this interpretation, $\Delta \hat{s}$ could vanish and the [negative]{} measured value $\Delta s <0$ could be entirely due to a [positive]{} value of $\Delta G$. If intrinsic gluons were responsible for $\bar{s}s$ production during the $\bar{p}p$ annihilation via the perturbative vertex $$\L_{QCD} \sim \bar{\Psi}\gamma_{\mu}\Psi\ G^{\mu}, \label{QCD}$$ the $\bar{s}s$ pair would be produced in a spin-triplet state, as inferred from the $\bar{\Lambda}\Lambda$ spin correlations. [However]{}, if $\Delta G > 0$ as suggested in the gluon interpretation of the proton spin puzzle, the depolarization should be [positive]{}: $D_{nn}>0$ (see Fig.4). We note however that the more conventional quark model based on effective vector exchange also predicts positive depolarization [@pos1]. 0.3cm Thus the measurement of the depolarization in the $\bar{\Lambda}\Lambda$ process could serve as an interesting test of the dynamics responsible for the apparently “anomalous" decomposition of the proton spin. 0.5cm We thank M. Karliner and M.G. Sapozhnikov for enjoyable collaboration at the initial stage of this work, E.M. Henley and L. Wilets for useful discussions, and K. Röhrich for his interest in this work and stimulating suggestions. The work of M.A. was supported in part by the U.S. Department of Energy and by the National Science Foundation, grant no. PHY-9223618. D.K. acknowledges support of the German Research Ministry (BMFT) under the Contract 06 BI 721. [999]{} The PS185 Collaboration, P. Barnes et al., Nucl.Phys. (1991) 575; Phys.Lett. (1989) 432; Phys.Lett. (1987) 249. M.A. Alberg, E.M. Henley, L. Wilets and P.D. Kunz, Nucl.Phys. (1993) 365;\ M. Kohno and W. Weise, Phys.Lett. (1988) 584; (1986) 15;\ S. Furui and A. Faessler, Nucl.Phys. (1987) 669;\ M. Burkardt and M. Dillig, Phys.Rev. (1988) 1362. F. Tabakin and R.A. Eisenstein, Phys.Rev. (1985) 1857;\ M. Kohno and W. Weise, [op. cit.]{};\ R.G.E. Timmermans, T.A. Rijken and J.J. de Swart, Phys.Rev. (1992) 2288;\ P. LaFrance, B. Loiseau and R. Vinh Mau, Phys.Lett. (1988) 317; Nucl.Phys. (1991) 557;\ J. Haidenbauer et al., Phys.Rev. (1992) 931; (1992) 2158. J. Gasser, H. Leutwyler and M.E. Sainio, Phys.Lett. (1991) 252; 260. The EMC Collaboration, J. Ashman et al., Phys.Lett. (1988) 364.\ The EMC Collaboration, J. Ashman et al., Nucl.Phys. (1989) 1.\ The NMC Collaboration, P. Amaudruz et al., Phys.Lett. (1992) 159.\ The SMC Collaboration, B. Adeva et al., Phys.Lett. (1993) 533.\ The E142 Collaboration, P.L. Anthony et al., Phys.Rev.Lett. (1993) 959.\ The E143 Collaboration, K. Abe et al., SLAC-PUB-6508. The ASTERIX Collaboration, J. Reifenrother et al., Phys.Lett. (1991) 299;\ The Crystal Barrel Collaboration, M.A. Faessler, Phys.At.Nuc. (1994) 1693;\ The OBELIX Collaboration, V.G. Ableev et al., Phys.At.Nuc. (1994) 1716; Phys.Lett. (1994) 237;\ The JETSET Collaboration, L. Bertolotto et al., Nucl.Phys. (1993) 27c. J. Ellis, E. Gabathuler and M. Karliner, Phys.Lett. (1989) 173. J. Ellis, M. Karliner, D.E. Kharzeev and M.G. Sapozhnikov, Preprint CERN-TH.7326/94, hep-ph/9412334. J. Ellis and M. Karliner, Phys.Lett. (1995) 397. G. Altarelli and G. Ross, Phys.Lett. (1988) 391;\ A.V. Efremov and O.V. Teryaev, Dubna report, JINR-E2-88-287, 1988;\ R.D. Carlitz, J.D. Collins and A.H. Mueller, Phys.Lett. (1988) 229. The PS185 Collaboration, K. Röhrich et al., CERN/SPSLC 95-13, Proposal P287, 1995. M.A. Alberg, E.M. Henley, L. Wilets and P.D. Kunz, Phys.At.Nuc. (1994) 1608. J. Haidenbauer et al., Phys.Lett. (1992) 223; Phys.Rev. (1992) 2158. S.J. Brodsky, J. Ellis and M. Karliner, Phys.Lett. (1988) 309. M.A. Shifman, A.I. Vainshtein and V.I. Zakharov, Nucl.Phys. (1979) 385; 448; 519. B.L. Ioffe, Nucl.Phys. (1981) 317 \[Erratum: (1981) 591\];\ L.J. Reinders, H. Rubinstein and S. Yazaki, Phys.Rep. (1985) 1. A. Manohar and H. Georgi, Nucl.Phys. (1984) 189;\ E.J. Eichten, I. Hinchliffe and C. Quigg, Phys.Rev. (1992) 2269;\ T.P. Cheng and L.-F. Li, Preprint CMU-HEP-94-30, hep-ph/9410345. 0.3cm Fig.1 a,b) Strong correlation between light valence quarks and vacuum strange antiquarks in the spin–singlet pseudoscalar channel induces a spin–triplet $\bar{s}s$ component of the proton wave function aligned opposite to the proton spin. (In all figures the direction of spin quantization is taken normal to the plane of the quark line diagrams.) 0.3cm Fig.2 a) The emission of a $K^+$ meson turns the light quark into a strange quark with the opposite spin orientation. b) Dissociation of the $K$ meson leads to formation of an $\bar{s}s$ component with [net]{} polarization opposite to the proton spin, but the $\bar{s}s$ pair can be in either a spin–triplet or a spin–singlet state with [a priori*]{} statistical weights. 0.3cm Fig.3 The $\bar{p}p\to\bar{\Lambda}\Lambda$ process viewed as the dissociation of a spin–triplet $\bar{s}s$ pair from the initial state proton (or antiproton) wave function into a $\bar{\Lambda}\Lambda$ state. The spin of the produced $\Lambda$ is always opposite to the spin of the initial proton. 0.3cm Fig.4 The $\bar{p}p\to\bar{\Lambda}\Lambda$ process viewed as the dissociation of a polarized gluon from the initial state proton (or antiproton) wave function into a spin–triplet $\bar{s}s$ state. The spin of the produced $\Lambda$ is parallel to the spin of the gluon, which is in turn parallel to the spin of the initial proton. [^1]: E-mail: [email protected] [^2]: E-mail: [email protected] [^3]: E-mail: [email protected] [^4]: Studies of initial– and final–state interactions suggest that these do not affect significantly the simple polarization arguments we present here and elsewhere in this paper. Calculations for several different initial- and final- state interactions show changes in $D_{nn}$ of order $30\%$, averaged over scattering angle. However, the distinction in sign between the quark models and meson-exchange models persists, i.e. $D_{nn}$ is positive for quark models and negative for meson-exchange models [@pos1],[@neg].	{ "pile_set_name": "ArXiv" }
--- abstract: 'Most of the existing work on automatic facial expression analysis focuses on discrete emotion recognition, or facial action unit detection. However, facial expressions do not always fall neatly into pre-defined semantic categories. Also, the similarity between expressions measured in the action unit space need not correspond to how humans perceive expression similarity. Different from previous work, our goal is to describe facial expressions in a continuous fashion using a compact embedding space that mimics human visual preferences. To achieve this goal, we collect a large-scale faces-in-the-wild dataset with human annotations in the form: Expressions A and B are visually more similar when compared to expression C, and use this dataset to train a neural network that produces a compact (16-dimensional) expression embedding. We experimentally demonstrate that the learned embedding can be successfully used for various applications such as expression retrieval, photo album summarization, and emotion recognition. We also show that the embedding learned using the proposed dataset performs better than several other embeddings learned using existing emotion or action unit datasets.' author: - \| Raviteja Vemulapalli\ Google AI\ [[email protected]]{} - \| Aseem Agarwala\ Google AI\ [[email protected]]{} title: A Compact Embedding for Facial Expression Similarity --- Introduction ============ Automatic facial expression analysis has received significant attention from the computer vision community due to its numerous applications such as emotion prediction, expression retrieval (Figure \[fig:example\]), photo album summarization, candid portrait selection [@Candid], etc. Most of the existing work [@EmotionSurvey; @AUSurvey] focuses on recognizing discrete emotions or action units defined by the Facial Action Coding System (FACS) [@FACS]. However, facial expressions do not always fit neatly into semantic boxes, and there could be significant variations in the expression within the same semantic category. For example, smiles can come in many subtle variations, from shy smiles, to nervous smiles, to laughter. Also, not every human-recognizable facial expression has a name. In general, the space of facial expressions can be viewed as a continuous, multi-dimensional space. In this work, we focus on learning a compact, language-free, subject-independent, and continuous expression embedding space that mimics human visual preferences. If humans consider two expressions to be visually more similar when compared to a third one, then the distance between these two expressions in the embedding space should be smaller than their distances from the third expression. To learn such an embedding we collect a new dataset, referred to as the Facial Expression Comparison (FEC) dataset, that consists of around 500K expression triplets generated using 156K face images, along with annotations that specify which two expressions in each triplet are most similar to each other. To the best of our knowledge, this is the first large-scale face dataset with expression comparison annotations. This dataset can be downloaded from <https://ai.google/tools/datasets/google-facial-expression/>. ![Expression retrieval results for embeddings learned using the proposed dataset (top) and an existing emotion classification dataset (bottom).[]{data-label="fig:example"}](images/kissy_face_comparison.jpg){width="45.00000%" height="20.00000%"} We show that a compact (16-dimensional) expression embedding space can be learned by training a deep network with the proposed FEC dataset using triplet loss [@FaceNet]. Based on the distances in the learned embedding space, we are able to predict the most similar pair in a triplet with an accuracy of 81.8% when evaluated on a held-out validation set. The accuracy of median human rater is 87.5% on this validation set, and the accuracy of random selection is 33.3%. We also show that the embedding learned using the FEC dataset performs better than several other embeddings learned using existing emotion or action unit datasets. We experimentally demonstrate that the expression embedding learned using the FEC dataset can be successfully used for various applications such as expression retrieval, photo album summarization, and emotion recognition. Our contributions ----------------- - We introduce the FEC dataset, which is the first large-scale face dataset with expression comparison annotations. This dataset is now available to public. - We experimentally demonstrate that a 16-dimensional expression embedding learned by training a deep neural network with the FEC dataset can be successfully used for several expression-based applications. - We show that the embedding learned using the FEC dataset performs better than several other embeddings learned using existing emotion or action unit datasets. Related work ============ Most of the existing research in the area of automatic facial expression analysis focuses on the following three topics: (i) Categorical model: Assigning discrete emotion category labels, (ii) FACS model: Detecting the presence/absence (and the strength) of various action units defined by FACS [@FACS], and (iii) Dimensional model: Describing emotions using two or three dimensional models such as valence-arousal [@ValAro], pleasure-arousal-dominance [@PAD], etc. Summarizing the vast amount of existing research on these topics is beyond the scope of this paper and we refer the readers to [@Aff-wild; @EmotionSurvey; @AUSurvey] for recent surveys on these topics.\ Expression datasets: Several facial expression datasets have been created in the past that consist of face images labeled with discrete emotion categories [@Emotionnet; @afew-7; @Afew; @Sfew; @Fer2013; @Multipie; @Raf-db; @CK+; @AffectNet; @Fer-Wild; @MMI; @ExpW; @Oulu-casia], facial action units [@Emotionnet; @CK+; @Disfa; @AM-FED; @MMI], and strengths of valence and arousal [@Deap; @Aff-wild; @afew-va; @AffectNet; @Reloca]. While these datasets played a significant role in the advancement of automatic facial expression analysis in terms of emotion recognition, action unit detection and valence-arousal estimation, they are not the best fit for learning a compact expression embedding space that mimics human visual preferences.\ Expression embedding: A neural network was trained in [@TripletEmbedding] using an emotion classification dataset and category label-based triplet loss [@FaceNet] to produce a 128-dimensional embedding, which was combined with an LSTM-based network for animating three basic expressions. Emotion labels do not provide information about within-class variations and hence a network trained with label-based triplets may not encode fine-grained expression information. The proposed FEC dataset addresses this issue by including expression comparison annotations for within-class triplets. A self-supervised approach was proposed in [@EmbeddingBMVC18] to learn a 256-dimensional facial attribute embedding by watching videos, and the learned embedding was used for multiple tasks such as head pose estimation, facial landmarks prediction, and emotion recognition by training an additional classification or regression layer using labeled training data. However, as reported in [@EmbeddingBMVC18], its performance is worse than existing approaches on these tasks. Different from [@EmbeddingBMVC18], we follow a fully-supervised approach for learning a compact (16-dimensional) expression embedding.\ Triplet loss-based representation learning: Several existing works have used triplet-based loss functions for learning image representations. While majority of them use category label-based triplets [@HTL; @TCLoss; @PersonReId; @DRDL; @FaceNet; @LiftingLoss; @AngularLoss; @BinaryEmb], some existing works [@ImgSimJMLR; @DeepRanking] have focused on learning fine-grained representations. While [@DeepRanking] used a similarity measure computed using several existing feature representations to generate groundtruth annotations for the triplets, [@ImgSimJMLR] used text-image relevance based on Google image search to annotate the triplets. Different from these approaches, we use human raters to annotate the triplets. Also, none of these works focus on facial expressions. ![image](images/zinta_smile.jpg){width="\textwidth" height="\textwidth"} ![image](images/pinto_smile.jpg){width="\textwidth" height="\textwidth"} ![image](images/male_happy.jpg){width="\textwidth" height="\textwidth"} ![image](images/federer_angry.jpg){width="\textwidth" height="\textwidth"} ![image](images/happy_surprise.jpg){width="\textwidth" height="\textwidth"} ![image](images/female_surprise.jpg){width="\textwidth" height="\textwidth"} ![image](images/kajal_laugh.jpg){width="\textwidth" height="\textwidth"} ![image](images/smith_angry.jpg){width="\textwidth" height="\textwidth"} ![image](images/contempt_male.jpg){width="\textwidth" height="\textwidth"} Facial expression comparison dataset {#sec::motivation} ==================================== In this section, we introduce the FEC dataset, which is a large-scale faces-in-the-wild dataset with expression comparison annotations provided by human raters. To the best of our knowledge, there is no such publicly-available expression comparison dataset. Most of the existing expression datasets are either annotated with emotion labels, or facial action units, or strengths of valence and arousal. One may think that we could generate comparison annotations for the existing datasets using the available emotion or action unit labels. However, there are several issues with such an approach: - Emotion labels do not provide information about within-class variations and hence we cannot generate comparison annotations within a class. For example, while all the expressions in Figure \[fig:triplets\](a) fall into the Happiness category, the left and middle expressions are visually more similar when compared to the right expression. Such within-class comparisons are important to learn a fine-grained expression representation. - Due to within-class variations and between-class similarities, two expressions from the same category need not be visually more similar when compared to an expression from a different category. For example, while the middle and right expressions in Figure \[fig:triplets\](b) belong to the Surprise category, the middle expression is visually more similar to the left expression which belongs to the Anger category. - It is difficult to predict the visual similarity relationships between expressions from three different emotion categories by using labels. For example, while the three expressions in Figure \[fig:triplets\](c) belong to three different categories, the left and middle expressions are visually more similar when compared to the right expression. Such comparisons are useful for learning an embedding that can model long-range visual similarity relationships between different categories. - It is unclear how the difference in the strengths of action units between two expressions could be converted into a distance function that mimics visual preferences. ![image](images/Faceexp_network_new.png){width="\textwidth"} Dataset {#sec::dataset} ------- Each sample in the FEC dataset consists of a face image triplet $(I_1, I_2, I_3)$ along with a label $L \in \{1, 2, 3\}$ that indicates which two images in the triplet form the most similar pair in terms of facial expression. For example, $L=1$ means $I_2$ and $I_3$ are visually more similar when compared to $I_1$. Note that these triplets do not have a notion of anchor, and each triplet provides two annotations: $I_2$ is closer to $I_3$ than $I_1$, and $I_3$ is closer to $I_2$ than $I_1$. This is different from the commonly-used triplet annotation [@FaceNet; @LMNN] that consists of an anchor, a positive image and a negative image. Also, in this dataset, an image A can be (relatively) closer to another image B in one triplet and (relatively) farther from the same image B in another triplet. This is different from the triplets generated using category labels [@FaceNet], in which any two images will either form a similar pair or a dissimilar pair in all the triplets they appear in. -- --------------- ----------- ----------- ------------- --------- -- One-class Two-class Three-class All Strong 115,544 124,665 117,540 357,749 Strong + Weak 137,266 138,034 132,435 407,735 All 152,674 150,234 146,235 449,143 Strong 13,046 14,607 13,941 41,594 Strong + Weak 15,411 15,908 15,404 46,723 All 17,059 17,107 16,894 51,060 Strong 128,590 139,272 131,481 399,343 Strong + Weak 152,677 153,942 147,839 454,458 All 169,733 167,341 163,129 500,203 -- --------------- ----------- ----------- ------------- --------- -- The triplets in the FEC dataset were generated by sampling images from a partially-labeled [^1] internal face dataset in which each face image has one or more of the following emotion labels [@emotions-pnas; @emotions-cognitive]: Amusement, Anger, Awe, Boredom, Concentration, Confusion, Contemplation, Contempt, Contentment, Desire, Disappointment, Disgust, Distress, Doubt, Ecstasy, Elation, Embarrassment, Fear, Interest, Love, Neutral, Pain, Pride, Realization, Relief, Sadness, Shame, Surprise, Sympathy, and Triumph. To reduce the effect of category-bias, we sampled the images such that all these categories are (roughly) equally represented in the triplet dataset. Each triplet was annotated by six human raters, and the raters were instructed to focus only on expressions ignoring other factors such as identity, gender, ethnicity, pose and age. A total of 40 raters participated in the process, each annotating a subset of the entire dataset. Based on the existing emotion labels, each triplet in this dataset can be categorized into one of the following types [^2]: - One-class triplets: All the three images share a category label, see Figure \[fig:triplets\](a). These triplets are useful for learning a fine-grained expression representation. - Two-class triplets: Only two images share a category label and the third image belongs to a different category, see Figure \[fig:triplets\](b). As explained in Section \[sec::motivation\], images sharing a category label need not form the (visually) most similar pair in these triplets. - Three-class triplets: None of the images share a common category label, see Figure \[fig:triplets\](c). These triplets are useful for learning long-range visual similarity relationships between different categories. One-class triplets are relatively the most difficult ones since the expressions could be very close to each other, and two-class triplets are relatively the easiest ones since the images sharing a label could potentially be different from the remaining image (though not always). While there are other possible types of triplets based on other label combinations (for example, $I_1$, $I_2$ sharing a label, and $I_2$, $I_3$ sharing another label), we prioritized the above three types while collecting the dataset as the other types could be confusing for the raters. Extending the dataset to include the other types will be considered in the future. Table \[tab:dataset\] shows the number of triplets in this dataset along with the number of faces used to generate the triplets. The dataset is further divided into training (90%) and test (10%) sets, and we encourage the users of this dataset to use the training set for training their algorithms and the test set to validate them.\ Annotation agreement: Each triplet in this dataset was annotated by six raters. For a triplet, we say that the raters agree strongly if at least two-thirds of them voted for the maximum-voted label, and agree weakly if there is a unique maximum-voted label and half of the raters voted for it. The number of such triplets for each type are shown in Table \[tab:dataset\]. Raters agree strongly for about 80% of the triplets suggesting that humans have a well-defined notion of visual expression similarity. Facial expression embedding network {#sec:network} =================================== In the recent past, the performance of face recognition systems has improved significantly [@MegaFaceLeaderboard; @MSCeleb1MLeaderboard; @MegaFace1; @MegaFace2] in part due to the availability of large-scale (several million data samples) training datasets such as MS-Celeb-1M [@MSCeleb1M], MegaFace [@MegaFace2], SFC [@DeepFace] and Google-Face [@FaceNet]. Neural networks trained on these large-scale datasets see images with significant variations along different dimensions such as lighting, pose, age, gender, ethnicity, etc. during training. Compared to these large-scale face datasets, our facial expression comparison dataset is significantly smaller (just 130K training faces). Hence, in order to leverage the power of a large training set, we build our facial expression embedding network using the pre-trained FaceNet proposed in [@FaceNet], see Figure \[fig:network\]. We use the NN2 version of pre-trained FaceNet [@FaceNet] up to the inception (4e) block [^3] whose output is a $7 \times 7$ feature map with 1024 channels. This feature map is processed by a DenseNet which consists of a $1\times 1$ convolution layer (512 filters) followed by a Dense block [@DenseNet] with 5 layers and growth rate of 64. The output of DenseNet is passed to a $7\times 7$ average pooling layer followed by a fully connected (FC) layer with 512 hidden units and an embedding layer (a linear FC layer + $\ell_2$ normalization layer). Batch normalization [@BatchNorm] and ReLu6 [@Relu6] activation function are used in the DenseNet and the first FC layer. We also use dropout for regularization. The input to our network is an aligned (rotated to undo roll transformation and scaled to maintain an inter-ocular distance of 55 pixels) $224\times 224$ face image $I$, and the output is a $d$-dimensional embedding $e_{I}$ of unit $\ell_2$ norm. Triplet loss function --------------------- For training the embedding network using the proposed FEC dataset, we use a triplet loss function that encourages the distance between the two images that form the most similar pair to be smaller than the distances of these two images from the third image. For a triplet $(I_1, I_2, I_3)$ with the most similar pair $(I_1, I_2)$, the loss function is given by $$\begin{aligned} l(I_1, I_2, I_3) &= max(0, \\|e_{I_1} - e_{I_2} \\|_2^2 - \\|e_{I_1} - e_{I_3} \\|_2^2 + \delta)\\ &+ max(0, \\|e_{I_1} - e_{I_2} \\|_2^2 - \\|e_{I_2} - e_{I_3} \\|_2^2 + \delta), \end{aligned}$$ where $\delta$ is a small margin. Experiments =========== In this section, we demonstrate the usefulness of the expression embedding learned from the proposed FEC dataset for various applications such as expression retrieval, photo album summarization, and emotion classification. In all our experiments, we only use the triplets with strong rater agreement for both training and evaluation. We also tried using the triplets with weak rater agreement for training, but the results did not improve (see Section \[sec::comparison\]). In the rest of the paper, we refer to the proposed expression embedding network trained on the proposed FEC dataset as FECNet. Comparative approaches ---------------------- Most of the existing large-scale expression datasets focus on the task of classification. One can train a classification network with such a dataset, and use the output of the final or penultimate layer as an expression embedding. Here, we train two networks: AFFNet-CL for emotion recognition using the AffectNet dataset [@AffectNet], and FACSNet-CL for facial action unit detection using the DISFA dataset [@Disfa]. AffectNet is a large-scale faces-in-the-wild dataset manually labeled with eight emotion categories. This dataset has around 288K training and 4K validation images. DISFA is a widely-used spontaneous facial actions dataset manually labeled with the presence/absence of 12 action units [^4]. This dataset has around 260K images, out of which 212K images are used for training and 48K images are used for validation. We create four expression embeddings using these two classification networks: - AFFNet-CL-P and AFFNet-CL-F: Penultimate and final layer outputs of AFFNet-CL. - FACSNet-CL-P and FACSNet-CL-F: Penultimate and final layer outputs of FACSNet-CL. Another way to learn an embedding using a classification dataset is to train an embedding network with a category label-based triplet loss similar to [@FaceNet]. So, we also train an embedding network (referred to as AFFNet-TL) on AffectNet dataset using triplet loss. For a fair comparison, the input and architecture of all the networks are chosen to be same as FECNet (Figure \[fig:network\]) except that the embedding layer is replaced by a softmax classifier for AFFNet-CL and separate binary classifiers for FACSNet-CL. Training and validation ----------------------- We define triplet prediction accuracy as the percentage of triplets for which the distance (in the embedding space) between the visually most similar pair is less than the distances of these two images from the third. As for the validation measure during training, we use triplet prediction accuracy on the FEC test set for FECNet, (following [@Fabnet]) average area under ROC curve (AUC-ROC) on the AffectNet validation set for AFFNet-CL and AFFNet-TL [^5], and (following [@FAUECCV18; @AUDandAlign]) average F1-score on the DISFA validation set for FACSNet-CL. For all the networks, the parameters of the FaceNet layers were kept fixed and the newly-added DenseNet and FC layers were trained starting from Xavier initialization [@Xavier] using Adam optimizer [@Adam] with a learning rate of $5\times10^{-4}$ and dropout of 0.5. FECNet was trained on the FEC training set with mini-batches of 90 samples (30 triplets from each of the triplet types) for 50K iterations, AFFNet-CL and AFFNet-TL were trained on the AffectNet training set with mini-batches of 128 samples (16 samples from each of the eight emotion categories) for 10K iterations, and FACSNet-CL was trained on the DISFA training set with mini-batches of 130 samples (at least 10 positive samples for each action unit and 10 samples with no action units) for 20K iterations. For training FECNet, the value of margin $\delta$ was set to 0.1 for one-class triplets, and 0.2 for two-class and three-class triplets. For training AFFNet-TL, the loss margin was set to 0.2 and the embedding dimensionality was set to 16. All the hyper-parameter values were chosen based on the corresponding validation measures. Average human performance ------------------------- To estimate how good humans are at identifying the most similar pair in a triplet, we computed the triplet prediction accuracy values for individual raters based on how often they agree with the maximum-voted label. Figure \[fig:human-accuracy\] shows the accuracy values for all the 30 raters who contributed to the FEC test set annotations. The mean and median values are 86.2% and 87.5%. Dimensionality of the FECNet embedding -------------------------------------- While we want to represent facial expressions in a continuous fashion using an embedding, it is unclear how many dimensions should be used for the embedding space. To answer this question, we trained FECNet for different values of the output dimensionality. Figure \[fig:dimensions\] shows how the triplet prediction accuracy on the FEC test set varies with the dimensionality of the embedding space. The accuracy increases till 16 dimensions and drops slightly after that. Based on these results, we choose 16-dimensions to represent the expression embedding space (referred to as FECNet-16d). Figure \[fig:dimensions\] also shows the median rater accuracy. Using 16 dimensions, the proposed FECNet is able to achieve an accuracy of 81.8%, which is fairly close to the median rater accuracy (87.5%). Note that the triplet prediction accuracy of random choice is 33.3%. -------------- ---------- ---------- -------- $\ell_1$ $\ell_2$ Cosine FACSNet-CL-F 47.1 47.1 40.7 FACSNet-CL-P 45.3 44.2 48.3 AFFNet-CL-F 49.0 47.7 49.0 AFFNet-CL-P 52.4 51.6 53.3 AFFNet-TL - 49.6 - FECNet-16d - 81.8 - -------------- ---------- ---------- -------- : Triplet prediction accuracy on the FEC test set.[]{data-label="tab:affnet-embeddings"} Comparison of different embeddings {#sec::comparison} ---------------------------------- Table \[tab:affnet-embeddings\] shows the triplet prediction accuracy of various embeddings on the FEC test set using different distance functions. Among all the AFFNet and FACSNet embeddings, the combination of AFFNet-CL-P and cosine distance gives the best accuracy, and hence, we use this combination for comparison with FECNet-16d in the rest of the experiments. It is worth noting that the proposed FECNet-16d (81.8%) performs significantly better than the best competing approach (AFFNet-CL-P + Cosine distance; 53.3%). We also trained FECNet-16d by adding the triplets with weak rater agreement to the training set, but the test accuracy dropped from 81.8% to 80.5%. Triplet type AFFNet-CL-P FECNet-16d Median rater -------------- ------------- ------------ -------------- One-class 49.2 77.1 85.3 Two-class 59.8 85.1 89.3 Three-class 50.4 82.6 87.2 All triplets 53.3 81.8 87.5 : Triplet prediction accuracy for different types of triplets in the FEC test set.[]{data-label="fig:triplet-type-performance"} Performance for different triplet types --------------------------------------- Table \[fig:triplet-type-performance\] shows the triplet prediction accuracy of median rater, FECNet-16d and AFFNet-CL-P for each triplet type in the FEC test set. As expected, the performance is best (85.1%) for two-class triplets, which are relatively the easiest ones, and is lowest (77.1%) for one-class triplets, which are relatively the most difficult ones. Visualization of the FECNet embedding space ------------------------------------------- ![2D visualization of the FECNet-16d embedding space using t-SNE [@tSNE].[]{data-label="fig:tSNE"}](images/tSNE.png) Figure \[fig:tSNE\] shows a 2D t-SNE [@tSNE] visualization of the learned FECNet-16d embedding space using the AffectNet validation set. The amount of overlap between two categories in this figure roughly tells us about the extent of visual similarity between them. For example, fear and surprise have a high overlap indicating that they could be confused easily, and both of them have a very low overlap with contempt indicating that they are visually very distinct from contempt. Also, the spread of a category in this figure tells us about the visual diversity within that category. For example, happiness category maps to three distinct regions indicating that there are three visually distinct modes within this category. See Figure \[fig:tsne\_faces\] for a visualization of the face images that fall into different regions in Figure \[fig:tSNE\]. Album BO CB DT GB HC JL JC KM LJ LS --------------------------- ----- ----- ----- ----- ------ ----- ----- ------ ----- ----- FECNet-16d vs AFFNet-CL-P 5-2 9-1 5-1 9-0 10-0 9-0 7-1 10-0 1-4 1-6 Applications ------------ ### Image retrieval We can perform expression-based image retrieval by using nearest neighbor search in the expression embedding space. To compare the retrieval performance of FECNet-16d and AFFNet-CL-P embeddings, we use a query set consisting of 25 face images and a database (CelebA dataset [@CelebA]) consisting of 200K face images. For each query, we retrieved $N$ nearest neighbors ($N$ varied from 1 to 10) using both the embeddings and ranked the $2N$ retrieved images based on how close they are to the query as judged by ten human raters. Since ranking all $2N$ images at once is difficult for human raters, we asked them to rank two images at a time. In each pairwise ranking, the winner and looser get a score of +1 and -1, respectively. If it is a tie, i.e., the two images get equal number of rater votes, then both of them get a score of zero. We obtained such pairwise ranking scores for all pairs and converted them into a global ranking based on the overall scores. For numerical evaluation, we use rank-difference metric, defined as the average difference in the ranks of images retrieved by AFFNet-CL-P and FECNet-16d embeddings, respectively, divided by the number of retrieved images $N$. Positive value of this rank-difference metric indicates that FECNet-16d embedding is better than AFFNet-CL-P embedding. The lowest value for this metric is $-1$, corresponding to the case when all the AFFNet-CL-P retrieval results are ranked lower than all the FECNet-16d retrieval results, and the highest value is $+1$, corresponding to the case when all the FECNet-16d retrieval results are ranked lower than all the AFFNet-CL-P retrieval results. Figure \[fig:rank-difference\] shows the rank-difference metric for different values of $N$. Positive value of the metric for all values of $N$ clearly indicates that the proposed FECNet-16d embedding produces better matches compared to the AFFNet-CL-P embedding. Figure \[fig:retrieval\_results\] shows the top-5 retrieved images for some of the queries. The overall results of the proposed FECNet-16d embedding are clearly better than the results of AFFNet-CL-P embedding. Specifically, the FECNet-16d embedding pays attention to finer details such as teeth-not-visible (first query), eyes-closed (second and third queries) and looking straight (fourth query). See Figures \[fig:retrieval-first\] to \[fig:retrieval-last\] for additional retrieval results. ### Photo album summarization In this task, we are interested in summarizing the diverse expression content present in a given photo album using a fixed number of images. Expression embedding can be used for this task by combining it with a clustering algorithm. For evaluation, we created ten photo albums (100-200 images in each album) by downloading images of ten celebrities using Google image search. For each album, we ran hierarchical agglomerative clustering [^6] (10 clusters) with FECNet-16d and AFFNet-CL-P embeddings, and used the images that are closest to cluster centers for generating the summaries. We showed these two summaries to ten human raters and asked them which one is better. The raters were also allowed to choose the difficult-to-decide option. Table \[tab:clustering\_results\] shows the number of votes received by both the embeddings for all the albums. Humans prefer the summaries generated by the proposed FECNet-16d embedding for eight out of ten albums. Figures \[fig:summary1\] and \[fig:summary2\] show the summary results for all the albums. We can see that the expression content is more diverse in the summaries produced by the FECNet-16d embedding for most of the albums. Approach Neutral Happiness Sadness Surprise Fear Disgust Anger Contempt Average ------------------------------- ---------- ----------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- AFFNet-CL 84.6 96.5 90.7 88.5 90.2 85.2 88.3 85.0 88.6 AFFNet-TL 85.9 96.0 89.2 88.5 89.6 83.7 87.9 82.6 87.9 FECNet-16d + K-NN 83.3 94.9 78.0 83.0 84.5 79.3 78.7 81.2 82.9 AlexNet [@AffectNet] - - - - - - - - 82.0 VGG-Face descriptor [@Fabnet] 75.9 92.2 80.5 81.4 82.3 81.4 81.2 77.1 81.5 FAb-Net [@Fabnet] 72.3 90.4 70.9 78.6 77.8 72.5 76.4 72.2 76.4 ![image](images/Retrieval_results.jpg){width="\textwidth" height="50.00000%"} \[fig:retrieval\_results\] ### Emotion classification The proposed FECNet-16d embedding can be used for emotion classification by combining it with K-Nearest Neighbor (K-NN) classifier. Figure \[fig:classification\] shows the average AUC-ROC of the FECNet-16d embedding on the AffectNet validation set as a function of the number of neighbors used. The performance increases up to 200 neighbors and then remains stable. Table \[tab:affnet-classification\] compares the classification performance of the FECNet-16d embedding (using 200 neighbors) with other approaches. Note that AFFNet-CL and AFFNet-TL have the same architecture as FECNet-16d and are specifically trained for classification using AffectNet training data. Hence, as expected, they perform a bit better than FECNet-16d. However, despite not being trained for classification, the FECNet-16d embedding outperforms AlexNet and VGG-Face based classifiers, demonstrating that it is well-suited for classification. Conclusions and Future Work =========================== In this work, we presented the first large-scale facial expression comparison dataset annotated by human raters, and learned a compact (16-dimensional) facial expression embedding using this dataset with triplet loss. The embedding learned using this dataset performs better than various other embeddings learned using existing emotion and action units datasets. We experimentally demonstrated the usefulness of the proposed embedding for various applications such as expression retrieval, photo album summarization, and emotion classification. Another interesting application of the FECNet embedding is hard-negative mining for expression classification. Since FECNet is trained using human visual preferences, negative samples that are close to the positive samples in the FECNet embedding space can be considered as hard negatives while training a classification model. We plan to explore this further in our future work. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Gautam Prasad, Ting Liu, Brendan Jou, Alan Cowen, Florian Schroff and Hartwig Adam from Google for their support and suggestions during the data collection process. ![image](images/tsne_faces.jpg) \[fig:tsne\_faces\] ![image](images/expnet_kate.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_kate.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_jim.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_jim.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_jennifer.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_jennifer.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_cardi.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_cardi.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_bush.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_bush.jpg){width="8cm" height="3.5cm"} \[fig:summary1\] ![image](images/expnet_hillary.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_hillary.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_obama.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_obama.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_lilly.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_lilly.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_lebron.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_lebron.jpg){width="8cm" height="3.5cm"} ![image](images/expnet_trump.jpg){width="8cm" height="3.5cm"} ![image](images/affnet_trump.jpg){width="8cm" height="3.5cm"} \[fig:summary2\] ![image](images/q21_2.jpg){width="85.00000%" height="85.00000%"} ![image](images/q9_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q20_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q12_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q34_1.jpg){width="85.00000%" height="85.00000%"} \[fig:retrieval-first\] ![image](images/q28_2.jpg){width="85.00000%" height="85.00000%"} ![image](images/q29_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q5_3.jpg){width="85.00000%" height="85.00000%"} ![image](images/q13_4.jpg){width="85.00000%" height="85.00000%"} ![image](images/q2_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q16_2.jpg){width="85.00000%" height="85.00000%"} ![image](images/q18_2.jpg){width="85.00000%" height="85.00000%"} ![image](images/q23_3.jpg){width="85.00000%" height="85.00000%"} ![image](images/q22_3.jpg){width="85.00000%" height="85.00000%"} ![image](images/q25_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q25_2.jpg){width="85.00000%" height="85.00000%"} ![image](images/q27_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q6_3.jpg){width="85.00000%" height="85.00000%"} ![image](images/q1_1.jpg){width="85.00000%" height="85.00000%"} ![image](images/q10_2.jpg){width="85.00000%" height="85.00000%"} \[fig:retrieval-last\] [^1]: The images in this dataset are not exhaustively labeled, i.e., an image may not have all the labels that are applicable to it. [^2]: The images in the dataset (from which we sampled the faces) were not exhaustively labeled, and hence, a triplet classified as a two/three-class triplet based on the existing labels may not be be a two/three-class triplet if the images had been exhaustively labeled. [^3]: We also experimented with features from inception 4d, 5a and 5b blocks, and features from 4e block performed the best. [^4]: The frames with action unit intensities greater than 2 are treated as positives and the remaining are treated as negatives. [^5]: Nearest neighbor classifier with 800 neighbors is used. [^6]: Cosine distance and maximum linkage were used.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The probability of quantum relocation of a human body, at a given distance, is estimated using two different methods, giving comparable results. Not only the obtained values for the probabilities are inconceivably small, but assumptions of a sci-fi nature are also necessary to ensure that they are not identical to zero. The notions of ‘non-spatiality’ and ‘superselection rule’ are also briefly discussed.' author: - Massimiliano Sassoli de Bianchi title: 'On the quantum “self-teleportation” probability of a human body' --- Introduction ============ Recently a science fiction writer asked me the following question:\ What is the probability for an individual to suddenly vanish from one place and, one second after, reappear in another predetermined place, tens of kilometers away, according to the laws of quantum physics?\ He also told me that a famous physicist (he had forgotten the name) used to pose such question to his students, so that it necessarily had to be a simple textbook problem. His interest in this question was that the protagonist of his story had to take advantage of this probability, no matter how infinitesimal (he was equipped with a futurist amplifier of probabilities), to “transfer” all of a sudden his body to a considerable spatial distance. Inspired by his curious quiz, which only apparently is a textbook one, I will try in this article to offer a few elements of clarification about some important concepts of quantum physics, in particular the concept of non-spatiality, which I will illustrate by means of a simple metaphor. I will also provide two different estimates of the teleportation probability in question, on the basis of a number of simplifying assumptions, some of which will necessarily be of a sci-fi nature. Despite these assumptions, the values I will obtain are so small that they are almost impossible to conceive. Not to create any misunderstanding, let me assert very clearly, from the beginning, that the probability that in normal conditions an individual would disappear from one place and be teleported to another place is, according to the today known laws of quantum physics (and the author’s personal understanding of them), exactly equal to zero! In fact, as I will explain, and until evidence to the contrary, ordinary macroscopic bodies (such as our human bodies), in standard environmental conditions (for instance of temperature and pressure), do not obey the quantum laws. But before proceeding in my discussion, I have to face a little problem of terminology. The term “teleportation” is used in quantum physics to denote a very specific class of phenomena that have nothing to do with the nature of the question addressed to me by the science fiction writer [@Bennet1993]. These phenomena describe the possibility of carrying information from one place to another, in ways that allow the construction of an exact duplicate of a given physical entity. This construction can be obtained only on the condition that the system of origin (i.e., the system that is to be duplicated) is altered, if not destroyed, in the process, since a well-known theorem, called the no-cloning theorem, forbids to create a perfect clone of a quantum entity [@Wootters1982; @Dieks1982]. Apart from this difficulty, the quantum teleportation, understood in the usual sense mentioned above, requires the preparation of special pairs of non-separated (entangled) systems that have to connect the two spatial regions between which the teleportation is to be produced, which for this reason is also called entanglement-assisted teleportation. In other words, this form of quantum teleportation requires the presence of technological apparatuses tailored to the specificities of the entities to be teleported, and the execution of a series of operations that wil produce their destruction and reconstitution in the place of destination. This is not a spontaneous process, associated with probabilities, but a determinative process, which requires a specific technology to be implemented. Let me add that in the quantum teleportation only the information about the entity is transported, so as to allow its reconstruction, while nothing material is actually moved (except the carriers of information along an ordinary communication channel); and even though experiments of entanglement-assisted teleportation have already been successfully carried out (the current record is a teleportation over a distance of 143 km, between the Canary Islands of La Palma and Tenerife [@Ma2012]), these remain so far limited to individual microscopic entities and finite-dimensional physical observables, such as the polarization of a photon, or the spin of an electron. Having said that, and in order to avoid misunderstandings, I will use in the following the term “quantum self-teleportation,” or more simply self-teleportation, to designate a hypothetical process of spatial relocation of a physical entity, to distinguish it from the aforementioned quantum teleportation assisted by entanglement. As previously emphasized, self-teleportation does not seem to be possible for macroscopic bodies, in standard conditions, as they only obey the laws of classical physics. Therefore, some additional sci-fi-like assumptions will be needed to explore this possibility and provide an estimate of the probability of such event, for a macroscopic entity like the body of a human being. Non-spatiality ============== I will start by explaining a little better why a macroscopic body cannot behave like a microscopic entity. It is important to observe that macroscopic bodies, such as human bodies, or whatever ordinary objects, like rocks, grains of sand, etc., are spatial entities. This means that they evolve while remaining within the so-called 3-dimensional Euclidean space. To clarify what I mean by this, I will use a simple metaphor. Imagine a swimmer in a pool. The pool’s water corresponds to the 3-dimensional physical space, and the swimmer in it represents a macroscopic entity. If she wants to move from one point to another of the pool, that is, from one point in space to another point in space, she can only do so by swimming, and of course, due to the viscosity of water, the speed of her movement will be limited: she will not be able to exceed a determined maximum speed, which we can assume to be, say, of $2$ m/s. Thus, if we assume that the swimmer is located near the trampoline, and she wants to reach a point located at the center of the pool, say $10$ m away, this will take her $5$ s, if she can travel at the maximum possible speed. Imagine now a child on the trampoline. In this metaphor the child is a microscopic entity, located outside of the pool’s water, that is, outside the ordinary 3-dimensional physical space. Indeed, microscopic entities, when not organized into macroscopic aggregates, or when not interacting with macroscopic entities, are typically non-spatial entities [@Aerts1998; @Aerts1999; @Sassoli2011; @Sassoli2012; @Sassoli2013; @AertsSassoli2014; @Sassoli2015], not belonging to the water of the pool. The child, as a non-spatial entity, “moves” through another “space,” which in a sense is adherent to our physical space, and which in our metaphor is represented by the layer of air above the pool; and since the viscosity of air is lower than that of water, he will be able to do so with greater effective speed than the swimmer. Suppose that the limit speed in the air, for the child, is of $10$ m/s, and that he is actually running at that speed on the trampoline, while in the process of diving. He will then be able to pass from the region of the trampoline to the region of the center of the pool in about $1$ s, which is something the swimmer is obviously unable to do. The interesting thing is that from the perspective of the swimmer, it is as if the diving child would appear out of nowhere in the middle of the pool, because he was not moving through the water, as the swimmer is forced to do, but through the air, which corresponds to a different layer of reality, of a non-ordinary kind, that we cannot directly perceive using our ordinary perceptual tools. I hope it is clear to everyone that a swimmer who is immersed in the water of the pool will never be able to move from one point to another as a diver (who is outside of it) can do. Similarly, a macroscopic body (the swimmer in our metaphor), being forced to move while remaining in the 3-dimensional physical space, will never be able to mimic the behavior of a microscopic entity, which instead is almost always outside of it (unless of course we would find a way to bring it out from that water, which is the sci-fi hypothesis we will have later on to consider). There are several ways to infer the mysterious non-spatiality of the microscopic entities. The simplest is to take seriously the uncertainty principle of Heisenberg. In fact, according to it, it is not possible to simultaneously determine both the position and momentum of a microscopic entity. Therefore, it is not possible to solve the equations of motion (which require as an input both quantities), and as a consequence it is not possible to determine the spatial trajectory of the entity in question. This impossibility does not arise from the fact that we would lack some crucial information about the state of the entity (as the no-go theorems about hidden-variable theories illustrate [@Neumann1932; @Bell1966; @Gleason1957; @Jauch1963; @Kochen1967; @Gudder1970]), which if we would possess would allow us to determine its trajectory: it is an impossibility of a fundamental, irreducible nature, which forces us to acknowledge that such trajectory in space does not exist, and since it does not exists, we must also abandon the idea that a microscopic entity would be always present in the 3-dimensional space. In the words of the previous metaphor, a microscopic entity is essentially a diver, not a swimmer, and if you look for a diver you will find him almost always on the trampoline, or in the air, and not in the water. On the other hand, the human body, which is a macroscopic entity, is a genuine spatial entity, that is, a swimmer, not a diver, who cannot disappear from space as if by magic, only to reappear in another region of the same; certainly not in normal conditions, and according to the known laws of physics. Le me add a further terminological clarification. In the scientific literature the term non-spatiality is much less used than that of non-locality. However, both terms express the same idea. In fact, all that is stably present in our three-dimensional space is necessarily local, that is, locally present in it, in actual and not in potential terms (also an extended object, like a cloud, is a local object, as it possesses local actual properties). Therefore, what is not present in a local sense is in fact not present at all (which doesn’t mean it doesn’t exist), and consequently the concepts of non-locality and non-spatiality are intimately related. Now, as I tried to illustrate with the metaphor of the pool, reality is layered, and one of these layers is that in which the microscopic entities live: it is a layer that could be called prespatial (and which in a sense is also pretemporal). In adherence to this prespatial layer (represented in the metaphor as the layer of air above the pool), lies our ordinary spatial layer (the water in the pool), inside which macroscopic entities usually evolve, like the objects of our daily lives, and our human bodies. Of course, the pool metaphor is only to be understood as a very crude allegorical simplification. The non-spatial or prespatial layer is a non-ordinary reality whose dimensionality is much higher than the three dimensions of our ordinary space, or the four dimensions of spacetime, and in general it could even be considered to be infinite-dimensional (as infinite-dimensional is in general the Hilbert state-space of a quantum entity). This cannot be represented in the too simple pool metaphor, in which the dimensionality of the air region above the pool, and of the water region inside the pool, is the same. Also, the region of contact between the spatial and prespatial layers is much more articulate and intricate than what the metaphor suggests, and certainly the non-spatial (or pre-spatial) entities cannot be represented as simple corpuscular entities. Wave-packet spreading ===================== The problem that we need first to consider is the evolution of the probability of presence (in space) of a microscopic entity, such as a single atom, when it evolves freely, i.e., when no external forces or other entities (microscopic or macroscopic) interact with it (apart the measuring system). The term “probability of presence” should be understood in the sense of the probability with which the microscopic entity in question lends itself to the creation of a spatial localization, in a given region of space $R$, at a given time $t$, through its interaction with a measuring apparatus [@Aerts1999; @Sassoli2012]. In quantum theory, this probability is given by the squared modulus $\|\psi_t({\bf x})\|^2$ of the wave function (or wave packet) $\psi_t({\bf x})$ (describing the state of the entity in question, at time $t$), integrated over the spatial region $R$, that is: $${\cal P}_t(R)=\int_R d{\bf x}\, \|\psi_t ({\bf x})\|^2.$$ What I am now going to do is to estimate the width of such wave packet in the simplest case of a hydrogen atom, which is the first and simplest element of the famous Mendeleev’s periodic table. To determine the wave function of a hydrogen atom, it is useful to express the problem in the so-called variables of the center of mass and relative movement. In doing so, I will neglect for simplicity the description of the spins of the electron and proton. Without going into the details of this procedure, which can be found in any textbook of quantum mechanics, we can observe that due to this change of variables it is possible to transform the problem of two interacting bodies (electron + proton) into an effective, simpler problem, of two bodies that evolve independently of each other, and whose equations can therefore be solved separately. The first body corresponds to the evolution of the center of mass of the system, and is equivalent to the evolution of a free entity (an entity evolving in the absence of any interaction) of total mass $M=M_e+M_p$, where $M_e$ and $M_p$ are the masses of the electron and proton, respectively. The second body corresponds instead to the evolution of a entity of (reduced) mass $\mu = M_e M_p/M$, which moves in the presence of a Coulombian central force field. The solutions of the Schrödinger equation associated with the first problem are the so-called plane waves, which cover a continuum of possible energies, from zero to infinity (we speak in this case of a continuous spectrum). The solutions of the problem with the central force field are instead associated with discrete energy values, given by the well-known formula: $E_n= -E_I/n^2$, where $n=1,2,\dots$, and $E_I\approx 13.6\, eV\approx 22 \times 10^{-19}$ J is the ionization energy of the hydrogen atom. One speaks in this case of a discrete spectrum, corresponding to the known (emission and absorption) spectral lines, observed experimentally. Now, as regards the possibility of acquiring different positions in space, what really matters is the movement of the center of mass of the hydrogen atom, which, as previously mentioned, evolves according to free evolution. What we are interested in is to calculate the spatial spreading of the wave packet associated with the center of mass variable, since such spreading will provide us a good estimate of the probability of observing the hydrogen atom at a certain distance from the place where it was initially observed, say at time $t = 0$ s. The spatial spreading of the wave packet at time $t$ can be estimated by calculating the so-called standard deviation $\Delta Q_t$ of the position observable ${\bf Q}$ (associated with the center of mass), which by definition is given by the square root: $\Delta Q_t =\sqrt{\langle {\bf Q}^2\rangle_t- \langle {\bf Q}\rangle_t^2}$, where $\langle\cdots\rangle_t$ denotes the quantum average relative to the state of the center of mass entity. Using Ehrenfest theorem, with which one can calculate the average values of quantum observables, one can show (following a little long but not difficult calculation) that by judiciously choosing the origin of the time axis the spreading $\Delta Q_t$ of the center of mass wave packet, at time $t$, is given by: $$\Delta Q_t = \sqrt{{\Delta P_0^2\over M^2} \, t^2 + \Delta Q_0^2}, \label{eq1}$$ where $\Delta Q_0$ is the spatial spreading at time $t=0$, and $\Delta P_0$ the spreading with respect to the momentum observable ${\bf P}$, at time $t=0$. For the initial width $\Delta Q_0$ of the packet, we can choose the typical value of the Bohr radius (which in the semiclassical model of the Danish physicist corresponds to the radius of the innermost electron), i.e., about $5.3 \times 10^{-11}$ m ($0.53$ angstroms). For the value of $\Delta P_0$, we can instead consider a dispersion which is compatible with the energy of the ground state of the hydrogen atom, i.e., such that ${\Delta P_0^2\over 2M}\approx E_I$. Considering that the total mass is: $M \approx M_p \approx 1.67 \times 10^{-27}$ kg, we have $\Delta P_0 \approx 8.6 \times 10^{-23}$ Js/m, which is compatible with Heisenberg’s principle, as is clear that: $\Delta Q_0\Delta P_0 \approx 45.6 \times 10^{-34}~{\rm Js}\approx 43\, \hbar > \hbar / 2$ ($\hbar \approx 1.05 \times 10^{-34}$ Js). Inserting the above values in (\[eq1\]), and observing that the second term in the square root is negligible compared to the first, we thus obtain $\Delta Q_t \approx 5.1 \times 10^4\, t$ m/s, that is: $$t\approx 0.2\times 10^{−4}\,\, \Delta Q_t ~{\rm s/m}. \label{eq1bis}$$ This last expression tells us the time we roughly need to wait for the center of mass of the wave packet of the hydrogen atom to reach the spatial spreading $\Delta Q_t$. Let us consider some specific values. To obtain a spreading of $5$ km, that is, of $5 \times 10^3$ m, we have to wait about $10^{-1}$ s, i.e., a tenth of a second. In 1 s, instead, the packet will have reached a width of about $50$ km, while in $10$ s its approximate width will be of $500$ km, and so forth. In other words, the effective speed with which the radial dimension of the center of mass wave packet grows, is approximately $50$ km/s, that is $180,000$ km/h, which is a speed of all respect, and corresponds, in our previous metaphor, to the maximum speed of the diver (from our ordinary spatial perspective this is however only a potential speed, and certainly not an actual speed, as it is not associated with a body moving through our ordinary space). Disassembling the body ====================== Summarizing, for a hydrogen atom we have determined the approximate behavior of that part of the wave function which describes the potential spatial localization of its center of mass. In doing so, we have ignored for simplicity the relative motion between the proton of the nucleus and the orbital electron, as well as their spins. More precisely, we have calculated how the width of the center of mass wave function varies over time in (configuration) space, due to the so-called phenomenon of the spreading of the wave function, which can be understood as being a consequence of Heisenberg’s uncertainty principle. What is important to understand is that the domain in which the wave function is sensibly different from zero corresponds to the spatial region within which the atom in question has a chance of being detected. So, if the hydrogen atom, at time $t= 0$, was localized in a sphere whose radius is approximately equal to the Bohr radius, i.e., $r_0 = 5.3 \times 10^{-11}$ m (that is, its probability of presence in that sphere, at time $t=0$ s, is equal to $1$), what we have determined is that after for example $1$ s, that localization radius will have approximately grown to about $50~{\rm km} = 5 \times 10^4$ m. What we are interested in is to estimate the probability with which we can detect the atom not in any location of this macro-sphere of $50$ km of radius, but in a predetermined sub-region of it. Indeed, if we later want to extrapolate our reasoning to an entire macroscopic structure, it is necessary that every atom forming the structure will re-locate in a very specific place in relation to all the other atoms of the structure, so as to reconstitute it in every detail. So, let us suppose that this sub-region corresponds to a micro-sphere whose radius is equal to the Bohr radius $r_0$. To estimate the above probability I will make an additional simplifying assumption. The wave function being not a constant function, the probability of presence will vary according to the location of the micro-sphere within the macro-sphere. However, since we are only interested in estimating a rough order of magnitude, we can assume that the wave function is a step-function, only taking two values: a constant non-zero value inside the macro-sphere, and a zero value outside of it. With this simplification, we have everything we need to complete our estimation. For this, we have to remember that to calculate a probability of presence we have to integrate the squared module of the wave function over the region of interest. Considering the step-function hypothesis, this means that the probability that we seek will be proportional to the relative volume of the micro-sphere compared to the volume of the macro-sphere. More exactly, given that the volume of a sphere is proportional to its radius to the cube, we obtain for the probability ${\cal P}$ that the hydrogen atom in question will be detected, after $1$ s, in a predetermined micro-sphere of radius $r_0$, within the macro-sphere of radius $r= 50$ km, the following order of magnitude: $${\cal P}\approx {r_0^3\over r^3}\approx \left({5.3\times 10^{-11} \,{\rm m}\over 5\times 10^4 \,{\rm m}}\right)^3\approx 10^{-45}. \label{eq2}$$ This is undoubtedly a very small number, with 45 zeros after the decimal point! And of course, we can easily do the same calculation for larger macro-spheres, i.e., waiting more time than just a second. For example, if we wait $10$ s, the radius of the macro-sphere will increase by a further factor of 10 (from $50$ km to $500$ km), and consequently the estimated value of the probability ${\cal P}$ will decrease from $10^{-45}$ to $10^{-48}$, i.e., by a factor of a thousand, and so on. Now that we have obtained an estimate of the probability of quantum self-teleportation of a hydrogen atom from an initial micro-sphere to a given final micro-sphere, we must consider the case of an entire macroscopic body, like that of a human being of planet Earth, which we can assume to have a mass of $100$ kg. Here of course we have to face the already mentioned problem that a macroscopic body has the property of spatiality, and therefore cannot be conveniently described by a wave function (more will be said about this in the next section). But suppose that for some reasons, unknown to us, all the interatomic bonds suddenly cease to exist, so that in an instant all the atoms that form the human body in question become separate and independent from each other, bringing them back to the prespatial layer of our physical reality. On the basis of this sci-fi hypothesis, each individual atom of the body structure can be conveniently described by the laws of quantum mechanics, and we can apply the previous calculation to each one of them. In fact, this is not really true, as is clear that all these atoms will constantly be bombarded by the countless entities present in the environment, in particular the thermal photons, so that we also need to assume that the sci-fi process of disassembly of the human body is able to induce a perfect isolation of the different atomic constituents from all the other entities (micro and macro) present in the environment. To keep the discussion as simple as possible, we further assume that the body structure is constituted solely by hydrogen atoms, and since the mass of a hydrogen atom is about $1.67\times 10^{-27}$ kg, a human of $100$ kg, if constituted only by hydrogen atoms, would contain approximately a number $N$ of them given by: $$N\approx {100\, {\rm kg}\over 1.67 \times 10^{-27}\, {\rm kg}}\approx ≈ 4.2 \times 10^{28}. \label{eq3}$$ Each of these atoms will have to individually re-locate in a specific region of space, to reconstitute the entire body structure, with no errors. Therefore, the (estimated) probability ${\cal P}_{\rm body}$ of self-teleportation of the overall body structure will be given by the product of the self-teleportation probabilities ${\cal P}$ of every single atom contained in that structure. If the body would be formed only by two atoms, that is, $N= 2$, the probability would be ${\cal P}_{\rm body} \approx {\cal P}^2\approx 10^{-45\times 2} = 10^{-90}$. With three atoms, the probability would become: ${\cal P}_{\rm body} \approx {\cal P}^3\approx 10^{-45\times 3} = 10^{-135}$. Therefore, with $N= 10^{28}$ atoms, we obtain: $${\cal P}_{\rm body} \approx {\cal P}^{10^{28}}\approx 10^{-4.5\times 10^{29}}. \label{eq4}$$ Let us reflect for a moment on the amazing infinitesimality of this number. To write it in decimal, non-scientific notation, we must use more than $10^{29}$ zeros, i.e., more than one hundred billion billion billion zeros! If we write with a printer on paper ten zeros per second, to write the entire number will take us more than $10^{28}$ seconds, that is, more than $10^{21}$ years, which is about a hundred thousand billion times the assumed age of the known universe (according to today cosmological theories)! In other terms, the value we have obtained for ${\cal P}_{\rm body}$, although not strictly equal to zero, is nevertheless so small that we have no point of comparison to be able to understand it. Yet, we have probably overestimated it. In fact, we have assumed that for a reason unknown to us all the atoms of the human body will suddenly disassemble and become non-spatial entities, so allowing their individual wave packets to spread. But we have also neglected the problem of the relative motion between the different atomic constituents, equating the individual atoms to free elementary-like particles. Furthermore, we have neglected the spin variables of the different atomic constituents. Also, we have assumed that the environment in which the different atomic components evolve, once disassembled, corresponds to an effective absolute vacuum, otherwise the associated wave-packets cannot be considered to evolve freely, and that each atom is able to evolve without interacting with all the others, before regaining a specific spatial location. In addition to that, we have hypothesized that when the various atoms reappear in the relative positions they occupied before being disassembled, the entire macroscopic structure will be able to reconstitute, without any particular inconvenience. Taking into account all these assumptions would of course further reduce the value of ${\cal P}_{\rm body}$, by a factor which is very difficult, if not impossible, to evaluate. But that’s not all. There is another “sci-fi miracle,” which is implicit in our reasoning, perhaps even more amazing than that of the disassembly of the initial structure (which is possible to relax; see the next section). This second miracle has to do with the different hydrogen atoms being simultaneously “drawn” back into space. Let me explain. An elementary entity, such as a proton, an electron, or an entire hydrogen atom, spends most of its time in a non-spatial (non-local) condition, unless it is incorporated into a macroscopic structure. Now, although there is no consensus on this among physicists, many agree that a microscopic entity is unable to acquire a precise spatial localization spontaneously, as this can only be done by interacting with a macroscopic material structure, like for instance that forming a measurement apparatus. Experimental physicists are undoubtedly able to build detection apparatuses allowing microscopic entities to temporarily acquire a spatial localization, in specific places, and even though these apparatuses could in principle localize in space a certain number of microscopic entities at a time, as far as I know a device which can localize an entire macroscopic structure doesn’t exist, and perhaps is not even conceivable. The possibility remains, of course, that the process of spatial localization could occur even in the absence of macroscopic structures playing the role of detection devices, as is suggested in some interpretations of quantum theory, like the so-called objective collapse theories [@Ghirardi2011], the transactional interpretation of quantum physics [@Ruth2013], and others, the discussion of which, however, would go beyond the scope of the present article. Among the factors that we have not taken into account in the estimation of ${\cal P}_{\rm body}$, there are of course also those that could slightly increase the value of the probability. For example, we have implicitly assumed that all the atomic components have to re-localize at exactly the same instant. However, nothing prevents us from admitting a small time-delay in the localization process of the individual atoms, which, if sufficiently small, may not affect the correct re-assembly of the entire body macrostructure. But it is unlikely that considerations of this kind would be able to significantly change the infinitesimality of ${\cal P}_{\rm body}$. Cooling down the body ===================== At this point some readers may rightly object that we don’t really needed our “disassembling sci-fi hypothesis,” as what generally makes a macroscopic object like our human body behave classically, i.e., spatially and locally, is just the fact that it is immersed in a thermal environment, i.e., that it is constantly subjected to the random collisions of countless microscopic entities, in particular photons, and that the overall effect of these innumerable interactions is that of producing its continuous “collapse into space,” which would be essentially the reason why it would behave differently from a “pure” quantum entity, like an electron. To use once more our metaphor, this bombardment is what would force the body to remain inside the water of the pool, preventing its owner from becoming a diver. So, one could object that, to allow the body to quit the “spatial pool,” and temporarily become a non-spatial entity, it would be sufficient to shield it from the external thermal environment, so that there would be no need for having it first disassembled into smaller atomic fragments and then recombined, which is the operation that apparently produced the inconceivable infinitesimality of the self-teleportation probability, as each of the $10^{28}$ fragments needed to relocalize in a predetermined place, within a sphere of $50$ km of radius. This is a pertinent objection that I’m now going to explore. This objection, by the way, could appear to be in contradiction with what I have just stated above, at the end of the last section, regarding the lack of an apparatus that could objectify (spatialize) a whole macroscopic structure. If our standard terrestrial environment is able to keep a macroscopic body into space, then wouldn’t be that same environment the measuring apparatus that is able to achieve the required goal of producing the collapse – the objectification – of an entire macroscopic object? If this is true, then it would be sufficient to isolate an ordinary object to obtain its automatic de-localization (i.e., its de-spatialization). However, we cannot expect this to work, as the object is also in contact with another environment: its own internal one. If the body is sufficiently large, as is certainly the case of a human body (but also of a speck of dust, and of much smaller entities), then the mutual interactions of its constituents can also have an influence in determining its overall classical (spatial) versus quantum (non-spatial) behavior. The reason for this is easy to explain. In the case of the hydrogen atom, we were able to separate the wave function relative to the center of mass from that associated with the relative motion. In this way, the center of mass was described by a free evolving wave function. With a macroscopic body, we may want to do the same, i.e., to separate the wave function describing the center of mass from the contribution coming from the different movements of all its constituents, relative to that center and to each others. Here we can consider the ensemble of these constituents as an entity playing the role of a measuring apparatus with respect to the “center of mass entity,” so that the latter would be constantly subjected to a measurement process, thus producing its classical behavior. Therefore, to describe the center of mass by means of a free evolving wave function, the evolution of the body’s center of mass needs to decouple from that of its internal degrees of freedom, and this can reasonably be done only if the body is cooled down to extremely low temperatures. How low? Well, we can say, remaining here necessarily vague, low enough to avoid any exchange of energy between the center of mass degree of freedom and the degrees of freedom associated with the internal relative movements [@Sun2001]. In the previous section we have assumed that by some sci-fi action the body was all of a sudden disassembled (and each constituent isolated from one another, and the environment). This was an assumption of simplicity, as in this way we were able to use the well-known factorization of the wave-function of a two-body system, which of course is much harder to obtain in general for a macroscopic body. We can however replace the “disassembling sci-fi hypothesis” with the requirement that not only the body in question will have to evolve in a perfect vacuum (no thermal bombardment), but also that it will be cooled down instantaneously to temperatures almost equal to the absolute zero. In other terms, we now replace the “disassembling sci-fi hypothesis” by a “freezing sci-fi hypothesis.” The advantage is that cooling down a body seems an operation less impossible to achieve than disassembling it into its atomic fragments, without destroying them, but also, and more importantly for our idealized discussion, the entire structure of the body will be preserved in this way, which hopefully will increase the value of the self-teleportation probability. Of course, also in this case we will have to assume that the body is additionally isolated from the environmental thermal bombardment. So, let us assume, as we did before, that the mass of the body is $10^2$ kg. By assumption, since the internal and external environments have now been made totally silent, and cannot anymore play the role of generalized detector instruments with respect to the wave function of the body’s center of mass, we can consider that the latter is described by a free evolving wave packet. To further simplify the discussion, we assume that at time $t=0$, the wave packet is approximately Gaussian (this is a reasonable assumption, considering that the probability density of any non-Gaussian wave packet becomes approximately Gaussian as it spreads [@Mita2007]), which means that the inequality in Heisenberg’s uncertainty principle is approximately an equality: $\Delta Q_0\Delta P_0 \approx {\hbar\over 2}$. As we did with the hydrogen atom, we take the standard deviation of the center of mass position observable $\Delta Q_0$ to be equal to the Bohr radius. Therefore: $$\Delta P_0\approx {\hbar\over 2 \Delta Q_0}\approx {1.05 \times 10^{-34}\, {\rm J}\, {\rm s} \over 2 \times 5.3 \times 10^{-11}\, {\rm m}}\approx 10^{-24}\, {\rm kg}\, {\rm m}\, {\rm s}^{-1}. \label{eq5}$$ Inserting this value into (\[eq1\]), we find that the spreading $\Delta Q_t$ of the center of mass wave packet at time $t$ is given by: $$\Delta Q_t \approx \sqrt{(10^{-52} \,{\rm m}^2 \,{\rm s}^{-2})\, t^2 + 3\times 10^{-21}\,{\rm m}^2}. \label{eq6}$$ Now, the original question of the sci-fi writer was to have the body disappearing from one place and reappearing tens of kilometers away. Considering, as we did before, a distance of $50$ km, we have to set $\Delta Q_t = 50~{\rm km} = 5\times 10^{4}~{\rm m}$ in the above equation. If we do so and solve for $t$, we find the value $t\approx 5\times 10^{30}$ s. Considering that 1 year corresponds to $3.154\times 10^7$ s, we obtain that the center of mass wave packet of the macroscopic body will reach a width of $50$ km after approx. $1.6\times 10^{23}$ years, which is approximately ten million billion times the assumed age of the known universe! Here we see an important difference between the wave packet spreading of a hydrogen atom, who was extremely fast, and the wave packet spreading of a macroscopic body, which is inconceivably slow. But to answer the question of the sci-fi writer, we certainly cannot wait so long, because he explicitly asked the self-teleportation to happen in a matter of seconds. Also, even in case we would accept to wait for so long, we may have a problem with the “expiration date” of our universe! And anyhow, without having to freeze the body and the environment, using any classical means of transport through space (the “swimming modality”) would be in this case much more effective to travel the distance of $50$ km. On the other hand, we can also say that, although the width of the wave packet almost doesn’t increase as time passes by, even if we perfectly confine the position of center of mass in a given small region of space, at time $t=0$ s, a fraction of a second after its wave function will have acquired an infinite tail. The value of this tail will be infinitesimally small, but nevertheless different from zero. So, let us estimate this value, at a distance of $50$ km, after exactly one second of free evolution. The body’s center of mass wave packet can be written as the product of three identical Gaussian factors: $\|\psi_t({\bf x})\|^2= \|\psi_t(x_1)\|^2 \|\psi_t(x_2)\|^2\|\psi_t(x_3)\|^2$, where: $$\|\psi_t(x_i)\|^2=\sqrt{2\over\pi a^2}{1\over \sqrt{1 + {4\hbar^2 t^2 \over M^2a^4}}}\exp\left[-{{2a^2\left(x_i-{\hbar k_i\over M}t \right)^2}\over a^4 + {4\hbar^2 t^2 \over M^2}}\right],\quad i=1,2,3. \label{gaussian}$$ To evaluate $\|\psi_t({\bf x})\|^2$, at time $t=1$ s, we can set $k_1=k_2=k_3=0$ (the body is at rest at time $t=0$ s), $x_2=x_3=0$ m, $x_1\approx 5\times 10^4$ m, $M=10^2$ kg, and $a = 2r_0 \approx 10^{-10}$ m. This implies that $\|\psi_1(x_2)\|^2=\|\psi_1(x_3)\|^2\approx 10^{10}$, and $\|\psi_1(x)\|^2\approx 10^{10}\times e^{-5\times 10^{29}}$. Multiplying the probability density $\|\psi_t({\bf x})\|^2$ by the volume of the micro-sphere of Bohr radius in which we want the center of mass to relocate, at time $t=1$ s, we thus obtain that: $${\cal P}_{\rm body} \approx e^{-5\times 10^{29}}\approx 10^{-2.2\times 10^{29}}. \label{eq4-bis}$$ Comparing (\[eq4-bis\]) with (\[eq4\]), we observe that have obtained a self-teleportation probability of the same order of magnitude. This means that, quite surprisingly, even if we avoid the sci-fi procedure of disassembling the macroscopic body, which as we have seen was responsible for the inconceivable infinitesimality of the obtained probability, and replace it with a procedure of total internal freezing, thus preserving the structural integrity of the body, a similar inconceivably infinitesimal self-teleportation probability is obtained, this time because of the extreme slowness of the spreading of the macroscopic wave function and the extreme infinitesimality of its long-distance tails. Superselection rules ==================== Considering my above pessimistic analysis, I’ll leave it to the science fiction writer the task of finding a convincing sci-fi solution, not violating too many physical laws at the same time, allowing the hero of his story to teleport himself and accomplish his mission, whatever it is. As for me, let me offer a final thought. According to quantum theory, and the phenomenon of the spreading of the wave packet, a hydrogen atom, if left to evolve freely, will quickly acquire a truly gigantic size, apparently in contradiction with what is usually observed. Also, when considering the spectrum of energies of a hydrogen atom, in addition to the discrete energy values, associated with the relative electron-proton movement, we also have to consider the continuous energy values associated with the translational degrees of freedom of the center of mass. The spectrum of the total energy of the atom is thus given by the sum of these two energy spectra. But the sum of a discrete spectrum and a continuous spectrum produces a continuous spectrum, apparently in contradiction with the spectral lines experimentally observed. In short, without further precautions, the application of the Schrödinger equation to the problem of the hydrogen atom does not allow to obtain results in agreement with the experimental observation, that is, in agreement with the fact that atoms do not usually possess macroscopic sizes, nor spectra of a continuous nature. To solve this problem, one can make use of the notion of superselection rules [@Streater1964]. Rules of this kind restrict the physically realizable states, and when associated with a given observable, they prevent considering states that would be a superposition of states associated with different values of this observable, as these superpositions would be in disagreement with the experimental data. In other terms, the existence of superselection rules indicates that the structure of the state space is not strictly Hilbertian (as linearity would not apply for all states), but more general. An example of superselection rule is that associated with the observable determining whether the infamous Schrödinger’s cat is alive or dead. If $\psi_A$ is the wavefunction describing the alive cat, and $\psi_D$ the wave function describing the dead cat, then, as far as we know, the wave function $\psi = \psi_A + \psi_D$ obtained by superposing these two wave functions does not describe a physically realizable state. This means that there is a superselection rule on the “life observable” of the cat, which forbids the superposition of wave functions characterized by different values of this observable. In the case of the hydrogen atom, if we want to obtain values for its energy spectrum in agreement with the experimental data, it is necessary to consider the position and momentum of its center of mass as variables of a classical kind, associated with superselection rules, and same thing if we want to correctly describe its observed non-macroscopic size. Of course, the reasons for this inhibition of quantum superpositions and the associated classical behavior of certain observables can be multiple, and will generally depend on the specificities of the environment in which the entity in question is immersed. So, determining what are the classical observables and what the quantum ones, in a given context, is a problem not necessarily easy to solve, and there is no unique recipe for this: in some contexts, certain observables will behave classically, while in other contexts they will behave quantum mechanically, and still in others their behavior will be semiclassical, or semiquantum, that is, in between these two regimes. But then, if the center of mass of the hydrogen atom is the expression of a superselection rule, goodbye self-teleportation! On the other hand, if we consider it as quantum observable, goodbye agreement with many experimental data. But as I said, to determine the classical or quantum nature of an observable it is necessary to take into account the specificities of the experimental context. When an atom is incorporated into a macroscopic material structure, or undergoes continuous interactions with countless microscopic entities and force fields present in the environment, it usually undergoes a process of de-synchronization of its wave function, able to transform certain quantum observables into classical ones. That’s why, in the beginning of this article, I have argued that, strictly speaking, quantum self-teleportation would be impossible. More precisely, it is impossible (${\cal P}_{\rm body}=0$) if we consider the standard environment in which we humans evolve, which makes our bodies, and the objects with which we interact in our everyday life, classical entities. What I’m here suggesting is that the classical or quantum nature of a physical entity is not an intrinsic feature of the same, but a contextual one: in some contexts certain entities will behave as quantum entities (when subjected to certain observational processes), and in other contexts they will behave, instead, classically. These considerations open to an important reflection, which can be summarized in the following question:\ Is the physical reality fundamentally quantum?\ The majority of physicists seem to believe so, that is, to think that quantum theory would be more fundamental than classical theory, and that a classical behavior would always emerge from a quantum substrate, when certain circumstances are met. However, a different view is also possible. For instance, one can consider that our physical reality is neither classical nor quantum, but genuinely hybrid, that is, a complex combination of these two aspects. In other words, the physical entities forming our reality would generally be quantum-like, i.e., they would be entities potentially manifesting both aspects, the classical and the quantum aspects, depending on their state and the nature of the experimental questions we address to them. According to this view, supported by some very general (operational) approaches to the foundations of physical theories, especially that of the so-called Geneva-Brussels School of Quantum Mechanics (nowadays mainly active in Belgium, at the Center Leo Apostel, led by the Belgian physicist Diederik Aerts), the classical regime and the quantum regime would correspond to very specific limit cases of more general situations [@Aerts1998; @Aerts1999; @Sassoli2013; @Aerts2014d]. More precisely, the classical regime would be associated with experimental situations where all fluctuations can be controlled, so that all observational processes are predictable in advance. On the other hand, the quantum regime would be associated with experimental contexts in which the fluctuations are maximal (and uniform), so producing a situation of maximum lack of knowledge. In between these two regimes, intermediate, hybrid regimes can also exist, neither purely classical nor strictly quantum, that physicists have just begun to investigate and that seem to provide a more general and realistic model for the description of the countless physical entities in interaction with their multiple environments. C. H. Bennett et al., “Teleporting an Unknown Quantum State via Dual Classical and Einstein-Podolsky-Rosen Channels,” Phys. Rev. Lett., 70, 1895–1899 (1993). W. Wootters and W. Zurek, “A Single Quantum Cannot be Cloned," Nature 299: 802–803 (1982). D. Dieks, “Communication by EPR devices,” Physics Letters A 92(6): 271–272 (1982). X. Ma, T. Herbst, T. Scheidl, D. Wang, S. Kropatschek, W. Naylor, B. Wittmann, A. Mech, J. Kofler, E. Anisimova, V. Makarov, T. Jennewein, R. Ursin and A. Zeilinger, “Quantum teleportation over 143 kilometres using active feed-forward,” Nature 489, 269–273 (2012). D. Aerts, “The entity and modern physics: the creation discovery view of reality,” In: Interpreting Bodies: Classical and Quantum Objects in Modern Physics, edited by Elena Castellani (pp. 223–257). Princeton Unversity Press, Princeton (1998). D. Aerts, “The Stuff the World is Made of: Physics and Reality,” In: The White Book of ‘Einstein Meets Magritte’, edited by Diederik Aerts, Jan Broekaert and Ernest Mathijs (pp. 129–183). Kluwer Academic Publishers, Dordrecht (1999). M. Sassoli de Bianchi, “Ephemeral Properties and the Illusion of Microscopic Particles,” Foundations of Science 16, 393–409 (2011). M. Sassoli de Bianchi, “From Permanence to Total Availability: A Quantum Conceptual Upgrade,” Foundations of Science 17, 223–244 (2012). M. Sassoli de Bianchi, “The $\delta$-Quantum Machine, the $k$-Model, and the Non-ordinary Spatiality of Quantum Entities,” Foundations of Science 18, 11–41 (2013). D. Aerts and M. Sassoli de Bianchi, “The extended Bloch representation of quantum mechanics and the hidden-measurement solution to the measurement problem,” Annals of Physics 351, 975–1025 (2014). M. Sassoli de Bianchi, “God May Not Play Dice, But Human Observers Surely Do,” Foundations of Science 20, 77–105 (2015). J. Von Neumann, “Grundlehren,” Math. Wiss. XXXVIII (1932). J. S. Bell, “On the Problem of Hidden Variables in Quantum Mechanics,” Rev. Mod. Phys. 38, 447–452 (1966). A. M. Gleason, “Measures on the closed subspaces of a Hilbert space,” J. Math. Mech. 6, 885–893 (1957). J. M. Jauch and C. Piron, “Can hidden variables be excluded in quantum mechanics?,” Helv. Phys. Acta 36, 827–837 (1963). S. Kochen and E. P. Specker, “The problem of hidden variables in quantum mechanics,” J. Math. Mech. 17, 59–87 (1967). S. P. Gudder, “On Hidden-Variable Theories,” J. Math. Phys 11, 431–436 (1970). G. C. Ghirardi, “Collapse Theories,” The Stanford Encyclopedia of Philosophy (Winter 2011 Edition), Edward N. Zalta (ed.), URL = <http://plato.stanford.edu/archives/win2011/entries/qm-collapse/>. R. E. Kastner, The Transactional Interpretation of Quantum Mechanics, Cambridge University Press, Cambridge (2013). C. P. Sun, X. F. Liu, D. L. Zhou and S. X. Yu, “Localization of a macroscopic object induced by the factorization of internal adiabatic motion,” Eur. Phys. J. D 17, 85–92 (2001). R. F. Streater and A. S. Wightman, PCT, spin and statistics, and all that, W. A. Benjamin, Inc., New York (1964). K. Mita, “Dispersion of non-Gaussian free particle wave packets,” Am. J. Phys. 75, 950–952 (2007). D. Aerts, “Quantum Theory and Human Perception of the Macro-World,” Front. Psychol. 5, article 554 (2014).	{ "pile_set_name": "ArXiv" }
--- abstract: 'The magnetic breakout model has been widely used to explain solar eruptive activities. Here, we apply it to explain successive filament eruptions occurred in a quadrupolar magnetic source region. Based on the high temporal and spatial resolution, multi-wavelengths observations taken by the Atmospheric Imaging Assembly (AIA) on board the [Solar Dynamic Observatory]{} ([SDO]{}), we find some signatures that support the occurrence of breakout-like external reconnection just before the start of the successive filament eruptions. Furthermore, the extrapolated three-dimensional coronal field also reveals that the magnetic topology above the quadrupolar source region resembles that of the breakout model. We propose a possible mechanism within the framework of the breakout model to interpret the successive filament eruptions, in which the so-called magnetic implosion mechanism is firstly introduced to be the physical linkage of successive filament eruptions. We conclude that the structural properties of coronal fields are important for producing successive filament eruptions.' --- Introduction ============ Filament (prominence) eruption is one of the most spectacular, large-scale activity on the Sun, which often associates with solar flare and coronal mass ejection (CME). The eruption of a filament can severely impact the solar-terrestrial environment and human activities; and the study of these phenomena has developed into a new discipline dubbed space weather. However, the physical mechanism of filament eruption is still not well understood, even though extensive observational and theoretical works have been made in recent decades. Generally speaking, a filament eruption always starts from a closed magnetic system in quasi-static equilibrium, in which the upward magnetic pressure force of the low-lying sheared field is balanced by the downward tension force of the overlying field. When the eruption begins, the equilibrium is destroyed catastrophically, and part of the non-potential magnetic flux and the plasma are expelled violently from the Sun. Given different magnetic environments, the eruption of a filament can be failed, partial, and complete eruptions (e.g., [@gilb01; @liu09; @shen11]). Sometimes, several filaments far from each other or resided in one complex active region can erupt successively within a short time period. A key question of successive eruptions is whether they are physically connected or not. It seems that the answer is positive, and the connection is often of a magnetic nature (e.g., [@jian08; @jian11; @toro11; @shen12; @tito12; @lync13; @schr11; @schr13]). Currently, solar physicists have developed a number of models for interpreting filament/CME eruptions. Among various models, the magnetic breakout model assumes a large-scale quadrupolar field configuration; the core field is increasingly sheared by photospheric motions, which is surrounded by an overlying antiparallel loop system ([@anti98; @anti99]). Naturally, a null point is formed between the core and the overlying loop system. This model can be used to interpret many eruptions which occur in complex multipolar active regions (e.g., [@aula00; @maia03; @shen12]). Here, we apply the magnetic breakout model to explain successive filament eruptions which occurred in a quadrupolar magnetic source region, and propose a possible physical linkage between the filament eruptions. Results & Interpretation ======================== ![Topology structure of the breakout mode. The red (P1 and P2) and blue (N1 and N2) patches indicate the positive and negative polarities of the magnetic field, respectively. The field direction is indicated by a series of arrows. The gray patches represent the filaments (F1, F2, and F3) confined by the three low-lying lobes.[]{data-label="fig1"}](fig1.eps){width="80.00000%"} The basic magnetic topology of the breakout model is shown in Figure.\[fig1\]. It can be seen that the four magnetic poles (P1, N1, P2, and N2) are connected by three low-lying lobes and one overlying loop system; and a coronal null resides inbetween the middle lobe and the overlying antiparallel loop system. In addition, one can assume the existence of a filament under each lobe (F1, F2, and F3). In such a configuration, a small disturbance to the system could lead to the eruption of the whole system. Typically, there are two types of disturbance to the system. The first type is that the disturbance acts on F2, which will lead to the rising of F2 and the middle lobe, and further results in the external reconnection around the null point, which will removes the confining field of F2 to the lateral lobes and thereby reduce (increase) the confining capacity of the middle (lateral) lobe. Therefore, this type of disturbance often lead to failed eruptions of F1 and F3, while the eruption of F2 should be a successful one. The second type is that the disturbance acts on F1 or F2, which will not lead to any reconnection, and therefore no filament eruption occurs. Here, we present another type of filament eruption in a solar breakout event, in which successive partial and full filament eruptions are involved. ![Schematic demonstrating the successive filament eruptions. (a) The initial magnetic configuration. (b) The rising of F2, formation of CS1 and the external reconnection. (c) Formation of CS2 (CS3) underneath (above) F2 (CS3). (d) Reconnections in CS2 and CS3, and the production of the nearly simultaneous CMEs. The arrow pointing to F2 represent the disturbance. The red dotted lines indicate locations of the current sheets, while the reconnection sites are labeled by red “X” symbols. The yellow lines represent the field lines to be reconnected, while the green lines are reconnected ones.[]{data-label="fig2"}](fig2.eps){width="80.00000%"} The detailed analysis of the event can be found in [@shen12]. Here we just give a brief summery of the results. As shown in Fig.\[fig2\](a), we find that the extrapolated coronal field above the magnetic source region is of the topology of the breakout model, and the two filaments are located below the middle and the left lobes respectively. The initiation of the successive filament eruptions started from a small mass ejection, which directly interacted with the southern part of F2 and thus resulted in the slow rise of this filament. The slow rise of F2 lasted for about 23 minutes and a speed of 8 km/s. During this period, some signatures for breakout-like external reconnection were observed. For example, two brightening patches at both sides of F1, the appearance of bright loops and a weak hard X-ray source above F1. After the slow rising phase, F2 was quickly accelerated to 102 km/s, and finally, it erupted successfully and caused a CME. The activation of F1 started around the end of F2’s slow rising phase, which erupted with strong writhing motions. When F1 reached its maximum height, the eruption of a blob-like structure was observed above the filament. In the meantime, F1 began to fall back to the solar surface. These results indicate that the eruption of F1 should be a typical partial flux rope eruption. According to the model proposed by [@gilb01], the reconnection site should be located above the filament. We interpret the observations using the breakout model as shown in Fig.\[fig2\]. Panel (a) presents basic magnetic topology. Due to the disturbance introduced by a small plasma ejection, F2 slowly rises, expanding the middle lobe, which will result in the external magnetic reconnection within the current sheet formed around the coronal null point (see CS1 in panel (b)). According to the magnetic implosion mechanism proposed by [@huds00], the magnetic pressure around the reconnection site will decrease due to energy released during the energy conversion process in coronal transients such as flares. The reduction of magnetic pressure will lead to the contraction of the overlying loop system and the expansion of the low-lying lobes. In addition, the strong writhing of F1 indicates that the eruption of this filament was driven by the kink instability. The reduction of the magnetic tension force of the left lobe facilitates triggering the kink instability within F1. Hence, the magnetic implosion could be a possible physical linkage between the successive filament eruptions within the framework of the breakout model. As the rising of F1 and F2, new current sheet CS2 (CS3) will form underneath (above) F2 (F1). The reconnection within CS2 (CS3) will lead to the successful (partial) eruption of F2 (F1), and the CME (blob) (see panels (c) and (d)). In this model, we can expect two simultaneous CMEs. Summary ======= Based on multi-wavelengths observations, we propose an interpretation for the successive eruptions of two filaments in a solar breakout event. We first introduce the magnetic implosion mechanism to be the physical linkage of the successive filament eruptions. The observations of both the pre-eruptive signatures and the extrapolated three-dimensional coronal fields are in good agreement with the breakout model. Our scenario presented in this article implies the occurrence of nearly simultaneous CMEs. Therefore, this interpretation is important for the forecast of space weather. It should be noted that the breakout scenario is a possible explanation for the observations. We do not intend to exclude other possibilities. In any case, the structural properties of coronal fields are important for producing successive filament eruptions.\ [Acknowledgements]{} This work is supported by the Western Light Youth Project of Chinese Academy of Sciences (CAS), the CAS open research programs (KLSA201204, DMS2012KT008). Y. Shen thank the financial support for young researches to participate the 300th symposium (IAUS300: Nature of Prominences and their role in Space Weather) of the International Astronomical Union. Antiochos, S. K. 1998, [ApJ]{}, 502, L181 Antiochos, S. K., DeVore, C. R., & Klimchuk, J. A. 1999, [ApJ]{}, 510, 485 Aulanier, G., DeLuca, E. E., Antiochos, K. S., McMullen, R. A. & Golub, L. 2000, [ApJ]{}, 540, 1126 Gilbert, H. R., Holzer, T. E., Low, B. C., & Burkepile, J. T. 2001, [ApJ]{}, 549, 1221 Hudson, H. S. 2000, [ApJ]{}, 531, L75 Jiang, Y., Shen, Y., Yi, B., Yang, J., & Wang, J. 2008, [ApJ]{}, 677, 699 Jiang, Y., Yang, J., Hong, J., Bi, Y., & Zheng, R., 2011, [ApJ]{}, 738, 179 Liu, Y., Su, J., Xu, Z., Lin, H., Shibata, K. et al. 2009, [ApJ]{}, 696, L70 Lynch, B. J., & Edmondson, J. K. 2013, [ApJ]{}, 764, 87 Maia, D., Aulanier, G., Wang, S. J., et al. 2003, [A&A]{}, 405, 313 Plunkett, S. P., Vourlidas, A., $\rm \check{S}$imberov$\rm \acute{a}$, S. et al. 2000, [Sol. Phys.]{}, 194, 371 Schrijver, C. J., Title, A. M., Yeates, A. R., & DeRosa, M. L. 2013, [ApJ]{}, 773, 93 Schrijver, C. J., & Title, A. M. 2011, [[J. Geophys. Res.]{}]{}, 116, A04108 Shen, Y., Liu, Y., & Liu, R. 2011, [Res. Astron. Astrophys.]{}, 11, 594 2012, [ApJ]{}, 750, 12 Titov, V. S., Mikic, Z., T$\rm \ddot{o}r\ddot{o}$k, T., Linker, J. A., Panasenco. O. 2012, [ApJ]{}, 759, 70 T$\rm \ddot{o}r\ddot{o}$k, T., Panasenco, O., Titov, V. S., et al. 2011, [ApJ]{}, 739, L63	{ "pile_set_name": "ArXiv" }
--- author: - 'P. Coelho' - 'B. Barbuy' - 'M.-N. Perrin' - 'T. Idiart' - 'R. P. Schiavon' - 'S. Ortolani' - 'E. Bica' date: 'Received 30 January 2001; accepted 11 June 2001' title: 'Membership of 23 stars towards the bulge globular clusters NGC 6528 and NGC 6553 [^1]' --- Introduction ============ The study of the stellar populations in the Galactic bulge is very important to constrain possible models of galaxy formation. In particular, the determination of the metallicities and abundance ratios of bulge stars, either from the field or in clusters, provides key information to help decide among the possible scenarios for the history of chemical enrichment of the Galaxy. There are few previous studies of radial velocities and metallicity estimations of bulge stars from low resolution spectra. In the analysis of 400 field bulge stars by Sadler et al. (1996), metallicities and \[Mg/Fe\] values were estimated. Minniti (1995a,b) studied the membership of stars towards 7 bulge globular clusters. The best studied among the bulge globular clusters are NGC 6528 ($\alpha_{2000}$ = 18$^{\rm h}$04$^{\rm m}$49.6$^{\rm s}$, $\delta_{2000}$ = -30$^{\rm o}$03’20.8“) and NGC 6553 ($\alpha_{2000}$ = 18$^{\rm h}$09$^{\rm m}$15.7$^{\rm s}$, $\delta_{2000}$ = -25$^{\rm o}$54’27.9”). Ortolani et al. (1995) have shown that, besides being old, these clusters have luminosity functions which are very similar to that of Baade’s Window, which indicates that they belong to the same stellar population. NGC 6528 is located in the Baade Window, at a distance d$_{\odot}$ = 7.83 kpc from the Sun, and NGC 6553 is relatively close to the Sun, at a distance d$_{\odot}$ = 5.1 kpc (Barbuy et al. 1998). As they are both located in crowded fields, the measurement of radial velocities of individual stars is of crucial importance for the determination of their membership in the clusters. Both clusters are known to be metal-rich. However, there is no consensus in the literature regarding their detailed metal abundances. Recently, Barbuy et al. (1999) analysed high resolution spectra of two giant stars of NGC 6553. An Iron abundance of \[Fe/H\] = -0.55$\pm$0.2 and abundance ratios \[Na/Fe\] $\approx$ \[Al/Fe\] $\approx$ \[Ti/Fe\] $\approx$ +0.5, \[Mg/Fe\] $\approx$ \[Si/Fe\] $\approx$ \[Ca/Fe\] $\approx$ +0.3 were derived. These ratios imply an overall metallicity \[Z/Z$_{\rm \odot}$\] $\approx$ -0.1. Cohen et al. (1999), analysing high resolution spectra of five red horizontal branch stars, obtained a mean metallicity \[Fe/H\] = -0.16 and an excess of the $\alpha$-element calcium to iron of about 0.3 dex, which imply an overall metallicity \[Z/Z$_{\rm \odot}$\] $\approx$ +0.1. Metal abundances of these clusters are also discussed in Barbuy et al. (1999) and Barbuy (2000). In view of the disagreement between previous determinations of \[Fe/H\], it is important that abundance estimations be extended to a larger number of stars of both clusters. In this paper, we determine radial velocities, effective temperatures, gravities and estimations of metallicities \[Fe/H\] based on low resolution spectra for 23 stars towards NGC 6553 and NGC 6528, and verify their membership in these clusters. In Sect. 2 the observations are described. The radial velocities derived are presented in Sect. 3. In Sect. 4 the stellar parameters are derived, and synthetic spectra are compared to observations to estimate metallicities. Concluding remarks are given in Sect. 5. Observations ============ Low resolution spectra of individual stars of NGC 6528 and NGC 6553 were obtained in 1992 August and 1994 June, at the 1.5m ESO telescope at ESO (La Silla). The Boller & Chivens spectrograph was employed. In 1992 August the Thompson CCD \# 18 with 1024$\times$1024 pixels, with a pixel size of 19 $\mu$m was used. A resolution of $\Delta\lambda$ $\sim$ 8 [Å]{} and a spectral coverage of $\lambda\lambda$ 4800-8800 [Å]{} were achieved. In 1994 June, the Ford Aerospace FA 2048 L, frontside illuminated, uncoated CCD detector (ESO \# 24) with 2048$\times$2048 pixels and pixel size 15$\times$15 $\mu$m was used. The grating \# 27 resulted in a spectral resolution $\Delta\lambda$ $\sim$ 4 [Å]{} and a spectral coverage in the range $\lambda\lambda$ 4800-7550 [Å]{}. The log of observations is provided in Table 1. The stars are identified according to the charts by Hartwick (1975) for NGC 6553 and van den Bergh & Younger (1979) for NGC 6528. Spectra of a given star were co-added by weighting their S/N ratios; the final S/N are indicated in Table 1. [lccccc]{} Star & V & V-I & Exp. (s) & Date & S/N\ I 1 & 16.10 & 1.93 & 5400 & 06.08.92 & 190\ I 2 & 15.73 &2.59 & 1800 & 16.06.94 & 25\ & & & 5400 & 06.08.92 & 75\ I 5 & 15.37 &2.22 & 2700 & 17.06.94 & 20\ I 6 & 15.89 &3.54 & 2700 & 17.06.94 & 40\ II 8 & 15.71 &2.19 & 2100 & 17.06.94 & 20\ II 14 & 15.76 &3.47 & 2700 & 17.06.94 & 10\ I 23 & 17.19 &1.70 & 1800 & 16.06.94 & 10\ I 24 & 16.89 &1.66 & 4500 & 17.06.94 & 20\ I 25 & 16.11 &2.09 & 4500 & 16,17.06.94 & 40\ I 27 & 15.90 &3.08 & 1800 & 16.06.94 & 20\ & & & 5400 & 06.08.92 & 100\ I 36 & 16.41 &1.98 & 2100 & 17.06.94 & 20\ II 39 & 15.88 &2.30 & 2100 & 17.06.94 & 30\ I 40 & 15.93 &2.08 & 2100 & 17.06.94 & 20\ I 42 & 16.42 &2.15 & 2100 & 17.06.94 & 30\ II 70 & 15.85 &2.36 & 1800 & 16.06.94 & 20\ III 2 & 16.89 &1.95 & 1800 & 16.06.94 & 20\ III 3 & 15.82 &2.41 & 3600 & 14,16.06.94 & 25\ & & & 5400 & 07.08.92 & 100\ III 17& 15.36 &3.01 & 1800 & 14.06.94 & 30\ II 51 & 15.48 &2.54 & 1800 & 17.06.94 & 10\ & & & 5400 & 07.08.92 & 80\ II 52 & 16.84 &1.93 & 1800 & 17.06.94 & 40\ & & & 5400 & 07.08.92 & 230\ II 85 & 15.52 &2.51 & 3600 & 14,16.06.94 & 55\ II 94 & 15.44 &3.38 & 1800 & 17.06.94 & 20\ II 95 & 15.73 &2.64 & 1800 & 17.06.94 & 35\ \ In Figs. 1 and 2 are shown the $V$ vs. $V-I$ Colour-Magnitude Diagrams of NGC 6528 and NGC 6553 using data obtained with the Hubble Space Telescope (Ortolani et al. 1995) where the sample stars are identified. ![ [Top]{}: HST Colour-magnitude diagram of NGC 6528, where the observed stars are indicated as filled circles; [Bottom]{}: A map of the cluster scanned from van den Bergh & Younger (1979).[]{data-label="n6528 cmd"}](1084f1a.eps "fig:"){width="8.2cm"} ![ [Top]{}: HST Colour-magnitude diagram of NGC 6528, where the observed stars are indicated as filled circles; [Bottom]{}: A map of the cluster scanned from van den Bergh & Younger (1979).[]{data-label="n6528 cmd"}](1084f1b.eps "fig:"){width="8.5cm"} ![ [Top]{}: HST Colour-magnitude diagram of NGC 6553, where the observed stars are indicated as filled circles; [Bottom]{}: A map of the cluster scanned from Hartwick (1975). []{data-label="n6553 cmd"}](1084f2a.eps "fig:"){width="8.2cm"} ![ [Top]{}: HST Colour-magnitude diagram of NGC 6553, where the observed stars are indicated as filled circles; [Bottom]{}: A map of the cluster scanned from Hartwick (1975). []{data-label="n6553 cmd"}](1084f2b.eps "fig:"){width="8.2cm"} Radial velocities ================= The radial velocities were determined by means of three methods, as explained below and reported in Table 2. The observed radial velocities derived were transformed to heliocentric values using the observation dates given in Table 1. \(a) A Fourier cross-correlation was applied on the program spectra relative to selected template spectra. As templates, 12 G, K and M stars were selected from the Jacoby et al.’s (1984) library, which have approximately the same spectral resolution (4.5 [Å]{}) of the sample spectra. The templates adopted were the ones closest in spectral type to each of the program stars. The spectra of both sample and template stars were normalized and different regions in the spectra were defined in order to give highest peaks of cross-correlation for each considered template. The results obtained with this method are given in column 2 of Table 2. The r.m.s. of the values derived with each template spectrum is of the order of 15 km s$^{-1}$. A systematic effect was identified for the coolest stars, since all of them appeared to show lower velocities when compared to the hotter stars of the same cluster and these values were not considered. \(b) Mean shifts between the observed wavelengths of identified absorption lines and laboratory wavelengths were measured (column 3 of Table 2). The r.m.s. of the values derived is of the order of 15 km s$^{-1}$. \(c) the code HALO (Cayrel et al. 1991) derives radial velocities by comparing the observed spectrum to a grid of synthetic spectra, using a cross-correlation technique. The grid of synthetic spectra available (Barbuy et al. 2001) does not contain stars cooler than T$_{\rm eff}$ $<$ 4000 K, and for this reason the errors should be higher for velocities of stars cooler than T$_{\rm eff}$ $<$ 3700 K in which TiO bands are pronounced. [lccc]{} Star & 3 v$_{\rm r}$ (km s$^{-1}$)\ & method (a) & method (b) & method (c)\ I 1 & — & 262 & 174\ I 2 & 224 & 262 & 208\ I 5 & — & 261 & 224\ I 6 & 232 & 246 & 225\ II 8 & 289 & 283 & 271\ II 14& — & 264 & 227\ I 23 & 221 & 251 & —\ I 24 & 238 & 220 & 202\ I 25 & 265 & 257 & 231\ I 27 & — & 237 & 188\ I 36 & 244 & 249 & 229\ II 39& 30 & 17 & 7\ I 40 & 246 & 236 & 244\ I 42 & 197 & 235 & 212\ II 70& 230 & 263 & —\ III 2 & -2 & -7 & -25\ III 3 & 5 & 17 & 18\ III 17& 3 & 8 & -19\ II 51 & — & 60 & —\ II 52 & -16 &-28 & -33\ II 85 & 35 & 56 & -4\ II 94 &-42 &-56 & -63\ II 95 & 11 &-12 & -9\ Histograms of radial velocities of individual stars (coolest stars excluded) corresponding to each method were built. Gaussian curves were fitted to each histogram, from which the radial velocity corresponding to each method was derived, as reported in Table 3 together with values from the literature. An example of this procedure is presented in Fig. \[hist vr\] for the cross-correlation technique using IRAF. The final radial velocities adopted for the clusters correspond to the mean of the values derived from the three methods. ![Histograms of radial velocities obtained for the stars with the cross-correlation technique using IRAF, where the gaussian fits are presented. The deviant point in the histogram of NGC 6528 is the star II 39, which is probably a non-member.[]{data-label="hist vr"}](1084f3.eps){width="8.2cm"} [ccc]{} 2 v$_{\rm r}$ ( km s$^{-1}$ ) & Reference\ NGC 6528 & NGC 6553 &\ 218 & 6.4 & 1,2\ 164.8 & -24.5 & 3\ 208 & 48 & 4\ 212 & 8.4 & 5\ 189 & — & 6\ 160 & -5 & 7\ 143 & -12 & 8\ 236 (21) & 7 (14)& 9\ 248 (11) & 1 (16) & 10\ 217 (9) & -10 (13) & 11\ \ References to Table: 1 Barbuy et al. (1999); 2 Barbuy (2000); 3 Harris (1996); 4 Minniti (1995) (mean values excluding non-member stars); 5 Rutledge et al. (1997); 6 Armandroff & Zinn (1988); 7 Zinn & West (1984); 8 Zinn (1985); 9 present paper (cross-correlation method with IRAF; 10 present paper (mean wavelength shift with IRAF); 11 present paper (code HALO by a cross-correlation method) Stellar parameters ================== Temperatures ------------ The effective temperatures were estimated from $B-V$, $V-I$, $V-K$ and $J-K$ colours, based on the colour vs. T$_{\rm eff}$ calibrations by Bessell et al. (1998), which in turn are based on NMARCS models by Plez et al. (1992) and their grid extensions. These effective temperatures are listed in Table 5. V and I colours were obtained with the Hubble Space Telescope (Ortolani et al. 1995) and J and K colours were obtained with the detector IRAC2 at the 2.2m telescope of ESO (Guarnieri et al. 1998). For NGC 6528 the colour excesses adopted were E(V$-$I) = 0.68 and E(B$-$V) = 0.52 (Barbuy et al. 1998). For NGC 6553 E(V$-$I) = 0.95 and E(B$-$V) = 0.7 were adopted (Guarnieri et al. 1998). The $V-K$ and $J-K$ colours were dereddened assuming E(V-K)/E(B-V) = 2.744 and E(J$-$K)/E(B$-$V)=0.527 (Rieke & Lebofsky 1985). An independent method for the derivation of temperatures was based on calibrations of equivalent widths of TiO bands. The indices as defined in Table \[inTiO\] were measured on the grid of synthetic spectra by Schiavon & Barbuy (1999) in the range of parameters 3000 $\leq T_{\rm eff} \leq$ 5000 K, $-$0.5 $\leq$ log g $\leq$ 2.5 and \[Fe/H\] = $-$0.3. These indices are shown in Fig. \[calib TiO\] for a resolution of $\Delta\lambda$ = 8 [Å]{}. Polynomial curves of the form T$_{\rm eff}$ = f(W(TiO)) were derived and applied to the indices measured in the sample stars. The TiO indices are strongly sensitive to temperature for $T_{\rm eff} \leq$ 3800 K as illustrated in Fig. \[diftem\]. For $T_{\rm eff} \geq$ 4000 K a degeneracy appears due to the fact that TiO bands are not present at these higher temperatures. ![ TiO indices measured on the synthetic spectra as a function of effective temperatures. These measurements correspond to spectra convolved with FWHM = 8 ${\rm \AA}$[]{data-label="calib TiO"}](1084f4.eps){width="8.2cm"} ![ Spectra of three individual stars of NGC 6528 with different temperatures. It is clear that the T$_{\rm eff}$ = 3350 K star has more pronounced TiO molecular bands with respect to the T$_{\rm eff}$ = 4100 K one.[]{data-label="diftem"}](1084f5.eps){width="8.2cm"} A more general polynomial of the form log W(TiO) = ([a + b log T$_{\rm eff}$ + c log $g$ + d \[Fe/H\] + e (log T$_{\rm eff}$)$^{2}$ + f \[Fe/H\]$^{2}$ + g \[Fe/H\] log T$_{\rm eff}$]{}), valid in the range 2500 $\leq$ T$_{\rm eff}$ $\leq$ 5000 K, -0.5 $\leq$ log g $\leq$ 2.5 and -0.5 $\leq$ Fe/H $\leq$ 0 was derived. The coefficients of the formula above are shown in Table \[coeffs\] for convolutions of $\Delta\lambda$(FWHM) = 4 [Å]{} and 8 [Å]{}. [lccc]{} Index & Blue continuum & Bandpass & Red continuum\ TiO2 & 6033.6-6050.6 & 6300.0-6455.0 & 6525.0-6538.0\ TiO3 & 6525.0-6539.0 & 6617.6-6860.0 & 7036.0-7046.6\ TiO4 & 7036.0-7046.6 & 7053.0-7163.0 & 7534.2-7546.8\ [l@c@c@c@c@c@c@c@c@]{} &\ Star & $V-I$ & $B-V$ & $J-K$ & $V-K$ & TiO2 & TiO3 & TiO4 & Final\ \ I 1 & 4305 & 4449 & — & — & — & — & — & 4400\ I 2 & 3673 & 3883 & — & — & 3922 & 3698 & 3506 & 3700\ I 5 & — & — & — & — & 3468 & 3053 & — & 3250\ I 6 & 3451 & 3890 & — & — & 3612 & 3585 & 3405 & 3550\ II 8 & 3963 & 3992 & — & — & 3776 & 4049 & 3789 & 3950\ II 14 & — & — & — & — & 3468 & 3053 & — & 3250\ I 23 & 4783 & 4770 & — & — & — & — & — & 4800\ I 24 & 4883 & 4707 & — & — & — & — & — & 4800\ I 25 & 4074 & 4092 & — & — & — & — & — & 4100\ I 27 & — & — & — & — & 3673 & — & 3001 & 3350\ I 36 & 4218 & 4216 & — & — & — & — & — & 4200\ I 40 & 4084 & 4137 & — & — & — & — & — & 4100\ I 42 & 3999 & 4149 & — & — & — & — & — & 4050\ II 70 & 3810 & 3951 & — & — & — & — & — & 3900\ \ III 2 & 4828 & — & 4809 & 4559 & — & 3842 & 3776 & 3800\ III 3 & 4015 & — & 4221 & 3939 & 3782 & 3800 & 3775 & 3800\ III 17 & 3606 & — & 3966 & 3611 & 4052 & 3698 & 3503 & 3750\ II 51 & 3885 & — & 3966 & 3856 & 3281 & 3206 & — & 3250\ II 52 & — & — & — & — & 4241 & 3703 & 3752 & 3900\ II 85 & 3912 & — & 4140 & 3835 & — & — & — & 3950\ II 94 & 3506 & — & 3984 & 3432 & 4802 & 4408 & 3832 & 3650\ II 95 & — & — & — & — & — & — & — & 6500\ The temperatures obtained and final values adopted are reported in Table 5. Photometric temperatures were adopted for stars for which the TiO temperature T$_{\rm TiO}$ $\geq$ 3800 K, whereas for stars with T$_{\rm TiO}$ $<$ 3800 K the mean of TiO temperatures were adopted. [l@r@r@r@r@r@r@]{} Coefficient & 2 TiO2 & 2 TiO3 & 2 TiO4\ $\Delta\lambda$ & 4[Å]{} & 8[Å]{} & 4[Å]{} & 8[Å]{} & 4[Å]{} & 8[Å]{}\ a (constant) & 81.49 & 81.89 & 55.70 & 56.51 & -175.11 & -158.38\ b (log(T$_{\rm eff}$)) & -39.44 & -39.71 & -22.93 & -23.53 & 105.33 & 95.92\ c (log $g$) & -0.04 & -0.04 & -0.04 & -0.05 & -0.06 & -0.06\ d (\[Fe/H\]) & 9.28 & 9.14 & 3.40 & 3.21 & -7.42 & -5.10\ e (log(T$_{\rm eff}$)$^2$) & 4.75 & 4.79 & 2.17 & 2.27 & -15.66 & -14.34\ f (\[Fe/H\]$^2$) & -0.47 & -0.45 & -0.40 & -0.38 & -0.45 & -0.08\ g (\[Fe/H\]$\times$ & & & & &\ log(T$_{\rm eff}$)) & -2.56 & -2.51 & -0.87 & -0.82 & 2.16 & 1.54\ $\chi$$^2$ & 0.92 & 0.92 & 0.90 & 0.90 & 0.97 & 0.96\ Gravities --------- Gravities were derived using the classical relation log g$_$= 4.44 + 4log T$_$/T$_{\odot}$ + 0.4(M$_{\rm bol}$-M$_{\rm bol \odot}$) + log M$_$/M$_{\odot}$, adopting T$_{\odot}$ = 5770 K, M$_$ = 0.8 M$_{\odot}$ and M$_{\rm bol \odot}$ = 4.74 cf. Bessell et al. (1998). For deriving M$_{\rm bol }$ we used the distance modulus adopting a total extinction A$_{\rm V}$ = 2.43 for NGC 6553 and A$_{\rm V}$ = 1.8 for NGC 6528 (Barbuy et al. 1998). The bolometric magnitude corrections were taken from Bessell et al. (1998). The resulting M$_{\rm bol }$ and gravities are given in Table 7. Taking into consideration the errors due to uncertainties in T$_{\rm eff}$ and M$_{bol}$ the final error in log g is estimated to be of $\pm$0.5 dex. Metallicities ------------- Spectrum synthesis calculations were used to fit the observed spectra. The calculations of synthetic spectra were carried out using the code described in Barbuy et al. (2000) where molecular lines of MgH A$^2$$\Pi$-X$^2$$\Sigma$, CH A$^2$$\Delta$-X$^2$$\Pi$, CN A$^2$$\Pi$-X$^2$$\Sigma$, C$_2$ Swan A$^3$$\Pi$-X$^3$$\Pi$ and TiO $\alpha$ C$^3$$\Delta$-X$^3$$\Delta$, $\gamma$ A$^3$$\Phi$-X$^3$$\Delta$ and $\gamma$’ B$^3$$\Pi$-X$^3$$\Delta$ systems are taken into account. For atomic lines the laboratory oscillator strengths by Fuhr et al. (1988), Martin et al. (1988), Wiese et al. (1969), and laboratory values compiled by McWilliam & Rich (1994) were adopted whenever available, otherwise they were taken from fits to the solar spectrum (see discussion in Barbuy et al. 1999). ATLAS9 and NMARCS models were employed. A grid of models using the ATLAS9 code (Kurúcz 1993) was created adopting a mixing length parameter $\alpha$ = 0.5 (see Barbuy et al. 2001). NMARCS photospheric models for giants by Plez et al. (1992) and their unpublished extended grids were employed (see more details in Schiavon & Barbuy 1999). The metallicities were obtained based on two methods, both using synthetic spectra: \(i) The observed spectra were compared to synthetic spectra in the range $\lambda\lambda$ 5000–7500 ${\rm \AA}$. The metallicities were estimated by interpolating between synthetic spectra of \[Fe/H\]= 0.0, –0.3, –0.5 and –0.6, in all cases assuming \[Mg/Fe\] = +0.3, and the temperatures and gravities determined in Sects. 4.1 and 4.2. \(ii) Comparisons with a grid of synthetic spectra in the wavelength region $\lambda\lambda$ 4600-5600 [Å]{}, using the differences method as described in Cayrel et al. (1991) and Barbuy et al. (2001), are carried out. In this method, the observed spectrum is divided by a reference synthetic spectrum. The resulting signal can be expressed as a linear combination of variations in temperature, gravity and metallicity. In conjunction with the grid of synthetic spectra, it is possible to establish the differences in T$_{\rm eff}$, log g and \[Fe/H\] between the program star and the reference synthetic spectrum through a perturbation method. The grid covers the range 4000 $\leq$ T$_{\rm eff}$ $\leq$ 7000 K, 0.0 $\leq$ log g $\leq$ 5.0, -3.0 $\leq$ \[Fe/H\] $\leq$ +0.3, and \[Mg/Fe\] = 0.0 and +0.4. Fig. \[bx6528-1\] shows the fit to NGC 6528 I-1. ![ Analysis of NGC 6528 I-1 employing the code HALO: [Top:]{} observed spectrum (dotted line) and synthetic spectrum (solid line) computed with T$_{\rm eff}$ = 4400 K, log $g$ = 1.7, \[Fe/H\] = -0.4 and \[Mg/Fe\] = +0.30 (FWHM = 8 [Å]{}); [Bottom:]{} Residual flux relative to the template synthetic spectrum. []{data-label="bx6528-1"}](1084f6.eps){width="8.2cm"} In Table 7 are listed the temperatures, gravities, \[Fe/H\] and \[Mg/Fe\] obtained with methods (i) and (ii). Note that method (ii) tends to give lower metallicities relative to method (i). This may be due to limitations of the grid of synthetic spectra, which is being extended to cover wider ranges of parameters. The stars NGC 6528 I-5, NGC 6553 III-2 and II-95 are probable non-members, given that their atmospheric parameters are incompatible with their location in the Colour-Magnitude Diagrams of the clusters. The star II-51 appears to be too cool (Table 7) with respect to its CMD locus (Fig. 2). However, considering that it could be a red variable, this star is tentatively classified as a possible member. Mean metallicities of \[Fe/H\] = -0.5 for NGC 6528 and \[Fe/H\] = -0.7 for NGC 6553 were estimated from gaussian fits to the histograms of metallicities given in Table 7 (typical standard deviations are of 0.2 dex). These metallicities, together with the magnesium excess of \[Mg/Fe\] $\approx$ +0.3, and assuming that all other $\alpha$ elements show an excess of +0.3 dex relative to iron, result in overall metallicities of \[Z/Z$_{\rm \odot}$\] = -0.25 and -0.45 for NGC 6528 and NGC 6553 respectively. [lccccccccc]{} Star & M$_{\rm bol}$ & $T_{\rm eff} $ & log g & \[Fe/H\] &\[Mg/Fe\] & Method\ \ I 1 & -0.8 & 4400 & 1.7 & -0.4 & 0.3 & i\ & & 4400 & 1.7 & -0.4 & 0.3 & ii\ I 2 & -2.0 & 3700 & 0.9 & -0.5 & 0.3 & i\ I 5 & – & 3250 & 0.7 & -0.5 & 0.3 & i\ I 6 & -3.0 & 3550 & 0.4 & -0.6 & 0.3 & i\ II 8 & -1.5 & 4000 & 1.2 & -0.5 & 0.3 & i\ & & 4000 & 0.5 & -0.6 & 0.3 & ii\ II 14 &-3.1 & 3250 & 0.2 & -0.5 & 0.3 & i\ I 23 & 0.6 & 4800 & 2.4 & 0 & 0.3 & i\ & & 4800 & 2.4 & -0.4 & 0.2 & ii\ I 24 & 0.4 & 4800 & 2.3 & 0 & 0.3 & i\ & & 4800 & 2.3 & -0.3 & 0.1 & ii\ I 25 & -1.0 & 4100 & 1.4 & -0.6 & 0.3 & i\ & & 4100 & 1.4 & -1.1 & 0.2 & ii\ I 27 & -2.5 & 3350 & 0.5 & -0.3 & 0.3 & i\ I 36 & -0.6 & 4200 & 1.7 & -0.6 & 0.3 & i\ & & 4250 & 1.0 & -0.8 & 0.4 & ii\ I 40 & -1.2 & 4100 & 1.4 & -0.4 & 0.3 & i\ & & 4100 & 1.4 & -0.7 & 0.3 & ii\ I 42 & -0.8 & 4050 & 1.5 & -0.4 & 0.3 & i\ & & 4050 & 1.5 & -1.2 & 0.2 & ii\ II 70 & -1.6 & 3900 & 1.1 & -0.6 & 0.3 & i\ \ III 2 & – & 3800 & 2.0 & -0.7 & 0.3 & i\ III 3 & -1.1 & 3800 & 1.3 & -0.7 & 0.3 & i\ III 17& -2.3 & 3750 & 0.8 & -0.7 & 0.3 & i\ II 51 & – & 3250 & 0.8 & -0.4 & 0.3 & i\ II 52 & 0.6 & 3900 & 2.0 & -0.6 & 0.3 & i\ II 85 & -1.5 & 3950 & 1.2 & -0.6 & 0.3 & i\ II 94 & -2.7 & 3650 & 0.6 & -1.1 & 0.3 & i\ II 95 & – & 6500 & 2.7 & -0.4: & — & ii\ Conclusions =========== The study of individual stars in globular clusters along their evolutionary stages is of prime importance for an improved understanding of stellar evolution. Low resolution spectroscopy provides a means for the study of a large number of stars. In the present work we have measured radial velocities and estimated metallicities in 23 stars towards the globular clusters NGC 6528 and NGC 6553, which allows us to identify member stars. We also obtained their atmospheric properties to a first approximation. This is an important step before applying efforts to obtain high resolution spectroscopy with 8m class telescopes. The method presented here is also of interest for last generation multi-object instruments such as VLT-VIMOS. The stars were analysed by comparisons between their observed spectra and a grid of synthetic spectra. TiO equivalent widths were used to estimate effective temperatures of stars cooler than $T_{\rm eff}$ $\leq$ 3800 K and a calibration of equivalent widths of TiO bands as a function of atmospheric parameters is presented. Mean values of heliocentric radial velocities of v$_{\rm r}$ = 234 km s$^{-1}$ for NGC 6528 and v$_{\rm r}$ = -1 km s$^{-1}$ for NGC 6553 are derived. Regarding membership, among the 23 stars observed we concluded that 4 of them are probable non-members. These are: NGC 6528 II-39, non-member due to a deviant radial velocity, and NGC 6528 I-5, NGC 6553 III-2 and II-95, non-members due to incompatibilities of atmospheric parameters vs. location in the Colour-Magnitude Diagrams. NGC 6553 II-51 could be a non-member, or a red variable for which the spectrum was taken during a cool phase. The basic stellar parameters derived show the interesting result that there is a trend for member giants of NGC 6528 to be more metal-poor than the two Horizontal Branch stars NGC 6528 I-23 and I-24, thus reproducing the discrepancy found between analysis of NGC 6553 giants by Barbuy et al. (1999) and Horizontal Branch stars by Cohen et al. (1999). Given the errors involved in the analysis of low resolution spectra, these results have to be checked with high resolution spectra, and further studies of this discrepancy will be possible only with a homogeneous analysis of stars ranging from the red giant branch to the HB, and also employing different sets of model atmospheres all along the evolutionary sequence. In summary, we obtained for NGC 6528 and NGC 6553 metallicities of \[Fe/H\] = -0.5 $\pm$ 0.3 and \[Fe/H\] = -0.7 $\pm$ 0.3. Using \[Mg/Fe\] $\approx$ +0.3, and assuming that other $\alpha$ elements show the same excess of +0.3 dex relative to iron, the results are \[Z/Z$_{\rm \odot}$\] = -0.25 and -0.45 for NGC 6528 and NGC 6553 respectively. We are grateful to A. Milone for having carried out part of the observations. We acknowledge partial financial support from CNPq and Fapesp. P. Coelho and T. Idiart acknowledge respectively the Fapesp Master fellowship n$^{\rm o}$ 98/07492-4, and Post-Doc fellowship n$^{\rm o}$ 97/13083-7. RPS acknowledges support provided by the National Science Foundation through grant GF-1002-99 and from the Association of Universities for Research in Astronomy, Inc., under NSF cooperative agreement AST 96-1361. Armandroff T.E., Zinn R., 1988, AJ 96, 92 Barbuy B., Bica E., Ortolani S., 1998, A&A 333, 117 Barbuy B., 2000, in [The Chemical Evolution of the Milky Way: Stars vs. Clusters]{}, Eds. F. Matteucci, Kluwer Acad. Pub., in press Barbuy B., Renzini A., Ortolani S., Bica E., Guarnieri M.D., 1999, A&A 341, 539 Barbuy B., Perrin M.-N., Katz D., Cayrel R., Spite M., van ’t Veer-Menneret C., 2001, A&A, submitted Bessell M.S., Castelli F., Plez B., 1998, A&A 337, 321 Carretta E., Gratton R.G., 1997, A&AS 121, 95 Cohen J.G., Gratton R.G., Behr B.B., Carretta E., 1999, ApJ 523, 739 Fuhr J.R., Martin G.A., Wiese W.L., 1988, [Atomic Transition Probabilities: Iron through Nickel]{}, Journal of Physical and Chemical Reference Data, vol. 17, no. 4 Guarnieri M.D., Ortolani S., Montegriffo P., Renzini A., Barbuy B., Bica E., Moneti A., 1998, A&A 331, 70 Cayrel R., Perrin M.-N., Barbuy B., Buser R., 1991, A&A 247, 108 Harris W.E., 1996, AJ 112, 1487 Hartwick F.D.A., 1975, PASP 87, 77 Jacoby G.H., Hunter D.A., Christian C.A., 1984, ApJ 419, 592 Kurúcz, R., 1993, CD-ROM 18 Martin G.A., Fuhr J.R., Wiese W.L., 1988, [Atomic Transition Probabilities: Scandium through Manganese]{}, Journal of Physical and Chemical Reference Data, vol. 17, no. 3 McWilliam A., Rich R.M., 1994, ApJS 91, 749 (MR) Minniti D., 1995a, A&A 303, 468 Minniti D., 1995b, A&AS 113, 299 Schiavon R.P., Barbuy B., 1999, ApJ 510, 934 Ortolani S., Renzini A., Gilmozzi R., Marconi G., Barbuy B., Bica E., Rich R.M., 1995, Nature 377, 701 Plez B., Brett J.M., Nordlund ${\rm \AA}$, 1992, A&A 256, 551 Rieke G.H., Lebofsky M.J., 1985, ApJ 288, 618 Rutledge G.A., Hesser J.E., Stetson P.B., Mateo M., Simard L., Bolte M., Friel E.D., Copin Y., 1997, PASP 109, 883 Sadler E., Rich R.M., Terndrup D.M., 1996, AJ 112, 171 van den Bergh S., Younger F., 1979, AJ 84, 1305 Wiese W.L., Martin G.A., Fuhr J.R., 1969, [Atomic Transition Probabilities: Sodium through Calcium]{}, NSRDS-NBS 22 Zinn R., 1985, ApJ 293, 424 Zinn R., West, M.J., 1984, ApJS 55, 45 [^1]: Observations collected at the European Southern Observatory - ESO, Chile	{ "pile_set_name": "ArXiv" }
and -3.2mm A unified evolutionary scheme that includes post-AGB systems, barium stars, symbiotics, and related systems, explaining their similarites as well as differences. Can we construct it? We compare these various classes of objects in order to construct a consistent picture. Special attention is given to the comparison of the barium pollution and symbiotic phenomena. Finally, we outline a ‘transient torus’ evolutionary scenario that makes use of the various observational and theoretical hints and aims at explaining the observed characteristics of the relevant systems. stars: AGB and post-AGB, stars: binaries \[Sect:intro\] The term “after-AGB binaries” will be used in this paper to refer to binary systems in which at least [one of the components]{} has gone through the AGB phase, as distinguished from the term “post-AGB”, commonly used to denote the short transition phase between AGB and PN core (CSPN) stages of (single or binary) stellar evolution. In many of such after-AGB systems, the mass transfer from an AGB star has left its mark on the companion, enhancing its abundances with the products of the AGB nucleosynthesis, most remarkably C, F, and s-process elements (see Habing & Olofsson 2003 for a recent review). This pollution will manifest itself later in the companion’s evolution. An exemplary case of after-AGB systems are barium stars: G-K type giants remarkable for their overabundances of Ba (McClure et al. 1980). Related families include Abell-35 subclass of PNe (Bond et al. 1993), barium dwarfs (including the so called WIRRing stars, Jeffries & Stevens 1996), subgiant and giant CH stars, extrinsic S stars and d’-type yellow symbiotics. But not all of the after-AGBs need to be s-process rich. The post-AGB binaries are an interesting case, as they are all by definition after-AGBs: some of them do exhibit s-process enhancement while others do not (Van Winckel 2003, also this volume). Red s-type symbiotic stars (SyS) with massive white dwarf companions ($M_{\rm WD}>0.5 M_{\odot}$), another member of the after-AGB group, also do not exhibit s-process enhancement (Jorissen 2003a). Not much in this respect can be said about most binary CSPNe, as the unevolved companion is usually too faint to be seen. Finally, some of the cataclysmic variables (CV) with massive white dwarfs should also belong to the after-AGB family. \[Sect:links\] Peculiar red giants are a characteristic part of the after-AGB family and are in many respects closely related to symbiotic stars. The links between them have been reviewed previously (Jorissen 2003a, Jorissen et al. 2005); here we provide an updated discussion based on Fig. 1 which displays the various types of SyS and peculiar red giants in a metallicity – spectral-type plane. The vertical axis corresponds to metallicity which impacts (i) the taxonomy of the classes (CH giants for instance – box 1 – are the halo-equivalent of the disk barium stars – box 6); (ii) the efficiency of heavy-element synthesis (Clayton 1988), and (iii) the location of evolutionary tracks in the HR diagram (hence the correspondence between spectral type and evolutionary status, like the onset of TP-AGB, will depend on metallicity). Fig. 1 therefore considers three different metallicity ranges: (i) \[Fe/H\]$<-1$, corresponding to the halo population; (ii) $-1 \la [Fe/H] \la 0$, or disk metallicity; (iii) \[Fe/H\]$\ga 0$, solar and super-solar metallicities. The horizontal axis in Fig. 1 displays spectral type, which roughly corresponds to an evolutionary sequence at a given metallicity, passing from the red giant branch (RGB) to the thermally-pulsing AGB (TP-AGB) phases (between these two phases is the core He-burning phase which is hardly distinguishable from the lower RGB; CH giants probably belong to that phase). Symbiotic activity is expected in the middle of this sequence, because (i) at the left end, stars (like CH) are not luminous enough to experience a mass loss sufficient to power symbiotic activity; (ii) at the right end, the stars with the barium syndrome need not be binaries (above the TP-AGB luminosity threshold, heavy-elements are synthesized in the stellar interior and dredge-up to the surface, so that “intrinsic Ba” (or S) stars occupy the rightmost boxes – 4 and 8 – of Fig. 1), and hence need not exhibit any symbiotic activity. Those evolved giants which [are]{} members of binary systems (like Mira Ceti – box 9) will of course exhibit symbiotic activity. It is noteworthy that late M giants are inexistent in a halo population (hence the crossed box 5), because evolutionary tracks are bluer as compared to higher metallicities. Examples of such very evolved (relatively warm) stars in a halo population (box 4) include CS 30322-023 (Masseron et al. 2006) and V Ari (Van Eck et al. 2003). In the halo population, the most interesting issue is to understand the origin of the difference between yellow s-type SyS (box 2) and metal-deficient Ba stars (box 3): why do the latter not exhibit any symbiotic activity? The reason seems to reside in a difference in their orbital period distributions: yellow s-type SyS represent the short-period tail of the distribution (Fig. 2), where for a given mass-loss rate accretion will be more efficient, and hence will trigger symbiotic activity. [r]{}\[0pt\][68mm]{} In the intermediate-metallicity regime, the situation is quite clear: as the star evolves to the right along the spectral sequence, its luminosity (hence mass loss) increases, and for binary stars along that sequence, the symbiotic character will become stronger. The real issue here is to understand the origin of the difference between boxes 7 and 10/11: why are red s- (and d-)type SyS never exhibiting the barium syndrome, despite very similar locations in the HR diagram and similar orbital-period distributions (Fig. 2)? In former discussions of this issue (e.g. Jorissen 2003a), it had been suggested that binary stars with the barium syndrome (box 7) and SyS without it (box 10) differ in their metallicities. It is known that at high metallicities heavy-element synthesis is less efficient (Clayton 1988). However, there is so far [no evidence for red symbiotic stars being on average more metal-rich than barium or extrinsic S stars.]{} Schild et al. (1992) and Schmidt & Mikołajewska (2003) have compared the carbon abundances of SyS and normal giants, and found absolutely no difference, thus confirming at the same time the absence of any signature of internal nucleosynthesis and dredge-ups, or of pollution through mass transfer. Other solutions to this puzzle may be suggested, such as (i) s- and d-type SyS are not intrinsic barium stars, because they are not TP-AGB stars; (ii) neither are they extrinsic barium stars, because their companion never went through the TP-AGB phase, either because it is a He WD or because it is a main sequence star. Regarding item (i), it is very likely indeed that red s-type SyS are not TP-AGB stars, since they rather involve early M giants. The situation is less clear for d-type SyS, as they involve Miras which are often claimed to be TP-AGB stars. Nevertheless, many Miras do not exhibit signatures of heavy-element nucleosynthesis (they are not carbon stars and lack lines from the unstable element Tc; Little et al. 1987). Regarding item (ii), the possibility for hot companions to SyS to be He WDs is the most appealing since (a) the eccentricities observed for symbiotic systems are much smaller than those observed in pre-mass-transfer systems (M giants in the period range 200 – 1000 d have eccentricities up to 0.3; Jorissen et al. 2004), thus suggesting that mass transfer [ has taken place]{} in these systems; (b) the mass distribution of the hot components peaks between 0.4 and 0.5 $M_\odot$ (Mikołajewska 2003 and this volume), as expected for He WDs. Of course, an alternative explanation – like a main sequence companion – needs to be found for those SyS companions with masses exceeding 0.5 $M_\odot$ (T CrB, FG Ser, FN Sgr, AR Pav, V1329 Cyg; Mikołajewska 2003). Although a main sequence companion is quite unlikely in recurrent or symbiotic novae like T CrB and V1329 Cyg, the situation regarding the nature of symbiotic-star companions for non-nova systems is far from being settled, as mentioned by Mikołajewska (2003) while answering a question by one of the authors at the La Palma symbiotic-star conference: [ the question of whether \[the companion to CI Cyg, Z And, FN Sgr\] is a disk-accreting main-sequence star or a quasi-steady hydrogen-burning white dwarf is open so long as we have no good theory to distinguish between these possibilities]{}. Indeed, the nature of the companion to CI Cyg has changed over the years, from main-sequence accretor (Kenyon & Webbink 1984; Kenyon et al. 1991; Mikołajewska & Kenyon 1992) to hot and luminous stellar source powered by thermonuclear burning (Mikołajewska 2003)! The same move from main-sequence to white-dwarf accretor holds true for AR Pav (Kenyon & Webbink 1984,Quiroga et al. 2002). \[Sect:solar\_Z\] The evolutionary status of the rare set of yellow d’ SyS (box 12), which were all shown to be of solar metallicity, has recently been clarified (Jorissen et al. 2005) with the realisation that in these systems, the companion is [intrinsically]{} hot (because it recently evolved off the AGB), rather than being powered by accretion or nuclear burning. Several arguments support this claim: (i) d’ SyS host G-type giants whose mass loss is not strong enough to heat the companion through accretion and/or nuclear burning; (ii) the cool dust observed in d’ SyS (Schmid & Nussbaumer 1993) is a relic from the mass lost by the AGB star; (iii) the optical nebulae observed in d’ SyS are most likely genuine planetary nebulae (PN) rather than the nebulae associated with the ionized wind of the cool component (Corradi et al. 1999). d’ SyS often appear in PN catalogues. AS 201 for instance actually hosts [two]{} nebulae (Schwarz 1991): a large fossil planetary nebula detected by direct imaging, and a small nebula formed in the wind of the current cool component; (iv) rapid rotation is a common property of the cool components of d’ SyS (see Table 1 of Jorissen et al. 2005). It has likely been caused by spin accretion from the former AGB wind like in WIRRing systems (Jeffries & Stevens 1996; Jorissen 2003b). The fact that the cool star has not yet been slowed down by magnetic braking is another indication that the mass transfer occurred fairly recently (Theuns et al. 1996). Corradi & Schwarz (1997) obtained 4000 y for the age of the nebula around AS 201, and 40000 y for V417 Cen. The possible existence, in box 13, of binary systems of nearly solar metallicity with orbital properties typical of barium systems, but not exhibiting the barium syndrome, is still controversial, as discussed by Jorissen (2003b). \[Sect:orbital\] Intense AGB mass loss/transfer is not only important for chemical abundances; it does also influence the orbital properties of after-AGB systems. Four binary evolution processes are usually invoked when describing the after-AGB systems formation: (i) tidal interactions, (ii) wind accretion, including tidally enhanced winds (Companion-Reinforced Attrition Process or CRAP, Eggleton 1986), (iii) stable Roche-lobe overflow (RLOF), and (iv) common envelope (CE) evolution. Alas, current evolutionary computations fail to reproduce the correct ranges of orbital periods, eccentricities and s-process enhancement levels (e.g. Pols et al. 2003; Frankowski 2004). The basic reason for this problem is quite simple, as nicely put by Iben & Tutukov (1996): [as a result of CE interaction, initially close systems become closer and, because of wind mass loss, initially wide systems become wider. \[...\] most known symbiotic systems belong to a rare population on the borderline between initially close and wide binaries.]{} The models do not produce eccentric systems with periods below $\sim2000$ – 3000d and all systems below $\sim$1000d enter a CE and undergo a dramatic orbital shrinkage. The observed after-AGB systems with intermediate periods (100 – 2000d) have somehow avoided the catastrophic outcome of a CE, but the theoretical concepts proposed so far are not satisfactory in explaining this fact: (i) the inclusion of tidal forces affects only the detached evolution and does not improve the final results; (ii) CRAP does allow for slightly shorter final periods in detached evolution but still not below 2000d and any stronger effect would prevent TP-AGB, thus impeding s-process; (iii) stable RLOF occurs only for a narrow range of initial parameters; (iv) lowered binding energy of the AGB envelope due to the inclusion of ionisation energy as proposed e.g. by Han et al. (1994) is problematic (Harpaz 1998); (v) CE formalism based on angular-momentum instead of on energy (Nelemans et al. 2000) is promising, however, for the moment it lacks physical explanation. Also puzzling is high eccentricity (up to $e\!=\!0.4$) at periods down to 300d observed among post-AGB binaries, and to a lesser extent Ba stars and extrinsic S stars. The most promising explanation here is eccentricity pumping by a circumbinary disk (Waelkens et al. 1996; Artymowicz et al. 1991). Another suggestion is periastron mass loss eccentricity pumping (Soker 2000) but this mechanism can operate only for wide (detached on the AGB) systems. We suggest that these conundrums are part of a bigger puzzle together with the following observational and theoretical hints. First, some of the young after-AGB objects exhibit combined RS CVn and Ba star properties: X-rays, H$_\alpha$ emission and fast rotation combined with Ba enhancement and long orbital periods. The list consists of Ba stars 56 Peg (Frankowski & Jorissen 2006), HD 165141 (Jorissen et al. 1996), d’ symbiotics (Jorissen et al. 2005, see also the discussion in Sect. 2.3), WIRRing stars, (Jeffries & Stevens 1996), and Abell-35 CSPNe (Thévenin & Jasniewicz 1997). They form a strong evidence for fast rotation in young after-AGB systems, supposedly due to spin accretion from wind (Jeffries & Stevens 1996; Jorissen 2003b). Second, post-AGB systems, the youngest among the after-AGB family, are known to possess circumbinary disks (Van Winckel 2003). Dusty circumbinary disks, tori and bipolar outflows are common among bipolar and ring-like PNe, and have also been observed in some AGB stars, notably $\pi^1$ Gru (Sahai 1992) and V Hya (Knapp et al. 1999). The latter object is also remarkable for having fast rotation velocity (6-16 km s$^{-1}$) and a long secondary photometric period ($\sim 6200$d, in addition to the radial pulsation period of 530d), possibly due to a binary companion. Another notable factor is that dust formation and radiation-driven wind cause reshaping of Roche equipotentials and reduction of the effective gravity of the mass-losing star (Jorissen 2003b; Schuerman 1972; Frankowski & Tylenda 2001). \[Sect:scenario\] Gathering the observational and theoretical constraints described above, we propose a ‘transient torus’ scenario for explaining the observed orbital periods and eccentricities of “after AGB” binaries. This scenario can be divided into four phases, schematically represented in Fig. 3: 1\. Wind accretion. The system is well detached and the companion accretes mass and angular momentum from the giant’s wind. Spin accretion is especially efficient, proceeding through an accretion disk formed around the companion (Theuns et al. 1996; Mastrodemos & Morris 1998). Orbital evolution proceeds roughly as in spherically-symmetric wind case (Jeans mode), i.e., $a(M_1+M_2) =$ const and the eccentricity stays almost constant. 2\. (Near) RLOF with substantial $L_2/L_3$ outflow. Tidal forces and evolutionary expansion of the giant bring it closer to its Roche lobe. The outflow becomes concentrated in the direction to the companion, which happens even before the actual Roche-lobe filling (e.g. Frankowski & Tylenda 2001). The matter is ’funnelled’ through the vicinity of $L_1$. 3. Formation of a circumbinary torus. Matter escaping through the vicinity of $L_2$ (or $L_3$, after mass ratio reversal) forms a spiral around the system. But after one orbital period every portion of ejecta becomes shadowed from the giant by the newly ejected matter and ceases being accelerated outwards by the radiation pressure on dust. Part of the older ejecta gravitationally falls back onto the binary and collides with the new stream. A thick circumbinary torus is formed. 4\. Formation of a Keplerian circumbinary disk. The torus drags angular momentum from the binary and at the same time it is slowly pushed outwards by the radiation pressure on dust. The leftovers become a Keplerian disk. Only small part of the ejecta is pulled into Keplerian motion, so the angular momentum removal from the central binary is moderate and the orbital period can stay as long as a few hundred days. Point 2. in this sequence deserves particular consideration. At this stage the companion resides within the wind acceleration zone which is governed by the dust condensation radius, $R_{\rm cond}$. On TP-AGB $R_{\rm cond}$ is 2–5 $R_{}$ (Gail & Sedlmayr 1988). [r]{}\[0pt\][67mm]{} Thus the matter flowing preferentially through $L_1$ in the direction of the companion is still moving slowly in the vicinity of the companion (which favors higher accretion rate) and is still feeling an outward acceleration due to radiation pressure on dust (so the [modified]{} Roche potential is in force and a dynamical mass [transfer]{} leading to a CE can be avoided). Mass [loss]{} from the binary can proceed on a dynamical time scale for some part of this phase without an ensuing CE. These effects do not play a role for non-dusty winds, thus not changing the classical CE at RGB and E-AGB, leading to pre-CV and CV systems, as required for explaining those close binary populations. -1mm P., [Clarke]{} C. J., [Lubow]{} S. H., [Pringle]{} J. E. 1991, ApJ, 370, 35 H. E., [Ciardullo]{} R., [Meakes]{} M. G. 1993, in [Planetary nebulae]{}, IAU Symp. 155, eds. R. Weinberger & A. Acker, Kluwer Academic Publishers, Dordrecht, p. 397 D. D. 1988, MNRAS, 234, 1 R., [Schwarz]{} H. E. 1997, in [Physical Processes in Symbiotic Binaries and Related Systems]{}, ed. J. Mikołajewska, Copernicus Foundation for Polish Astronomy, Warsaw, p. 147 R. L. M., [Brandi]{} E., [Ferrer]{} O. E., [Schwarz]{} H. E. 1999, A&A, 343, 841 P. P. 1986, in J. [Trümper]{}, W. H. G. [Lewin]{}, W. [Brinkmann]{} (eds.), [The evolution of galactic X-ray binaries]{}, Reidel, Dordrecht, p. 87 A. 2004, PhD thesis, N. Copernicus Astronomical Center, Warsaw A., [Jorissen]{} A. 2006, Obs., 126, 25 A., [Tylenda R.]{} 2001, A&A, 367, 513 H.-P., [Sedlmayr]{} E. 1988, A&A, 206, 153 H. J., Olofsson, H. 2003, [Asymptotic Giant Branch Stars]{}, Springer Verlag, New York Z., [Podsiadlowski]{} P., [Eggleton]{} P. P. 1994, MNRAS, 270, 121 A. 1998, ApJ, 498, 293 I. Jr., [Tutukov]{} A. V. 1996, ApJS, 105, 145 R. D., [Stevens]{} I. R. 1996, MNRAS, 279, 180 A. 2003a, in R. L. M. [Corradi]{}, J. [Mikołajewska]{}, T. J. [Mahoney]{} (eds.), [Symbiotic stars probing stellar evolution]{}, ASP Conf. Ser., 303, 25 A. 2003b, in H. [Habing]{}, H. [Olofsson]{} (eds.), [Asymptotic Giant Branch Stars]{}, Springer Verlag, New York, p. 461 A., [Famaey]{} B., [Dedecker]{} M., [Pourbaix]{} D., [Mayor]{} M., [Udry]{} S. 2004, Rev. Mex. Astron. Astrof. Conf. Ser. 21, 71–72 A., [Schmitt]{} J. H. M. M., [Carquillat]{} J. M., [Ginestet]{} N., [Bickert]{} K. F. 1996, A&A, 306, 467 A., [Za[č]{}s]{} L., [Udry]{} S., [Lindgren]{} H., [Musaev]{} F. A. 2005, A&A, 441, 1135 S. J., [Oliversen]{} N. A., [Mikołajewska]{} J., [Mikolajewski]{} M., [Stencel]{} R. E., [Garcia]{} M. R., [Anderson]{} C. M. 1991, AJ, 101, 637 S. J., [Webbink]{} R. F. 1984, ApJ, 279, 252 G. R., [Dobrovolsky]{} S. I., [Ivezić]{} Z., [Young]{} K., [Crosas]{} M., [Mattei]{} J. A., [Rupen]{} M. P. 1999, A&A, 351, 97 S. J., [Little-Marenin]{} I. R., [Bauer]{} W. H. 1987, AJ, 94, 981 T., [Van Eck]{} S., [Famaey]{} B., [Goriely]{} S., [Plez]{} B., [Siess]{} L., [Beers]{} T., [Primas]{} F., [Jorissen]{} A. 2006, A&A, 455, 1059 N., [Morris]{} M. 1998, ApJ, 497, 303 R. D., [Fletcher]{} J. M., [Nemec]{} J. M. 1980, ApJ, 238, 35 J. 2003, in R. L. M. [Corradi]{}, J. [Mikołajewska]{}, T. J. [Mahoney]{} (eds.), [Symbiotic stars probing stellar evolution]{}, ASP Conf. Ser., 303, 9 J., [Kenyon]{} S. J. 1992, MNRAS, 177 G., [Verbunt]{} F., [Yungelson]{} L. R., [Portegies Zwart]{} S. F. 2000, A&A, 1011 O. R., [Karakas]{} A. I., [Lattanzio]{} J. C., [Tout]{} C. A. 2003 in R. L. M. [Corradi]{}, J. [Mikołajewska]{}, T. J. [Mahoney]{} (eds.), [Symbiotic stars probing stellar evolution]{}, ASP Conf. Ser., 303, 290 C., [Miko[ł]{}ajewska]{} J., [Brandi]{} E., [Ferrer]{} O., [Garc[í]{}a]{} L. 2002, A&A, 387, 139 R. 1992, A&A, 253, 33 H., [Boyle]{} S. J., [Schmid]{} H. M. 1992, MNRAS, 258, 95 H. M., [Nussbaumer]{} H. 1993, A&A, 268, 159 M., [Mikołajewska]{} J. 2003, in R. [Corradi]{}, J. [Mikołajewska]{}, T. J. [Mahoney]{} (eds.), [Symbiotic stars probing stellar evolution*]{}, ASP Conf. Ser., 303, 163 D. W. 1972, Ap&SS, 19, 351 H. E. 1991, A&A, 243, 469 N. 2000, A&A, 357, 557 T., [Boffin]{} H. M. J., [Jorissen]{} A. 1996, MNRAS, 280, 1264 F., [Jasniewicz]{} G. 1997, A&A, 320, 913 S., [Goriely]{} S., [Jorissen]{} A., [Plez]{} B. 2003, A&A, 404, 291 H. 2003, ARA&A, 41, 391 C., [Van Winckel]{} H., [Waters]{} L. B. F. M., [Bakker]{} E. J. 1996, A&A, 314, 17	{ "pile_set_name": "ArXiv" }
--- abstract: 'I discuss some problems connected with the high precision study of neutrino oscillations. In the general case of $n$-neutrino mixing I derive a convenient expression for transition probability in which only independent terms (and mass-squared differences) enter. For three-neutrino mixing I discuss a problem of a definition of a large (atmospheric) neutrino mass-squared difference. I comment also possibilities to reveal the character of neutrino mass spectrum in future reactor neutrino experiments.' --- [Some comments on high precision study of neutrino oscillations]{} S. M. Bilenky [Joint Institute for Nuclear Research, Dubna, R-141980, Russia\ ]{} [TRIUMF 4004 Wesbrook Mall, Vancouver BC, V6T 2A3 Canada\ ]{} Introduction ============ The observation of neutrino oscillations in the atmospheric Super-Kamiokande [@Fukuda:1998mi], solar SNO [@Ahmad:2002jz], reactor KamLAND [@Araki:2004mb] and solar neutrino oscillation experiments [@Cleveland:1998nv; @Altmann:2005ix; @Abdurashitov:2002nt] is one of the most important recent discovery in the particle physics. Small neutrino masses, many orders of magnitude smaller than masses of other fundamental fermions, are an evidence of a beyond the Standard Model physics. One of the plausible scenario which allow to explain the smallness of neutrino masses is based on the assumption that small (Majorana) neutrino masses are generated by the lepton-number violating dimension five effective Lagrangian [@Weinberg:1979sa]. In this case neutrino masses are suppressed with respect to masses of leptons and quarks by the ratio of the electroweak scale $v=(\sqrt{2}G_{F})^{-1/2}\simeq 246$ GeV and a scale $\Lambda\gg v$ of a new lepton number-violating physics. Neutrino oscillation data can be described by the three-neutrino mixing $$\label{mix} \nu_{lL}(x)=\sum^{3}_{i=1}U_{li}~\nu_{iL}(x).\quad (l=e,\mu,\tau)$$ Here $\nu_{i}(x)$ is the field of neutrinos (Dirac or Majorana) with mass $m_{i}$ and $U$ is the unitary $3\times3$ PMNS [@Pontecorvo:1957cp; @Pontecorvo:1957qd; @Maki:1962mu] mixing matrix. In the framework of the three-neutrino mixing neutrino oscillations are characterized by two neutrino mass-squared differences $\Delta m^{2}_{23}$ and $\Delta m^{2}_{12}$, three mixing angles $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and one $CP$ phase $\delta$. From the analysis of the data of neutrino oscillation experiments it was established that $\Delta m^{2}_{12}\ll \Delta m^{2}_{23}$, mixing angles $\theta_{23}$ and $\theta_{12}$ are large and mixing angle $\theta_{13}$ is small. The first information about the angle $\theta_{13}$ was obtained from the reactor CHOOZ experiment [@Apollonio:1999ae] in which only the upper bound $\sin^{2}2\theta_{13}\leq 1\cdot 10^{-1}$ was found. First data of neutrino oscillation experiments were described by expressions for neutrino transition probabilities in the leading approximation which was based on the assumption that $\sin^{2}\theta_{13}=0$. In this approximation oscillations in atmospheric and KamLAND (solar) regions are decoupled (see [@Bilenky:1998dt]): in the atmospheric region (atmospheric and long-baseline accelerator neutrino oscillation experiments) neutrino oscillations are two-neutrino $\nu_{\mu}\rightleftarrows\nu_{\tau} $ oscillations, in the solar region (the reactor KamLAND experiment) neutrino oscillations are $\bar\nu_{e}\rightleftarrows\bar\nu_{\mu,\tau} $ oscillations. From analysis of the atmospheric and long-baseline accelerator oscillation experiments parameters $\Delta m^{2}_{23}$ and $\sin^{2}2\theta_{23}$ were determined. From analysis of the data of the KamLAND and solar experiments another two neutrino oscillation parameters $\Delta m^{2}_{12}$ and $\sin^{2}2\theta_{12}$ were inferred. In the leading approximation the character of the neutrino mass spectrum and such important effect of the three-neutrino mixing as $CP$ violation in the lepton sector can not be revealed. With the measurement of the mixing angle $\theta_{13}$ in the reactor Daya Bay [@An:2013zwz], RENO [@Ahn:2012nd] and Double CHOOZ [@Abe:2013sxa] experiments the situation with the study of neutrino oscillations drastically changed. The investigation of neutrino oscillations entered into high precision era, era of measurements of small, beyond the leading approximation effects which could allow to determine the character of the neutrino mass spectrum and to measure $CP$ phase $\delta$. In this paper for the general case of the $n$-neutrino mixing we will derive a convenient expression for the neutrino transition probability in vacuum in which only independent terms (and mass-squared differences) enter. In different papers large (atmospheric) neutrino mass-squared difference is determined differently. Difference between different definitions is small (a few %) but in the era of precision measurements apparently it is desirable to have one unified definition. The expression for transition probability we will present here provides natural framework for introduction of two independent neutrino mass-squared differences in the case of the three-neutrino mixing. Determination of the character of the neutrino mass spectrum is one of the major aim of future reactor neutrino experiments JUNO [@Li:2014qca] and RENO-50 [@Kim:2014rfa]. On the basis of the proposed expression for the transition probability I will comment this possibility. General expression for neutrino transition probability in vacuum ================================================================ For the general case of the neutrino mixing $$\label{genmix} \nu_{\alpha L}(x)=\sum^{3+n_{s}}_{i=1}U_{\alpha i}~\nu_{iL}(x)\quad (\alpha=e, \mu, \tau, s_{1},...s_{n_{s}})$$ we will derive here an expression for $\nu_{\alpha}\to \nu_{\alpha'}$ transition probability alternative to the standard one. Here $n_{s}$ is the number of sterile neutrino fields, $U$ is an unitary $(3+n_{s})\times (3+n_{s})$ mixing matrix, $\nu_i(x)$ is the field of neutrino with mass $m_{i}$. From (\[genmix\]) and Heisenberg uncertainty relation it follows that normalized states of flavor $\nu_{e}, \nu_{\mu}, \nu_{\tau}$ and sterile $\nu_{s_{1}}, \nu_{s_{2}},...$ neutrinos are described by [coherent superpositions]{} of the states of neutrinos with definite masses (see, for example, [@Bilenky:1987ty; @Bilenky:1998dt; @Bilenky:2001yh]) $$\label{mixedstate} \|\nu_\alpha\rangle = \sum^{n}_{i=1} U_{\alpha i}^* \,~ \|\nu_i\rangle.$$ Here $\|\nu_i\rangle$ is the state of the left-handed neutrino with mass $m_{i}$, momentum $\vec{p}$ and energy $E_{i}=\sqrt{p^{2}+m^{2}_{i}}\simeq E+ \frac{m^{2}_{i}}{2E}$ ($E=p$ is the energy of neutrino at $m_{i}\to 0$). If at $t=0$ flavor neutrino $\nu_\alpha$ is produced, at the time $t$ we have $$\label{mixedstate2} \|\nu_\alpha\rangle_{t}= \sum_{\alpha'}\|\nu_{\alpha'}\rangle\langle\nu_{\alpha'}\| e^{-iH_{0}t}\|\nu_\alpha\rangle= \sum_{\alpha'}\|\nu_{\alpha'}\rangle (\sum_{i}U_{\alpha'i}, e^{-iE_{i}t}U^{}_{\alpha i})$$ where $H_{0}$ is the free Hamiltonian. From (\[mixedstate2\]) for the normalized probability of the $\nu_\alpha\to \nu_{\alpha'}$ transition we find the following expression $$\label{standard} P( \nu_\alpha\to \nu_{\alpha'})=\|\sum_{i}U_{\alpha'i}, e^{-iE_{i}t}U^{}_{\alpha i}\|^{2}=\sum_{i}\|U_{\alpha' i}\|^{2} \|U_{\alpha i}\|^{2}+ 2~\sum_{i>k}\mathrm{Re}(U_{\alpha'i} U^{}_{\alpha i}U^{}_{\alpha'k} U_{\alpha k} e^{-2i\Delta _{ki}}).$$ Here $$\label{phase} \Delta _{ki}= \frac{\Delta m^{2}_{ki}L}{4E},$$ where $\Delta m^{2}_{ki}= m^{2}_{k}- m^{2}_{i}$ and $L\simeq t$ is the neutrino source-detector distance. Taking into account the unitarity of the mixing matrix $U$ for the first term of the probability (\[standard\]) we have $$\label{standard1} \sum_{i}\|U_{\alpha' i}\|^{2} \|U_{\alpha i}\|^{2}=\delta _{\alpha' \alpha}- 2~\sum_{i>k}\mathrm{Re}(U_{\alpha'i} U^{}_{\alpha i}U^{}_{\alpha'k} U_{\alpha k}).$$ From (\[standard\]) and (\[standard1\]) for the ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha'}}}}}$ transition probability we obtain the following [standard expression]{} (see [@GonzalezGarcia:2007ib; @Giunti:2007ry; @Bilenky:2010zza]) $$\begin{aligned} \label{standard4} P(\bar \nu_\alpha\to \bar\nu_{\alpha'})&=&\delta _{\alpha' \alpha} -4~\sum_{i>k}\mathrm{Re}~(U_{\alpha'i} U^{}_{\alpha i}U^{}_{\alpha'k} U_{\alpha k})\sin^{2}\Delta _{ki}\nonumber\\ &\pm& 2~\sum_{i>k}\mathrm{Im}~(U_{\alpha'i} U^{}_{\alpha i}U^{}_{\alpha'k} U_{\alpha k})\sin 2\Delta _{ki}.\end{aligned}$$ Let us stress that not all quantities in (\[standard4\]) are independent. For example, in the case of the three-neutrino mixing three mass-squared differences (\[standard4\]) are connected by the relation $\Delta m^{2}_{13}=\Delta m^{2}_{12} +\Delta m^{2}_{23}$. For $\alpha'\neq \alpha$ the quantities in the last term of (\[standard4\]) are connected by the relations $\mathrm{Im}~(U_{\alpha'2}U^{}_{\alpha 2}U^{}_{\alpha'1} U_{\alpha 1})=\mathrm{Im}~(U_{\alpha'3}U^{}_{\alpha 3}U^{}_{\alpha'2} U_{\alpha 2})=-\mathrm{Im}~(U_{\alpha'3}U^{}_{\alpha 3}U^{}_{\alpha'1} U_{\alpha 1})$ which follow from the unitarity of the mixing matrix (see [@Giunti:2007ry; @Bilenky:2010zza]). We will obtain here a simple expression for the neutrino transition probability in vacuum in which - we will take into account that there is one arbitrary common phase in the transition amplitude, - we will use the unitarity of the mixing matrix in the transition amplitude. We have $$\begin{aligned} \label{Genexp1} P(\nu_{\alpha}\to \nu_{\alpha'})&=&\|\sum_{i}U_{\alpha'i}, e^{-i(E_{i}-E_{p})t}U^{}_{\alpha i}\|^{2}=\|\delta_{\alpha'\alpha}+ \sum_{i\neq p}U_{\alpha'i}~(e^{-2i\Delta _{pi}}-1)~ U^{}_{\alpha i}\|^{2}\nonumber\\&=&\|\delta_{\alpha'\alpha}-2i\sum_{i\neq p}U_{\alpha'i}~ U^{}_{\alpha i}e^{-i\Delta_{pi}}\sin\Delta_{pi}\|^{2},\end{aligned}$$ where $p$ is an arbitrary fixed index. From (\[Genexp1\]) we find $$\begin{aligned} &&P(\nu_{\alpha}\to \nu_{\alpha'})=\delta_{\alpha'\alpha} -4\sum_{i\neq p}\|U_{\alpha i}\|^{2}(\delta_{\alpha' \alpha } - \|U_{\alpha' i}\|^{2})\sin^{2}\Delta_{pi} \nonumber\\ &&+8~\sum_{i>k;i,k\neq p}\mathrm{Re}(U_{\alpha' i}U^{}_{\alpha i}U_{\alpha' k}^{}U_{\alpha k}~e^{-i(\Delta_{pi}-\Delta_{pk})})\sin\Delta_{pi}\sin\Delta_{pk}. \label{Genexp3}\end{aligned}$$ Finally, we obtain the following general expression for ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha'}}}}}$ transition probability [@Bilenky:2012zp] $$\begin{aligned} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha'}}}}}) =\delta_{\alpha' \alpha } -4\sum_{i\neq p}\|U_{\alpha i}\|^{2}(\delta_{\alpha' \alpha } - \|U_{\alpha' i}\|^{2})\sin^{2}\Delta_{pi}\nonumber\\ +8~\sum_{i>k;i,k\neq p }\mathrm{Re}~(U_{\alpha' i}U^{}_{\alpha i}U^{}_{\alpha' k}U_{\alpha k})\cos(\Delta_{pi}-\Delta_{pk})\sin\Delta_{pi}\sin\Delta_{pk}\nonumber\\ \pm 8~\sum_{i>k;i,k\neq p}\mathrm{Im}~(U_{\alpha' i}U^{}_{\alpha i}U^{}_{\alpha' k}U_{\alpha k})\sin(\Delta_{pi}-\Delta_{pk})\sin\Delta_{pi}\sin\Delta_{pk}, \label{Genexp4}\end{aligned}$$ where sign + (-) refers to $\nu_{\alpha}\to \nu_{\alpha'}$ ($\bar\nu_{\alpha}\to \bar\nu_{\alpha'}$) transition. In (\[Genexp4\]) only independent terms (and mass-squared differences) enter. For example, for the three-neutrino mixing there are only two independent mass-squared differences and one $i>k$ term in the transition probability (because $i,k\neq p$). Three-neutrino oscillations =========================== Atmospheric neutrino mass-squared difference? Flavor neutrino transition probability ------------------------------------------------------------------------------------ From analysis of the neutrino oscillation data it follows that one mass-squared difference (atmospheric) is much larger than the other one (solar). Two neutrino mass spectra are possible in such a situation [^1] 1. Neutrino spectrum with small mass-squared difference between lightest neutrinos (Normal spectrum, NS) $m_{1}<m_{2}<m_{3},\quad \Delta m_{12}^{2}\ll \Delta m_{23}^{2}$ 2. Neutrino spectrum with small mass-squared difference between heaviest neutrinos (Inverted spectrum, IS) $m_{3}<m_{1}<m_{2},\quad \Delta m_{12}^{2}\ll \|\Delta m_{13}^{2}\|$ There are only two possibilities to introduce small (solar) $\Delta m_{S}^{2}$ and large (atmospheric) $\Delta m_{A}^{2}$ mass-squared differences in the framework of the approach we are advocating [^2]. 1. $$\label{NNS} \mathrm{NS.}~~~\Delta m_{21}^{2}=-\Delta m_{S}^{2},\quad \Delta m_{23}^{2}=\Delta m_{A}^{2} ~~(p=2)$$ $$\label{IIS} \mathrm{IS.}~~~ \Delta m_{12}^{2}=\Delta m_{S}^{2},\quad \Delta m_{13}^{2}=-\Delta m_{A}^{2}~~ (p=1)$$ 2. $$\label{NNNS} \mathrm{NS.}~~~ \Delta m_{12}^{2}=\Delta m_{S}^{2},\quad \Delta m_{13}^{2}=\Delta m_{A}^{2}~~(p=1)$$ $$\label{IIIS} \mathrm{IS.}~~~ \Delta m_{21}^{2}=-\Delta m_{S}^{2},\quad \Delta m_{23}^{2}=-\Delta m_{A}^{2}~~ (p=2)$$ In all papers on neutrino oscillations mixing angles, $CP$ phase and solar mass-squared difference are determined in same way. However, atmospheric mass-squared difference in different papers is determined differently. For example, (in terms of parameters introduced in 1.) 1. The Bari group determines large neutrino mass-squared difference as follows (see [@Capozzi:2013psa]) $$\label{3numix4} \Delta m^{2}=\frac{1}{2}\| \Delta m_{13}^{2}+\Delta m_{23}^{2}\|=\Delta m_{A}^{2}+\frac{1}{2}\Delta m_{S}^{2}$$ 2. The NuFit group determines the atmospheric mass-squared difference as in 2. (see [@Gonzalez-Garcia:2014bfa]) $$\label{3numix5} \Delta m_{13}^{2}=\Delta m_{A}^{2}+\Delta m_{S}^{2},~~~(NS),~~~ \Delta m_{23}^{2}=-(\Delta m_{A}^{2}+\Delta m_{S}^{2})~~~(IS).$$ 3. In the T2K paper [@Abe:2014ugx] the atmospheric mass-squared difference is determined as in 1. 4. In the MINOS paper [@Adamson:2014vgd] large mass-squared difference is determined as $\|\Delta m_{23}^{2}\|$ for both mass spectra. It is obvious, however, that $\Delta m_{23}^{2}$ for NS and $\|\Delta m_{23}^{2}\|$ for IS are different quantities. The difference between different “atmospheric neutrino mass-squared differences” is a few %. It is determined by the ratio $\frac{ \Delta m_{S}^{2}}{ \Delta m_{A}^{2}}\simeq 3\cdot 10^{-2}$ and can not be neglected in the precision era. Apparently one definition is desirable. We will choose here the option 1. From (\[Genexp4\]) in the case of normal and inverted neutrino mass spectra we have respectively $$\begin{aligned} &&P^{NS}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l'}}}}}) =\delta_{l' l } -4\|U_{l 3}\|^{2}(\delta_{l' l} - \|U_{l' 3}\|^{2})\sin^{2}\Delta_{A}\nonumber\\&&-4\|U_{l 1}\|^{2}(\delta_{l' l} - \|U_{l' 1}\|^{2})\sin^{2}\Delta_{S} -8~\mathrm{Re}~(U_{l' 3}U^{}_{l 3}U^{}_{l' 1}U_{l 1})\cos(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}\nonumber\\ &&\mp 8~\mathrm{Im}~(U_{l' 3}U^{}_{l 3}U^{}_{l' 1}U_{l 1})\sin(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}, \label{Genexp5}\end{aligned}$$ and $$\begin{aligned} &&P^{IS}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l'}}}}}) =\delta_{l' l } -4\|U_{l 3}\|^{2}(\delta_{l' l } - \|U_{l' 3}\|^{2})\sin^{2}\Delta_{A}\nonumber\\&&-4\|U_{l 2}\|^{2}(\delta_{l' l} - \|U_{l' 2}\|^{2})\sin^{2}\Delta_{S} -8~\mathrm{Re}~(U_{l' 3}U^{}_{l 3}U^{}_{l' 2}U_{l 2})\cos(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}\nonumber\\ &&\pm 8~\mathrm{Im}~(U_{l' 3}U^{}_{l 3}U^{}_{l' 2}U_{l 2})\sin(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}. \label{Genexp6}\end{aligned}$$ Here $$\label{Delta} \Delta_{A,S}=\frac{\Delta m^{2}_{A,S}L}{4E},$$ Thus, transition probabilities depend on “extreme values” of the elements of neutrino mixing matrix: $U_{l'1(3)}$ and $U_{l1(3)}$ in the NS case ($p=2$) and $U_{l'2(3)}$ and $U_{l2(3)}$ in the IS case ($p=1$). Difference in signs of the last terms of (\[Genexp5\]) and (\[Genexp6\]) is connected with signs in (\[NNS\]) and (\[IIS\]). If $CP$ is violated in the lepton sector in this case we have $$\label{1CP} P(\nu_{l}\to \nu_{l'})\neq P(\bar\nu_{l}\to \bar\nu_{l'})\quad (l'\neq l)$$ Let us determine the $CP$ asymmetry $$\label{CPasymm} A^{CP}_{l' l }=P(\nu_{l}\to \nu_{l'})-P(\bar\nu_{l}\to \bar\nu_{l'})$$ The $CP$ asymmetry satisfies the following general conditions $$\label{CPasymm1} A^{CP}_{l' l }=-A^{CP}_{l l'}.$$ and $$\label{CPasymm2} \sum_{l'}A^{CP}_{l' l }=0.$$ The first condition follows from the relation $$\label{CPasymm3} P(\nu_{l}\to \nu_{l'})= P(\bar\nu_{l'}\to \bar\nu_{l})$$ which is a consequence of the $CPT$ invariance. The second condition follows from the conservation of the probability $$\label{CPasymm4} \sum_{l'}P(\nu_{l}\to \nu_{l'})=\sum_{l'}P(\bar\nu_{l}\to \bar\nu_{l'})=1.$$ From (\[CPasymm1\]) and (\[CPasymm2\]) it follows that in the case of the three-neutrino mixing $CP$ asymmetries in different flavor channels are connected by the following relations [@Bilenky:1981hf] $$\label{CPasymm5} A^{CP}_{\mu e }=A^{CP}_{e \tau}=-A^{CP}_{\mu \tau }.$$ From (\[Genexp5\]) in the case NS we have $$\label{CPNH} A^{CP}_{l' l }=-16~\mathrm{Im}~U_{l' 3}U^{}_{l 3}U^{}_{l' 1}U_{l 1}\sin(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}.$$ For IS from (\[Genexp6\]) we find $$\label{CPIH} A^{CP}_{l' l }=16~\mathrm{Im}~U_{l' 3}U^{}_{l 3}U^{}_{l' 2}U_{l 2}\sin(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}.$$ In the next subsections we will present expressions for transition probabilities which are of experimental interest. For that we will use the standard parametrization of the PMNS mixing matrix $$\begin{aligned} U=\left(\begin{array}{ccc}c_{13}c_{12}&c_{13}s_{12}&s_{13}e^{-i\delta}\\ -c_{23}s_{12}-s_{23}c_{12}s_{13}e^{i\delta}& c_{23}c_{12}-s_{23}s_{12}s_{13}e^{i\delta}&c_{13}s_{23}\\ s_{23}s_{12}-c_{23}c_{12}s_{13}e^{i\delta}& -s_{23}c_{12}-c_{23}s_{12}s_{13}e^{i\delta}&c_{13}c_{23} \end{array}\right). \label{unitmixU1}\end{aligned}$$ Here $c_{12}=\cos\theta_{12}$, $s_{12}=\sin\theta_{12}$ etc. $\bar\nu_{e}\to \bar\nu_{e}$ survival probability ------------------------------------------------- Expressions for the three-neutrino $\bar\nu_{e}$ survival probabilities are important for analysis of the data of the reactor neutrino experiments. From (\[Genexp5\]) and (\[Genexp6\]) for normal and inverted mass ordering we have respectively $$\begin{aligned} &&P^{\mathrm{NS}}(\bar\nu_{e}\to \bar\nu_{e})=1- 4~\|U_{e 3}\|^{2}(1-\|U_{e3}\|^{2})~ \sin^{2}\Delta_{A} \nonumber\\ &&- 4~\|U_{e 1}\|^{2}(1-\|U_{e 1}\|^{2})~ \sin^{2}\Delta_{S} \nonumber\\ &&-8~\|U_{e 3}\|^{2}\|U_{e 1}\|^{2}~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3nue1}\end{aligned}$$ and $$\begin{aligned} &&P^{\mathrm{IS}}(\bar\nu_{e}\to \bar\nu_{e})=1- 4~\|U_{e 3}\|^{2}(1-\|U_{e3}\|^{2})~ \sin^{2}\Delta_{A} \nonumber\\ &&- 4~\|U_{e 2}\|^{2}(1-\|U_{e 2}\|^{2})~ \sin^{2}\Delta_{S} \nonumber\\ &&-8~\|U_{e 3}\|^{2}\|U_{e 2}\|^{2}~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3nue2}\end{aligned}$$ Using the standard parametrization of the PMNS mixing matrix (\[unitmixU1\]) from (\[Genexp5\]) and (\[Genexp6\]) for NS and IS we have respectively $$\begin{aligned} &&P^{\mathrm{NS}}(\bar\nu_{e}\to \bar\nu_{e})=1- \sin^{2}2\theta_{13} \sin^{2}\Delta_{A} \nonumber\\ &&- (\sin^{2}2\theta_{12}c^{2}_{13}+\sin^{2}2\theta_{13}c^{4}_{12}) ~ \sin^{2}\Delta_{S} \nonumber\\ &&-2\sin^{2}2\theta_{13}c^{2}_{12} ~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3nue4}\end{aligned}$$ and $$\begin{aligned} &&P^{\mathrm{IS}}(\bar\nu_{e}\to \bar\nu_{e})=1- \sin^{2}2\theta_{13} \sin^{2}\Delta_{A} \nonumber\\ &&- (\sin^{2}2\theta_{12}c^{2}_{13}+\sin^{2}2\theta_{13}s^{4}_{12}) ~ \sin^{2}\Delta_{S} \nonumber\\ &&-2\sin^{2}2\theta_{13}s^{2}_{12} ~\cos(\Delta_{A}+\Delta_{S})\sin\Delta_{A}\sin\Delta_{S}.\label{3nue5}\end{aligned}$$ Notice that $P^{\mathrm{IS}}(\bar\nu_{e}\to \bar\nu_{e})$ can be obtained from $P^{\mathrm{NS}}(\bar\nu_{e}\to \bar\nu_{e})$ by the change $c^{2}_{12}\to s^{2}_{12}$. $\nu_{\mu}\to \nu_{e}$ ($\bar\nu_{\mu}\to \bar\nu_{e}$) appearance probability ------------------------------------------------------------------------------ Vacuum three-neutrino expressions for ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}}$ transition probabilities are important for analysis of the data of long baseline accelerator experiments in which matter effects are negligible. From (\[Genexp5\]) and (\[Genexp6\]) we have $$\begin{aligned} &&P^{\mathrm{NS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})= 4~\|U_{e 3}\|^{2}\|U_{\mu 3}\|^{2}~ \sin^{2}\Delta_{A} \nonumber\\ &&+4~\|U_{e 1}\|^{2}\|U_{\mu 1}\|^{2}~ \sin^{2}\Delta_{S}\nonumber\\ &&-8~\mathrm{Re}~(U_{e 3}U^{}_{\mu3}U_{e 1}^{}U_{\mu 1})~\cos(\Delta _{A}+\Delta _{S}) \sin\Delta _{A}\sin\Delta _{S}\nonumber\\ &&\mp 8~\mathrm{Im}~(U_{e 3}U^{}_{\mu 3}U_{e 1}^{}U_{\mu 1})~\sin(\Delta _{A}+\Delta _{S}) \sin\Delta _{A}\sin\Delta _{S} .\label{3numue1}\end{aligned}$$ and $$\begin{aligned} &&P^{\mathrm{IS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})= 4~\|U_{e 3}\|^{2}\|U_{\mu 3}\|^{2}~ \sin^{2}\Delta_{A} \nonumber\\ &&+4~\|U_{e 2}\|^{2}\|U_{\mu 2}\|^{2}~ \sin^{2}\Delta_{S}\nonumber\\ &&-8~\mathrm{Re}~(U_{e 3}U^{}_{\mu3}U_{e 2}^{}U_{\mu 2})~\cos(\Delta _{A}+\Delta _{S}) \sin\Delta _{A}\sin\Delta _{S}\nonumber\\ &&\pm 8~\mathrm{Im}~(U_{e 3}U^{}_{\mu 3}U_{e 2}^{}U_{\mu 2})~\sin(\Delta _{A}+\Delta _{S}) \sin\Delta _{A}\sin\Delta _{S} .\label{3numue2}\end{aligned}$$ Using the standard parameterization of the PMNS mixing matrix in the case of NS we have $$\begin{aligned} &&P^{\mathrm{NS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})=\sin^{2}2\theta_{13}s^{2}_{23}\sin^{2}\Delta_{A} \nonumber\\&&+(\sin^{2}2\theta_{12}c^{2}_{13}c^{2}_{23}+ \sin^{2}2\theta_{13}c^{4}_{12} s^{2}_{23}+Kc^{2}_{12}\cos\delta) \sin^{2}\Delta_{S}\nonumber\\&&+(2\sin^{2}2\theta_{13}s^{2}_{23}c^{2}_{12}+K\cos\delta) ~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}\nonumber\\ &&\mp 8 J_{CP}~\sin(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}. \label{3numue3}\end{aligned}$$ Here $$\label{3numue4} K=\sin2\theta_{12}\sin2\theta_{13}\sin2\theta_{23}c_{13}.$$ and $$\label{Jarskog} J_{CP}=\frac{1}{8}K\sin\delta$$ is the Jarlskog invariant [@Jarlskog:1985ht]. In the case of the inverted neutrino mass spectrum we find $$\begin{aligned} &&P^{\mathrm{IS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})=\sin^{2}2\theta_{13}s^{2}_{23}\sin^{2}\Delta_{A}\nonumber\\&&+(\sin^{2}2\theta_{12}c^{2}_{13} c^{2}_{23}+ \sin^{2}2\theta_{13} s^{4}_{12}s^{2}_{23}-K s^{2}_{12}\cos\delta) \sin^{2}\Delta_{S} \nonumber\\&&+(2\sin^{2}2\theta_{13}s^{2}_{23}s^{2}_{12}-K\cos\delta) ~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}\nonumber\\ &&\mp 8 J_{CP}~\sin(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}. \label{3numue5}\end{aligned}$$ For the $CP$ asymmetry in the case of NS (IS) we have $$\label{CPassym} A^{CP}_{e\mu} = -16 J_{CP}~\sin(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.$$ $\nu_{\mu}\to\nu_{\mu}$ ($\bar\nu_{\mu}\to \bar\nu_{\mu}$) survival probability ------------------------------------------------------------------------------- From (\[Genexp4\]) for ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to{\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}$ survival probability in the case of the normal and inverted mass ordering we have correspondingly $$\begin{aligned} &&P^{\mathrm{NS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}})=1- 4~\|U_{\mu 3}\|^{2}(1-\|U_{\mu3}\|^{2})~ \sin^{2}\Delta_{A} \nonumber\\ &&- 4~\|U_{\mu 1}\|^{2}(1-\|U_{\mu 1}\|^{2})~ \sin^{2}\Delta_{S} \nonumber\\ &&-8~\|U_{\mu 3}\|^{2}\|U_{\mu 1}\|^{2}~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3mumu}\end{aligned}$$ and $$\begin{aligned} &&P^{\mathrm{IS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}})=1- 4~\|U_{\mu 3}\|^{2}(1-\|U_{\mu 3}\|^{2})~ \sin^{2}\Delta_{A} \nonumber\\ &&- 4~\|U_{\mu 2}\|^{2}(1-\|U_{\mu 2}\|^{2})~ \sin^{2}\Delta_{S} \nonumber\\ &&-8~\|U_{\mu 3}\|^{2}\|U_{\mu 2}\|^{2}~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3mumu2}\end{aligned}$$ Using standard parametrization of the PMNS matrix we find $$\begin{aligned} &&P^{\mathrm{NS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}})=1- (\sin^{2}2\theta_{23}c^{2}_{13}+\sin^{2}2\theta_{13}s^{4}_{23}) ~ \sin^{2}\Delta_{A} \nonumber\\ &&-4(c^{2}_{23}s^{2}_{12}+s^{2}_{23}c^{2}_{12}s^{2}_{13} +\frac{K\cos\delta}{4c^{2}_{13}})(1-c^{2}_{23}s^{2}_{12}-s^{2}_{23}c^{2}_{12}s^{2}_{13}-\frac{K\cos\delta}{4c^{2}_{13}}) \sin^{2}\Delta_{S}\nonumber\\ &&-2(\sin^{2}2\theta_{23}c^{2}_{13}s^{2}_{12} + \sin^{2}2\theta_{13}c^{2}_{12}s^{4}_{23} \nonumber\\&&+Ks^{2}_{23}\cos\delta)~\cos(\Delta_{A}+\Delta_{A})\sin\Delta_{A}\sin\Delta_{S}.\label{3mumua}\end{aligned}$$ and $$\begin{aligned} &&P^{\mathrm{IS}}({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}})=1- (\sin^{2}2\theta_{23} c^{2}_{13}+\sin^{2}2\theta_{13}s^{4}_{23}) ~ \sin^{2}\Delta_{A} \nonumber\\ &&-4(c^{2}_{23}c^{2}_{12}+s^{2}_{23}s^{2}_{12}s^{2}_{13} -\frac{K\cos\delta}{4c^{2}_{13}})(1-c^{2}_{23}c^{2}_{12}-s^{2}_{23}s^{2}_{12}s^{2}_{13}+\frac{K\cos\delta}{4c^{2}_{13}}) \sin^{2}\Delta_{S}\nonumber\\ &&-2(\sin^{2}2\theta_{23}c^{2}_{13}c^{2}_{12} + \sin^{2}2\theta_{13}s^{2}_{12}s^{4}_{23} \nonumber\\&&-Ks^{2}_{23}\cos\delta)~\cos(\Delta_{A}+\Delta_{A})\sin\Delta_{A}\sin\Delta_{S}. \label{3mumu}\end{aligned}$$ where $K$ is given by the relation (\[3numue4\]). Notice that in the case of long baseline experiments with $\Delta_{A}\simeq 1$ (MINOS, T2K) the term proportional to $\sin^{2}\Delta_{S}$ gives very small contribution to the probability ($\sin^{2}\Delta_{S}\simeq 10^{-3}$). A comment on the possibility to reveal the character of neutrino mass spectrum in future reactor experiments ------------------------------------------------------------------------------------------------------------ Dependence on the neutrino mass ordering of the probability of reactor $\nu_{e}$’s to survive was noticed in the paper [@Bilenky:2001jq] in which reactor CHOOZ data were analyzed in the framework of three neutrino mixing. A reactor experiment with reactor-detector distance 20-30 km which could reveal the character of neutrino mass spectrum was proposed in [@Petcov:2001sy; @Choubey:2003qx]. Later in numerous papers a possibility to determine the neutrino mass ordering in a intermediate baseline reactor experiment ($\sim$ 50 km) was analyzed in details (see [@Li:2013zyd] and references therein). Two reactor experiments JUNO [@Li:2014qca] and RENO-50 [@Kim:2014rfa], in which the neutrino mass ordering is planned to be determined, are at preparation at present. The $\bar\nu_{e}\to \bar\nu_{e}$ survival probability (expressions (\[Genexp5\]) and (\[Genexp6\]) ) can be written in the form $$\begin{aligned} &&P(\bar\nu_{e}\to \bar\nu_{e})=1- \sin^{2}2\theta_{13} \sin^{2}\Delta_{A} \nonumber\\ &&- 4~X(1-X)~ \sin^{2}\Delta_{S} \nonumber\\ &&-8\sin^{2}\theta_{13} X~\cos(\Delta_{A}+\Delta_{S}) \sin\Delta_{A}\sin\Delta_{S}.\label{3nue6}\end{aligned}$$ In the case of the normal and inverted mass spectra we have respectively $$\label{3nue4} X=X_{NS}=\cos^{2}\theta_{13}\cos^{2}\theta_{12}$$ and $$\label{3nue5} X=X_{IS}=\cos^{2}\theta_{13}\sin^{2}\theta_{12}$$ From the fit of the data that will be obtained in the reactor JUNO experiment after six years of data taking the parameters $\Delta m^{2}_{S}$, $\Delta m^{2}_{A}$ and $\sin^{2}2\theta_{12}$ will be determined with accuracy better than 1% (see, for example, [@He:2013hla]). In the Daya Bay experiment the parameter $\sin^{2}\theta_{13}$ can be determined with accuracy $\sim$ 4%. Such a precision will, apparently, allow to distinguish the value $X\simeq 0.682$ (NS) from the value $X \simeq 0.295$ (IS) (we used best-fit values $\sin^{2}\theta_{12}=0.302,~~\sin^{2}\theta_{13}=0.0227$). Transitions of flavor neutrinos into sterile states =================================================== Data of atmospheric, solar, reactor and accelerator neutrino oscillation experiments are described by the three-neutrino mixing with two neutrino mass-squared differences $\Delta m_{S}^{2}\simeq 7.5 \cdot 10^{-5}~\mathrm{eV}^{2}$ and $\Delta m_{A}^{2}\simeq 2.4 \cdot 10^{-5}~\mathrm{eV}^{2}$. There exist, however, indications in favor of neutrino oscillations with mass-squared difference(s) about 1 $\mathrm{eV}^{2}$. These indications were obtained in following short baseline neutrino experiments (with $L$ ranging from a few meters to about 500 meters): 1. In the LSND experiment [@Aguilar:2001ty]. In this experiment neutrinos were produced in decays of $\pi^{+}$’s and $\mu^{+}$’s. Appearance of $\bar\nu_{e}$’s (presumably produced in the transition $\bar\nu_{\mu}\to \bar\nu_{e}$) were detected. In the MiniBooNE experiment [@Aguilar-Arevalo:2013pmq; @Conrad:2013mka]. In this experiment an excess of low energy $\nu_{e}$’s ( $\bar\nu_{e}'s$) was observed. 2. In the old reactor neutrino experiments. Data of these experiments were reanalyzed in [@Mention:2011rk]. In this new analysis recent calculations of the reactor neutrino flux [@Mueller:2011nm; @Huber:2011wv] was used. 3. In the calibration experiments, performed with radiative sources by the GALLEX [@Kaether:2010ag] and SAGE [@Abdurashitov:2009tn] collaborations. In these experiments a deficit of $\nu_{e}$’s was observed. In order to interpret these data in terms of neutrino oscillations it necessary to assume that in addition to the flavor neutrinos $\nu_{e},\nu_{\mu},\nu_{\tau}$ exist also sterile neutrinos. Let us consider first 3+1 scheme with three close neutrino masses $m_{i}$ ($i=1,2,3$) and forth mass $m_{4}$ separated from $m_{i}$ by about $ 1 \mathrm{eV}$ gap. We will choose $p=1$. In the region of $\frac{L}{E}$ sensitive to large neutrino mass-squared difference ($\frac{\Delta m_{14}^{2}L}{4E}\gtrsim 1$) we have $\Delta_{12}\simeq \Delta_{13}\simeq 0$. From (\[Genexp4\]) we find in this case $$\label{ster} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\alpha'}}}}})= \delta_{\alpha' \alpha } -4\|U_{\alpha 4}\|^{2}(\delta_{\alpha' \alpha } - \|U_{\alpha' 4}\|^{2})\sin^{2}\Delta_{14}.$$ From this expression for ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}}$ appearance probability and ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}}$ and ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}$ disappearance probabilities we have, respectively, the following expressions $$\label{transition2} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})=\sin^{2}2\theta_{e\mu}\sin^{2}\Delta_{14},$$ $$\label{transition3} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{e}}}}})=1-\sin^{2}2\theta_{e e}\sin^{2}\Delta_{14},$$ $$\label{transition4} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{\mu}}}}})=1-\sin^{2}2\theta_{\mu\mu}\sin^{2}\Delta_{14}.$$ Here $$\label{transition5} \sin^{2}2\theta_{e\mu}=4\|U_{e 4}\|^{2}\|U_{\mu 4}\|^{2},~\sin^{2}2\theta_{e e}=4\|U_{e 4}\|^{2}(1-\|U_{e 4}\|^{2}),~ \sin^{2}2\theta_{\mu\mu}=4\|U_{\mu 4}\|^{2}(1-\|U_{\mu 4}\|^{2}).$$ Notice that the global analysis of all short baseline neutrino data [@Giunti:2013aea; @Kopp:2013vaa] revealed inconsistency (tension) of existing short baseline data. Let us consider more complicated 3+2 scheme with 2 masses $m_{4}$ and $m_{5}$ separated from three close masses $m_{i}$ ($i=1,2,3$) by about 1 eV gaps. We will choose $p=1$. In the region of $\frac{L}{E}$ sensitive to large neutrino mass-squared differences $\Delta m^{2}_{14}$ and $\Delta m^{2}_{15}$ we have $\Delta_{12}\simeq \Delta_{13}\simeq 0$. From (\[Genexp4\]) we find the following expression for ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}$ ($l=e,\mu$) survival probability $$\begin{aligned} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}) &=&1 -4\|U_{l 4}\|^{2}(1 - \|U_{l 4}\|^{2})\sin^{2}\Delta_{14}-4\|U_{l 5}\|^{2}(1 - \|U_{l 5}\|^{2})\sin^{2}\Delta_{15}\nonumber\\ &+&8~\|U_{l 5}\|^{2}\|U_{l 4}\|^{2}\cos(\Delta_{15}-\Delta_{14})\sin\Delta_{15}\sin\Delta_{14}. \label{Ster5}\end{aligned}$$ For the probability of the transitions ${\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l'}}}}},~~l'\neq l$ we find $$\begin{aligned} P({\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l}}}}}\to {\ensuremath{\rlap {\kern-2.5pt\ensuremath {\overset{\scriptscriptstyle(-)}{\phantom{\nu}}}} {\ensuremath{{\nu}_{l'}}}}}) = 4\|U_{l' 4}\|^{2} \|U_{l 4}\|^{2}\sin^{2}\Delta_{14} +4\|U_{l' 5}\|^{2} \|U_{l 5}\|^{2}\sin^{2}\Delta_{15} \nonumber\\ +8~\mathrm{Re}~(U_{l' 5}U^{}_{l 5}U^{}_{l' 4}U_{l 4})\cos(\Delta_{15}-\Delta_{14}) \sin\Delta_{15}\sin\Delta_{14}\nonumber\\ \pm 8~\mathrm{Im}~(U_{l' 5}U^{}_{l 5}U^{}_{l' 4}U_{l 4})\sin(\Delta_{15}-\Delta_{14})\sin\Delta_{15}\sin\Delta_{14}. \label{Ster4}\end{aligned}$$ Conclusion ========== Discovery of neutrino oscillations is one of the most important recent discovery in the particle physics. After the first stage of investigation of this new phenomenon now with the measurement of the small parameter $\sin^{2}\theta_{13}\simeq 2.5\cdot 10^{-2}$ the era of precision study started. Such fundamental problems of neutrino masses and mixing as - what is the ordering of neutrino masses (normal or inverted), - what is the value of the $CP$ phase $\delta$, - what are precise values (with accuracies better than 1%) of other oscillation parameters, - is the number of massive neutrinos equal to the number of flavor neutrinos (three) or larger than three (are sterile neutrinos exist) are planned to be solved by future neutrino oscillation experiments. At the moment there is no consensus in definition of the large (atmospheric) neutrino mass-squared difference: in different experimental and theoretical papers this parameter is defined differently. Today it is not so important but with future precision different “atmospheric mass-squared differences” will distinguishable. I believe that universal definition must be accepted. In this paper for the general case of $n$-neutrino mixing we propose convenient expression for neutrino transition probability in vacuum in which the unitarity of the mixing matrix is fully employed and freedom of the common phase is used. As a result only independent quantities (including mass-squared differences) enter into expression for the transition probability. On the basis of the proposed expression I discuss the problem of the atmospheric neutrino mass-squared difference and comment a possibility to reveal the character of the neutrino mass spectrum in future reactor neutrino experiments. I thank A. Olshevskiy and C. Giunti for useful discussions. [100]{} Y. Fukuda [et al.]{}, Super-Kamiokande Collaboration , Phys. Rev.Lett. [81]{}(1998) 1562, arXiv: hep-ex/9807003. Q. R. Ahmad [et al.]{}, SNO Collaboration , Phys. Rev. Lett. [89]{} (2002) 011301, arXiv: nucl-ex/0204008. T. Araki [et al.]{}, KamLAND Collaboration, Phys. Rev. Lett. [94]{} (2005) 081801, arXiv:hep-ex/0406035. B. T. Cleveland [et al.]{}, Homestake Collaboration , Astrophys. J. [496]{} (1998) 505. M. Altmann [et al.]{}, GNO Collaboration , Phys. Lett. [B616]{} (2005) 174, arXiv: hep-ex/0504037. J. N. Abdurashitov [et al.]{}, SAGE Collaboration, J. Exp. Theor.Phys. [95]{}(2002) 181, arXiv: astro-ph/0204245. S. Weinberg S. Phys. Rev. Lett. [43]{} (1979) 1566. B. Pontecorvo, Sov. Phys. JETP [6]{} (1957) 429, \[Zh. Eksp. Teor. Fiz. 33, 549 (1957)\]. B. Pontecorvo, Sov. Phys. JETP [7]{} (1958) 172, \[Zh. Eksp. Teor. Fiz. 34, 247 (1958)\]. Z. Maki, M. Nakagawa and S. Sakata, Prog. Theor. Phys. [28]{} (1962) 870. M. Apollonio [et al.]{} \[CHOOZ Collaboration\], Phys. Lett. B [466]{} (1999) 415, hep-ex/9907037. S. M. Bilenky, C. Giunti and W. Grimus, Prog. Part. Nucl. Phys. [43]{} (1999) 1, hep-ph/9812360. F. An [et al.]{}, Daya Bay Collaboration , Phys.Rev.Lett. [112]{} (2014) 061801, arXiv:1310.6732 \[hep-ex\]. S.-B. Kim [et al.]{}, RENO Collaboration , Phys. Rev. Lett. [108]{} (2012) 191802, arXiv:1204.0626 \[hep-ex\]. Y. Abe [et al.]{}, Double Chooz Collaboration , Phys.Lett. [B723]{}, (2013) 66, arXiv:1301.2948 \[hep-ex\]. Y. F. Li, Int. J. Mod. Phys. Conf. Ser. [31]{} (2014) 1460300, arXiv:1402.6143 \[physics.ins-det\]. S. B. Kim, arXiv:1412.2199 \[hep-ex\]. S. M. Bilenky and S. T. Petcov, Rev. Mod. Phys. [59]{} (1987) 671 \[Erratum-ibid. [61]{} (1989) 169\] \[Erratum-ibid. [60]{} (1988) 575\]. S. M. Bilenky and C. Giunti, Int. J. Mod. Phys. A [16]{} (2001) 3931, hep-ph/0102320. M. C. Gonzalez-Garcia and M. Maltoni, Phys. Rept. [460]{} (2008) 1, arXiv:0704.1800 \[hep-ph\]. C. Giunti and C. W. Kim, Fundamentals of Neutrino Physics and Astrophysics, Oxford, UK: Univ. Pr. (2007) 710 p S. Bilenky, Introduction to the physics of massive and mixed neutrinos, Lect. Notes Phys. [817]{} (2010) 1. S. M. Bilenky, arXiv:1208.2497 \[hep-ph\]. F. Capozzi, E. Lisi and A. Marrone, Phys. Rev. D [89]{} (2014) 1, 013001, arXiv:1309.1638 \[hep-ph\]. M. C. Gonzalez-Garcia, M. Maltoni and T. Schwetz, JHEP [1411]{} (2014) 052, arXiv:1409.5439 \[hep-ph\]. K. Abe [et al.]{},T2K Collaboration, Phys. Rev. Lett. [112]{} (2014) 18, 181801, arXiv:1403.1532 \[hep-ex\]. P. Adamson [et al.]{}, MINOS Collaboration, Phys. Rev. Lett. [112]{} (2014) 191801, arXiv:1403.0867 \[hep-ex\]. S. M. Bilenky and F. Niedermayer, Sov. J. Nucl. Phys. [34]{} (1981) 606 \[Yad. Fiz. [34]{} (1981) 1091\]. C. Jarlskog, Phys. Rev. Lett. [55]{} (1985) 1039. S. M. Bilenky, D. Nicolo and S. T. Petcov, Phys. Lett. B [538]{} (2002) 77, hep-ph/0112216. M. He, arXiv:1310.7343 \[physics.ins-det\]. S. T. Petcov and M. Piai, Phys. Lett. B [533]{} (2002) 94, hep-ph/0112074. S. Choubey, S. T. Petcov and M. Piai, Phys. Rev. D [68]{} (2003) 113006, hep-ph/0306017. Y. F. Li, J. Cao, Y. Wang and L. Zhan, Phys. Rev. D [88]{} (2013) 013008, arXiv:1303.6733 \[hep-ex\]. A. A. Aguilar-Arevalo [et al.]{}, LSND Collaboration ( A. Aguilar-Arevalo [et al.]{}), Phys. Rev. D [64]{} (2001) 112007, hep-ex/0104049. A. A. Aguilar-Arevalo et al., MiniBooNE Collaboration , Phys.Rev.Lett. 110* (2013) 161801, arXiv:1303.2588 \[hep-ex\]. J. M. Conrad, W. C. Louis, M. H. Shaevitz, Ann.Rev.Nucl.Part.Sci. 63 (2013) 45, arXiv:1306.6494 \[hep-ex\]. G. Mention, M. Fechner, T. Lasserre, T. A. Mueller, D. Lhuillier, M. Cribier and A. Letourneau, Phys. Rev. D [83]{} (2011) 073006, arXiv:1101.2755 \[hep-ex\]. Th. A. Mueller et al., Phys. Rev. C83 (2011) 054615, arXiv:1101.2663 \[hep-ex\]. P. Huber, Phys. Rev. C84 (2011) 024617, arXiv:1106.0687 \[hep-ph\]. F. Kaether, W. Hampel, G. Heusser, J. Kiko and T. Kirsten, Phys. Lett. B [685]{} (2010) 47, arXiv:1001.2731 \[hep-ex\]. J. N. Abdurashitov [et al.]{}, SAGE Collaboration , Phys. Rev. C [80]{} (2009) 015807, arXiv:0901.2200 \[nucl-ex\]. C. Giunti, M. Laveder, Y. F. Li and H. W. Long, Phys. Rev. D [88]{} (2013) 073008, arXiv:1308.5288 \[hep-ph\]. J. Kopp, P. A. N. Machado, M. Maltoni and T. Schwetz, JHEP [1305]{} (2013) 050, arXiv:1303.3011 \[hep-ph\]. [^1]: Usually neutrinos with small mass-squared difference are called $\nu_{1}$ and $\nu_{2}$. It is assumed also that $m_{2}>m_{1}$, i.e. that $\Delta m^{2}_{12}>0$. [^2]: Notice that the first option corresponds to extraction of the phase connected with the intermediate neutrino mass in the expression (\[Genexp1\]) and second option corresponds to extraction of the phase connected with last mass.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Great attention is given to the first star formation and the epoch of reionization as main targets of planned large radio interferometries (e.g. Square Kilometre Array). Recently, it is claimed that the supersonic relative velocity between baryons and cold dark matter can suppress the abundance of first stars and impact the cosmological reionization process. Therefore, in order to compare observed results with theoretical predictions it is important to examine the effect of the supersonic relative motion on the small-scale structure formation. In this paper, we investigate this effect on the nonlinear structure formation in the context of the spherical collapse model in order to understand the fundamental physics in a simple configuration. We show the evolution of the dark matter sphere with the relative velocity by both using N-body simulations and numerically calculating the equation of motion for the dark matter mass shell. The effects of the relative motion in the spherical collapse model appear as the delay of the collapse time of dark matter halos and the decrease of the baryon mass fraction within the dark matter sphere. Based on these results, we provide the fitting formula of the critical density contrast for collapses with the relative motion effect and calculate the mass function of dark matter halos in the Press-Schechter formalism. As a result, the relative velocity decreases the abundance of dark matter halos whose mass is smaller than $10^8~M_\odot/h$.' author: - Shinsuke Asaba - Kiyotomo Ichiki - Hiroyuki Tashiro bibliography: - 'rvsc\_v2.bib' title: \| Effect of supersonic relative motion between baryons\ and dark matter on collapsed objects --- INTRODUCTION {#A} ============ The standard cosmological model, called the $\Lambda$CDM model, composed with two relativistic species (photons and neutrinos), two nonrelativistic matters (baryons and dark matter), and the energy having negative pressure (dark energy) with a nearly scale-invariant spectrum of curvature perturbations, has achieved great success in explaining large-scale cosmological observations, e.g., large-scale structure formation [@2014arXiv1411.1074A] and cosmic microwave background [@2015arXiv150201582P]. The theoretical description of small-scale structure formation is, however, still debatable. Understanding of small-scale structure formation at high redshifts is essential to study first stars and the epoch of reionization (EoR). One expects redshifted 21 cm lines from the hyperfine structure of hydrogen atoms as the powerful probe for the EoR and first stars [@2006PhR...433..181F; @2012RPPh...75h6901P] and the matter density underlying HI distribution constrains extended parameters of the $\Lambda$CDM model [@2013PhLB..718.1186O; @2014JCAP...09..014K; @2014PhRvD..90h3003S]. Currently, to probe such small-scale structure formation, there are many planed observations including Murchison Widefield Array [@MWA] and Square Kilometre Array (SKA) [@SKA]. Therefore, nowadays, the detailed studies on small-scale structure formation at high redshifts attract a lot of attention. Recently, Ref. [@2010PhRvD..82h3520T] reports the importance of the supersonic relative motion between dark matter and baryons on small-scale structure formation related to the EoR. This supersonic relative motion is originated from the difference in the motions between baryons and dark matter before recombination. Baryons before recombination are tightly coupled with photons by Thomson scattering. As a result, baryons and photons act as one fluid with the sound speed $\sim c/\sqrt3$ and have the velocity field associated with the acoustic oscillation. On the other hand, dark matter does not suffer from Thomson scattering and dark matter density fluctuations can grow gravitationally. Therefore, the relative motion between baryons and dark matter is induced. After recombination, baryons are fully decoupled with photons and the sound speed of baryons quickly drops to $\sim 6~{\rm km/s}$. Since the root mean square of the relative velocity reaches $\sim {30~\rm km/s}$ at that time, the relative velocity is about five times larger than the sound speed of baryons. Because the relative motion is highly supersonic, the effect on the structure formation could be significant. In particular, the abundance of small dark matter halos ($M{\lower.5ex\hbox{$\; \buildrel < \over \sim \;$}}10^7~M_\odot)$ is highly suppressed due to the supersonic relative motion. This effect has been intensively studied by many authors with N-body/smoothed particle hydrodynamics (SPH) simulations [@2011ApJ...730L...1S; @2012ApJ...747..128N; @2013ApJ...763...27N]. Therefore, according to the effect on the structure formation, the scenario of the cosmic reionization and the prediction of the 21 cm line signals from the EoR could be modified from those predicted in the conventional cosmological model. The recent relevant studies are reviewed in Ref. [@2014IJMPD..2330017F]. So far the effects of the relative motion on the structure formation have been studied in numerical simulations mainly, because these effects are complicated. However, a study with the analytical model is useful to obtain some insights into physics involved in complicated phenomena. Additionally the analysis of observation data with numerical simulations generally takes enormous time and, sometimes, it seems unrealistic. Therefore, modeling in a form which is easy to handle in analytic studies is highly required. In the study on the structure formation, the halo mass function is one of the interesting quantities. In particular, the Press-Schechter formalism with the sphere collapse model provides the mass function in the analytical form which relatively agrees well with the results of N-body simulations. In this paper, we revisit the effect of the supersonic relative motion on small-scale structure formation in the context of the spherical collapse model by both N-body simulations and a semianalytical way. The effect of the relative motion on the spherical collapse can be represented as the modification of the critical density contrast. We propose a fitting formula of the critical density contrast, as a function of the amplitude of the relative motion, the halo mass and the initial density fluctuation within dark matter halos. We also apply this fitting formula to evaluate the mass function of small dark matter halos based on the Press-Schechter formalism. This paper is organized as follows. In Sec. \[B\] we review the effect of supersonic relative motion on the perturbation theory by taking into account the background velocity of baryons and construct the spherical collapse model with two components. In Sec. \[C\] we describe the setup of our N-body simulation, and we show the results of the N-body simulations and check the reproducibility of the spherical collapse in Sec. \[Ca\]. Moreover we present the change of the collapse time by supersonic relative motion and the validity of the semianalytical model introduced in Sec. \[B\]. Section \[D\] is devoted to the discussion of the relative motion effect on dark matter halos. We discuss the modification of the baryon fraction in a dark matter halo by the relative motion, and, providing a fitting formula of the modified collapse time (i.e. the critical density contrast for the collapse). We show the suppression of the dark matter halo abundance around the EoR as an application of our results. Finally we summarize this paper in Sec. \[E\]. ANALYTICAL FORMALISM {#B} ==================== In this section, we show the effect of the supersonic relative motion between baryons and dark matter on the linear perturbation theory, and evaluate this effect on the nonlinear growth by adopting the spherical collapse model. Perturbation theory {#Ba} ------------------- Since the supersonic relative motion between baryons and dark matter has been analytically studied in the moving-background perturbation theory (MBPT) [@2010PhRvD..82h3520T], we first make a brief review of the MBPT. The MBPT introduces the background peculiar velocity which corresponds to the relative velocity between baryons and dark matter, i.e. $\vec{v}^{~\rm bg}=\vec{v}_{bc}$. According to the energy momentum conservation equation with homogeneous background densities of baryons and dark matter, the evolution of $\vec{v}_{bc}$ is in reverse proportion to the scale factor due to the cosmic expansion. In the MBPT, the first order energy momentum conservation equations after recombination are given by $$\begin{aligned} \frac{d\delta_c}{dt}=&-\theta_c,\notag\\ \frac{d\theta_c}{dt}=&-\frac{3H^2}{2}(\Omega_c\delta_c+\Omega_b\delta_b)-2H\theta_c,\notag\\ \frac{d\delta_b}{dt}=&-\frac{i}{a}\vec{v}_{bc}\cdot\vec{k}\delta_b-\theta_b,\notag\\ \frac{d\theta_b}{dt}=&-\frac{i}{a}\vec{v}_{bc}\cdot\vec{k}\theta_b-\frac{3H^2}{2}(\Omega_c\delta_c+\Omega_b\delta_b)-2H\theta_b+\frac{c_s^2k^2}{a^2}\delta_b,\label{eqpt}\end{aligned}$$ where $c_s$ is the sound velocity of the baryon fluid, the subscripts $c$ and $b$ denote cold dark matter and baryons respectively and $\theta=ia^{-1}\nabla\cdot\vec{v}$ represents the divergence of the peculiar velocity. In Eq. (\[eqpt\]) we take the frame where the background velocity of cold dark matter is absent. In other words, $\vec{v}^{\rm bg}_{b}=\vec{v}_{bc}$ and $\vec{v}^{\rm bg}_{c}=0$. For simplicity, we ignore perturbations of the sound velocity although they might affect the growth of the density fluctuation on small scales [@2005MNRAS.362.1047N; @2011MNRAS.418..906T; @2011MNRAS.416..232N]. Equation (\[eqpt\]) can be rewritten to the second order differential equations of the density fluctuations as $$\begin{aligned} \frac{d^2\delta_c}{dt^2}=&-2H\frac{d\delta_c}{dt}+\frac{3H^2}{2}(\Omega_c\delta_c+\Omega_b\delta_b),\notag\\ \frac{d^2\delta_b}{dt^2}=&-\left(2H+2i\mu v_{bc}\frac{k}{a}\right)\frac{d\delta_b}{dt}+\frac{3H^2}{2}(\Omega_c\delta_c+\Omega_b\delta_b)\notag\\ &-\left(c_s^2+\mu^2v_{bc}^2\right)\frac{k^2}{a^2}\delta_b,\label{eqdl}\end{aligned}$$ where $\mu=\vec{v}_{bc}\cdot \vec{k}/\|\vec{v}_{bc}\|\|\vec{k}\|$. Equation (\[eqdl\]) tells us that the relative motion prevents the growth of density fluctuations on small scales in the same way of the fluid pressure in the discussion of the Jeans instability. On large scales where the relative motion does not have the preferred direction, while the odd term of $\mu$ in the last equation of Eq. (\[eqdl\]) vanishes by averaging over all random directions of the relative motion, the third term of the right-hand side is enhanced due to the existence of the term with $v_{bc}^2$. As a result, the effective Jeans scale (the suppression scale) of Eq. (\[eqdl\]) becomes large due to the existence of the relative velocity. Since the relative velocity after recombination is roughly $\langle v_{bc}^2\rangle^{1/2}\sim 5c_s$, the suppression scale for the relative motion is $k_{bc}=aH/\langle v_{bc}^2\rangle^{1/2}\sim 40~h{\rm Mpc}^{-1}$. The corresponding mass scale for the suppression is $M_{bc}\sim 10^7~M_\odot/h$. However, when we consider a local sufficiently small patch, the odd term of $\mu$ cannot vanish in the patch. Instead, the relative motion in this patch can be assumed to be a homogeneous flow with one direction. In this case, when the relative velocity is larger than the Hubble flow, $\mu v_{bc}>k/aH$, the density fluctuations inside the patch start to grow exponentially due to the relative motion flow as shown in Eq. (\[eqdl\]). The perturbation theory is not valid in this case, and we need to consider the effect of the relative velocity on the nonlinear growth, e.g., in the spherical collapse model. Spherical collapse model {#Bb} ------------------------ The spherical collapse model is a simple analytical model to investigate the nonlinear evolution of an overdensity region. In this model, the evolution of the overdensity region is described as the motion of the constant density spheres. Let us consider the collapse of a mass shell inside which the mass is $M=4\pi x_i^3\bar{\rho}_i(1+\delta_i)/3$, where $x_i$ is the initial radius and $\delta_i$ is the initial density contrast within the sphere with the radius $x_i$ (hereafter the subscript $i$ represents the initial time value). The equation of motion (EoM) for the proper radius $x$ of a shell is written as $$\begin{aligned} \frac{d^2 x}{dt^2}=-\frac{GM}{x^2}. \label{eq1}\end{aligned}$$ Equation (\[eq1\]) can be solved analytically, and the solution is given by $$\begin{aligned} \tilde{t}=\frac{t}{t_i}&=\frac{3}{4\sqrt{1+\delta_i}}\left[1-\frac{(v_i/H_ix_i)^2}{1+\delta_i}\right]^{-3/2}(\theta-\sin\theta),\notag\\ \tilde{x}=\frac{x}{x_i}&=\frac{1}{2}\left[1-\frac{(v_i/H_ix_i)^2}{1+\delta_i}\right]^{-1}(1-\cos\theta), \label{eq3}\end{aligned}$$ where $t_i$ is the initial time and $v_i$ is the initial velocity. The solution of Eq. (\[eq3\]) depends only on $\delta_i$ and $v_i$. In order to keep the constant mass $dM/dt=0$, we give the initial velocity of the shells $$\begin{aligned} v_i=H_ix_i\left[1-\frac{\delta_i}{3(1+\delta_i)}\right], \label{eq4}\end{aligned}$$ where we assume the matter dominated era, $t\propto a^{3/2}$. The first term represents the Hubble flow and the second term corresponds to the peculiar velocity. According to Eq. (\[eq4\]), the solution, Eq. (\[eq3\]) depends on only the initial density fluctuation, $\delta_i$. Furthermore, we can obtain the critical density contrast that is the density contrast at the collapse time $\theta=2\pi$ in the linear perturbation theory, $$\begin{aligned} \delta_{\rm crit}=\frac{a(\theta=2\pi)}{a_i}\delta_i=\frac{3}{5}\left(\frac{3}{2}\pi\right)^{2/3}, \label{eq5}\end{aligned}$$ where we use the fact that the growth factor of the matter density perturbation is proportional to the scale factor in the matter dominated era. Next we consider the effect of the homogeneous supersonic relative motion on the nonlinear evolution in the spherical collapse model. A simple extension is to introduce the two kinds of mass shells for dark matter and baryons. Taking into account the supersonic relative motion, the baryon mass shells have the initial bulk velocity, because we take the frame where baryons have the homogeneous relative flow to dark matter. As a result, the collapsing of the baryon mass shells is not spherical and these mass shells are collapsing to the different position from the dark matter shells. However, the collapse of dark matter precedes the one of the baryons in a halo formation and we are interested in a baryon fraction within a collapsed dark matter halo. Therefore, we focus on the dark matter mass shell and, instead of following the evolution of the baryon mass shells, we introduce the baryon mass within the dark matter mass shell, $M_b$, and rewrite Eq. (\[eq1\]) to $$\begin{aligned} \frac{d^2 x_c}{dt^2}&=-\frac{G(M_{c,i}+M_b)}{x_c^2},\notag\\ M_{c,i}&=\frac{4\pi}{3} \bar{\rho}_{c,i}x_{c,i}^3(1+\delta_{c,i}),\notag\\ M_b&=\frac{4\pi}{3}\bar{\rho}_bx_c^3(1+\delta_b),\label{rveq}\end{aligned}$$ where $M_{c,i}$ is the mass within the shell with the initial radius at $x_{c,i}$, $x_c$ is the radius of the mass shell with $M_{c,i}$ at each time, and $M_b$ is the baryon mass in the dark matter shell at the radius $x_c$. As a shell collapses, the baryon mass $M_b$ inside the shell increases with the growth of $\delta_b$, namely the baryon collapsing. Although the density fluctuation of baryons without the relative motion catches up soon with that of dark matter, the relative motion prevents this process. Therefore, the effect of the relative motion is included through the evolution of $M_b$. However it is difficult to evaluate analytically $M_b$ with the relative motion. Thus in order to compute Eq. (\[rveq\]), we adopt $M_b$ obtained from the N-body simulation in the following section. We also compare the result based on the spherical collapse model with that from the full N-body simulations in the later section. N-BODY SIMULATION {#C} ================= Besides the analytical way mentioned in the previous section, we evaluate the effect of the supersonic motion between dark matter and baryons on the structure formation at high redshifts by using N-body simulations. In this section, we describe the setup of our N-body simulations. We perform N-body simulations with the public code Gadget-2 [@2005MNRAS.364.1105S]. In all N-body simulations, the cosmological parameters are set to $(\Omega_m,\ \Omega_\Lambda\ h)=(0.31,\ 0.69,\ 0.68)$ with $\Omega_b/\Omega_m\sim1/6$. The effect of the supersonic relative motion on the structure formation works after the decoupling between photons and baryons. Therefore, the initial redshift for the simulations is $z_i=1000$. Note that our results almost do not depend on the cosmological parameters because we are interested in the structure formation in the matter dominated era. For the initial distribution of the particles, we consider the spherical top-hat overdensity region in the isolated system. The simulation box has the uniform distribution of the particles with a uniform overdensity sphere. We set the box size to $L_{\rm Box}=200~{\rm kpc}/h$ and the radius of the overdensity sphere to $r_i=50~{\rm kpc}/h$. Note that we denote hereafter $x$ as the proper distance and $r$ as the comoving distance. In the box, the number of the uniform particles is $3\times 10^6$ and the initial density contrast of dark matter in the overdense sphere is $\delta_{c,i}=0.033$. Thus the mass of particles is $2\times 10^2~M_\odot/h$ and the mass of dark matter within the initial overdensity sphere is given by $M_c \sim 4\times 10^7~M_\odot/h$. Moreover we set the softening parameter to $\epsilon=0.1~{\rm kpc}/h$. We confirm that changes in these parameters do not affect our result qualitatively. According to the cosmological perturbation theory, the amplitude of baryon density fluctuations is 1% of that of dark matter density fluctuations with $k=100~{\rm Mpc^{-1}}$ at $z\sim 1000$. The initial fluctuations of baryons are negligible compared with those of dark matter at $z_i=1000$. Therefore, we assume that $\Omega_b/\Omega_m~(\sim1/6)$ of the uniform distributed particles is composed of baryons and there is no baryon fluctuation in the overdensity sphere. Figure \[inipos\] shows the initial configuration of the particles. The red dots represent the particle uniformly distributed in the box and the green dots are for the particles included in the overdense sphere. Figure \[inidel\] shows the initial density contrast of dark matter particles as a function of the radius from the center. In this figure, the red points indicate the values averaged over five realizations of our simulations and the error bars represent the shot noise caused by the finite particle number. The black line is the analytical prediction from our initial condition. This figure tells us that the density is constant in the top-hat sphere. Therefore we can convert the initial position of mass shell to the mass contained within each shell by using the relation $M_{c,i}=4\pi\bar{\rho}(1+\delta_i)x_i^3/3$. We set the initial velocity of dark matter given by only the second term of Eq. (\[eq4\]), because N-body simulations are performed in the comoving coordinate. In order to take into account the supersonic relative motion, we give the additional velocity to all baryons. The correlation of the supersonic relative velocity has the significant value on larger scale than scales of our interest that are smaller than Mpc. Therefore, we assume that all baryons in the simulations have the constant supersonic relative velocity $v_{bc}$ in one direction. In other words, in the simulation, the additional initial velocity for baryons is represented as $\vec{v}_{b,i}=(v_{bc},0,0)$. All terms related to the relative velocity in Eq. (\[eqdl\]) are proportional to $k v_{bc}$. This fact suggests that the effect of relative motion on the spherical collapse of dark matter halos also depends on a factor $v_{bc}k\propto v_{bc}/M_c^{1/3}$. Thus, instead of changing both the dark matter halo mass $M_c$ and the relative velocity $v_{\rm bc}$, we perform numerical simulations for different relative velocities ($5~{\rm km}$, $15~{\rm km}$, $30~{\rm km}$, $50~{\rm km}$, $100~{\rm km}$, $150~{\rm km}$, $200~{\rm km}$, $300~{\rm km}$, $500~{\rm km}$) with fixing the dark matter halo mass $M_c\sim 4\times 10^7~M_\odot/h$, in order to evaluate the dependence of the effect of the relative velocity on $M_c$ and $v_{bc}$. In the simulations, we use the periodic boundary condition. Therefore, when the relative velocity is higher than $200~{\rm km/s}$, baryon particles leaving the simulation box along the direction of the relative velocity reenter the box from the opposite direction due to the boundary condition. In this case, the distribution of reentering baryons is no longer homogeneous in the perpendicular direction to the relative velocity and, resultantly, the dependence on the boundary condition arises in the results. In order to remove this dependence, we make the perpendicular positions of baryons random when the baryon particles reenter the simulation box. ![The initial configuration of the particles. The green particles are contained within the top-hat sphere and the red points are otherwise.[]{data-label="inipos"}](inipos.eps){width=".36\textwidth"} ![The initial density contrast distribution of dark matter. The red points are the values averaged five realizations and the error bars show the shot noise caused by the number of particles contained within mass shells. The black line is the analytical prediction.[]{data-label="inidel"}](icmp.eps){width=".45\textwidth"} RESULT {#Ca} ====== In this section, we present the results of our N-body simulations. First we show the result of the reference model that is the case without the relative velocity. ![The evolutions of radii of mass shells containing $0.6M_c$ (red), $0.8M_c$ (green), and $M_c$ (blue). The solid black line is the analytical solution of spherical collapse model with $\delta_{m,i}=0.028$. The dashed black line is the analytical solution corrected in consideration of the periodic boundary condition with $\tilde{B}=4$.[]{data-label="sc"}](nsc.eps){width=".45\textwidth"} Figure \[sc\] shows the time evolutions of the radii of mass shells from the center. In our simulation, we determined the center of the collapsing shells by using the mean position of the particles contained initially within the top-hat overdensity sphere at each time step. In this figure, the red line represents the mass shell containing $0.6 M_c\ (r_i=42~{\rm kpc/h})$, and the green and blue lines are for that containing $0.8 M_c\ (r_i=46~{\rm kpc/h})$ and $M_c\ (r_i=50~{\rm kpc/h})$, respectively. Additionally the black line corresponds to the analytical solution of the spherical collapse model, Eq. (\[eq3\]), with $\delta_{m,i}=(\bar{\rho}_{c,i}\delta_{c,i}+\bar{\rho}_{b,i}\delta_{b,i})/(\bar{\rho}_{c,i}+\bar{\rho}_{b,i})=0.028$ and the velocity $v_i$ given by Eq. (\[eq4\]). Note that the shell evolution of the spherical collapse model depends on $\delta_{m,i}$ only. Therefore, in our initial condition where the overdensity sphere has the homogeneous density profile, the spherical collapse model predicts that all mass shells inside the overdensity sphere trace the black dashed line and collapse at the same time, independently on the mass contained by the mass shells. The evolutions of radii of the mass shells from N-body simulations agree with the analytic solution of the spherical collapse model before the turnaround time when the radius reaches the maximum. However, after the turnaround time, the results from N-body simulations deviate from the analytic solution. One of the reasons for this deviation is that the particles cannot be concentrated on the infinitesimal point in N-body simulations. Therefore, the particles in the simulations begin to be relaxed with each other after the turnaround time and the collapse is prevented. Furthermore the shot noise induces the substructures inside the collapsing sphere and the ejection of the particles from the mass shells, and resultantly causes the dispersion of the collapse time as discussed in Ref. [@2015MNRAS.446.1335W]. These effects of relaxation and shot noise lead the delay of the collapse and cause the deviation from the analytical solution after the turnaround time. Figure \[sc\] also shows that the outer mass shells collapse later than the inner ones, although the theoretical spherical collapse model claims that all mass shells collapse at the same time. The reason is the effect of the periodic boundary condition. In order to evaluate this effect simply, we consider the motion of a particle along the $x$-axis direction with the periodic boundary condition. Since we should take into account the gravitational force from the overdensity region in the other boxes due to the boundary condition, the EoM, Eq. (\[eq1\]), along the $x$-axis is corrected to $$\begin{aligned} \frac{d^2 \tilde{x}}{d\tilde{t}^2}=-\frac{2(1+\delta_i)}{9\tilde{x}^2}+\sum_{n=1}^\infty\left[\frac{2\delta_i}{9(na\tilde{B}-\tilde{x})^2}-\frac{2\delta_i}{9(na\tilde{B}+\tilde{x})^2}\right] \label{bcsc},\end{aligned}$$ where $\tilde{B}=a_iL_{\rm Box}/x_i$. When the position of the shell is close to the center of the simulation box at the initial time, $\tilde{B}$ becomes small and vice versa. Namely, the smaller $\tilde{B}$ is, the more efficient the boundary effect is. In our calculation, the most outer shell that contains the mass $M_c$ inside has $\tilde{B}=4$. The evolution for $\tilde{B}=4$ is plotted in the black dashed line in Fig. \[sc\]. ![The turnaround time (upper panel) and the reference collapse time (lower panel). The red points are the results of N-body simulations with the standard error measured five realizations, and the black lines are the analytical predictions. In the upper panel the shaded region shows the error caused from shot noise shown in Fig. \[inidel\]. The dashed line shows the prediction from the solutions of Eq. (\[bcsc\]) with $\tilde{B}$ converted from $r_i$.[]{data-label="tac"}](nrea.eps){width=".45\textwidth"} Figure \[tac\] shows the turnaround time (upper panel) and the reference collapse time (lower panel) in the reference model as functions of the initial radius of the mass shell. Here, the turnaround time and the reference collapse time are defined as the times when the radius of the mass shell becomes maximum and minimum, respectively. In this figure, the red points with the error bars represent the averages and the standard errors obtained from the five realizations of N-body simulations, and the black line corresponds to the theoretical predictions in the spherical collapse model. Moreover the dashed line shows the turnaround time obtained from Eq. (\[bcsc\]) which includes the effect of the boundary condition. One can find that the turnaround time is consistent with the theoretical prediction within $r_i=40~{\rm kpc/}h$. The one of the reasons why the outer mass shells turn around later is the effect of the periodic boundary condition discussed above. In the case where we take the boundary condition into account, the turnaround time matches the theoretical prediction within $r_i=43~{\rm kpc}/h$ that corresponds with $M=0.65M_c$. The difference between the turnaround time estimated from the outer shells than $r_i=43~{\rm kpc}/h$ and theoretical one is due to the ejections of particles. As we have mentioned, the reference collapse time in N-body simulations delays for all $r_i$, compared with the theoretical prediction. Additionally, similarly to the turnaround time, the deviation becomes large as $r_i$ increases. ![The evolutions of mass shell containing $0.6 M_c$ with supersonic relative velocities $v_{bc}=30~{\rm km}$ (red), $v_{bc}=50~{\rm km}$ (green), $v_{bc}=100~{\rm km}$ (blue), $v_{bc}=150~{\rm km/s}$ (magenta) and reference model (black). Each dashed line is the solution of Eq. (\[rveq\]) with the baryon density fluctuation derived from N-body simulations.[]{data-label="rvsc"}](nrvsc.eps){width=".45\textwidth"} Next we show how the supersonic relative motion affects the collapse in N-body simulations. Performing five realizations for different $v_{bc}$, we obtain the typical evolution of mass shells by averaging each realization. Figure \[rvsc\] represents the evolutions of the mass shells containing $0.6M_c$ with four different supersonic relative velocities, $v_{bc}=30~{\rm km/s}$ (in red), $v_{bc}=50~{\rm km/s}$ (in green), $v_{bc}=100~{\rm km/s}$ (in blue) and $v_{bc}=150~{\rm km/s}$ (in magenta). We can convert the results for different velocities with a fixed mass into those for different masses with a fixed velocity through the dependence of the relative velocity effect on $v_{bc}/M_c^{1/3}$ as mentioned above. For comparison, the corresponding evolution of the mass shell with $0.6M_c$ in the reference model ($v_{bc}=0~{\rm km/s}$) is plotted as the black solid line. As the relative velocity becomes large, the start of the collapse delays and the maximum radius increases. Therefore, we can conclude that the supersonic relative motion prevents the collapse. ![The turnaround time (upper panel) and the reference collapse time (lower panel) as the function of the initial radius of the mass shell with the supersonic relative velocity $v_{bc}=30~{\rm km}$ (red), $v_{bc}=50~{\rm km}$ (green), $v_{bc}=100~{\rm km}$ (blue), $v_{bc}=150~{\rm km/s}$ (magenta) and the reference model (black). In the upper panel the dashed lines are the turnaround time estimated from the solution of the semianalytical model.[]{data-label="da"}](nda.eps){width=".45\textwidth"} Additionally we show the solutions of Eq. (\[rveq\]) as the dashed lines in Fig. \[rvsc\]. Solving Eq. (\[rveq\]) numerically, we use the baryon fluctuation $\delta_b$ obtained from the particle data of the N-body simulations within $r_i\leq 42~{\rm kpc/h}$ corresponding to $0.6M_c$. We find that the semianalytical model agrees with the results of N-body simulations before the turnaround time. To illustrate the delay of the collapse due to the supersonic relative motion we plot the turnaround time and the reference collapse time in Fig. \[da\] for the different supersonic relative velocities. In this figure, both the turnaround and the reference collapse time are represented as the functions of the initial radius of the mass shell. We show additionally the turnaround times obtained from the solutions of Eq. (\[rveq\]) as the dashed lines in the top panel of Fig. \[da\]. The evaluations of the turnaround time from the semianalytical solutions are consistent with the turnaround times from N-body simulations within $r_i\sim40~{\rm kpc}/h$ that is same as the reference case. In the following discussions, we use twice the turnaround time as the collapse time. In this case, the scale factor at the collapse time is given by $a_{\rm col}=\left(2\tilde{t}_{\rm ta}\right)^{2/3}a_i$. DISCUSSION {#D} ========== In this section, we show the baryon fraction within the dark matter overdensity sphere. We also discuss the delay of the collapse time and provide the fitting formula of the critical density contrast with the relative motion between dark matter and baryons. Furthermore we present the modification of the halo mass function by taking account of the relative motion. Baryon fraction {#Db} --------------- First we consider the baryon fraction which represents the mass ratio between baryons and total matter ($f_b=M_b/M_m$) within the dark matter over-density sphere. The baryon fraction is important not only to estimate the effect of the supersonic relative motion on the collapse of dark matter spheres, but also to discuss the first star formation or observables related with baryons. We calculate the baryon fraction by counting baryon particles within the dark matter collapsing over-density sphere. ![The baryon fractions within the dark matter over-density sphere whose initial radius are $50~{\rm kpc}/h$ (solid lines) and $42~{\rm kpc}/h$ (dashed lines). The arrows show the reference collapse times.[]{data-label="bf"}](nbf.eps){width=".45\textwidth"} Figure \[bf\] shows the time evolutions of the baryon fraction. As the relative velocity increases, the baryon fraction becomes smaller. In the case with the nonzero relative velocity, the baryon overdensity region is no longer spherically symmetric and the peak position of baryon density is different from that of dark matter, depending on the amplitude of the relative velocity. However, such asymmetry of the baryon distribution does not affect the dark matter collapse well. As shown in Fig. \[rvsc\], the dark matter collapse in the N-body simulations is consistent with the spherical collapse model until the turnaround time. Therefore, we can infer that, in spite of the asymmetric distribution for baryons, the dark matter collapse remains spherical. Around the reference collapse time of the dark matter shells, the baryon fraction estimated within $r_i=42~{\rm kpc}/h$ starts to oscillate. This is mainly due to the difference of the density peak positions between baryons and dark matter. The baryon overdensity region is attracted by that of dark matter gravitationally and oscillates around. As time goes, the difference of the peak positions will be relaxed and the peak position of baryons is expected to overlap that of dark matter. Note that this result is based on the spherical collapse model which is an ideal isolated system. However, the actual collapse happens with many surrounding effects as shown in cosmological simulations. Therefore, to evaluate the baryon fraction properly, these effects could be not negligible. Delay of the halo formation {#Dc} --------------------------- The supersonic relative motion delays the collapse time of dark matter halos as shown in Fig. \[da\]. In the spherical collapse model without the relative velocity, the collapse time is dependent on only the initial density fluctuation. However, in the case of the nonzero relative velocity, the collapse time depends on the halo mass $M_c$ and the amplitude of the relative velocity. Additionally, the effect of the relative motion cumulatively becomes large for the small initial density contrast, because the small initial density contrast takes longer time to the collapse. Therefore, the change of the collapse time with the relative velocity is represented as a function of $M_c$, $v_{bc}$ and $\delta_{c,i}$ We define the correction of the scale factor at the collapse time related with the modification of the critical density contrast, as $$\begin{aligned} \mathcal{A}(M_c,v_{bc},\delta_{c,i})&\equiv\frac{a_{\rm col}(M_c,v_{bc},\delta_{c,i})-a_0(\delta_{c,i})}{a_0(\delta_{c,i})},\label{eq:col}\end{aligned}$$ where $a_0(\delta_{c,i})$ is the scale factor at the collapse time without the relative velocity between baryons and dark matter. ![The relative time difference for the collapse as a function of relative velocity $v_{bc}$ with the halo mass fixed to $M_c\sim 4\times 10^7~M_\odot/h$ and three initial density fluctuations $\delta_{c,i}=0.016$ (black), $\delta_{c,i}=0.033$ (rad) and $\delta_{c,i}=0.066$ (blue) evaluated from turnaround time. The shaded regions show the standard error region from the N-body simulation. The solid lines are the fitting formula Eq. (\[ff\]). The dashed line are results from the N-body simulations with the usual periodic boundary condition. The upper horizontal axis shows the mass converted with $v_{bc}=30~{\rm km/s}$.[]{data-label="rda"}](nlda.eps){width=".45\textwidth"} Figure \[rda\] shows the relative difference $\mathcal {A}$ as a function of the relative velocity with a fixed mass $M_c\sim 4\times 10^7~M_\odot/h$. We plot $\cal A$ for different three initial density contrasts estimated from the N-body simulations with our boundary condition (shaded regions) or the usual periodic boundary condition (dashed lines). We find that the effect of the supersonic relative velocity is negligible for velocities smaller than 50 km/s. The relative velocity is small so that baryons are captured gravitationally by the dark matter halo and accrete to the halo. Therefore, we can roughly estimate the threshold velocity as the circular velocity of the dark matter halo at the initial redshift, $$\begin{aligned} v_{\rm cir}=\sqrt{\frac{GM_c}{x_i}}\simeq 57~\left(\frac{M_c}{3.85\times 10^7~M_\odot}\right)^{1/3}~{\rm km/s}.\end{aligned}$$ This criterion can be also obtained from the condition that the third term related to the relative velocity, $\mu^2 k^2$, dominates in the right-hand side in the second equation of Eq. (\[eqdl\]). Therefore, when the relative velocity is larger than the criterion velocity, the relative motion prevents the collapse by the third term in Eq. (\[eqdl\]). Similarly, when the relative velocity is larger than $v_{\rm cir}$, the effect of the relative motion on the structure formation arises. The upper horizontal axis represents the corresponding mass in the case of a fixed velocity $v_{bc} = 30~{\rm km/s}$, which is converted through the $v_{bc}/M_c^{1/3}$-dependence of the relative motion effect. Thus one can find that the supersonic relative motion does not affect the formation of dark matter halos with mass larger than $M_c{\lower.5ex\hbox{$\; \buildrel > \over \sim \;$}}10^7~M_\odot/h$ for $v_{bc}=30{\rm km/s}$ which corresponds to the effective Jeans scale discussed in Sec. \[Ba\]. In the N-body simulation results with the usual periodic boundary condition, the modification becomes independent of the amplitude of the relative velocity in the case with $v_{bc}{\lower.5ex\hbox{$\; \buildrel > \over \sim \;$}}200~{\rm km/s}$ shown in Fig. \[rda\]. However this independence is due to the artificial condition of the simulations. When $v_{bc}{\lower.5ex\hbox{$\; \buildrel > \over \sim \;$}}200~{\rm km/s}$, all baryons can travel a distance larger than the simulation box size $L_{\rm Box}$ until $\tilde{t}\sim100$ and are attracted gravitationally twice by the collapsing dark matter sphere. Therefore, the effect of the relative velocity seems to be saturated in this relative velocity region. On the other hand, the results of the N-body simulations with our boundary condition shown as the shaded region in Fig. \[rda\] are not saturated. In order to verify the validity of the result of our N-body simulations, we show the collapse time and the baryon fraction in the limit of the large homogeneous relative velocity. One can easily imagine that, when the relative velocity is enough high ($v_{bc}=500~{\rm km/s}$), which corresponds to very small dark matter halos ($M_c\sim10^4 ~M_\odot/h$) in the case of $v_{bc}=30~{\rm km/s}$, baryons do not collapse along the relative velocity direction. In this limit, we can ignore the gravitational force from the dark matter halo on the baryon motion along the relative velocity direction. In other words, the gravitational collapsing of baryons occurs perpendicular to the relative velocity direction and does not along the parallel direction. Therefore, to evaluate the evolution of the density fluctuations, it is useful to consider the motion of baryons with the relative velocity in the comoving cylindrical coordinate system whose axis is parallel to the relative velocity motion. In this case, the EoM of baryons particles with the initial velocity $(v_{b\parallel},v_{\perp})=(v_{bc},0)$ is given as $$\begin{aligned} \frac{d^2\vec{r}_b}{dt^2}=&-2H\vec{v}_b-\frac{G\delta M_c}{a^3r_b^3}\vec{r}_b,\notag\\ \delta M_c=&\frac{4}{3}\pi \bar{\rho}_c[r_{c,i}^3(1+\delta_{c,i})-r_c^3]\notag\\ &\times \max[1,(r_b/r_c)^3],\label{lim}\end{aligned}$$ where $r_b$ is the radial component of $\vec{r}_b$ from the center of the dark matter sphere and $r_c$ is the radius of the overdensity sphere of dark matter. Note that we ignore the gravitational force of the baryon fluctuation in Eq. (\[lim\]), because $\delta M_b\ll \delta M_c$. We solve Eqs. (\[rveq\]) and (\[lim\]) numerically with different initial positions $\vec{r}_{b,i}$ chosen randomly. We calculate the resultant baryon density contrast in a collapsing spherical shell of dark matter by taking the average of $\delta_b=(r_{bi\perp}/r_{b\perp})^{2}-1$ for baryons inside the shell of dark matter, because the collapse of baryons is cylindrical. ![The baryon fractions within the dark matter halo ($r_i=42~{\rm kpc}/h$) with $M_c\sim 4\times 10^7 M_\odot/h$ and $v_{bc}=500~{\rm km/s}$ and solution of Eqs. (\[rveq\]) and (\[lim\]) with or without the periodic boundary condition (red or blue line). The each color shaded region show the 1$\sigma$ dispersion of the baryon fractions estimated from five realizations of N-body simulations. The vertical dashed lines show the turnaround times from N-body simulations.[]{data-label="del"}](nlbf.eps){width=".45\textwidth"} Figure \[del\] shows the evolution of the baryon fractions with $M_c\sim 4\times 10^7 M_\odot/h$ and $v_{bc}=500~{\rm km/s}$. The red solid line represents the solution obtained with the periodic boundary condition. In order to take into account the periodic boundary condition, we solve Eqs. (\[rveq\]) and (\[lim\]) with the assumption that the position of baryons, $\vec{r}_b$, is limited within the box size and baryons return into the box from the opposite side when they exit from one side of the box. For comparison, we also solve the equations without the boundary condition and plot the solution in the blue solid line. Baryons with the periodic boundary condition feel gravitational force stronger than without the boundary condition. Therefore, the collapse is faster with the boundary condition than without the boundary condition. Moreover, in Fig. \[del\], we show the results from N-body simulations. The red shaded region represents the standard error region from the N-body simulations with the periodic boundary condition, while the blue shaded region gives the standard error region for the N-body simulations with our boundary condition which is the periodic boundary condition with the position shuffling of the baryon particles reentering into the box. In Fig. \[del\], N-body simulations with our boundary condition is consistent with the numerically solution without the periodic boundary condition, while N-body simulations with the periodic boundary condition agrees with the solution with the periodic boundary condition. We find that, in the both boundary condition cases, the differences between the numerical solutions and N-body simulations arise around $\tilde{t}\simeq 100$. This is because, as the collapse proceeds, the baryon density in the N-body simulations grows as the spherical collapse rather than the cylindrical one. Therefore, the baryon fraction is larger in N-body simulations than in the numerical calculations. ![The collapse time with $M_c\sim 4\times 10^7 M_\odot/h$ and $v_{bc}=500~{\rm km/s}$ estimated by solving Eqs. (\[rveq\]) and (\[lim\]) with boundary condition (red) and without boundary condition (blue). The black solid line is corresponded the solution of Eqs. (\[rveq\]) with $\delta_b=0$ at any time. Points represent estimations from N-body simulations with $v_{bc}=500~{\rm km/s}$. The upper axis presents the collapse time from Eq. (\[eq3\]) with the total matter density fluctuation $\delta_{m,i}=\bar{\rho}_c\delta_{c,i}/(\bar{\rho}_c+\bar{\rho}_b)$.[]{data-label="limda"}](nlimda.eps){width=".45\textwidth"} In addition, we plot $\mathcal{A}$ as functions of the initial density fluctuations $\delta_{c,i}$ with $M_c\sim 4\times 10^7 M_\odot/h$ and $v_{bc} =500~{\rm km/s}$ in Fig. \[limda\]. The red and blue solid lines are obtained from the numerical calculations with and without the periodic boundary condition, respectively. In the case with the periodic boundary condition, we overestimate the gravitational force to collapse as mentioned above and, therefore, the delay of the collapse is not larger than in the case without the boundary condition. For comparison, we plot the black solid line which represents the results with the assumption that baryons cannot collapse. The difference from the black solid line represents the contribution due the collapse of the baryon component. When the relative velocity is large enough, baryons cannot collapse to the dark matter mass shell. Accordingly, as the relative velocity becomes large, the blue solid line shifts to the black line. We also plot the results of N-body simulations with the periodic boundary condition and our boundary condition as red and blue points with the standard error bars in Fig. \[limda\], respectively. As shown in Fig. \[del\], the numerical calculation with the periodic boundary condition agrees with N-body simulations with the periodic boundary condition, while the numerical calculation without the boundary condition is consistent with N-body simulations with our boundary condition. We remind you that our boundary condition is introduced to remove the artificial distribution of the reentering baryon particles due to the periodic boundary condition in N-body simulations. We conclude that the saturation in $\mathcal{A}$ from N-body simulations with the periodic boundary condition is caused by this artifact. The results from N-body simulations with our boundary condition present realistic phenomena. Based on the results of our N-body simulations, we find the fitting formula of $\cal A$ represented as the solid lines in Fig. \[rda\]. The fitting formula is given by $$\begin{aligned} \mathcal{A}(M_c,v_{bc},\delta_{c,i})&=\mathcal{A}_{\delta_b=0}(\delta_{c,i})\frac{\mathcal{B}^\nu(M_c,v_{bc},\delta_{c,i})}{\mathcal{B}^\nu(M_c,v_{bc},\delta_{c,i})+1},\notag\\ \mathcal{B}(M_c,v_{bc},\delta_{c,i})&=\frac{v_{bc}}{v_{\rm norm}(\delta_{c,i})}\left(\frac{M_c}{3.85\times 10^7~M_\odot/h}\right)^{-1/3},\notag\\ v_{\rm norm}(\delta_{c,i})&=a_v - b_v\delta_{c,i},\label{ff}\end{aligned}$$ where $\mathcal{A}_{\delta_b=0}$ is the solution of Eq. (\[rveq\]) with $\delta_b=0$ shown in Fig. \[limda\], and $\nu$, $a_v$ and $b_v>0$ are the fitting parameters. The velocity $v_{\rm norm}$ is the critical velocity for the collapse of baryons along the direction of the relative velocity. When $v_{\rm bc} \gg v_{\rm norm}$, the relative velocity is much larger than $v_{\rm circ}$ even at the collapse time and baryons does not collapse along the direction of the relative velocity as mentioned above. Since the collapse time becomes long with decreasing $\delta_i$ and $v_{\rm bc}$ is inversely proportional to the scale factor, $v_{\rm norm}$ increases as $\delta_i$ decreases. Thus we conclude that the delay is controlled by two critical velocity $v_{\rm circ}$ and $v_{\rm norm}$. Nevertheless it can be calculated numerically, we use the approximated function of $\mathcal{A}_{\delta_b=0}$, $$\begin{aligned} \mathcal{A}_{\delta_b=0}(\delta_{c,i})=\frac{\delta_{c,i}^{-0.146}-1.06}{4.12\times \delta_{c,i}^{0.648}+1.93}.\end{aligned}$$ We estimate the parameters by fitting simultaneously Eq. (\[ff\]) to $\mathcal{A}$ with using three different initial density fluctuations. Furthermore we perform a Fisher analysis and obtain the parameters as $$\begin{aligned} \nu=2.02\pm0.07,\ a_v=205\pm16,\ b_v=877\pm253,\end{aligned}$$ where the standard errors are estimated after marginalizing over other parameters. Since we sample the data for three different initial conditions, the parameter $b_v$ has a large error. However, we find that the form in Eq. (\[ff\]) fits well with the N-body simulation results. Mass function ------------- The delay of the halo formation due to the relative velocity modifies the abundance of dark matter halos. In this section, we evaluate the modification based on the Press-Schechter formalism. In the Press-Schechter formalism, the delay of the collapse is represented as the increase of the critical density contrast. Using the relative time difference $\cal A$, we can write the modified critical density contrast during the matter dominated era as $$\begin{aligned} \tilde{\delta}_{\rm crit}(M_c,v_{bc},\delta_{c,i})&=\delta_{\rm crit}\left[1+\mathcal {A}(M_c,v_{bc},\delta_{c,i})\right].\label{eq:delta}\end{aligned}$$ The critical density contrast depends on the relative velocity in the region where the collapses happens. Therefore, the modified halo mass distribution can be written with the probability distribution function of the amplitude of the relative velocity at the initial time $f(v_{bc})$ as $$\begin{aligned} \tilde{n}(M_c,z)=&\int dv_{bc}~f(v_{bc})\sqrt{\frac{2}{\pi}}\frac{\bar{\rho}(z)}{M_c}\frac{\tilde{\delta}_{\rm crit}(M_c,v_{bc},\delta_{c,i})}{\sigma(M_c,z)}\notag\\ &\times\left[\frac{d\ln\tilde{\delta}_{\rm crit}(M_c,v_{bc},\delta_{c,i})}{dM_c}-\frac{d\ln\sigma(M_c,z)}{dM_c}\right]\notag\\ &\times\exp\left[-\frac{\tilde{\delta}_{\rm crit}^2(M_c,v_{bc},\delta_{c,i})}{2\sigma^2(M_c,z)}\right]. \label{ps}\end{aligned}$$ Note that, because of the existence of the relative motion, the redshift of the collapse time depends not only on the initial density fluctuation of cold dark matter, but also on the halo mass $M_c$ and the amplitude of relative velocity $v_{bc}$. Therefore, in Eq. (\[ps\]), the mass derivative term of the critical density additionally arises because the mass dependence of the critical density contrast affects the hierarchical structure formation. We assume that the probability distribution $f(v_{bc})$ follows the Maxwell-Boltzmann distribution because the each component of relative velocity is independent and obey the same Gaussian distribution whose mean value is zero and the dispersion is $\sigma_v$; $$\begin{aligned} f(v_{bc}) dv_{bc} =4\pi v_{bc}^2 \left(\frac{3}{2\pi\sigma_v^2}\right)^{3/2}\exp\left(-\frac{3v_{bc}^2}{2\sigma_v^2}\right) dv_{bc},\end{aligned}$$ where we use $\sigma_v=28.8~{\rm km/s}$ according to the latest cosmological parameters from PLANCK paper and we use CAMB[@2000ApJ...538..473L] to calculate $\sigma(M_c,z)$. ![The ratio of mass function of the dark matter halo between with and without relative motion at four redshifts $z=10$ (red), $z=15$ (green) $z=20$ (blue) and $z=30$ (magenta).[]{data-label="mf"}](nzps.eps){width=".45\textwidth"} Figure \[mf\] shows the ratio between the mass function with and without the relative velocity. Here we plot the ratio at four different redshifts, $z=$10, 15, 20 and 30. These lines are evaluated by using the fitting formula Eqs. (\[ff\]) for $\mathcal{A}$ in Eq. (\[eq:delta\]). The suppression of the mass function due to the relative motion is more significant for smaller masses region and at higher redshifts. At $z = 30$, although the modification $\mathcal{A}$ is very small around $10^7 M_\odot/h\lesssim M_c \lesssim 10^9~M_\odot/h$, the suppression of the mass function is not negligible. This is because such massive halos at high redshifts are rare objects which satisfy $\tilde{\delta}_{\rm crit}\gg \sigma(M_c)$ in the Press-Schechter formalism. Therefore, the effect of the modification $\mathcal{A}$ appears exponentially in the Press-Schechter formalism, even if $\mathcal{A}$ is small. The ratio of the mass functions reaches the minimum at $M_c\sim 10^5~M_\odot/h$. On smaller scales than $M_c\sim 10^5~M_\odot/h$, the suppression of the mass function decreases because the mass derivative term of the critical density in Eq. (\[ps\]) increases around $M_c\sim 10^5~M_\odot/h$ as shown in Fig. \[rda\]. Figure \[mf\] also shows that the mass function is suppressed even around the EoR ($z\sim 10$). This is because the redshift dependence of the derivative terms in Eq. (\[ps\]) is weak. Although the suppression scales are consistent, the suppression of mass function around EoR is stronger than in other previous works. Especially Ref. [@2012ApJ...747..128N] reported that the suppression disappears at $z \sim 10$ in their SPH simulations. Of course, unlike the SPH simulations, we only simulate the gravitational force, ignoring the baryon physics. However, it is worth mentioning reasons of the difference between our calculation of the mass function and previous work in term of the gravitational growth. The first reason for this difference is that we do not include the environmental effects of the structure formation, e.g., an accretion of other density peaks of baryons neighboring the dark matter halo. These effects increase the baryon fraction within a dark matter halo and promotes the halo formation. The second reason is that the relative motion produces halos derived from offsetting baryon peaks [@2014ApJ...791L...8N]. This effect increases simply the number of halos. It is difficult to include this effect in the Press-Schechter formalism. We will be able to estimate a number of halos originated from the baryon peaks by applying the our simulations or our semianalytical model of Eq. (\[rveq\]). Finally, the initial condition for matter density and velocity fields is still debatable in the numerical simulations with the relative velocity. In our simulation, we use the initial condition that leads to the maximum delay of the collapse time. Thus the halo mass function in Fig. \[mf\] is calculated on the basis of the optimistic case where the baryons can escape most efficient from the dark matter halo. SUMMARY {#E} ======= The relative motion between baryons and dark matter plays an important role, particularly, in small-scale structure formation at high redshifts. We have studied their effect on the dark mater halo formation. We have evaluated the delay of the dark matter halo collapse due to the supersonic relative motion by using the cosmological N-body simulation. We have found that the delay of the collapse becomes large for a dark matter halo with $M_c\sim4\times 10^7~M_\odot/h$, when the relative velocity is larger than $v_{\rm cir}=57~{\rm km/s}$. In other words, the delay of the collapse happens when the relative velocity is larger than the typical circular velocity of the dark matter halo. Moreover, we have shown that the supersonic relative motion delays the fall of baryons into the potential well of the dark matter halo in the context of the spherical collapse model. We have also evaluated the baryon fraction $M_b/M_m$ of the dark matter halos with the supersonic relative motion by the N-body simulation. The baryon fraction becomes smaller as the amplitude of the relative motion increases. We have pointed out that, when the relative velocity is large enough to escape from the potential of dark matter halos, baryons can collapse only along the perpendicular direction of the relative velocity, like the cylindrical collapse. Furthermore we show the delay of the collapse time for dark matter halo by the relative motion depends on the initial density fluctuation within dark matter spheres, which determines the collapse time of the dark matter halo without relative motion. The smaller initial density fluctuation lead the longer time during which the supersonic relative motions affect the halo collapse. In consequence, the effect of the relative motion is more efficient on dark matter halos formed at later time. Finally we have estimated the suppression of the abundance of dark matter halos by supersonic relative motion in the context of the spherical collapse model. In the Press-Schechter formalism, the delay of the collapse increases the critical density contrast for the collapse. We have found the fitting formula of the critical density contrast depending on the halo mass, the initial density fluctuations and, the relative velocity. Using the fitting formula, we have calculated the mass function of dark matter halos. The relative motion decreases the mass function with mass smaller than $10^{8}~M_\odot /h$ before EoR. In particular, the abundance of halos with $M_c=10^5~M_\odot/h$ is suppressed by 80% at $z=30$ and a half at $z=10$. The delay of the dark matter halo collapse and the decrease of the baryon fraction in dark matter halos due to the relative motion can give the effect on first star formation and the reionization history [@2011ApJ...736..147G; @2011MNRAS.412L..40M; @2012MNRAS.424.1335F]. Such effect could impact the cosmological signals of the EoR including the CMB polarization [@2012PhRvD..85d3523F], the redshifted 21 cm lines [@2011arXiv1110.4659B; @2012ApJ...760....3M] and these cross-correlation [@2008MNRAS.389..469T]. Moreover, the relative motion between dark matter and baryons influences the large scale structure, e.g., baryon acoustic oscillation [@2011JCAP...07..018Y; @2015MNRAS.448....9S; @2013PhRvD..88j3520Y], and there is the challenging work detecting this effect by using the results of galaxy distribution from two independent galaxy spectroscopic survey [@2015arXiv150603900B]. Based on the results of this paper, we will investigate the effect of the relative motion on the cosmological signals probed by ongoing or planned observations. This work was supported by JSPS KAKENHI Grant No. 26-2667 (S.A.), No. 24340048 (K.I.) and No. 15K17646 (H.T.). H.T. also acknowledges the support by MEXT’s Program for Leading Graduate Schools PhD professional, “Gateway to Success in Frontier Asia”.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A Multiplicative-Exponential Linear Logic (MELL) proof-structure can be expanded into a set of resource proof-structures: its Taylor expansion. We introduce a new criterion characterizing those sets of resource proof-structures that are part of the Taylor expansion of some MELL proof-structure, through a rewriting system acting both on resource and MELL proof-structures.' author: - Giulio Guerrieri - Luc Pellissier - 'Lorenzo [Tortora de Falco]{}' bibliography: - 'Biblio.bib' title: 'Glueability of resource proof-structures: inverting the Taylor expansion (long version)' ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent studies in image classification have demonstrated a variety of techniques for improving the performance of Convolutional Neural Networks (CNNs). However, attempts to combine existing techniques to create a practical model are still uncommon. In this study, we carry out extensive experiments to validate that carefully assembling these techniques and applying them to a basic CNN model in combination can improve the accuracy and robustness of the model while minimizing the loss of throughput. For example, our proposed ResNet-50 shows an improvement in top-1 accuracy from 76.3% to 82.78%, and an mCE improvement from 76.0% to 48.9%, on the ImageNet ILSVRC2012 validation set. With these improvements, inference throughput only decreases from 536 to 312. The resulting model significantly outperforms state-of-the-art models with similar accuracy in terms of mCE and inference throughput. To verify the performance improvement in transfer learning, fine grained classification and image retrieval tasks were tested on several open datasets and showed that the improvement to backbone network performance boosted transfer learning performance significantly. Our approach achieved 1st place in the iFood Competition Fine-Grained Visual Recognition at CVPR 2019 [^1], and the source code and trained models will be made publicly available [^2].' author: - \| Jungkyu Lee, Taeryun Won, Kiho Hong\ \ Clova Vision, NAVER Corp.\ [{jungkyu.lee, lory.tail, kiho.hong}@navercorp.com]{} bibliography: - 'egbib.bib' title: \| Compounding the Performance Improvements of Assembled Techniques\ in a Convolutional Neural Network --- [UTF8]{}[mj]{} Introduction {#sec:introduction} ============ Since the introduction of AlexNet [@krizhevsky2012imagenet], many studies have mainly focused on designing new network architectures for image classification to increase accuracy. For example, new architectures such as Inception [@szegedy2015going], ResNet [@he2016deep], DenseNet [@huang2017densely], NASNet [@zoph2018learning], MNASNet [@tan2019mnasnet] and EfficientNet [@tan2019efficientnet] have been proposed. Inception introduced new modules into the network with convolution layers of different kernel sizes. ResNet utilized the concept of skip connection, and DenseNet added dense feature connections to boost the performance of the model. In addition, in the area of AutoML, network design was automatically decided to create models such as NASNet and MNASNet. EfficientNet proposes an efficient network by balancing the resolution, height, and width of the network. The resulting performance of EfficientNet for ImageNet top-1 accuracy was greatly improved relative to AlexNet. Unlike these studies which focus on designing new network architecture, He [@he2019bag] propose different approaches to improve model performance. They noted that performance can be improved not only through changes in the model structure, but also through other aspects of network training such as data preprocessing, learning rate decay, and parameter initialization. They also demonstrate that these minor “tricks” play a major part in boosting model performance when applied in combination. As a result of using these tricks, ImageNet validation top-1 accuracy of ResNet-50 improved from 75.3% to 79.29%. This improvement is hugely significant, and shows as much performance improvement as improvements to network design. Thus, it is of critical importance to combine these existing techniques. Inspired by their works, we conducted a more extensive and systematic study of assembling several CNN-related techniques into a single network. When considering the many techniques that have been introduced, we first divided the techniques into two categories: network tweaks and regularization. Network tweaks are methods that modify the CNN architectures to be more efficient. For example, a representative work is SENet [@hu2018squeeze]. Regularization is a method that prevents overfitting by increasing the training data through data augmentation processes such as AutoAugment [@cubuk2018autoaugment] and Mixup [@zhang2017mixup], or by limiting the complexity of the CNN with processes such as Dropout [@srivastava2014dropout], and DropBlock [@dai2018batch]. Furthermore, we systematically analyze the process of assembling these two types of techniques through extensive experiments and demonstrate that our approach leads to significant performance improvements. In addition to top-1 accuracy, mCE and throughput were used as performance indicators for combining these various techniques. Hendrycks [@hendrycks2019benchmarking] proposed mCE (mean corruption error), which is a measure of network robustness against input image corruption. Moreover, we used throughput (images/sec) instead of the commonly used measurement of FLOPS (floating point operations per second) because we observed that FLOPS is not proportional to the inference speed of the actual GPU device. Detailed experiments on the discrepancy between the FLOPS and the throughput of GPU devices are included in the Appendix \[sec:appendix-flops\_and\_throughput\]. Our contributions can be summarized as follows: 1. By organizing the existing CNN-related techniques for image classification, we find techniques that can be assembled into a single CNN. We then demonstrate that our resulting model surpasses the state-of-the-art models with similar accuracy in terms of mCE and inference throughput (Table \[tab:summary-model\]). 2. We provide detailed experimental results for the process of assembling CNN techniques and release the code for accessibility and reproducibility. Preliminaries {#sec:basemodel} ============= Before introducing our approach, we describe default experimental settings and evaluation metrics used in Sections \[sec:assembling\] and \[sec:Experiment Results\]. Training Procedure ------------------ We use the official TensorFlow [@abadi2016tensorflow] ResNet [^3] as base code. We use the ImageNet ILSVRC-2012 [@russakovsky2015imagenet] dataset, which has 1.3M training images and 1,000 classes. All models were trained on a single machine with 8 Nvidia Tesla P40 GPUs compatible with the CUDA 10 platform and cuDNN 7.6. TensorFlow version 1.14.0 was used. The techniques proposed by He [@he2019bag] are basically applied to all our models described in Section \[sec:assembling\]. We briefly describe the default hyperparameters and training techniques as follows. - Preprocessing In the training phase, a rectangular region is randomly cropped using a randomly sampled aspect ratio from 3/4 to 4/3, and the fraction of cropped area over whole image is randomly chosen from 5% to 100%. Then, the cropped region is resized as a 224x224 square image flipped horizontally with a random probability of 0.5. During validation, we first resize the shorter size of each image to 256 pixels while the aspect ratio is maintained. Next, we center crop the image to the 224x224 size and normalize the RGB channels, identically to training. - Hyperparameter We use 1,024 batch sizes for training. In our study, this is close to the maximum size that can be received on a single machine with 8 P40 GPUs. The initial learning rate is 0.4 and weight decay is set to 0.0001. The default number of training epochs is 120, but some techniques require a different number of epochs, which is given explicitly when necessary. Stochastic gradient descent with momentum 0.9 is used as the optimizer. - Learning rate warmup If a large batch size is set, using a high learning rate may result in numerical instability. Goyal [@goyal2017accurate] proposes a warmup strategy that linearly increases the learning rate from 0 to the initial learning rate at warm-up periods set to first 5 epochs. - Zero $\boldsymbol{\gamma}$ We initialize $\gamma=0$ for all batch-norm layers that sit at the end of a residual block. Therefore, all the residual blocks return only their shortcut branch result in the early stages of training. It is easier to train by creating an effect that shrinks the entire layer at the initial stage. - [Mixed-precision floating point]{} We only use mixed-precision floating point in the training phase because mixed-precision accelerates the overall training speed if the GPU supports it [@micikevicius2017mixed]. In our study, training speed for mixed-precision is 1.2 times faster than FP32 on an Nvidia P40 GPU, and is twice as fast as FP32 on an Nvidia V100 GPU. However, this does not result in improvement in top-1 accuracy. - [Cosine learning rate decay (cosine)]{} The cosine decay [@loshchilov2016sgdr] starts at a low rate from the beginning of training, and then drops to a large rate in the middle and again at a small rate in the end. Evaluation Metrics ------------------ The selection of metrics used to measure the performance of the model is important because it indicates the direction in which the model is developed. We use the following three metrics as key indicators of model performance. The top-1 is a measure of classification accuracy on the ImageNet ILSVRC-2012 [@russakovsky2015imagenet] validation dataset. The validation dataset consists of 50,000 images of 1,000 classes. Throughput is defined as how many images are processed per second on the GPU device. We measured inference throughput for an Nvidia P40 1 GPU. For comparison with other models, we used FP32 instead of FP16 in our experiments, using a batch size of 64. The mean corruption error (mCE) was proposed by Hendrycks [@hendrycks2019benchmarking] to measure the performance of the classification model on corrupted images. Method {#sec:assembling} ====== In this section, we introduce various network tweaks and regularization techniques to be assembled, and describe the details of the implementation. We also perform preliminary experiments to study the effect of different parameter choices. Network Tweaks {#sec:network-tweak} -------------- ![ Assembling techniques into ResNet-50. We apply network tweaks such as ResNet-D, SK, Anti-alias, DropBlock, and BigLittleNet to ResNet. In more detail, ResNet-D and SK apply to all blocks in all stages. Downsampling with anti-aliasing only applies to the downsampling block from Stage 2 to Stage 4. DropBlock only applies all blocks in Stage 3 and Stage 4. Little-Branch from BigLittleNet uses one residual block with smaller width. []{data-label="fig:architecture"}](resnet_paper_architecture.eps){width="1\linewidth"} Figure \[fig:architecture\] shows the overall flow of our final ResNet-50 model. Various network tweaks were applied to vanilla ResNet. The network tweaks we used are as follows. ResNet-D is a minor adjustment to the vanilla ResNet network architecture model proposed by He [@he2019bag]. It is known to work well in practice and has little impact to computational cost [@he2019bag]. Three changes were added to the ResNet model, as illustrated in Figure \[fig:resnet-d-tweaks\]. First, the stride sizes of the first two convolutions have been switched. (Blue in Figure \[fig:resnet-d\]). Second, a 2×2 average pooling layer has been added with a stride of 2 before the convolution (Green). Last, a large 7x7 convolution has been replaced with three smaller 3x3 convolutions in Stem layer (Red). We examine two tweaks in relation to channel attention. First, Squeeze and Excitation (SE) network [@hu2018squeeze] focuses on enhancing the representational capacity of the network by modeling channel-wise relationships. SE eliminates spatial information by global pooling to get channel information only, and then two fully connected layers in this module learn the correlation between channels. Second, Selective Kernel (SK) [@li2019selective] is used, is inspired by the fact that the receptive sizes of neurons in the human visual cortex are different from each other. SK unit has multiple branches with different kernel sizes, and all branches are fused using softmax attention. ![Modified SK Unit. We use one 3x3 kernel with doubled output channel size instead of 5x5 and 3x3 kernels.[]{data-label="fig:modified-sk"}](resnet_paper_sk.eps){width="1\linewidth"} The original SK generates multiple paths with 3x3 and 5x5 convolutions, but we instead use two 3x3 convolutions to split the given feature map. This is because two convolutions of the same kernel size can be replaced by a convolution with twice as many channels, thereby lowering the inference cost. Figure \[fig:modified-sk\] shows an SK unit that replaces two branches with one convolution operation. Table \[tab:channelattention\] shows the results for different configurations of channel attention. Compared with SK, SE has higher throughput but lower accuracy (C1 in Table \[tab:channelattention\]). Comparing R50+SK$^\dagger$ (C3) to R50+SK with 3x3 and 5x5 kernels (C2), the top-1 accuracy only differs by 0.08% (78.00 and 77.92), but the throughput is significantly different (326 and 382). Considering the trade-off between accuracy and throughput, we used one 3x3 kernel with doubled channel size instead of 3x3 and 5x5 kernels. Comparing C3 and C4, we see that changing the setting of reduction ratio $r$ for SK units from 2 to 16 yields a large degradation of top-1 accuracy relative to throughput improvement. Applying both SE and SK (C5) not only decreases accuracy by 0.42 (from 77.92 to 77.50), but also decreases inference throughput by 37 (from 382 to 345). For a better trade-off between top-1 accuracy and throughput, R50+SK$^\dagger$ is preferred. CNN models for image classification are known to be very vulnerable to small amounts of distortion [@xie2019adversarial]. Zhang [@zhang2019shiftinvar] proposes AA to improve the shift-equivariance of deep networks. The max-pooling is commonly viewed as a competing downsampling strategy, and is inherently composed of two operations. The first operation is to densely evaluate the max operator and second operation is naive subsampling [@zhang2019shiftinvar]. AA is proposed as a low-pass filter between them to achieve practical anti-aliasing with any existing strided layer such as strided-conv. In the original paper, AA applies to max-pooling, projection-conv, and strided-conv of ResNet. In addition, the smoothing factor can be adjusted by changing the blur kernel filter size, where a larger filter size results in increased blur. Table \[tab:aa\] shows the experimental results for AA. We observed that reducing the filter size from 5 to 3 maintains the top-1 accuracy while increasing inference throughput (A1,2 in Table \[tab:aa\]). However, removing the AA applied to the projection-conv does not affect the accuracy (A3). We also observe that applying AA to max-pooling degrades throughput significantly (A1-3). Finally, we apply AA only to strided-conv in our model (Green in Figure \[fig:architecture\]). BigLittleNet [@chen2018big] applies multiple branches with different resolutions while aiming at reducing computational cost and increasing accuracy. The Big-Branch has the same structure as the baseline model and operates at a low image resolution, whereas the Little-Branch reduces the convolutional layers and operates at same image resolution as the baseline model. BigLittleNet has two hyperparameters, $\alpha$ and $\beta$, which adjust the width and depth of the Little-Branch, respectively. We use $\alpha$=$2$ and $\beta$=$4$ for ResNet-50 and use $\alpha$=$1$ and $\beta$=$2$ for ResNet-152. The left small branch in Figure \[fig:architecture\] represents the Little-Branch. The Little-Branch has one residual block and is smaller in width than the main Big-Branch. Regularization {#sec:regularization} -------------- #### AutoAugment (Autoaug) AutoAugment [@cubuk2018autoaugment] is a data augmentation procedure which learns augmentation strategies from data. It uses reinforcement learning to choose a sequence of image augmentation operations with the best accuracy by searching in a discrete search space of their probability of application and magnitude. We borrow the augmentation policy found by Autoaug on ImageNet ILSVRC-2012 [^4]. Mixup [@zhang2017mixup] creates one example by interpolating two examples of the training set for data augmentation. Neural networks are known to memorize training data rather than generalizing from the data [@zhang2016understanding]. As a result, the neural network produces unexpected outputs when it encounters data that differs from the distribution of the training set. Mixup mitigates these problems by showing the neural network interpolated examples of filling empty space on the feature space of the training dataset. Mixup has two types of implementation. The first type uses two mini batches to create a mixed mini batch. this type of implementation is suggested in the original paper [@zhang2017mixup]. The second type uses a single mini batch to create the mixed mini batch by mixing the mini batch with a shuffled clone of itself. The second type of implementation uses less CPU resources because only one mini batch needs to be preprocessed to create one mixed mini batch. However, experiments show that the second type of implementation reduces top-1 accuracy (Table \[tab:mixup\]). Therefore, in later experiments, we use the first type of implementation. We set the Mixup hyperparameter $\alpha$ to 0.2. Dropout [@srivastava2014dropout] is a popular technique for regularizing deep neural networks. It prevents the network from over-fitting the training set by dropping neurons at random. However, Dropout does not work well for extremely deep networks such as ResNet [@ghiasi2018dropblock]. DropBlock [@ghiasi2018dropblock] can remove specific semantic information by dropping a continuous region of activations. Thus, it is efficient for the regularization of very deep networks. We borrow the same DropBlock setting used in the original paper. We apply DropBlock to Stages 3 and 4 for ResNet-50 and linearly decay the $keep\_prob$ hyperparameter from 1.0 to 0.9 during training. In the classification problem, class labels are expressed as one hot encoding. If CNN trains to minimize cross entropy with this hard one hot encoding target, the logits of the last fully connected layer of CNN grow to infinity, which leads to over-fitting [@he2019bag]. Label smoothing [@pereyra2017regularizing] suppresses infinite output and prevents over-fitting. We set the label smoothing factor $\epsilon$ to 0.1. Knowledge Distillation [@hinton2015distilling] is a technique for transferring knowledge from one neural network (teacher) to another (student). Teacher models are often complex but cumbersome models with high accuracy, and a weak but light student model can improve its own accuracy by mimicking a teacher model. The $T$ hyperparameter of KD was said to be optimal when set to $2$ or $3$ in the originally paper [@hinton2015distilling], but we use $T$=$1$ for our model. Because our model uses Mixup and KD techniques together, the teacher network should also be applied to Mixup. This leads to better performance at lower temperatures because the teacher’s signal itself is already smoothed by the Mixup (Table \[tab:kd\]). We used AmoebaNet-A as a teacher with 83.9% of ImageNet validation top-1 accuracy. Experiment Results {#sec:Experiment Results} ================== Ablation Study {#sec:network-tweaks-result} -------------- In this section we will describe ablation experiments for assembling the individual network tweaks covered in Section \[sec:network-tweak\] to find better model. The results are shown in Table \[tab:model-tweak\]. Adding ResNet-D to the baseline model improves top-1 accuracy by 0.5% (from 76.87 to 77.37) (M1 in Table \[tab:model-tweak\]), and adding SK tweaks improves accuracy by 1.46% (from 77.37 to 78.83) (M2). In Table \[tab:channelattention\], We show that the accuracy is increased by 1.62% when SK is independently applied to ResNet (from 76.30 to 77.92). Stacking ResNet-D and SK increases the top-1 accuracy gains almost in equal measure to the sum of the performance gains of applying ResNet-D and SK separately. The results show that the two tweaks can improve performance independently with little effect on each other. Applying BL to R50D+SK improves top-1 accuracy by 0.44% (from 79.27 to 78.83) (M3). To achieve higher accuracy while maintaining throughput similar to that of the R50D+SK, we use a 256x256 image resolution for inference, whereas we use 224x224 image resolution for training. Applying AA to the R50D+SK+BL improves top-1 accuracy by 0.12% (from 79.27 to 79.39) and decreases throughput by 47 (from 359 to 312) (M4). Because AA is a network structure designed for robustness to image distortion, the top-1 accuracy does not reliably determine the AA effect. The effect on AA is further shown in the next section, wherein mCE is introduced to evaluate models. The ablation study described in Table \[tab:assemble\] shows the impact of assembling the techniques described in Section \[sec:regularization\]. We stack the network tweaks and regularizations alternately to balance the performance effects. The regularization techniques increase both accuracy and mCE, but the performance improvement effect of mCE is greater than the improvement of accuracy (E2,3,5,7,11). For example, applying Mixup, DropBlock, KD, and Autoaug individually improves top1/mCE 0.75%/6.08%, 0.69%/1.84%, 0.29%/1.26%, and 0.09%/4.14% respectively. It can be seen that regularization help to make CNNs more robust to image distortions. Adding SE improves top-1 accuracy by 0.61% and improves mCE by 3.71% (E4). SE having a greater effect on mCE enhancement than top-1 accuracy is similar to the result of the regularization techniques. We confirm that channel attention is also helpful for robustness to image distortion. Replacing SE with SK improves performance by 1.0% and 4.3% for the top-1 and mCE, respectively (E6). In Table \[tab:channelattention\], when SE is changed to SK without regularization, the accuracy increases by 0.5%. Compared to SK without regularization, replacing SE with SK with regularization leads to nearly double the accuracy improvement. This means that SK is more complementary for regularization techniques than SE. Changing the epochs from 270 to 600 improves performance (E8). Because data augmentation and regularization are stacked, they have a stronger effect of regularization, so longer epochs seems to yield better generalization performance. BL shows a high performance improvement not only on top-1, but also on mCE, and without inference throughput loss (E9). AA also shows higher performance gain in mCE relative to top-1 (E10), which agrees with AA being used as a network tweak to make the CNN robust for image translations as claimed in [@zhang2019shiftinvar]. The assembled model of all the techniques described so far has a top-1 accuracy of 82.78% and an mCE of 48.89%. This final model is listed in Table \[tab:assemble\] as E11, and we call this model Assemble-ResNet-50. We also experiment with ResNet-152 for comparison as E12, we call this model Assemble-ResNet-152. Transfer Learning: FGVC {#sec:transfer-fgvc} ----------------------- In this section, we will investigate whether these improvements discussed so far can help with transfer learning. Before that, we first need to analyze the contribution of transfer learning for each technique. To do this, we performed an ablation study on the Food-101 [@bossard2014food] dataset, which is the largest public fine-grained visual classification (FGVC) dataset. The basic experiment setup and hyperparameters that differ from the backbone training are: - Initial learning rate reduced from $0.1$ to $0.01$. - Weight decay is set to 0.001. - Momentum for BN is set to $\max(1 - 10/s, 0.9)$. - Keep probability of DropBlock starts at $0.9$ and decreases linearly to $0.7$ at the end of training - The training epoch is set differently for each dataset and is indicated in Appendix \[sec:appendix-fgvc\_dataset\]. As shown in Table \[tab:food101\], stacking network tweaks and regularization techniques have steadily improved both top-1 accuracy and mCE for the transfer learning task on the Food-101 dataset. In particular, comparing the experiments F4-F8 with experiments F9-F13 (in Table \[tab:food101\]) shows the effect of regularization on the backbone. We use the same network structure in F4-F13, but for F9-F13, they have regularization such as Mixup, DropBlock, KD and Autoaug on the backbone. This regularization of the backbone gives performance improvements for top-1 accuracy as expected. On the other hand, the aspect of mCE performance differed from the top-1 accuracy. Without regularization during fine-tuning such as in F4 and F9, the backbone with regularization leads to better mCE performance than backbone without regularization. However, adding regularization during fine-tuning narrows the mCE performance gap (F5-8 and F10-13). For convenience, we call the final F13 model in Table \[tab:food101\] as Assemble-ResNet-FGVC-50. We also have evaluated Assemble-ResNet-FGVC-50 in Table \[tab:food101\] on the following datasets: Stanford Cars [@krause20133d], CUB-200-201l [@wah2011caltech], Oxford 102 Flowers [@nilsback2008automated], Oxford-IIIT Pets [@parkhi2012cats], FGVC-Aircraft [@maji2013fine], and Food-101 [@bossard2014food]. The statistics for each dataset are as shown in Appendix \[sec:appendix-fgvc\_statistic\]. We borrow the same training settings from Kornblith [@kornblith2019better] and fine-tuned new datasets from Assemble-ResNet-50 ImageNet checkpoint. Table \[tab:fgvc\] shows the transfer learning performance. Compared to EfficientNet [@tan2019efficientnet] and AmoebaNet-B [@huang2019gpipe] which are state-of-the-art model for image classification tasks. Our Assemble-ResNet-FGVC-50 model achieves comparable accuracy with 20x faster inference throughput than the existing state-of-the-art models. Transfer Learning: Image Retrieval {#sec:transfer-ir} ---------------------------------- We also conducted an ablation study on three standard fine-grained image retrieval (IR) datasets: Stanford Online Products (SOP) [@song2016deep], CUB200 [@wah2011caltech] and CARS196 [@krause20133d]. We borrow the zero-shot data split protocol from [@song2016deep]. The basic experiment setup and hyperparameters are as follows. - Image preprocessing resizes to 224x224 without maintaining aspect ratio with probability 0.5 and resize to 256x256 and random crop to 224x224 with probability 0.5. - Data augmentation includes random flip with 0.5 probability. - Momentum for BN is set to $\max(1 - 10/s, 0.9)$. - Weight decay is set to 0.0005. - The training epoch, batch size, learning rate decay and assembling configuration is set differently for each dataset. We will describe the settings in the Appendix \[sec:appendix-fgir\_dataset\]. On top of that, cosine-softmax based losses were used for image retrieval. In this work, we use ArcFace [@deng2019arcface] loss with a margin of 0.3 and use generalized mean-pooling (GeM) [@radenovic2018fine] for a pooling method without performing downsampling at Stage 4 of backbone networks because it has better performance for the image retrieval task. In the case of SOP, the degree of the effect was examined by an ablation study with the results listed in Table \[tab:sop\]. In contradiction to our results on FGVC, the particular combination of network tweaks and regularization that worked well on the SOP dataset were different from that for FGVC datasets. Comparing S2-4, we see that BL and AA did not work well on the SOP dataset. Of the regularizers, DropBlock works well, but Autoaug do not improve the recall@1 performance (S2 and S5,6). Nevertheless, in the best configuration, there was a significant performance improvement of 3.0% compared to the baseline ResNet-50. The recall at 1 for image retrieval datasets (recall@1) are reported in Table \[tab:fgir\]. There is also a significant performance improvement on CUB200 and CARS196 datasets. Conclusion {#sec:conclusion} ========== In this paper, we show that assembling various techniques for CNNs to single convolutional networks leads to improvements of top-1 accuracy and mCE on the ImageNet ILSVRC2012 validation dataset. Synergistic effects have been achieved by using a variety of network tweaks and regularization techniques together in a single network. Our approach has also improved performance consistently on transfer learning such as FGVC and image retrieval tasks. More excitingly, our network is not frozen, but is still evolving, and can be further developed with future research. For example, we already are planning to reassemble various new studies such as AugMix [@hendrycks2019augmix] and ECA-net [@wang2019ecanet], which were recently published. We expect that there will be further improvements if we change the vanilla backbone to a more powerful backbone such as EfficientNet [@tan2019efficientnet] instead of ResNet, which we leave as future works. FLOPS and throughput {#sec:appendix-flops_and_throughput} ==================== We observe in several experiments that FLOPS is not proportional to the inference speed of the actual GPU. FGVC Configuration {#sec:appendix-fgvc_dataset} ================== We use the same hyperparameters for as all datasets as possible for transfer learning. The different parameters settings for each dataset are described in Table \[tab:fgvc-dataset\_configuration\]. FGVC Datasets {#sec:appendix-fgvc_statistic} ============= IR Configuration {#sec:appendix-fgir_dataset} ================ IR uses a different regularization for each dataset. We use the same hyperparameters for as all datasets as possible for transfer learning. The parameters set differently for each dataset are described in Table \[tab:fgir-dataset\_configuration\]. [^1]: https://www.kaggle.com/c/ifood-2019-fgvc6/leaderboard [^2]: https://github.com/clovaai/assembled-cnn [^3]: https://github.com/tensorflow/models [^4]: https://github.com/tensorflow/models/tree/master/research/autoaugment	{ "pile_set_name": "ArXiv" }
--- abstract: \| This is the first in a series of at least 4 articles. We will study Carlson’s $<_1$-relation in the whole class of ordinals and later we will link it with ordinals $\alpha \leqslant \left\| \Pi^1_1 - {\ensuremath{\operatorname{CA}}}_0 \right\|$. The main motivation to study $<_1$ are the works of T. Carlson and G. Wilken. The first version $\prec_1$ of $<_1$ was used by Carlson as a tool to show Reinhardt’s conjecture: The Strong Mechanistic Thesis is consistent with Epistemic Arithmetic (see [[@Carlson3]]{}); moreover, Carlson showed a characterization of $\varepsilon_0$ in terms of $\prec_1$ (see [[@Carlson1]]{}) and indeed, set up a different approach to ordinal notation systems based on these ideas (see [[@Carlson2]]{}). $<_1$ is a binary relation in the class of ordinals and in it’s original form, $\alpha <_1 \beta$ asserts that the structure $(\alpha, <, +, <_1)$ is a $\Sigma_1$-substructure of $(\beta, <, +, <_1)$. Here, instead of the original definition of $<_1$, an equivalent, model-theoretical notion (see appendix) is consider as the fundamental notion: $\alpha <_1 \beta$ means $\alpha < \beta$ and the following assertion: for any finite subset $Z$ of $\beta$, there exists an ($<, +, <_1$)-embedding $h : Z \longrightarrow \alpha$ with $h \left\|_{Z \cap \alpha} = {\ensuremath{\operatorname{Id}}}_{Z \cap \alpha} \right.$ (see definition \[Definition\_of\_<less>=\_1\]). Moreover, $\alpha \leqslant_1 \beta$ stands for $\alpha = \beta$ or $\alpha <_1 \beta$. The study of $<_1$, as done here, is then a study of (a sort of) isomorphisms between the finite subsets of an ordinal. In this introductory article we will study the (canonical) isomorphisms $\left\{ g \left( 0, \alpha, \beta \right) \| \alpha, \beta \in \mathbbm{P} \right\}$, the $<^0$-relation and it’s cofinality properties and see how it is that $<_1$ induces, through all of these notions, thinner $\kappa$-club classes of ordinals. In comming articles it will be shown the complete generalization of these ideas to the thinnest $\kappa$-club classes induced by $<_1$. author: - Parménides García Cornejo title: 'Carlson’s $<_1$-relation on the class of addtive principal ordinals' --- Basic conventions used throughout this work =========================================== We use the standard logical and set theoretical symbols in it’s standard way: $\wedge, \vee, \Longrightarrow, \Longleftrightarrow, \forall, \exists, \neg, \emptyset, \cup, \cap, \subset, =, \in$, etc. By $B \subset_{{\ensuremath{\operatorname{fin}}}} A$ we mean $B$ is a finite subset of $A$. $h : A \longrightarrow B$ denotes that $h$ is a functional with domain $A$ and codomain $B$. For a functional $h : A \longrightarrow B$ and $C \subset A$, we define $h [C] {:=}\{h (x) \| x \in C\}$. For a functional $h : A \longrightarrow B$, we denote ${\ensuremath{\operatorname{Dom}}} h {:=}A$ and ${\ensuremath{\operatorname{Im}}} h {:=}h [A]$. By ${\ensuremath{\operatorname{OR}}}$ we denote the class of ordinals. $0, 1, 2$,... denote, as usual, the finite ordinals. $\omega$ denotes the first infinite ordinal. ${\ensuremath{\operatorname{Lim}}}$ denotes the class of limit ordinals. $\mathbbm{P}$ denotes the class of additive principal ordinals. $\mathbbm{E}$ denotes the class of epsilon numbers. $<$, $+$, $\lambda x. \omega^x$ denote the usual order, the usual addition and the usual $\omega$-base-exponentiation in the ordinals, respectively. For an ordinal $\alpha \in {\ensuremath{\operatorname{OR}}}$, $\varepsilon_{\alpha}$ denotes the $\alpha$-th epsilon number. $\min A$ denotes the minimum element of $A$ (with respect to the order $<$). $\max A$ denotes the maximum element of $A$ (with respect to $<$). In case $\exists \alpha \in {\ensuremath{\operatorname{OR}}} .A \subset \alpha$, then $\sup A$ denotes the minimal upper bound of $A$ with respect to $<$ (the supremum of $A$). ${\ensuremath{\operatorname{Lim}}} A {:=}{\ensuremath{\operatorname{Lim}}} (A) {:=}\{\alpha \in {\ensuremath{\operatorname{OR}}} \| \alpha = \sup (A \cap \alpha)\}$. By $(\xi_i)_{i \in I} \subset A$ we mean $(\xi_i)_{i \in I}$ is a sequence of elements of $A$. Given an ordinal $\alpha \in {\ensuremath{\operatorname{OR}}}$ and a sequence $(\xi_i)_{i \in I} \subset {\ensuremath{\operatorname{OR}}}$, we say that $(\xi_i)_{i \in I}$ is cofinal in $\alpha$ whenever $I \subset {\ensuremath{\operatorname{OR}}}$, $\forall i \in I \forall j \in I.i \leqslant j \Longrightarrow \xi_i \leqslant \xi_j$, $\forall i \in I \exists j \in I.i < j \wedge \xi_i < \xi_j$ and $\sup \{\xi_i \| i \in I\} = \alpha$. By $\xi_i\underset{cof}{{{\lhook\joinrel\relbar\joinrel\rightarrow}}}\alpha$ we mean that the sequence $(\xi_i)_{i \in I}$ is cofinal in $\alpha$. Whenever we write $\alpha =_{{\ensuremath{\operatorname{CNF}}}} \omega^{A_1} a_1 + \ldots + \omega^{A_n} a_n$, we mean that $\omega^{A_1} a_1 + \ldots + \omega^{A_n} a_n$ is the cantor normal form of $\alpha$, that is: $\alpha = \omega^{A_1} a_1 + \ldots + \omega^{A_n} a_n$, $a_1, \ldots, a_n \in \omega \backslash \{0\}$, $A_1, \ldots, A_n \in {\ensuremath{\operatorname{OR}}}$ and $A_1 > \ldots > A_n$. Given two ordinals $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$ with $\alpha \leqslant \beta$, we denote: $[\alpha, \beta] {:=}\{\sigma \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant \sigma \leqslant \beta\}$ $[\alpha, \beta) {:=}\{\sigma \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant \sigma < \beta\}$ $(\alpha, \beta] {:=}\{\sigma \in {\ensuremath{\operatorname{OR}}} \| \alpha < \sigma \leqslant \beta\}$ $(\alpha, \beta) {:=}\{\sigma \in {\ensuremath{\operatorname{OR}}} \| \alpha < \sigma < \beta\}$ Given $\alpha \in \mathbb{E}$, we denote by $\alpha^+$ or by $\alpha (+^1)$ to $\min \{e \in \mathbbm{E}\| \alpha < e\}$. For a set $A$, $\|A\|$ denotes the cardinality of $A$; the only one exception to this convention is when we denote as $\| {\ensuremath{\operatorname{ID}}}_n \|$ and $\| \Pi^1_1$-${\ensuremath{\operatorname{CA}}}_0 \|$ to the proof theoretic ordinals of the theories ${\ensuremath{\operatorname{ID}}}_n$ and $\Pi^1_1$-${\ensuremath{\operatorname{CA}}}_0$ respectively. The $<_1$-relation ================== Our purpose is to study the (binary) relation $<_1$ defined by recursion on the ordinals as follows \[Definition\_of\_<less>=\_1\]Let $\beta \in {\ensuremath{\operatorname{OR}}}$ be arbitrary and suppose $\alpha' <_1 \beta'$ has already been defined for any $\beta' \in \beta \cap {\ensuremath{\operatorname{OR}}}$ and for any $\alpha' \in {\ensuremath{\operatorname{OR}}}$. Let $\alpha \in {\ensuremath{\operatorname{OR}}}$ be arbitrary.Then\ $\alpha <_1 \beta$ $: \Longleftrightarrow$ $\alpha < \beta$ and $\forall Z \subset_{{\ensuremath{\operatorname{fin}}}} \beta \exists \tilde{Z} \subset_{{\ensuremath{\operatorname{fin}}}} \alpha . \exists h$ such that: (i) $h : (Z, +, <, <_1) \longrightarrow ( \tilde{Z}, +, <, <_1)$ is an isomorphism, that is: + $h : Z \longrightarrow \tilde{Z}$ is a bijection. + For any $a_1, a_2 \in Z$ $\bullet$ $a_1 + a_2 \in Z \Longleftrightarrow h (a_1) + h (a_2) \in \tilde{Z}$ $\bullet$ If $a_1 + a_2 \in Z$, then $h (a_1 + a_2) = h (a_1) + h (a_2)$. + For any $a_1, a_2 \in Z$, $\bullet$ $a_1 < a_2 \Longleftrightarrow h (a_1) < h (a_n)$. $\bullet$ $a_1 <_1 a_2 \Longleftrightarrow h (a_1) <_1 h (a_n)$. (ii) $h\|_{Z \cap \alpha} = {\ensuremath{\operatorname{Id}}} \|_{Z \cap \alpha}$, where ${\ensuremath{\operatorname{Id}}} \|_{Z \cap \alpha} : Z \cap \alpha \longrightarrow Z \cap \alpha$ is the identity function. By $\alpha \leqslant_1 \beta$ we mean that $\alpha <_1 \beta$ or $\alpha = \beta$. Moreover, to make our notation simpler, we will write $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}} \|_{\alpha}$ instead of $h\|_{Z \cap \alpha} = {\ensuremath{\operatorname{Id}}} \|_{Z \cap \alpha}$. We will eventually use functions $f : Z \longrightarrow \tilde{Z}$ that are $\lambda x. \omega^x$-isomorphisms; of course, by this we mean the analogous situation as the one we had with $+$ above:\ For any $a \in Z$, $\bullet$ $\omega^a \in Z \Longleftrightarrow f (\omega^{\alpha}) \in \tilde{Z}$ $\bullet$ If $\omega^a \in Z$, then $f (\omega^a) = \omega^{f (a)}$. Some of the most basic properties that $\leqslant_1$ satisfies are the following \[connectedness\_and\_continuity\]Let $\alpha, \beta, \gamma \in {\ensuremath{\operatorname{OR}}}$. $\alpha \leqslant_1 \beta \Longrightarrow \{x \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant_1 x \leqslant \beta\} = [\alpha, \beta]$. Let $(\xi_i)_{i \in I} \subset {\ensuremath{\operatorname{OR}}}$ be a sequence such that $\xi_i \underset{{\ensuremath{\operatorname{cof}}}}{{{\lhook\joinrel\relbar\joinrel\rightarrow}}} \beta$. Then\ $[\forall i \in I. \alpha \leqslant_1 \xi_i] \Longrightarrow \alpha \leqslant_1 \beta$. $\alpha \leqslant_1 \beta \leqslant_1 \gamma \Longrightarrow \alpha \leqslant_1 \gamma$. Let $(\xi_i)_{i \in I} \subset {\ensuremath{\operatorname{OR}}}$ be a sequence such that $\xi_i \underset{{\ensuremath{\operatorname{cof}}}}{{{\lhook\joinrel\relbar\joinrel\rightarrow}}} \beta$. Then\ $[\exists i_0 \in I. \alpha \nless_1 \xi_{i_0} \wedge \alpha < \xi_{i_0}] \Longrightarrow \alpha \nless_1 \beta$. The proofs of $a)$, $b)$ and $c)$ follow direct from definition \[Definition\_of\_<less>=\_1\]. Moreover, $d)$ follows easily from $a)$. We call $\leqslant_1$[[-connectedness]{}]{} (or just connectedness) to the property $a$) of previous proposition \[connectedness\_and\_continuity\]; moreover, we call $\leqslant_1$[[-continuity]{}]{} (or just continuity) and $\leqslant_1$[[-transitivity]{}]{} (or just transitivity) to the properties $b$) and $c$) (respectively) of the same proposition. We will make use of the three of them over and over along all our work. \[existence\_of\_m(alpha)\]Let $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$ with $\alpha < \beta$ and $\alpha \nless_1 \beta$. Then there exists $\gamma \in [\alpha, \beta)$ such that $\{x \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant_1 x\} = [\alpha, \gamma]$. $\{x \in {\ensuremath{\operatorname{OR}}} \| \alpha < x, \alpha \nless_1 x\} = [\gamma + 1, \infty)$. For any $\sigma > \gamma$, $\gamma \nless_1 \sigma$. Let $k {:=}\min \{r \in {\ensuremath{\operatorname{OR}}} \|r > \alpha \nless_1 r\}$. Then $k \leqslant \beta$. Moreover, since $\forall \sigma \in [\alpha, k) . \alpha \leqslant_1 \sigma$, then $k$ must be a successor (otherwise, by $\leqslant_1$-continuity would follow $\alpha <_1 k$). So $k = \gamma + 1 \leqslant \beta$ for some $\gamma \in {\ensuremath{\operatorname{OR}}}$ and therefore $\{x \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant_1 x\} = [\alpha, \gamma]$. This shows $a$). On the other hand, note that for any $\sigma \geqslant k$, it is not possible that $\alpha \leqslant_1 \sigma$ (otherwise, by $\leqslant_1$-connectedness, one gets the contradiction $\alpha <_1 k$). This proves $b$). Finally, observe it is not possible that for some $\sigma > \gamma$, $\gamma <_1 \sigma$, otherwise, from $\alpha \leqslant_1 \gamma \leqslant_1 \sigma$ and $\leqslant_1$-transitivity follows $\alpha <_1 \sigma$, which is contradictory with $b)$ (because $\sigma \geqslant k = \gamma + 1$). For an ordinal $\alpha$, the ordinal $\gamma$ referred in previous proposition \[existence\_of\_m(alpha)\] will be very important for the rest of our work. Because of that we make the following \[def\_max<less>=\_1\_reach\](The maximum $\leqslant_1$-reach of an ordinal). Let $\alpha \in {\ensuremath{\operatorname{OR}}}$. We define\ $m (\alpha) {:=}\left\{ \begin{array}{l} \max \{\xi \in {\ensuremath{\operatorname{OR}}} \| \alpha \leqslant_1 \xi\} \text{ {\ensuremath{\operatorname{iff}}} {\ensuremath{\operatorname{there}}} {\ensuremath{\operatorname{is}}} } \beta\in{\ensuremath{\operatorname{OR}}} \text{{\ensuremath{\operatorname{with}}} } \alpha<\beta \text{ {\ensuremath{\operatorname{and}}} } \alpha\nless_1\beta\\ \infty \text{ {\ensuremath{\operatorname{otherwise}}}, {\ensuremath{\operatorname{that}}} {\ensuremath{\operatorname{is}}}, } \forall \beta \in {\ensuremath{\operatorname{OR.}}} \alpha < \beta \Longrightarrow \alpha <_1 \beta \end{array} \right .$ Note that when $m (\alpha) \in {\ensuremath{\operatorname{OR}}}$, then it is the only one $\gamma \in {\ensuremath{\operatorname{OR}}}$ satisfying $\alpha \leqslant_1 \gamma$ and $\alpha \nleqslant_1 \gamma + 1$. Because of this [[we call]{}]{} ${\ensuremath{\boldsymbol{m (\alpha)}}}$ [[the maximum]{}]{} ${\ensuremath{\boldsymbol{\leqslant_1}}}$[[-reach of $\alpha$.]{}]{} Characterization of the ordinals $\alpha$ such that $\alpha <_1 \alpha + 1$ =============================================================== Up to this moment we do not know whether there are ordinals $\alpha, \beta$ such that $\alpha <_1 \beta$; however, in such a case, since $\alpha < \alpha + 1 \leqslant \beta$, then by $\leqslant_1$-connectedness we would conclude that the relation $\alpha <_1 \alpha + 1$ must hold. This shows that the simplest nontrivial case when we can expect that something of the form $\alpha <_1 \beta$ holds is for $\beta = \alpha + 1$. Then, for this simplest case, what should $\alpha$ satisfy?. The answer to this question is the purpose of this subsection. \[prop1.02.11.2008\]Let $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$, $\alpha =_{{\ensuremath{\operatorname{CNF}}}} \omega^{\alpha_1} a_1 + \ldots + \omega^{\alpha_n} a_n$, with $n \geqslant 2$ or $a_1 \geqslant 2$. Moreover, suppose $\alpha < \beta$. Then $\alpha \nless_1 \beta$. Case $n \geqslant 2$.\ Since $\alpha < \beta$, then $\{\omega^{\alpha_1} a_1, \ldots, \omega^{\alpha_n} a_n \} \subset \alpha \cap \beta$, but $\beta \ni \omega^{\alpha_1} a_1 + \ldots + \omega^{\alpha_n} a_n = \alpha {\not\in}\alpha$, and so there is no $+$-isomorphism $h : Z \rightarrow \tilde{Z}$ from $Z {:=}\{\omega^{\alpha_1} a_1, \ldots, \omega^{\alpha_n} a_n, \alpha\} \subset_{{\ensuremath{\operatorname{fin}}}} \beta$ in some $\tilde{Z} \subset_{{\ensuremath{\operatorname{fin}}}} \alpha$ such that $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}} \|_{\alpha}$, since any of such isomorphisms should accomplish\ $h (\omega^{\alpha_1} a_1 + \ldots + \omega^{\alpha_n} a_n) = h (\omega^{\alpha_1} a_1) + \ldots + h (\omega^{\alpha_n} a_n) = \alpha {\not\in}\alpha$. The same argument works for the case $n = 1, a_1 \geqslant 2$. \[corollary1.02.11.2008\]Let $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$. If $\alpha <_1 \beta$, then $\alpha =_{{\ensuremath{\operatorname{CNF}}}} \omega^{\gamma} \in \mathbbm{P} \subset {\ensuremath{\operatorname{Lim}}}$, for some $\gamma \in {\ensuremath{\operatorname{OR}}}, \gamma > 0$. Direct from previous proposition \[prop1.02.11.2008\]. The only left cases are $\alpha = 0$ or $\alpha = 1$ but for those cases it is very easy to see that $\alpha \nless_1 \alpha + 1$, since $\alpha + 1$ has $\alpha + 1$ elements and $\alpha$ has only $\alpha$ elements, and so for those cases $\alpha \nless_1 \beta$ for any $\beta > \alpha$. If $\alpha = \omega^n$, $n \in \omega$, then $\alpha \nless_1 \alpha + 1$. Not hard. But we will give a more general proof of this fact in the next propositions. Let $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$. If $\alpha <_1 \beta$, then $\alpha =_{{\ensuremath{\operatorname{CNF}}}} \omega^{\gamma}$ for some $\gamma \in {\ensuremath{\operatorname{OR}}}, \gamma \geqslant \omega$. From previous proposition and previous corollary. (This will be proved in the next three propositions in a more general way). \[prop.exist.maximum\]Let $\alpha \in {\ensuremath{\operatorname{OR}}}$, $1 < \alpha \in {\ensuremath{\operatorname{Lim}}}$. Suppose $\alpha \cap \mathbbm{P}$ is not confinal in $\alpha$. Then $M {:=}\max (\mathbbm{P} \cap \alpha)$ exists. Since $\mathbbm{P}$ is a closed class of ordinals, then $\sup (\mathbbm{P} \cap \alpha) \in \mathbbm{P} \cap \alpha$. So $M = \sup (\mathbbm{P} \cap \alpha)$. \[prop0.03.11.2008\]Let $\alpha, p \in {\ensuremath{\operatorname{OR}}}$, $1 < \alpha <_1 p + 1$, with $p \in \mathbbm{P}$ an additive principal number. Then: \(i) $\alpha \cap \mathbbm{P}$ is confinal in $\alpha$. \(ii) $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P} \subset \mathbbm{P}$, (or equivalently, (ii’) $\alpha = \omega^{\gamma}$, for $\gamma \in {\ensuremath{\operatorname{Lim}}}$.) $(i)$. By corollary \[corollary1.02.11.2008\] we know $\alpha \in {\ensuremath{\operatorname{Lim}}}$. Now, suppose $\alpha \cap \mathbbm{P}$ is not confinal in $\alpha$. Then by previous proposition \[prop.exist.maximum\], let $M {:=}\max \alpha \cap \mathbbm{P} \in \alpha$. Then $M + p = p$, but on the other hand, $\forall \gamma \in \alpha .M + \gamma > \gamma$. Therefore, for\ $Z {:=}\{M, p\} \subset_{{\ensuremath{\operatorname{fin}}}} p + 1$ and for any $\tilde{Z} \subset \alpha$ there is no $+$-isomorphism $h : Z \rightarrow \tilde{Z}$, such that $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}} \|_{\alpha}$, since any such function would satisfy\ $h (p) = h (M + p) = h (M) + h (p) = M + h (p) > h (p)$ (Contradiction!). Thus $\alpha \cap \mathbbm{P}$ is confinal in $\alpha$. $(i i)$. Clear from $(i)$. \[alpha<less>\_1beta\_implies\_alpha\_in\_Lim(P)\]Let $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$ such that $\alpha <_1 \beta$. Then $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$. From corollary \[corollary1.02.11.2008\] we have that $\alpha <_1 \beta$ implies $\alpha \in \mathbbm{P}$. Moreover, from $\alpha <_1 \beta$ we know $\alpha < \alpha + 1 \leqslant \beta$ and then $\alpha <_1 \alpha + 1$ by $<_1$-connectedness. Finally, from $\alpha <_1 \alpha + 1$, $\alpha \in \mathbbm{P}$ and the previous proposition \[prop0.03.11.2008\], $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$. \[characterization\_of\_alpha<less>\_1alpha+1\]Let $\alpha \in {\ensuremath{\operatorname{OR}}}$. The following are equivalent: $\alpha <_1 \alpha + 1$ $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$ $\alpha = \omega^{\gamma}$ for some $\gamma \in {\ensuremath{\operatorname{Lim}}}$. $\alpha = \omega^{\gamma}$ and $\gamma =_{{\ensuremath{\operatorname{CNF}}}} \omega^{A_1} a_1 + \ldots + \omega^{A_n} a_n$ with $A_n \neq 0$. The proof of $b) \Longleftrightarrow c) \Longleftrightarrow d)$ is a standard fact about ordinals. $a) \Longrightarrow b)$ is previous corollary \[alpha<less>\_1beta\_implies\_alpha\_in\_Lim(P)\]. So let’s prove $b) \Longrightarrow a)$. Let $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$. Take $B \subset_{{\ensuremath{\operatorname{fin}}}} \alpha + 1$. If $\alpha {\not\in}B$, then $l : B \longrightarrow \alpha$, $l (x) {:=}x$ is an\ $(<, <_1, +)$-isomorphism such that $l\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. So suppose $B = \{a_0 < \ldots < a_n = \alpha\}$ for some natural number $n$. Let $A {:=}\{m (a) \|a \in (B \cap \alpha) \wedge m (a) < \alpha\}$. Since $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$ and $A$ is finite, then there exists $\rho \in (a_{n - 1}, \alpha) \cap (\max A, \alpha) \cap \mathbbm{P}$. Let $h : B \longrightarrow h [B] \subset \alpha$ be the function\ $h (x) {:=}\left\{ \begin{array}{l} x \text{ {\ensuremath{\operatorname{iff}}} } x < \alpha\\ \rho \text{ {\ensuremath{\operatorname{otherwise}}}} \end{array} \right.$. It is clear that $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. We assure that $h$ is an $(<, <_1, +)$-isomorphism. The details are left to the reader. The ordinals $\alpha$ satisfying $\alpha <_1 t$, for some $t \in [\alpha, \alpha \omega)$. ================================================================ We have seen previously that the “solutions of the $<_1$-inequality” $x <_1 x + 1$ are the elements of ${\ensuremath{\operatorname{Lim}}} \mathbbm{P}$. It is natural then to ask himself about the solutions of $x <_1 x + 2$ or of $x <_1 x + \omega$. In general, this question can be informally stated as: What are the solutions of $x <_1 \beta$, where “we pick $\beta$ as big as we can”?. The descriptions of such solutions in a certain way is a main purpose of this work: we will describe them as certain classes of ordinals obtained by certain thinning procedure. The rest of this article is devoted to our investigations concerning this question for $x \in \mathbbm{P}$ and $\beta \in [x, x \omega]$. We will introduce various concepts that at the first sight may look somewhat artificial; however, these concepts and the way to use them is just “the most basic realization” of the general tools and methodology shown in comming articles that will allow us to understand the $<_1$-relation in the whole class of ordinals. Class(0) -------- Let ${\ensuremath{\operatorname{Class}}} (0) {:=}\mathbbm{P}$. For $\alpha, \beta \in {\ensuremath{\operatorname{OR}}}$, let\ $- \alpha + \beta {:=}\left\{ \begin{array}{l} \text{{\ensuremath{\operatorname{the}}} {\ensuremath{\operatorname{only}}} {\ensuremath{\operatorname{one}}} {\ensuremath{\operatorname{ordinal}}} } \sigma \text{ {\ensuremath{\operatorname{such}}} {\ensuremath{\operatorname{that}}} } \alpha + \sigma = \beta \text{ {\ensuremath{\operatorname{iff}}} } \alpha \leqslant \beta\\ \\ - 1 \text{ {\ensuremath{\operatorname{otherwise}}}} \end{array} \right.$ Let $\alpha, c \in {\ensuremath{\operatorname{Class}}} (0)$ with $\alpha \leqslant c$. We define $g (0, \alpha, c) : \alpha \omega \longrightarrow c \omega$ as:\ $g (0, \alpha, c) (x) {:=}x$ iff $x < \alpha$.\ $g (0, \alpha, c) (x) {:=}c n + l$ iff $x \in [\alpha n, \alpha n + \alpha) \wedge x = \alpha + l$ for some $l \in \alpha$. Moreover, we define $g (0, c, \alpha) {:=}g (0, \alpha, c)^{- 1}$. \[g(0,a,c)\_prop1\]Let $\alpha, c \in {\ensuremath{\operatorname{Class}}} (0)$. Then ${\ensuremath{\operatorname{Dom}}} g (0, \alpha, c) = (\alpha \cap c) \cup \bigcup_{n \in [1, \omega)} \{t \in [\alpha n, \alpha n + \alpha) \| - \alpha n + t < c\}$. ${\ensuremath{\operatorname{Im}}} g (0, \alpha, c) = (\alpha \cap c) \cap \bigcup_{n \in [1, \omega)} \{t \in [c n, c n + c) \| - c n + t < \alpha\}$. $g (0, \alpha, c) : {\ensuremath{\operatorname{Dom}}} g (0, \alpha, c) \longrightarrow {\ensuremath{\operatorname{Im}}} g (0, \alpha, c)$ is an $(<, +)$-isomorphism and $g (0, \alpha, c) \|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. Left to the reader. \[g(0,a,c)\|\_(a,a2)\_is\_iso\]Let $\alpha, c \in {\ensuremath{\operatorname{Class}}} (0)$ and $X {:=}(\alpha \cap c) \cup \bigcup_{n \in [1, \omega)} \{t \in [\alpha n, \alpha n + \alpha) \| - \alpha n + t < c\}$. Then the function $H : (\alpha, \alpha \omega) \cap X \longrightarrow H [(\alpha, \alpha \omega) \cap X] \subset (c, c \omega)$, $H (x) {:=}g (0, \alpha, c) (x)$ is an\ $(<, <_1, +)$-isomorphism. Let $\alpha, c$, $X$ and $H$ be as stated. By previous proposition \[g(0,a,c)\_prop1\] follows easily that $H$ is an $(<, +)$-isomorphism. Moreover, $H$ is also an $<_1$-isomorphism because by proposition \[characterization\_of\_alpha<less>\_1alpha+1\] and $<_1$-connectedness it follows that $\forall a, b \in (\alpha, \alpha \omega) .a \nless_1 b$ and $\forall a, b \in (c, c \omega) .a \nless_1 b$. \[def\_T(0,alpha,t)\]Consider $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in \alpha \omega$.\ We define $T (0, \alpha, t) {:=}\left\{ \begin{array}{l} \{t\} \text{ {\ensuremath{\operatorname{iff}}} } t < \alpha\\ \\ \{t, - \alpha n + t\} \text{ {\ensuremath{\operatorname{iff}}} } t \in [\alpha n, \alpha n + \alpha) \text{{\ensuremath{\operatorname{for}}} {\ensuremath{\operatorname{some}}}} n \in [1, \omega) . \end{array} \right.$ \[Domg(0,a,t)\_and\_T(0,a,t)\]$\forall \alpha, c \in {\ensuremath{\operatorname{Class}}} (0) . \forall t \in \alpha \omega .t \in {\ensuremath{\operatorname{Dom}}} (g (0, \alpha, c)) \Longleftrightarrow T (0, \alpha, t) \cap \alpha \subset c$ Direct from definition \[def\_T(0,alpha,t)\] and proposition \[g(0,a,c)\_prop1\]. Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega]$. By $\alpha <^0 t$ we mean $\alpha < t$ $\forall B \subset_{{\ensuremath{\operatorname{fin}}}} t. \exists \delta \in {\ensuremath{\operatorname{Class}}} (0) \cap \alpha$ such that $( \bigcup_{t \in B} T (0, \alpha, t) \cap \alpha) \subset \delta$; The function $h : B \longrightarrow h [B]$ defined as $h (x) {:=}g (0, \alpha, \delta) (x)$ is an $(<, <_1, +)$-isomorphism with $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. As usual, $\alpha \leqslant^0$ just means $\alpha <^0 t$ or $\alpha = t$. Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$, $(\xi_i)_{i \in I} \subset [\alpha, \alpha \omega] \ni \beta, \gamma$. Then $\alpha \leqslant^0 \beta \Longrightarrow \alpha \leqslant_1 \beta$. If $\alpha \leqslant \beta \leqslant \gamma \wedge \alpha \leqslant^0 \gamma$ then $\alpha \leqslant^0 \beta$. ($\leqslant^0$-connectedness) If $\forall i \in I. \alpha \leqslant^0 \xi_i \wedge \xi_i \underset{{\ensuremath{\operatorname{cof}}}}{{{\lhook\joinrel\relbar\joinrel\rightarrow}}} \beta$ then $\alpha \leqslant^0 \beta$. ($\leqslant^0$-continuity) Left to the reader. \[<less>\^0\_1st\_cof.\_prop.\](First fundamental cofinality property of $<^0$).\ Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega)$.\ Then $\alpha <^0 t + 1 \Longrightarrow \alpha \in {\ensuremath{\operatorname{Lim}}} \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$. Let $\alpha$, $t$ be as stated. Suppose $\alpha <^0 t + 1$. [[*(\1)*]{}]{} Let $\gamma \in \alpha$ be arbitrary and consider $B_{\gamma} {:=}\{\gamma, \alpha, t\} \subset_{{\ensuremath{\operatorname{fin}}}} t + 1$. By (\1) there exists $\delta_{\gamma} \in \alpha \cap {\ensuremath{\operatorname{Class}}} (0)$ such that $( \bigcup_{q \in B} T (0, \alpha, q) \cap \alpha) \subset \delta_{\gamma}$ and the function $h : B \longrightarrow h [B] \subset \alpha$, $h (x) {:=}g (0, \alpha, \delta_{\gamma}) (x)$ is an ($<, <_1, +$)-isomorphism with $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. In particular, note: 1\. $\gamma < \delta_{\gamma}$ because $\gamma \in ( \bigcup_{q \in B} T (0, \alpha, q) \cap \alpha) \subset \delta_{\gamma}$. 2\. $\delta_{\gamma} = g (0, \alpha, \delta_{\gamma}) (\alpha) \leqslant_1 g (0, \alpha, \delta_{\gamma}) (t)$ because $T (0, \alpha, t) \cap \alpha \subset \delta_{\gamma}$ and $\alpha \leqslant_1 t \Longleftrightarrow h (\alpha) \leqslant_1 h (t)$. Since the previous was done for arbitrary $\gamma < \alpha$, 1 and 2 show that\ $\forall \gamma \in \alpha \exists \delta_{\gamma} \in \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| \gamma < \beta \wedge T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$. Thus\ $\alpha \in {\ensuremath{\operatorname{Lim}}} \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$. \[<less>\^0\_2nd\_cof.\_prop.\](Second fundamental cofinality property of $<^0$).\ Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega)$.\ Then $\alpha <^0 t + 1 \Longleftarrow \alpha \in {\ensuremath{\operatorname{Lim}}} \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$. Let $\alpha$, $t$ be as stated. Suppose $\alpha \in {\ensuremath{\operatorname{Lim}}} \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$. [[*(\1)]{}]{} We prove by induction: $\forall s \in [\alpha, t + 1] . \alpha \leqslant^0 s$. [[(\2)]{}]{} Let $s \in [\alpha, t + 1]$ and suppose $\forall q \in s \cap [\alpha, t + 1] . \alpha \leqslant^0 q$. [[(IH)]{}]{} Case $s = \alpha$. Then clearly (\2) holds. Case $s \in {\ensuremath{\operatorname{Lim}}} \cap (\alpha, t + 1]$. Since by our (IH) $\forall q \in s \cap [\alpha, t + 1] . \alpha \leqslant^0 q$, then $\alpha \leqslant^0 s$ follows by $\leqslant^0$-continuity. Suppose $s = l + 1 \in (\alpha, t + 1]$. Let $B \subset_{{\ensuremath{\operatorname{fin}}}} l + 1$ be arbitrary. Consider $A {:=}\{\alpha, l\} \cup \{m (a) \| a \in B \cap \alpha \wedge m (a) < \alpha\}$. Then the set $\bigcup_{q \in B \cup A} T (0, \alpha, q) \cap \alpha$ is finite and then, by (\1), there is some $\delta \in {\ensuremath{\operatorname{Class}}} (0) \cap \alpha$ such that $( \bigcup_{q \in B \cup A} T (0, \alpha, q) \cap \alpha) \subset \delta \wedge \delta \leqslant_1 g (0, \alpha, \delta) (t)$. [[(\3)*]{}]{} Consider the function $h : B \longrightarrow h [B] \subset \alpha$ defined as $h (x) {:=}g (0, \alpha, \delta) (x)$. From (\3) and propositions \[Domg(0,a,t)\_and\_T(0,a,t)\] we know that $h$ is well defined; moreover, from proposition \[g(0,a,c)\_prop1\] it follows that $h$ is an ($<, +$)-isomorphism with $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. [[*(\4)]{}]{} Before showing that $h$ is an $<_1$-isomorphism, we do two observations: Let $b \in B$ with $b \geqslant \alpha$. Then $\alpha \leqslant b \leqslant l$, which, together with $\alpha\underset{\text{{\ensuremath{\operatorname{by (IH)}}}}}{\leqslant^0}l$, imply by $\leqslant^0$-connectedness that $\alpha \leqslant^0 b$; subsequently, $\alpha \leqslant_1 b$. This shows $\forall b \in B. \alpha \leqslant b \Longrightarrow \alpha \leqslant_1 b$ [[(\5)]{}]{} Let $b \in B$ with $b \geqslant \alpha$. Then $\alpha \leqslant b \leqslant t$ implies\ $\delta = g(0, \alpha, \delta) (\alpha)\underset{g(0, \alpha, \delta)\text{ strictly increasing} }{\leqslant}g(0, \alpha, \delta)(b)$\ $\underset{g (0,\alpha,\delta) \text{ strictly increasing}}{\leqslant}g (0, \alpha, \delta)(t)$; the latter together with\ $\delta\underset{\text{ by (3)}}{<_1}$ $g (0, \alpha, \delta) (t)$ imply by $\leqslant_1$-connectedness that\ $g (0, \alpha, \delta) (\alpha) = \delta \leqslant_1 g (0, \alpha, \delta) (b)$. All this shows $\forall b \in B. \alpha \leqslant b \Longrightarrow \delta \leqslant_1 g (0, \alpha, \delta) (b)$ [[*(\6)]{}]{}. Now we show that $h$ is an $<_1$-isomorphism. [[(\7)]{}]{} Let $a, b \in B$ with $a < b$. [[Case]{}]{} ${\ensuremath{\boldsymbol{\alpha < a < b}}}$. Then $a <_1 b$ $\underset{\text{by proposition } \ref{g(0,a,c)\|_(a,a2)_is_iso}}{\Longleftrightarrow}$ $h (a) = g (0, \alpha, \delta) (a) <_1 g (0, \alpha, \delta) (b) = h (b)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{a = \alpha < b}}}$. By (\5) and (\6) we have that $\alpha <_1 b$ and $h (\alpha) = g (0, \alpha, \delta) (\alpha) = \delta <_1 g (0, \alpha, \delta) (b) = h (b)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{a, b < \alpha}}}$. Then $a <_1 b$ $\underset{\text{by (4)}}{\Longleftrightarrow}$ $a = h (a) <_1 b = h (b)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{a < \alpha \leqslant b}}}$. $a <_1 b$ $\underset{\text{by } \leqslant_1\text{-connectedness and (5)}}{\Longrightarrow}$ $a <_1 \alpha \leqslant_1 b$ $\underset{\text{by proposition }\ref{g(0,a,c)_prop1} \text{ and by (6)}}{\Longrightarrow}$\ $a = g (0, \alpha, \delta) (a) < g (0, \alpha, \delta) (\alpha) = \delta < \alpha \wedge a <_1 \alpha \wedge \delta \leqslant_1 g (0, \alpha, \delta) (b)$ $\underset{\text{by } \leqslant_1\text{-connectedness}}{\Longrightarrow}$\ $a = g (0, \alpha, \delta) (a) <_1 g (0, \alpha, \delta) (\alpha) = \delta \wedge \delta \leqslant_1 g (0, \alpha, \delta) (b)$ $\underset{\text{by } \leqslant_1\text{-transitivity} }{\Longrightarrow}$\ $h (a) = g (0, \alpha, \delta) (a) <_1 g (0, \alpha, \delta) (b) = h (b)$. $a \nless_1 b \Longrightarrow a \nless_1 \alpha$ (because $a <_1 \alpha$ implies, using (\5), that $a <_1 b$), that is, $a \in B \cap \alpha$ with $m (a) < \alpha$. Then, $m (a)$ $\underset{\text{by (3)}}{<}$ $\delta = g (0, \alpha, \delta) (\alpha)$ $\underset{g (0, \alpha, \delta) \text{ is strictly increasing}}{\leqslant}$ $g (0, \alpha, \delta) (b)$, that is, $h (\alpha) = a \nless_1 g (0, \alpha, \delta) (b) = h (b)$. The previous shows that (\7) holds. In fact, (4\) and (7\) show that (2\) also holds for the case $s = l + 1 \subset (\alpha, t + 1]$ and with this we have concluded the proof of (\2). Hence, the proposition holds. The idea now is that $<_1$ and $<^0$ have something to do with each other. The relation between $<_1$ and $<^0$ is very direct (see next proposition \[prop.\_<less>\^0<less>==<gtr><less>\_1\]); however, when we introduce ${\ensuremath{\operatorname{Class}}} (1)$ (or in general ${\ensuremath{\operatorname{Class}}} (n)$ for $n \in [1, n]$), the way to relate $<_1$ with a relation $<^1$ (or in general $<^n$ for $n \in [1, n]$) will be much harder and will be done through the covering theorem. So, said in other words, the covering theorem for ${\ensuremath{\operatorname{Class}}} (0)$ is trivial and therefore we can prove the next proposition \[prop.\_<less>\^0<less>==<gtr><less>\_1\] without anymore preparations. \[prop.\_<less>\^0<less>==<gtr><less>\_1\]Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega)$. Then $\alpha <^0 t + 1 \Longleftrightarrow \alpha <_1 t + 1$ \ $\Longrightarrow )$. Clear by the definition of $<^0$. $\Longleftarrow )$. Suppose $\alpha <_1 t + 1$. [[(\1)*]{}]{} Note (\1) and proposition \[characterization\_of\_alpha<less>\_1alpha+1\] imply that $\alpha \in {\ensuremath{\operatorname{Lim}}} \mathbbm{P}$ [[*(\2)*]{}]{}. Case $t = \alpha$. Let $B \subset_{{\ensuremath{\operatorname{fin}}}} t + 1 = \alpha + 1$ be arbitrary. Since $B \cap \alpha$ is finite and (2\) holds, then there exists $\delta \in \mathbbm{P}$ such that $B \cap \alpha \subset \delta$. This way, note\ $( \bigcup_{t \in B} T (0, \alpha, t) \cap \alpha) \subset B \cap \alpha \subset \beta$, and then, by proposition \[Domg(0,a,t)\_and\_T(0,a,t)\], the function $h : B \longrightarrow h [B] \subset \alpha$, $h (x) {:=}g (0, \alpha, \delta) (x)$ is well defined. Finally, note that from propositions \[g(0,a,c)\_prop1\] and \[g(0,a,c)\|\_(a,a2)\_is\_iso\] it follows that the function $h$ is an ($<, <_1, +$)-isomorphism with $h\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. Case $t > \alpha$. Let $B \subset_{{\ensuremath{\operatorname{fin}}}} t + 1$ be arbitrary. Consider\ $C {:=}B \cup \{\alpha, 1, \alpha + 1\} \cup \{\alpha m, l, \alpha m + l \| \alpha n + l \in B \wedge m \in [1, n] \wedge l \in [0, \alpha)\} \subset_{{\ensuremath{\operatorname{fin}}}} t + 1$. So, by (\1), there exists $k : C \longrightarrow k [C] \subset \alpha$ an ($<, <_1, +$)-isomorphism with $k\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. [[(\3)]{}]{} Then: 1\. $\alpha <_1 \alpha + 1 \Longleftrightarrow k (\alpha) <_1 k (\alpha + 1) = k (\alpha) + k (1) = k (\alpha) + 1$, that is,\ $k(\alpha)$ $\underset{\text{proposition } \ref{characterization_of_alpha<less>_1alpha+1}}{\in}$ ${\ensuremath{\operatorname{Lim}}} \mathbbm{P}$. 2\. $\forall s \in C \cap \alpha .s < \alpha \Longleftrightarrow s = k (s) < k (\alpha)$ 3\. $\forall n \in [1, \omega) \forall s \in C \cap [\alpha n, \alpha n + \alpha) . - \alpha n + s < \alpha \Longleftrightarrow - \alpha n + s = k (- \alpha n + s) < k (\alpha) \}$ From 1, 2 and 3 follows that $\delta {:=}k (\alpha) \in {\ensuremath{\operatorname{Class}}} (0) \cap \alpha$, $( \bigcup_{t \in C} T (0, \alpha, t) \cap \alpha)$ $\underset{\text{propositions } \ref{g(0,a,c)_prop1} \text{ and } \ref{Domg(0,a,t)_and_T(0,a,t)}}{\subset}$ $\delta$ and that the function $H : C \longrightarrow H [C] \subset \alpha$, $H (x) {:=}g (0, \alpha, \delta) (x)$ is well defined. Moreover, by propositions \[g(0,a,c)\_prop1\] it follows that $H$ is an $(<, +)$-isomorphism with $H\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. [[(\4)]{}]{} Now we show that $H$ is also an $<_1$-isomorphism. [[(\5)]{}]{} Let $a, b \in C$ with $a < b$. [[Case*]{}]{} ${\ensuremath{\boldsymbol{a = \alpha \wedge b \in [\alpha n, \alpha n + \alpha)}}} {{\textbf{for some}}}$ ${\ensuremath{\boldsymbol{n \in [1, \omega)}}}{{\textbf{.}}}$ Then $\alpha <_1 t + 1$ and $\alpha < b < t + 1$ imply by $\leqslant_1$-connectedness that $\alpha <_1 b$.\ On the other hand, note $H (\alpha) = k (\alpha)$ $\underset{\text{by (3)}}{<_1}$ $k (b) = k (\alpha n + ( - \alpha n + b))$ $\underset{\text{by (3)}}{=}$\ $k (\alpha n) + k (- \alpha n + b) = k (\alpha) n + (- \alpha n + b) = H (\alpha) n + H (- \alpha n + b)$ $\underset{\text{by (4)}}{=}$\ $H (\alpha n) + H (- \alpha n + b)$ $\underset{\text{by (4)}}{=}$ $H (\alpha n + ( - \alpha n + b)) = H (b)$. [[(\6)]{}]{} [[Case]{}]{} ${\ensuremath{\boldsymbol{a, b < \alpha}}}$. Then $a <_1 b \Longleftrightarrow a = H (a) <_1 b = H (b)$. [[Case*]{}]{} ${\ensuremath{\boldsymbol{a < \alpha \leqslant b}}}$. Then $a <_1 b$ $\underset{\leqslant_1\text{-connectedness and } \leqslant_1\text{-transitivity}}{\Longleftrightarrow}$ $a <_1 \alpha \leqslant_1 b$ $\underset{\text{by (3) and (6)}}{\Longleftrightarrow}$ $a = H (a) = k (a) <_1 k (\alpha) = H (\alpha) \leqslant_1 k (b) = H (b)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{\alpha < a < b}}}$. Then $a <_1 b$ $\underset{\text{by proposition } \ref{g(0,a,c)\|_(a,a2)_is_iso}}{\Longleftrightarrow}$ $H (\alpha) <_1 H (b)$. The previous shows that (\5) holds. Finally, from (\4), (\5) and the fact that $B \subset C$ we conclude, by proposition \[iso.restriction\] in the appendices section, that the function $H\|_B : B \longrightarrow H\|_B [B] \subset \alpha$, $H\|_B (x) = g (0, \alpha, \delta) (x)$ is an\ ($<, <_1, +$)-isomorphism with $H\|_{\alpha} = {\ensuremath{\operatorname{Id}}}_{\alpha}$. All the previous shows that $\alpha <^0 t + 1$. \[cor\_<less>\^0\_equivalences\]Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega)$. The following are equivalent: $\alpha <^0 t + 1$ $\alpha <_1 t + 1$ $\alpha \in {\ensuremath{\operatorname{Lim}}} \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t)\}$ Direct from previous propositions \[prop.\_<less>\^0<less>==<gtr><less>\_1\], \[<less>\^0\_1st\_cof.\_prop.\] and \[<less>\^0\_2nd\_cof.\_prop.\]. A hierarchy induced by $<_1$ and the intervals $[\omega^{\gamma}, \omega^{\gamma + 1})$. ----------------------------------------------------------------- In this subsection we show theorem \[Class(0)\_Hierarchy\] which is our way to link “solutions of the conditions $\alpha <_1 t + 1$, with $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $t \in [\alpha, \alpha \omega)$” (what below is defined as the $G^0 (t)$ sets) with a thinning procedure (the sets $A^0 (t)$, also defined below). After that, we will see that, for $\alpha = \kappa$ a regular non-countable ordinal, the set of “solutions of the condition $\kappa <_1 t + 1$” is club in $\kappa$. By recursion on $([\omega, \infty), <)$, we define\ $A^0 : [\omega, \infty) \longrightarrow {\ensuremath{\operatorname{Subclasses}}} ({\ensuremath{\operatorname{OR}}})$ in the following way: Let $t \in [\omega, \infty)$ be arbitrary. Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ be such that $t \in [\alpha, \alpha \omega)$. Then\ $A^0 (t) {:=}\left\{ \begin{array}{l} ({\ensuremath{\operatorname{LimClass}}} (0)) \cap (\alpha + 1) \text{ {\ensuremath{\operatorname{iff}}} } t = \alpha\\ {\ensuremath{\operatorname{Lim}}} A^0 (l + 1) \text{ {\ensuremath{\operatorname{iff}}} } t = l + 1\\ {\ensuremath{\operatorname{Lim}}} \{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \| T (0, \alpha, t) \cap \alpha \subset r \wedge \\ \hspace{5ex} r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\} \text{ {\ensuremath{\operatorname{iff}}} } t \in [\alpha, \alpha \omega) \cap {\ensuremath{\operatorname{Lim}}} \end{array} \right.$ $= \left\{ \begin{array}{l} ({\ensuremath{\operatorname{LimClass}}} (0)) \cap (\alpha + 1) \text{ {\ensuremath{\operatorname{iff}}} } t = \alpha\\ {\ensuremath{\operatorname{Lim}}} A^0 (l + 1) \text{ {\ensuremath{\operatorname{iff}}} } t = l + 1\\ {\ensuremath{\operatorname{Lim}}} \{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \|- \alpha n + t < r \wedge \\ r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\} \text{ {\ensuremath{\operatorname{iff}}} } \left\{ \begin{array}{l} t \in [\alpha n, \alpha n + \alpha) \cap {\ensuremath{\operatorname{Lim}}}\\ \text{{\ensuremath{\operatorname{for}}} {\ensuremath{\operatorname{some}}} } n \in [1, \omega) \end{array} \right. \end{array} \right.$ On the other hand, we define $G^0 : [\omega, \infty) \longrightarrow {\ensuremath{\operatorname{Subclasses}}} ({\ensuremath{\operatorname{OR}}})$ as follows: Let $t \in [\omega, \infty)$ be arbitrary. Let $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $n \in [1, \omega)$ be such that $t \in [\alpha n, \alpha n + \alpha)$. Then\ $G^0 (t) {:=}\{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \leqslant \alpha \wedge \beta \leqslant^0 g (0, \alpha, \beta) (t) + 1\}$ $= \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + t < \beta \leqslant \alpha \wedge \beta \leqslant^0 g (0, \alpha, \beta) (t) + 1\}$ $=$, by proposition \[prop.\_<less>\^0<less>==<gtr><less>\_1\], $= \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + t < \beta \leqslant \alpha \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t) + 1\}$. $= \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, t) \cap \alpha \subset \beta \leqslant \alpha \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t) + 1\}$ \[Class(0)\_Hierarchy\]$\forall t \in [\omega, \infty) .G^0 (t) = A^0 (t)$. We show $\forall t \in [\omega, \infty) .G^0 (t) = A^0 (t)$ by induction on $([\omega, \infty), <)$. Let $t \in [\omega, \infty)$ be arbitrary and consider $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ and $n \in [1, \omega)$ such that $t \in [\alpha n, \alpha n + \alpha)$. Suppose $\forall s \in t \cap [\omega, \infty) .G^0 (s) = A^0 (s)$. [[(IH)]{}]{} [[Case]{}]{} ${\ensuremath{\boldsymbol{t = \alpha}}}$. Then $G^0 (\alpha) = \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha + \alpha < \beta \leqslant \alpha \wedge \beta \leqslant_1 g (0, \alpha, \beta) (\alpha) + 1\} =$ $= \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| \alpha \geqslant \beta \leqslant_1 \beta + 1\}$ $\underset{\text{proposition } \ref{characterization_of_alpha<less>_1alpha+1}}{=}$ $({\ensuremath{\operatorname{Lim}}} {\ensuremath{\operatorname{Class}}} (0)) \cap (\alpha + 1) = A^0 (\alpha)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{t = l + 1}}}$ [[for some]{}]{} ${\ensuremath{\boldsymbol{l \in [\alpha n, \alpha n + \alpha)}}}$. Then $G^0 (l + 1) = \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \beta \leqslant \alpha \wedge \beta \leqslant_1 g (0, \alpha, \beta) (l + 1) + 1\}$ $\underset{{\ensuremath{\operatorname{corollary}}} \ref{cor_<less>^0_equivalences}}{=}$\ $\{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \beta \leqslant \alpha \wedge$\ $\beta \in {\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \beta n + g (0, \alpha, \beta) (l + 1) < \gamma \wedge \gamma \leqslant_1 g (0, \beta, \gamma) (g (0, \alpha, \beta) (l + 1))\} \}=$\ $\{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \beta \leqslant \alpha \wedge$\ $\beta \in {\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \beta n + (\beta n + ( - \alpha n + l + 1)) < \gamma \wedge$\ $\gamma \leqslant_1 \gamma n + (- \beta n + (\beta n + ( - \alpha n + l + 1)))\}\}=$\ $\{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \beta \leqslant \alpha \wedge$\ $\beta \in {\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \gamma \wedge \gamma \leqslant_1 \gamma n + ( - \alpha n + l + 1)\}\}=$\ ${\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 \gamma n + ( - \alpha n + l + 1)\}=$\ ${\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + (l + 1) < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (l + 1)\} =$\ ${\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + l < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (l) + 1\} =$\ Lim$G^0 (l)$ $\underset{\text{by (IH)}}{=}$ Lim$A^0 (l) = A^0 (l + 1)$. [[Case]{}]{} ${\ensuremath{\boldsymbol{\alpha < t \in [\alpha n, \alpha n + \alpha) \cap {\ensuremath{\operatorname{Lim}}}}}}$. In order to show $G^0 (t) = A^0 (t)$, we make some preparations first. Note\ $G^0 (t) = \{\beta \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + t < \beta \leqslant \alpha \wedge \beta \leqslant_1 g (0, \alpha, \beta) (t) + 1\} =$, as in the previous case, $= {\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + l < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (t)\}$. [[(\0)]{}]{} On the other hand, let’s show\ $\forall \xi \in {\ensuremath{\operatorname{Class}}} (0) . - \alpha n + t < \xi \leqslant \alpha \wedge \xi \leqslant_1 g (0, \alpha, \gamma) (t) \}\Longrightarrow$\ $ \xi \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \xi\}} A^0 (s)$ [[(\1)]{}]{} Let $\xi \in {\ensuremath{\operatorname{Class}}} (0)$ be such that $- \alpha n + t < \xi \leqslant \alpha \wedge \xi \leqslant_1 g (0, \alpha, \gamma) (t) \}$. [[(\2)*]{}]{}\ Let $s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \xi\}$ be arbitrary and let $m \in [1, n]$ be such that\ $s \in [\alpha m, \alpha m + \alpha)$. Then clearly $- \alpha m + s < \xi \leqslant \alpha$ and\ $\xi \leqslant \xi m + (- \alpha m + s + 1) \leqslant \xi n + (- \alpha n + t) = g (0, \alpha, \gamma) (t)$; the latter implies, by (\2) and $\leqslant_1$-connectedness, $\xi \leqslant_1 \xi m + (- \alpha m + s + 1) = (\xi m + ( - \alpha m + s)) + 1 = g (0, \alpha, \gamma) (s) + 1$. This shows $\xi \in \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha m + s < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (s) + 1\} = G^0 (s)$ $\underset{\text{by our (IH)}}{=}$ $A^0 (s)$ and since this was done for arbitrary $s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \xi\}$, it follows\ $\xi \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \xi\}} A^0 (s)$. Hence (\1) holds. Now we show $\{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + l < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (t)\} =$ $\{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \| - \alpha n + t < r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\}$ [[(\3)*]{}]{} Note from (\1) follows immediately that the contention $'' \subset''$ of (\3) holds. Let’s see that the contention $'' \supset''$ also holds: Let $\beta \in \{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \| - \alpha n + t < r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\}$ be arbitrary. Then $\beta \in {\ensuremath{\operatorname{Class}}} (0) \wedge - \alpha n + l < \beta \leqslant \alpha$ [[(\4)]{}]{} and\ $\beta \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \beta\}} A^0 (s)$ $\underset{\text{by (IH)}}{=}$ $\bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \beta\}} G^0 (s) =$\ $\bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset \beta\}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \alpha, s) \cap \alpha \subset \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (s)\}$ [[(\5)]{}]{}.\ This way, for the sequences $(\delta_s)_{s \in I}$ and $(\xi_s)_{s \in I}$ defined as\ $I : = \left\{ \begin{array}{l} (0, - \alpha n + t) \text{ {\ensuremath{\operatorname{iff}}} } t > \alpha n\\ (0, \beta) \text{ {\ensuremath{\operatorname{iff}}} } t = \alpha n \end{array} \right.$,\ $\delta_s {:=}\left\{ \begin{array}{l} \alpha n + s \text{ {\ensuremath{\operatorname{iff}}} } t > \alpha n\\ \alpha (n - 1) + s \text{ {\ensuremath{\operatorname{iff}}} } t = \alpha n \end{array} \right.$\ and\ $\xi_s {:=}\left\{ \begin{array}{l} \beta n + s \text{{\ensuremath{\operatorname{iff}}}} t > \alpha n\\ \beta (n - 1) + s \text{{\ensuremath{\operatorname{iff}}}} t = \alpha n \end{array} \right.$,\ we have that, by (\4) and (\5),\ $\forall s \in I.T (0, \alpha, \delta_s) \cap \alpha \subset \beta \leqslant_1 g (0, \alpha, \beta) (\delta_s) = \xi_s$ and\ $\xi_s$ $\underset{cof}{{{\lhook\joinrel\relbar\joinrel\rightarrow}}}$ $\left. \left\{ \begin{array}{l} \beta n + (- \alpha n + t) \text{ {\ensuremath{\operatorname{iff}}} } t > \alpha n\\ \beta n \text{ {\ensuremath{\operatorname{iff}}} } t = \alpha n \end{array} \right. \right\} = g (0, \alpha, \beta) (t)$. From all this and using $\leqslant_1$-continuity, we conclude $\alpha \geqslant \beta \in {\ensuremath{\operatorname{Class}}} (0) \wedge - \alpha n + t < \beta \leqslant_1 = g (0, \alpha, \beta) (t)$, that is, $\beta \in \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + t < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (t)\} = G^0 (t)$. Since this was done for arbitrary $\beta \in \{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \| - \alpha n + t < r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\}$, then $'' \supset''$ of (\3) also holds. Finally, it is now very easy to see that $G^0 (t) = A^0 (t)$ holds:\ $G^0 (t)$ $\underset{\text{by (0)}}{=}$ ${\ensuremath{\operatorname{Lim}}} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| - \alpha n + t < \gamma \leqslant \alpha \wedge \gamma \leqslant_1 g (0, \alpha, \gamma) (t)\}$ $\underset{\text{by (3)}}{=}$ $= {\ensuremath{\operatorname{Lim}}} \{r \in {\ensuremath{\operatorname{Class}}} (0) \cap (\alpha + 1) \| - \alpha n + t < r \in \bigcap_{s \in \{q \in (\alpha, t) \| T (0, \alpha, q) \cap \alpha \subset r\}} A^0 (s)\}$\ $= A^0 (t)$. Let $\kappa$ be a regular non-countable ordinal.\ Then $\forall t \in [\kappa, \kappa \omega) .A^0 (t)$ is closed unbounded in $\kappa$. By induction on $([\kappa, \kappa \omega), <)$. One needs to work a little bit with the usual properties of closed unbounded classes. As a final result here, we show that there are ordinals $\alpha \in {\ensuremath{\operatorname{Class}}} (0)$ such that $\alpha <_1 \alpha \omega$. Let $\kappa$ be a regular non-countable ordinal and $\alpha {:=}\min {\ensuremath{\operatorname{Class}}} (0) = \omega$. Then $\bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} A^0 (t) = \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \cap (\kappa + 1) \| \gamma <_1 \gamma \omega\}$. $\{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| \gamma <_1 \gamma \omega\}$ is closed unbounded in $\kappa$. Let $\kappa$ and $\alpha$ be as stated. 1.\ To show $\bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} A^0 (t) \subset \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \cap (\kappa + 1) \| \gamma \leqslant_1 \gamma \omega\}$. [[*(\0)*]{}]{} Let $\beta \in \bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} A^0 (t) = \bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} G^0 (t) =$\ $\bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \| T (0, \kappa, t) \cap \kappa \subset \gamma \leqslant \kappa \wedge \gamma \leqslant_1 g (0, \kappa, \gamma) (t) + 1\}$. Notice from this follows that $\forall n \in [1, \omega) .T (0, \kappa, \kappa n) \cap \kappa \subset \alpha \leqslant \beta \leqslant \kappa \wedge \beta \leqslant_1 g (0, \kappa, \beta) (\kappa n) + 1 = \beta n + 1$; therefore, since the sequence $(\beta n + 1)_{n \in [1, \omega)}$ is cofinal in $\beta \omega$, we get, by $\leqslant_1$-continuity, $\kappa \geqslant \beta \leqslant_1 \beta \omega$. Since this was done for arbitrary $\beta \in \bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} A^0 (t)$, then (\0) follows. To show\ $\bigcap_{t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha} A^0 (t) \supset \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \cap (\kappa + 1) \| \gamma \leqslant_1 \gamma \omega\}$. [[*(\1)]{}]{} Let $\beta \in \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \cap (\kappa + 1) \| \gamma \leqslant_1 \gamma \omega\}$. [[(\2)]{}]{} Let $t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha$ be arbitrary and let $n \in [1, \omega)$ be such that $t \in [\kappa n, \kappa n + \kappa)$. Then $T (0, \kappa, t) \cap \kappa =\{- \kappa n + t\} \subset \alpha \leqslant \beta \leqslant \beta n + (- \kappa n + t) + 1 < \beta (n + 1) < \beta \omega$ and then, by (\2) and $\leqslant_1$-connectedness, we get $T (0, \kappa, t) \cap \kappa \subset \beta \leqslant \kappa \wedge \beta \leqslant_1 \beta n + (- \kappa n + t) + 1 = g (0, \kappa, \beta) (t) + 1$, that is, $\beta \in G^0 (t) = A^0 (t)$. Since this was done for arbitrary $\beta \in \{\gamma \in {\ensuremath{\operatorname{Class}}} (0) \cap (\kappa + 1) \| \gamma \leqslant_1 \gamma \omega\}$ and for arbitrary $t \in [\kappa, \kappa \omega) \wedge T (0, \kappa, t) \cap \kappa \subset \alpha$, then we have shown that (\1) holds. Hence, by (\0) and (\1) the theorem holds. 2.\ Left to the reader. Appendix {#appendix .unnumbered} ======== The main goal of this appendix is to clarify how our definition of $\alpha <_1 \beta$ based on the notion of isomorfisms is equivalent to the assertion that $(\alpha, <, +, <_1)$ is a $\Sigma_1$-substructure of $(\beta, <, +, <_1)$. For this, it will be important the kind of language where one works. In the end, we will achieve our goal by showing theorem \[A<less>\_s\_1B\_equivalence\] which, given certain language $\mathcal{L}$ and corresponding structures $\mathcal{A}$ and $\mathcal{B}$ for it, characterizes when $\mathcal{A}$ is a $\Sigma_1$ substructure of $\mathcal{B}$. So let us first introduce all the notions that we need. The language $\mathcal{L}$ ========================== Sintax ------ In what follows, let us denote as $\bar{R}$, $\bar{f}$ and $\bar{c}$ to a finite set (or list) of relational, functional and constant symbols, respectively. As usual, we call the triad $\left\langle \bar{R}, \bar{f}, \bar{c} \right\rangle$ signature. The terms of $\mathcal{L}$ are build up based on an numerable set of individual variables $\left\{ w_1, w_2, \ldots \right\}$ and on the individual constant symbols $\bar{c}$ as follows (Atomic terms and terms). The atomic terms and terms of our language $\mathcal{L}$ with signature $\left\langle \bar{R}, \bar{f}, \bar{c} \right\rangle$ are defined as: [[Atomic terms]{}]{} - Every variable $w$ in $\left\{ w_1, w_2, \ldots \right\}$ is an atomic term. - Every constant $c$ in $\bar{c}$ is an atomic term. [[Terms]{}]{} - Every atomic term is a term - If $f$ is a functional symbol of arity $n$ and $s_1, \ldots, s_n$ are atomic terms, then $f \left( s_1, \ldots, s_n \right)$ is a term. On the other hand, the formulas of $\mathcal{L}$ are given in the following way (Atomic formulas and formulas). [[Atomic formulas]{}]{} - If $R$ is a relational constant symbol of arity $n$ and and $s_1, \ldots, s_n$ are terms, then $R \left( s_1, \ldots, s_n \right)$ is an atomic formula. - If $t_1$ is a term and $t_2$ is a constant or a variable, then $t_1 \approx t_2$ and $t_2 \approx t_1$ are atomic formulas. [[Formulas]{}]{} - Every atomic formula is a formula - Given the formulas $F_1$ and $F_2$ and the variable $w$, the following are formulas: $F_1 \vee F_2$, $F_1 \wedge F_2$, $F_1 \rightarrow F_2$, $\neg F_1$, $\exists w.F_1$, $\forall w.F_1$ Semantics --------- (Non-closed structures). Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ be a structure for our language $\mathcal{L}$ with the peculiarity that for a universe $U \supset A$, the functions $f^{\mathcal{A}} \in \bar{f}^{\mathcal{A}}$ have domain and codomain $U$. Our interst in these structures lies in the fact that, for $a \in A$, $f^{\mathcal{A}} \left( a \right)$ not necessarily belongs to $A$. We will call $\mathcal{A}$ a non-closed structure for the language $\mathcal{L}$. (On the other hand, for $R^{\mathcal{A}} \in \bar{R}^{\mathcal{A}}$ of arity $n$ and $c^{\mathcal{A}} \in \bar{c}^{\mathcal{A}}$, we require $c^{\mathcal{A}} \in A$ and $R^{\mathcal{A}} \subset A^n$). As usual, for a list of variables $\bar{w} = \left( w_1, \ldots, w_n \right)$ and a list of values $\bar{l} = \left( l_1, \ldots, l_n \right)$ in $A$, we denote as $[ \bar{w} {:=}\bar{l}] : \{w_1, \ldots, w_n \} \rightarrow A$ to the assignment of the variables $\bar{w}$ to the values $\bar{l}$ in $A$. Moreover, for a term $t$, we denote as $t [ \bar{w} {:=}\bar{l}]$ to the usual application of the assignment $[ \bar{w} {:=}\bar{l}]$ to the term $t$. [[Note this value might not lie in $A$.]{}]{} With respect to the satisfiability notions, we treat equality in $\mathcal{L}$ in the cannonical way: it has a fixed interpretation, namely, the identity. In general, we treat the satisfaction of a formula $\mathcal{L}$ by a non-closed structure exaclty in the same way as it is done with structures. Isomorphisms ============ Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures of a language $\mathcal{L}$. An isomorphism between them is a function $h : A \longrightarrow B$ such that: + $h$ is a bijection. + $h (c^{\mathcal{A}}) = c^{\mathcal{B}}$ for any individual constant symbol $c$. + For any functional symbol $f$ of arity $n$ and any $a_1, \ldots, a_n \in A$ $\bullet$ $f^{\mathcal{A}} (a_1, \ldots, a_n) \in A \Longleftrightarrow f^B (h (a_1), \ldots, h (a_n)) \in B$ $\bullet$ If $f^{\mathcal{A}} (a_1, \ldots, a_n) \in A$, then $h (f^{\mathcal{A}} (a_1, \ldots, a_n)) = f^{\mathcal{B}} (h (a_1), \ldots, (a_n))$. + For any $a_1, \ldots, a_n \in A$, and for any relational symbol $R$ of arity $n$, $R^{\mathcal{A}} (a_1, \ldots, a_n) \Longleftrightarrow R^{\mathcal{B}} (h (a_1), \ldots, h (a_n))$. \[h\^-1\_is\_iso\_too\]It is easy to see that in case $h : A \longrightarrow B$ is an isomorphism between $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$, then $h^{- 1} : B \longrightarrow A$ is an isomorphism between $\mathcal{B}= \langle B, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ and $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$. To link assignments of a non-closed structure $\mathcal{A}$ with assignments of another non-closed structure $\mathcal{B}$ which is isomorphic to the former, we introduce the following Let $\mathcal{A}$ be a non-closed structure for a language $\mathcal{L}$. Let $t$ be a term of $\mathcal{L}$ whose free variables are $w_1, \ldots, w_n$ and let $[ \bar{w} {:=}\bar{l}] : \{w_1, \ldots, w_n \} \rightarrow A$ be an assignment of the free variables of $t$ in $A$. We say that ${\ensuremath{\boldsymbol{[ \bar{w} {:=}\bar{l}]}}}$ [[evaluates]{}]{} ${\ensuremath{\boldsymbol{t}}}$ [[in]{}]{} ${\ensuremath{\boldsymbol{A}}}$ whenever $s [ \bar{w} {:=}\bar{l}] \in A$ for any subterm $s$ of $t$ (observe this means also that $t [ \bar{w} {:=}\bar{l}] \in A$). When $t [ \bar{w} {:=}\bar{l}] {\not\in}A$, but $s [ \bar{w} {:=}\bar{l}] \in A$ for any other subterm $s$ of $t$, we say that ${\ensuremath{\boldsymbol{[ \bar{w} {:=}\bar{l}]}}}$ [[quasi evaluates]{}]{} ${\ensuremath{\boldsymbol{t}}}$ [[in]{}]{} ${\ensuremath{\boldsymbol{A}}}$. \[remark\_all\_quasi\_evaluate\]Note that for any term $t$ of our language $\mathcal{L}$ and any non-closed structure $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$, any assignment $[ \bar{w} {:=}\bar{l}] : \{w_1, \ldots, w_n \} \rightarrow A$ of the free variables of $t$ in $A$ is an assignment that quasi evaluates $t$ in $A$. Given the previous notion, we can show \[Isom.Terminos\]Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures of the language $\mathcal{L}$ such that there is an isomorphism $h : A \rightarrow B$. Let $t$ be a term of $\mathcal{L}$ whose free variables are $w_1, \ldots, w_n$ and let $[ \bar{w} {:=}\bar{l}] : \{w_1, \ldots, w_n \} \rightarrow A$ be an assignment that evaluates $t$ in $A$.\ Then $[ \bar{w} {:=}\overline{h (l)}]$ evaluates $t$ in $\mathcal{B}$ and $h (t [ \bar{w} {:=}\bar{l}]) = t [ \bar{w} {:=}\overline{h (l)}]$. By induction over the terms of $\mathcal{L}$.\ + For $t$ a variable or a constant it is direct.\ + Suppose $t$ has the form $f (t_1, \ldots, t_n)$. Note $f (t_1, \ldots, t_n) [ \bar{w} {:=}\bar{l}] = f^{\mathcal{A}} (t_1 [ \bar{w} {:=}\bar{l}], \ldots, t_n [ \bar{w} {:=}\bar{l}]) \in A$ (because $[ \bar{w} {:=}\bar{l}]$ evaluates $f (t_1, \ldots, t_n)$) and since $h$ is an isomorphism, then\ $h (f (t_1, \ldots, t_n) [ \bar{w} {:=}\bar{l}]) =$\ $h (f^{\mathcal{A}} (t_1 [ \bar{w} {:=}\bar{l}], \ldots, t_n [ \bar{w} {:=}\bar{l}])) =$\ $f^{\mathcal{B}} (h (t_1 [ \bar{w} {:=}\bar{l}]), \ldots, h (t_n [ \bar{w} {:=}\bar{l}])) \in B$. (\1) But by induction hypothesis $[ \bar{w} {:=}\overline{h (l)}]$ evaluates $t_i$ in $\mathcal{B}$ for all $i \in \{1, \ldots, n\}$ and\ $h (t_i [ \bar{w} {:=}\bar{l}]) = t_i [ \bar{w} {:=}\overline{h (l)}]$; therefore from (\1) we have that\ $f^{\mathcal{B}} (t_1 [ \bar{w} {:=}\overline{h (l)}]), \ldots, t_n [ \bar{w} {:=}\overline{h (l)}]) = f^{\mathcal{B}} (h (t_1 [ \bar{w} {:=}\bar{l}]), \ldots, h (t_n [ \bar{w} {:=}\bar{l}])) \in B$. (\2) (\2) shows that $[ \bar{w} {:=}\overline{h (l)}]$ evaluates $f (t_1, \ldots, t_n)$ in $B$. Moreover, from (\1) and (\2) we get $h (f (t_1, \ldots, t_n) [ \bar{w} {:=}\bar{l}]) = f^{\mathcal{B}} (t_1 [ \bar{w} {:=}\overline{h (l)}]), \ldots, t_n [ \bar{w} {:=}\overline{h (l)}]) \underset{\text{{\ensuremath{\operatorname{by}}} {\ensuremath{\operatorname{definition}}}}}{=} f (t_1, \ldots, t_n) [ \bar{w} {:=}\overline{h (l)}]$. Isomorphisms and satisfiability ------------------------------- \[A\|=F\[w:=l\]\_iff\_B\|=F\[w:=hl\]\]Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures of $\mathcal{L}$ such that there is an isomorphism $h : A \rightarrow B$. Let $F$ be a formula without quantifiers of $\mathcal{L}$ and suppose $w_1, \ldots, w_n$ are all the free variables in $F$. Let $[ \bar{w} {:=}\bar{l}] : \{w_1, \ldots, w_n \} \rightarrow A$ be an assignment of the free variables of $F$ in $A$. Then $\mathcal{A} \vDash F [ \bar{w} {:=}\bar{l}] \Longleftrightarrow \mathcal{B} \vDash F [ \bar{w} {:=}\overline{h (l)}]$ Let $[ \bar{w} {:=}\bar{l}]$ be an assignment satisfying the assumptions of the theorem. We proceed by induction on the formula $F$. + F is an atomic formula $R (t_1, \ldots, t_n)$. To show $\mathcal{A} \vDash F [ \bar{w} {:=}\bar{l}] \Longrightarrow \mathcal{B} \vDash F [ \bar{w} {:=}\overline{h (l)}]$ (\0) $\mathcal{A} \vDash R (t_1, \ldots, t_n) [ \bar{w} {:=}\bar{l}] \Longleftrightarrow R^{\mathcal{A}} (t_1 [ \bar{w} {:=}\bar{l}], \ldots, t_n [ \bar{w} {:=}\bar{l}])$. (\0.1) But note (\0.1) means, in particular, that $t_1 [ \bar{w} {:=}\bar{l}], \ldots, t_n [ \bar{w} {:=}\bar{l}] \in A$; then, since by remark \[remark\_all\_quasi\_evaluate\] $[ \bar{w} {:=}\bar{l}]$ quasi evaluates all the terms appearing in $F$, it follows that $[ \bar{w} {:=}\bar{l}]$ evaluates $t_1, \ldots, t_n$ in $\mathcal{A}$. From this, (\0.1) and the fact that $h$ is an isomorphism we obtain $R^{\mathcal{B}} (h (t_1 [ \bar{w} {:=}\bar{l}]), \ldots, h (t_n [ \bar{w} {:=}\bar{l}])) \Longleftrightarrow$ (by lemma \[Isom.Terminos\]) $R^{\mathcal{B}} (t_1 [ \bar{w} {:=}\overline{h (l)}], \ldots, t_n [ \bar{w} {:=}\overline{h (l)}]) \Longleftrightarrow \mathcal{B} \vDash F [ \bar{w} {:=}\overline{h (l)}]$. This shows (\0). To show $\mathcal{A} \vDash F [ \bar{w} {:=}\bar{l}] \Longleftarrow \mathcal{B} \vDash F [ \bar{w} {:=}\overline{h (l)}]$ (\0.2) $\mathcal{B} \vDash R (t_1, \ldots, t_n) [ \bar{w} {:=}\overline{h \left( l \right)}] \Longleftrightarrow R^{\mathcal{B}} (t_1 [ \bar{w} {:=}\overline{h \left( l \right)}], \ldots, t_n [ \bar{w} {:=}\overline{h \left( l \right)}])$. (\0.3) Now, since by remark \[remark\_all\_quasi\_evaluate\] an arbitrary proper subterm $s$ of the terms $t_1, \ldots, t_n$ is evaluated by $[ \bar{w} {:=}\bar{l}]$ in $\mathcal{A}$, then by lemma \[Isom.Terminos\] we get that $s$ is evaluated by $[ \bar{w} {:=}\overline{h \left( l \right)}]$ in $\mathcal{B}$. By this and (\0.3) we conclude, just as in the previous case, that $[ \bar{w} {:=}\overline{h \left( l \right)}]$ evaluates $t_1, \ldots, t_n$ in $\mathcal{B}$. This, (\0.3) and the fact that $h^{- 1}$ is an isomorphism (by remark \[h\^-1\_is\_iso\_too\]), imply that $R^{\mathcal{A}} (h^{- 1} (t_1 [ \bar{w} {:=}\overline{h \left( l \right)}]), \ldots, h^{- 1} (t_n [ \bar{w} {:=}\overline{h \left( l \right)}])) \Longleftrightarrow$ (by lemma \[Isom.Terminos\]) $R^{\mathcal{A}} (t_1 [ \bar{w} {:=}\overline{h^{- 1} \left( h (l) \right)}], \ldots, t_n [ \bar{w} {:=}\overline{h^{- 1} \left( h (l) \right)}]) \Longleftrightarrow$ $R^{\mathcal{A}} (t_1 [ \bar{w} {:=}\bar{l}], \ldots, t_n [ \bar{w} {:=}\bar{l}]) \Longleftrightarrow \mathcal{A} \vDash F [ \bar{w} {:=}\bar{l}]$. This shows (\0.2). + $F$ is an atomic formula $t_1 \approx t_2$ with both $t_1$ and $t_2$ being either an individual constant or a variable. Then it follows very easily that $\mathcal{A} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\bar{l}] \Longleftrightarrow \mathcal{B} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\overline{h \left( l \right)}]$. + F is an atomic formula $t_1 \approx t_2$ with $t_2$ a constant or a variable and $t_1$ of the form $f \left( s_1, \ldots, s_q \right)$. Note that by remark \[remark\_all\_quasi\_evaluate\], $[ \bar{w} {:=}\bar{l}]$ evaluates $s_1, s_2, \ldots, s_q, t_2$ in $\mathcal{A}$. (\1)\ Moreover, by (\1) and lemma \[Isom.Terminos\], $[ \bar{w} {:=}\overline{h (l)}]$ evaluates $s_1,, \ldots, s_q, t_2$ in $\mathcal{B}$. (\1.2) To show $\mathcal{A} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\bar{l}] \Longrightarrow \mathcal{B} \vDash \left( t_1 \approx t_2 \right) [ \bar{w} {:=}\overline{h (l)}]$. (\2) $\mathcal{A} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\bar{l}] \Longleftrightarrow t_1 [ \bar{w} {:=}\bar{l}] = t_2 [ \bar{w} {:=}\bar{l}] \underset{\text{by (1)}}{\in} A \Longrightarrow$ $h (t_1 [ \bar{w} {:=}\bar{l}]) = h (t_2 [ \bar{w} {:=}\bar{l}]) \in B \Longrightarrow$ (by (\1), previous line and lemma \[Isom.Terminos\]) $t_1 [ \bar{w} {:=}\overline{h (l)}] = t_2 [ \bar{w} {:=}\overline{h (l)}] \Longrightarrow \mathcal{B} \vDash \left( t_1 \approx t_2 \right) [ \bar{w} {:=}\overline{h (l)}]$. This shows (\2). To show $\mathcal{A} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\bar{l}] \Longleftarrow \mathcal{B} \vDash \left( t_1 \approx t_2 \right) [ \bar{w} {:=}\overline{h (l)}]$. (\3) $\mathcal{B} \vDash \left( t_1 \approx t_2 \right) [ \bar{w} {:=}\overline{h (l)}] \Longrightarrow t_1 [ \bar{w} {:=}\overline{h (l)}] = t_2 [ \bar{w} {:=}\overline{h (l)}] \underset{\text{by (1.2)}}{\in} B \Longrightarrow$ $h^{- 1} \left( t_1 [ \bar{w} {:=}\overline{h (l)}] \right) = h^{- 1} \left( t_2 [ \bar{w} {:=}\overline{h (l)}] \right) \in A \Longrightarrow$ (by (\1.2), previous line, remark \[h\^-1\_is\_iso\_too\] and lemma \[Isom.Terminos\] used with the isomorphism $h^{- 1}$ and the assignment $[ \bar{w} {:=}\overline{h (l)}]$) $t_1 [ \bar{w} {:=}\bar{l}] = t_1 [ \bar{w} {:=}\overline{h^{- 1} \left( h (l) \right)}] = t_2 [ \bar{w} {:=}\overline{h^{- 1} \left( h (l) \right)}] = t_2 [ \bar{w} {:=}\bar{l}] \Longrightarrow$ $\mathcal{A} \vDash (t_1 \approx t_2) [ \bar{w} {:=}\bar{l}]$. This shows (\3). + $F$ is an atomic formula $t_2 \approx t_1$ with $t_2$ a constant or a variable and $t_1$ of the form $f \left( s_1, \ldots, s_q \right)$. The the proof is just as the previous case. + The case for the logical connectives follows immediatly by the induction hypothesis. Substructures and $\Sigma_1$ substructures ========================================== Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle \mathcal{B}, \bar{f}^{\mathcal{B}}, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures for $\mathcal{L}$. $\mathcal{A}$ is (non-closed) substructure of $\mathcal{B}$ iff $A \subset B$, $\bar{f}^{\mathcal{A}} = \bar{f}^{\mathcal{B}} \|_A$, $\bar{R}^{\mathcal{A}} \subset \bar{R}^{\mathcal{B}}$, and $\bar{c}^{\mathcal{A}} = \bar{c}^{\mathcal{B}}$. Let $F$ be a formula of $\mathcal{L}$ such that $s_1, \ldots, s_n$ are terms appearing in $F$. Let $t_1, \ldots, t_n$ be terms. Then we denote by $F \left\langle s_1 {:=}t_1, \ldots, s_n {:=}t_n \right\rangle$ to the formula obtained by the syntactical substitution, for any $i \in \left\{ 1, \ldots, n \right\}$, of all the ocurrences of the term $s_i$ by $t_i$ in $F$. Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ be a non-closed structure for $\mathcal{L}$ and $F$ a formula of $\mathcal{L}$. Suppose the terms $t_1, \ldots, t_n$ appear in $F$. For $l_1, \ldots, l_n \in A$ we define the formula $F \left\langle t_1 {:=}l_1, \ldots, t_n {:=}l_n \right\rangle$ which results by the syntactical substitution, for any $i \in \left\{ 1, \ldots, n \right\}$, of all the ocurrences of the term $t_i$ in $F$ by the constant $l_i$. Note that, formally, $F \left\langle t_1 {:=}l_1, \ldots, t_n {:=}l_n \right\rangle$ does not belong to the language $\mathcal{L}$. The idea is very simple: [[We]{}]{} just [[convey]{}]{} that, any time that we have a formula like $F$, we consider $F \left\langle t_1 {:=}l_1, \ldots, t_n {:=}l_n \right\rangle$ as a “formula of $\mathcal{L}$ with parameters $l_1, \ldots, l_n$”. Any of the parameters $l_i$ is simply an element of $A$ that behaves as a term with a fixed value under any assignment $\left[ \bar{w} {:=}\bar{e} \right]$, namely, $l_i \left[ \bar{w} {:=}\bar{e} \right] = l_i$. Similarly as in the previous definitions, consider the non-closed structures $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$, $\mathcal{B}= \langle B, \bar{f}^B, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ and a formula $F$ of $\mathcal{L}$ with parameters $l_1, \ldots, l_n \in A$. Moreover, for any $i \in \left\{ 1, \ldots, n \right\}$, let $d_i$ be either a term of $\mathcal{L}$ or a parameter $d_i \in B$. Then we denote as $F \left\langle l_1 {:=}d_1, \ldots, l_n {:=}d_n \right\rangle$ to the formula with parameters resulting by the syntactical substitution, for any $i \in \left\{ 1, \ldots, n \right\}$, of all the ocurrences of the parameter $l_i$ by $d_i$ in $F$. The reason of the previous (somewhat annoying) definitions is because we need them to ennunciate the main notion we want to characterize: Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^B, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures for our language $\mathcal{L}$. $\mathcal{A}$ is sigma one non-closed substructure of $\mathcal{B}$, which we abbreviate as usual $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$, if and only if the following two statments hold: $\mathcal{A}$ is (non-closed) substructure of $\mathcal{B}$; For any quantifier free formula $F$ with $n + m$ different free variables\ $x_1, \ldots, x_n, y_1, \ldots, y_m$ and for any $l_1, \ldots, l_m \in A$,\ $\mathcal{B} \vDash \exists x_1, \ldots x_n .F \left\langle y_1 {:=}l_1, \ldots, y_m {:=}l_m \right\rangle\Longleftrightarrow $\ $\mathcal{A} \vDash \exists x_1, \ldots x_n .F \left\langle y_1 {:=}l_1, \ldots, y_m {:=}l_m \right\rangle$. Note that 2. simply states that $\mathcal{B}$ is model of a $\Sigma_1$ sentence of $\mathcal{L}$ with parameters in $\mathcal{A}$ if and only if $\mathcal{A}$ is model of the same sentence. We can finally present the theorem that is our main interest: \[A<less>\_s\_1B\_equivalence\]Let $\mathcal{A}= \langle A, \bar{f}^{\mathcal{A}}, \bar{R}^{\mathcal{A}}, \bar{c}^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{f}^B, \bar{R}^{\mathcal{B}}, \bar{c}^{\mathcal{B}} \rangle$ be non-closed structures for our language $\mathcal{L}$. Then: $\mathcal{A}$ is a $\Sigma_1$ (non-closed) substructure of $\mathcal{B}$ (that is, $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$) $\Longleftrightarrow$ $\mathcal{A}$ is (non-closed) substructure of $\mathcal{B}$ and whenever $X$ is a finite subset of $A$ and $Y$ is a finite subset of $B \backslash A$, there exists a subset $\hat{Y} $ of $A$ and an isomorphism\ $h : X \cup Y \rightarrow X \cup \hat{Y}$ from the non-closed structure\ $\langle X \cup Y, \bar{f}^{\mathcal{B}} \|_{X \cup Y}, \bar{R}^{\mathcal{B}} \|_{X \cup Y}, \bar{c}^B \|_{X \cup Y} \rangle$ to the non-closed structure\ $\langle X \cup \hat{Y}, \bar{f}^{\mathcal{B}} \|_{X \cup \hat{Y}}, \bar{R}^{\mathcal{B}} \|_{X \cup \hat{Y}}, \bar{c}^B \|_{X \cup \hat{Y}} \rangle$ such that $h (x) = x$ for any $x \in X$. Let $\mathcal{L}$, $\mathcal{A}$ and $\mathcal{B}$ be as stated. Since $\mathcal{L}$ is of finite signature, then there exists a natural number $M \in \mathbbm{N}$ such that the arities of all the relational and functional symbols in $\mathcal{L}$ is less or equal to $M$. This way, for any natural numbers $n, m \in \left[ 1, M \right]$, let Rel$_n$ be the set of relational symbols of $\mathcal{L}$ of arity $n$ and let Func$_m$ be the set of functional symbols of $\mathcal{L}$ of arity $n$. Now we show the direction $\Longrightarrow$) of the theorem. [[*(\)]{}]{} Suppose $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$. [[(\0)]{}]{} So $\mathcal{A}$ is a substructure of $\mathcal{B}$ and we only have to prove the isomorphisms-related issue. Let $X \subset_{{\ensuremath{\operatorname{fin}}}} A$ and $Y \subset_{{\ensuremath{\operatorname{fin}}}} B \backslash A$ be arbitrary. Moreover, suppose $X = \left\{ b_1, \ldots, b_l \right\}$ and $Y = \left\{ c_1, \ldots, c_k \right\}$ for some $l, k \in \mathbbm{N}$. Consider the formula $\Gamma$ with parameters in $\mathcal{B}$ defined as $\underset{n \in \left[ 1, M \right] \wedge R \in {\ensuremath{\operatorname{Rel}}}_n \wedge R^{\beta} \left( q \right) {\ensuremath{\operatorname{for}}} q \in \left( X \cup Y \right)^n}{\bigwedge} R \left( q \right) \wedge$ $\underset{n \in \left[ 1, M \right] \wedge R \in {\ensuremath{\operatorname{Rel}}}_n \wedge \neg R^{\beta} \left( q \right) {\ensuremath{\operatorname{for}}} q \in \left( X \cup Y \right)^n}{\bigwedge} \neg R \left( q \right) \wedge$ $\underset{n \in \left[ 1, M \right] \wedge f \in {\ensuremath{\operatorname{Func}}}_n \wedge f^{\beta} \left( q \right) = t {\ensuremath{\operatorname{for}}} q \in \left( X \cup Y \right)^n, t \in X \cup Y}{\bigwedge} f (q) \approx t \wedge$ $\underset{n \in \left[ 1, M \right] \wedge f \in {\ensuremath{\operatorname{Func}}}_n \wedge f^{\beta} \left( q \right) \neq t {\ensuremath{\operatorname{for}}} q \in \left( X \cup Y \right)^n, t \in X \cup Y}{\bigwedge} \neg \left( f (q) \approx t \right)$. Let $z_1, \ldots, z_k$ be $k$ different variables of $\mathcal{L}$. Note $\mathcal{B} \vDash \Gamma$ and therefore $\mathcal{B} \vDash \exists z_1, \ldots z_k . \Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$. But $\Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$ is a $\Sigma_1$ sentence of $\mathcal{L}$ with parameters $b_1, \ldots, b_l$ in $\mathcal{A}$, and therefore, by our hypotheis (\0), $\mathcal{A} \vDash \exists z_1, \ldots z_k . \Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$, which means there is $\left( a_1, \ldots, a_k \right) \in A^k$ such that $\mathcal{A} \vDash \Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle \left[ z_1 {:=}a_1, \ldots, z_k {:=}a_k \right]$. [[*(\1)]{}]{} Let it be $\hat{Y} {:=}\left\{ a_1, \ldots, a_k \right\}$. We now show that the function $h : X \cup Y \xrightarrow[\begin{array}{l} X \ni x \longmapsto x\\ Y \ni c_i \longmapsto a_i \end{array}]{} X \cup \hat{Y}$ is the function we are looking for. $h$ is bijective. $h$ is injective because for $i, j \in \left[ 1, k \right]$, $i \neq j$, the formula $\neg \left( z_i \approx z_j \right)$ is a subformula of $\Gamma$. Moreover, from the definition of $h$ it is clearly surjective.\ Clearly $h \left( x \right) = x$ for any $x \in X$. Now, let $R \in {\ensuremath{\operatorname{Rel}}}_n$, $e_1, \ldots, e_n \in X \cup Y$ and $n \in \left[ 1, M \right]$. To show $R^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) \Longleftrightarrow R^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right)$. [[(\2)]{}]{} $\left. \Longrightarrow \right)$ $R^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) \Longrightarrow R^{\mathcal{B}} \left( e_1, \ldots, e_n \right) \Longrightarrow$ $R \left( e_1, \ldots, e_n \right)$ is\ subformula of $\Gamma$ $\Longrightarrow$ $R \left( e_1, \ldots, e_n \right) \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$ is subformula\ of $\Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$ ${\mathop{\Longrightarrow}\limits_{\text{by (1)}}^{}}$\ $\mathcal{A} \vDash R \left( e_1, \ldots, e_n \right) \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle \left[ z_1 {:=}a_1, \ldots, z_k {:=}a_k \right]$, but note that the latter is exactly the same as $\mathcal{A} \vDash R \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right)$ and so\ $R^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right)$. $\left. \Longleftarrow \right)$ We show $R^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) \Longrightarrow R^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right)$ by contrapositive. Suppose $R^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right)$ doesn’t hold. Then $\neg R \left( e_1, \ldots, e_n \right)$ is subformula of $\Gamma$ and then $\neg R \left( e_1, \ldots, e_n \right) \left\langle c_1 {:=}z_1, \ldots c_k {:=}z_k \right\rangle$ is subformula of $\Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$. Thus, by (\1),\ $\mathcal{A} \vDash \neg R \left( e_1, \ldots, e_n \right) \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle \left[ z_1 {:=}a_1, \ldots, z_k {:=}a_k \right]$ which is exactly the same as $\mathcal{A} \vDash \neg R \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right)$; this way,\ $R^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right)$ doesn’t hold. All the previous shows that (\2) holds. Now, let $f \in {\ensuremath{\operatorname{Func}}}_n$, $e_1, \ldots, e_n \in X \cup Y$ and $n \in \left[ 1, M \right]$. Let’s suppose $f^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) \in X \cup Y.$ [[*(\3)]{}]{} To show $f^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) \in X \cup \hat{Y}$ and $f^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) = h \left( f^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) \right)$. [[(\3.1)]{}]{} By (\3), $f^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) = f^{\mathcal{B}} \left( e_1, \ldots, e_n \right) = d$ for some $d \in X \cup Y$. Then the formula $f \left( e_1, \ldots, e_n \right) \approx d$ is a subformula of $\Gamma$ and therefore $\left( f \left( e_1, \ldots, e_n \right) \approx d \right) \left\langle c_1 {:=}z_1, \ldots c_k {:=}z_k \right\rangle$ is subformula of\ $\Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$. Then, by (\1),\ $\mathcal{A} \vDash \left( f \left( e_1, \ldots, e_n \right) \approx d \right) \left\langle c_1 {:=}z_1, \ldots c_k {:=}z_k \right\rangle \left[ z_1 {:=}a_1, \ldots, z_k {:=}a_k \right]$, which\ means $f^{\mathcal{A}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) = h \left( d \right) \in X \cup \hat{Y}$. Note the latter equality is exactly as the one in (\3.1), since $\mathcal{A}$ is substructure of $\mathcal{B}$. This shows the two assertions in (\3.1). Let’s suppose $f^{\mathcal{B}} \|_{X \cup Y} \left( e_1, \ldots, e_n \right) {\not\in}X \cup Y$. [[(\3.2)]{}]{} To show $f^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) {\not\in}X \cup \hat{Y}$. [[(\3.3)]{}]{} Let $d \in X \cup \hat{Y}$ be arbitrary. Similarly as before, (\3.2) implies that the formula $\neg \left( f \left( e_1, \ldots, e_n \right) \approx h^{- 1} \left( d \right) \right)$ is a subformula of $\Gamma$. So\ $\neg \left( f \left( e_1, \ldots, e_n \right) \approx h^{- 1} \left( d \right) \right) \left\langle c_1 {:=}z_1, \ldots c_k {:=}z_k \right\rangle$ is subformula of\ $\Gamma \left\langle c_1 {:=}z_1, \ldots, c_k {:=}z_k \right\rangle$. Then, by (\1),\ $\mathcal{A} \vDash \neg \left( f \left( e_1, \ldots, e_n \right) \approx h^{- 1} \left( d \right) \right) \left\langle c_1 {:=}z_1, \ldots c_k {:=}z_k \right\rangle \left[ z_1 {:=}a_1, \ldots, z_k {:=}a_k \right]$,\ which means $f^{\mathcal{B}} \|_{X \cup \hat{Y}} \left( h \left( e_1 \right), \ldots, h \left( e_n \right) \right) \neq h \left( h^{- 1} \left( d \right) \right) = d$. Since we have done this for arbitrary $d \in X \cup \hat{Y}$, we have shown (\3.3). All of the previous shows that $h$ is indeed an isomorphism with $h \left( x \right) = x$ for any $x \in X$ and therefore, we have shown (\). Now we show the direction $\Longleftarrow$) of the theorem. [[(\\)]{}]{} So assume the right hand side of the double implication\ asserting the theorem. [[(\4)]{}]{} We want to show that $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$. By hypothesis, $\mathcal{A}$ is a non-closed substructure of $\mathcal{B}$, so it is only left to show that: For any quantifier free formula $F$ with $n + m$ different free variables\ $u_1, \ldots, u_n, w_1, \ldots, w_m$ and for any $l_1, \ldots, l_m \in A$,\ $\mathcal{B} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle \Longleftrightarrow$\ $ \mathcal{A} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ [[(\5)]{}]{} So let’s show (\5). Let $F$ be an arbitrary quantifier free formula whose free (different to each other) variables are $u_1, \ldots, u_n, w_1, \ldots, w_m$ and let it be $l_1, \ldots, l_m \in A$. Clearly $\mathcal{A} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ implies\ $\mathcal{B} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ because $\mathcal{A}$ is a substructure of $\mathcal{B}$. So we actually only have to show that\ $\mathcal{B} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle \Longrightarrow$\ $\mathcal{A} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ [[ *(\6)*]{}]{} To show (\6). Suppose $\mathcal{B} \vDash \exists u_1 \ldots u_n .F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$. [[*(\7)]{}]{} Then there exist $e_1, \ldots, e_n \in B$ such that\ $\mathcal{B} \vDash F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle \left[ u_1 {:=}e_1, \ldots, u_n {:=}e_n \right]$. [[(\8)]{}]{} On the other hand, let $t_1, \ldots, t_q$ be all the terms and subterms appearing in $F \left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ such that $t_1, \ldots, t_q$ are evaluated in $B$ by $\left[ u_1 {:=}e_1, \ldots, u_n {:=}e_n \right]$. Moreover, for any $i \in \left\{ 1, \ldots, q \right\}$, let $b_i \in B$ be such that $t_i \left[ u_1 {:=}e_1, \ldots, u_n {:=}e_n \right] = b_i$. To make more manageable our notation, let’s abbreviate\ $\left\langle w_1 {:=}l_1, \ldots, w_m {:=}l_m \right\rangle$ and $\left[ u_1 {:=}e_1, \ldots, u_n {:=}e_n \right]$ as $\left\langle \bar{w} {:=}\bar{l} \right\rangle$ and $\left[ \bar{u} {:=}\bar{e} \right]$, respectively. Let $Y {:=}\left\{ b_i \|i \in \left\{ 1, \ldots, q \right\} \wedge b_i \in B \backslash A \right\} = \left\{ y_1, \ldots, y_j \right\}$ and\ $X {:=}\left\{ b_i \|i \in \left\{ 1, \ldots, q \right\} \wedge b_i \in A \right\} = \left\{ x_1, \ldots, x_k \right\}$. By (\4), there exists a set $\hat{Y} {:=}\left\{ \hat{y}_1, \ldots, \hat{y}_j \right\} \subset A$ such that the function $\underset{\begin{array}{l} x_i \longmapsto x_i\\ y_i \longmapsto \hat{y}_i \end{array}}{h : X \cup Y \longrightarrow X \cup \hat{Y}}$ is an isomorphism. [[*(\9)*]{}]{} Now observe that (\8) and our definition of $Y$ and $X$ imply that\ $X \cup Y \vDash F \left\langle \bar{w} {:=}\bar{l} \right\rangle \left[ \bar{u} {:=}\bar{e} \right]$, i.e., $X \cup Y \vDash F \left[ \bar{w} {:=}\bar{l}, \bar{u} {:=}\bar{e} \right]$ and therefore, by (\9) and theorem \[A\|=F\[w:=l\]\_iff\_B\|=F\[w:=hl\]\], $X \cup \hat{Y} \vDash F \left[ \bar{w} {:=}\overline{h \left( l \right)}, \bar{u} {:=}\overline{h \left( e \right)} \right]$; so, using that $\forall a \in Y \cap A.h \left( a \right) = a$, we get $X \cup \hat{Y} \vDash F \left\langle \bar{w} {:=}\bar{l} \right\rangle \left[ \bar{u} {:=}\overline{h \left( e \right)} \right]$. But then $X \cup \hat{Y} \vDash \exists u_1 \ldots u_n .F \left\langle \bar{w} {:=}\bar{l} \right\rangle$ and since $X \cup \hat{Y} \subset A$, we conclude $\mathcal{A} \vDash \exists u_1 \ldots u_n .F \left\langle \bar{w} {:=}\bar{l} \right\rangle$. All of the previous shows (\6). Therefore (\5) is also proven and subsequently (\\) has been proven too. This concludes the proof of the whole theorem. In previous theorem \[A<less>\_s\_1B\_equivalence\], it is necessary that $\mathcal{L}$ contains an equality symbol to show that $h : X \cup Y \rightarrow X \cup \hat{Y}$ is injective. To see this, consider a language $L_0$ with has only one binary relation $\sim$. Let $\mathcal{A}= \langle A, \sim^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \sim^{\mathcal{B}} \rangle$ be given as $A {:=}\left\{ 0 \right\}, \sim^{\mathcal{A}} {:=}\left\{ \left( 0, 0 \right) \right\}$, $B {:=}\left\{ 0, 1 \right\}$ and $\sim^{\mathcal{B}} {:=}\left\{ \left( 0, 0 \right), \left( 1, 1 \right), \left( 0, 1 \right), \left( 1, 0 \right) \right\}$. Then one can prove: $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$ Moreover, for $X {:=}\left\{ 0 \right\} \subset_{{\ensuremath{\operatorname{fin}}}} A$ and $Y {:=}\left\{ 1 \right\} \subset_{{\ensuremath{\operatorname{fin}}}} B \backslash A$ the following holds: There exist no $\hat{Y} \subset_{{\ensuremath{\operatorname{fin}}}} A$ and an isomorfism $h : X \cup Y \rightarrow X \cup \hat{Y}$ with $h \left( x \right) = x$ for any $x \in X$. For $\hat{Y} {:=}\emptyset \subset_{{\ensuremath{\operatorname{fin}}}} A$, the function $l : X \cup Y \xrightarrow[x \longmapsto 0]{} X \cup \hat{Y}$ is an homomorphism with $l \left( x \right) = x$ for any $x \in X$. A more classical version of theorem \[A<less>\_s\_1B\_equivalence\], where one does not have to deal with the hassles of considering non-closed structures, can be stated as follows: Let $\mathcal{A}= \langle A, \bar{R}^{\mathcal{A}}, =^{\mathcal{A}} \rangle$ and $\mathcal{B}= \langle B, \bar{R}^{\mathcal{B}}, =^{\mathcal{B}} \rangle$ be structures for a language $\mathcal{L}$ with equality symbol and with a finite number of relational symbols. Then: $\mathcal{A}$ is $\Sigma_1$ (non-closed) substructure of $\mathcal{B}$ (that is, $\mathcal{A} \prec_{\Sigma_1} \mathcal{B}$) $\Longleftrightarrow$ $\mathcal{A}$ is substructure of $\mathcal{B}$ and whenever $X$ is a finite subset of $A$ and $Y$ is a finite subset of $B \backslash A$, there exists a subset $\hat{Y} $ of $A$ and an isomorphism\ $h : X \cup Y \rightarrow X \cup \hat{Y}$ from $\left\langle X \cup Y, \bar{R}^{\mathcal{B}} \|_{X \cup Y} \right\rangle$ to $\left\langle X \cup \hat{Y}, \bar{R}^{\mathcal{B}} \|_{X \cup \hat{Y}} \right\rangle$ such that $h (x) = x$ for any $x \in X$. Finally, let us state a final proposition that is very useful while working with Carlson’s $<_1$-relation. \[iso.restriction\]Let $(C, \overline{R^{}}^C, \overline{f^{}}^C, \bar{c})$, $(Q, \overline{R^{}}^Q, \bar{f}^Q, \bar{q})$ be structures of a language $L$. Suppose\ $(B, \overline{R^{}}^B, \overline{f^{}}^B, \bar{b}) \subset (C, \overline{R^{}}^C, \overline{f^{}}^C, \bar{c})$, that is, $B \subset C$, $R^B = R^C \cap B^n$ for any n-ary relation $R^C$, $f^B = f^C \|_B$ for any function $f^C$ and any distinguished element $b$ of $B$ is a distinguished element of $C$. Suppose $h : (C, \overline{R^{}}^C, \overline{f^{}}^C, \bar{c}) \longrightarrow (h [C], \overline{R^{}}^{h [C]}, \overline{f^{}}^{h [C]}, \overline{h (c)}) \subset (Q, \overline{R^{}}^Q, \bar{f}^Q, \bar{q})$ is an isomorphism. Then $h\|_B : (B, \overline{R^{}}^B, \overline{f^{}}^B, \bar{b}) \longrightarrow (h [B], \overline{R^{}}^{h [B]}, \overline{f^{}}^{h [B]}, \overline{h (b)})$ is an isomorphism. For any $a_1, \ldots, a_n \in B$ and any relation $R^B$ we have $R^B (a_1, \ldots, a_n) \Longleftrightarrow R^C (a_1, \ldots, a_n) \Longleftrightarrow R^{h [C]} (h (a_1), \ldots, h (a_n)) \Longleftrightarrow R^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n))$. Clearly $b \in B$ is a distinguished element iff $h (b) = h\|_B (b) \in h [B]$ is a distinguished element. Let’s see that the operations behave also correctly (of course the problem is with the closure of such operations): Let $a_1, \ldots, a_n \in B$. Suppose $f^C (a_1, \ldots, a_n) = f^B (a_1, \ldots, a_n) \in B$. Then\ $f^{h [C]} (h (a_1), \ldots, h (a_n)) \in h [C]$ and $f^{h [C]} (h (a_1), \ldots, h (a_n)) = h (f^C (a_1, \ldots, a_n)) = h (f^B (a_1, \ldots, a_n))$. Clearly $h (a_1), \ldots, h (a_n) \in h [B] \subset h [C]$ and so from the previous equalities we have $f^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n)) = f^{h [C]} (h (a_1), \ldots, h (a_n)) = h (f^B (a_1, \ldots, a_n)) \in h [B]$. Now suppose $f^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n)) \in h [B]$. Then there exists $a \in B \subset C$ such that $h (a) = f^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n))$. [[(A)]{}]{} On the other hand, $f^{h [C]} (h (a_1), \ldots, h (a_n)) = f^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n)) \in h [B] \subset h [C]$; then $f^C (a_1, \ldots, a_n)) \in C$ and\ $h (f^C (a_1, \ldots, a_n)) = f^{h [C]} (h (a_1), \ldots, h (a_n)) = f^{h [B]} (h\|_B (a_1), \ldots, h\|_B (a_n))$. From this and (A) we have found that $h (f^C (a_1, \ldots, a_n)) = h (a)$ and therefore, since $h$ is bijective, $f^C (a_1, \ldots, a_n) = a \in B$. [1]{} W. Buchholz. A new system of proof-theoretic ordinal functions. [[[Annals of Pure and Applied Logic]{}]{}]{}, 1986. T. Carlson. [[[Knowledge, machines, and the consistency of Reinhardt’s strong mechanistic thesis]{}]{}]{}. Arch. Math. Logic, 1999. T. Carlson. Ordinal arithmetic and $\sum_1$-elementarity. [[[Archive for Mathematical Logic]{}]{}]{}, 1999. T. Carlson. Elementary patterns of resemblance. [[[Annals of Pure and Applied Logic]{}]{}]{}, 2001. G. Wilken. [[[The Bachmann-Howard structure in terms of $\sum_1$-elementarity]{}]{}]{}. Archive for Mathematical Logic, 2006. G. Wilken. [[[Ordinal Arithmetic based on Skolem hulling]{}]{}]{}. Annals of Pure and Applied Logic, 2007. G. Wilken. [[[$\sum_1$-elementarity and Skolem hull operators]{}*]{}]{}. Annals of Pure and Applied Logic, 2007.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The production of $W$ or $Z$ bosons in association with two jets is an important background to the Higgs boson search in vector-boson fusion at the LHC. The purely electroweak component of this background is dominated by vector-boson fusion, which exhibits kinematic distributions very similar to the Higgs boson signal. We consider the next-to-leading order QCD corrections to the electroweak production of $\ell \nul jj$ and $\ell^+\ell^- jj$ events at the LHC, within typical vector-boson fusion cuts. We show that the QCD corrections are modest, increasing the total cross sections by about 10%. Remaining scale uncertainties are below 2%. A fully-flexible next-to-leading order partonic Monte Carlo program allows to demonstrate these features for cross sections within typical vector-boson-fusion acceptance cuts. Modest corrections are also found for distributions.' --- 0.9cm QCD corrections to electroweak\ $\ell\nul jj$ and $\ell^+\ell^- jj$ production 0.5cm [ Carlo Oleari$^1$ and Dieter Zeppenfeld$^2$ ]{} .2cm [$^1$ [Department of Physics, University of Durham, South Road, Durham DH1 3LE, UK]{}]{}\ [$^2$ [Department of Physics, University of Wisconsin, Madison, WI 53706, USA ]{}]{}\ 1.3cm Introduction ============ Vector-boson fusion (VBF) processes have emerged as a particularly interesting class of scattering events from which one hopes to gain insight into the dynamics of electroweak symmetry breaking. The most prominent example is Higgs boson production via VBF, that is, the process $qq\,\to\, qqH$, which can be viewed as quark scattering via $t$-channel exchange of a weak boson, with the Higgs boson radiated off the $W$ or $Z$ propagator. Alternatively, one may view this process as two weak bosons fusing to form the Higgs boson. The kinematic characteristics of this process are very distinctive: two jets, in the forward and backward region of rapidity, with the Higgs boson decay products in the central region of the detector. This characteristic signature greatly helps to distinguish these $Hjj$ events from backgrounds. Higgs boson production via VBF has been studied intensively as a tool for Higgs boson discovery [@ATLAS; @CMS] and the measurement of Higgs boson couplings [@Zeppenfeld:2000td] in $pp$ collisions at the CERN Large Hadron Collider (LHC). Analogous to Higgs boson production via VBF, the electroweak production of a $W$ or $Z$ plus two jets, with the requirement that the weak boson is centrally produced and that the two jets are well separated in rapidity, will proceed with sizable cross section at the LHC[^1]. The decay leptons in $W \,\to\,\ell\nul$ and $Z\,\to\,\ell^+\ell^-$ lead to the final states $\ell \nul jj$ and $\ell^+\ell^-jj$ ($\ell=e,\mu,\tau$). These processes have already been considered in the literature at leading order (LO). To name but a few examples, they have been studied in the investigation of rapidity gaps at hadron colliders [@Chehime:1992ub; @Rainwater:1996ud; @Khoze:2002fa], as a probe of anomalous triple-gauge-boson couplings [@Baur:1993fv] or as a background to Higgs boson searches in VBF [@wbfhtautau; @wbfhtoww; @Eboli:2000ze]. In this last case, the $\ell\nul jj$ final state with an unidentified charged lepton, or $\nul\bar\nul jj$ events from $Z\to \nul\bar\nul$ decay, form a background to invisible Higgs boson decay (see e.g. Ref. [@Eboli:2000ze]). $\tau^+\tau^-jj$ events are a background to the decay $H\,\to\,\tau^+\tau^-$ [@wbfhtautau], and also to $H\,\to\, W^+ W^-$ when the $W$’s and the $\tau$’s decay leptonically [@wbfhtoww]. In these examples, off-shell corrections to $Z\,\to\,\tau^+\tau^-$ decay need to be included, since a Higgs boson mass in the range $114~{\rm GeV} < m_H < 200~{\rm GeV}$, well above the $Z$ peak, is favored by electroweak data [@Charlton:2001am]. While a LO analysis is perfectly adequate for exploratory investigation, precision measurements at the LHC require comparison with cross-section predictions which include higher-order QCD corrections. A poignant example is the extraction of Higgs boson couplings, where expected accuracies of the order of 10%, or even better [@Zeppenfeld:2000td], clearly require knowledge of the next-to-leading order (NLO) QCD corrections. In addition, one would like to exploit $W$ and $Z$ production, in VBF configurations, as calibration processes for Higgs boson production via VBF, namely as a tool to understand the tagging of forward jets or the distribution and veto of additional central jets in VBF (see e.g. Ref. [@Rainwater:1996ud; @Khoze:2002fa]). In fact, these processes share the same color structure: two colored quarks are scattered via the exchange of a colorless boson in the $t$-channel. The pattern of soft gluon radiation is then the same. Understanding the gap-survival probability in the known case of $W$ and $Z$ production can give insight on the gap survival for the case of Higgs boson production. The precision needed for Higgs boson studies and for the understanding of radiation patterns then requires the knowledge of NLO QCD corrections for $Wjj$ and $Zjj$ production as well. The NLO QCD corrections to the total $Hjj$ cross section from VBF has been known for many years [@Han:1992hr]. In a recent paper [@Figy:2003nv], we presented the calculation of these corrections in the form of a fully-flexible parton-level Monte Carlo program which allows the determination of NLO corrections to arbitrary (infrared-safe) distributions. Here, we extend this work and describe the calculation and first results for such corrections to $Wjj$ and $Zjj$ production in VBF configurations. To be precise, since the decaying weak bosons are spin-one particles, in order to retain all the possible angular correlations between the final state particles, we consider the electroweak processes $pp\,\to\, \ell^\pm \nul jjX$ and $pp\,\to\, \ell^ +\ell^-jjX$ at NLO. At LO, Feynman graphs for one such process, $uc\,\to\, dcW^+,W^+\,\to\, \ell^+\nul$, are shown in Fig. \[fig:feynBorn\]. Using the terminology introduced in [@Boudjema:1996qg], we consider bremsstrahlung (a, b, c), fusion (d) and multiperipheral (e, f) diagrams. We neglect diagrams corresponding to conversion, abelian and non-abelian annihilation, since these $q\bar q$ annihilation contributions are negligible when we impose VBF cuts, as explained in detail in Sec. \[sec:approximations\]. In the following, in order to use a shorthand notation, we will call processes such as the one depicted in Fig. \[fig:feynBorn\] “EW $Vjj$ production”, or VBF production of $W/Z$ plus two jets, since we consider these processes with the kinematic cuts typical for the selection of VBF (see Sec. \[sec:pheno\]). It should be understood that, in spite of this notation, multiperipheral diagrams like (e) and (f) are included, even though they cannot be represented as the production of a weak boson, followed by its decay into two leptons. The structure of the paper is as follows: in Sec. \[sec:calculation\], we outline the calculation of the tree-level diagrams, of real-emission contributions and of the virtual corrections. We dedicate Sec. \[sec:virtual\] to the discussion of the virtual contributions, with some of the analytical details relegated to Appendix \[sec:appendix\]. A list of checks which we have performed on our calculation concludes Sec. \[sec:calculation\]. Additional features of our Monte Carlo program, like the gauge invariant handling of finite $W$ and $Z$ widths, the inclusion of anomalous $WW\gamma$ and $WWZ$ couplings, the approximations with regard to crossed diagrams in the presence of identical quark flavors, the singularities for incoming photons and the choice of parameters, will be discussed in Sec. \[sec:MC\]. We then use this Monte Carlo program to present first results for EW $Vjj$ production at the LHC. Of particular concern is the scale dependence of the NLO results, which provides an estimate for the residual theoretical error of our cross-section calculations. We discuss the scale dependence and the size of the radiative corrections for various distributions in Sec. \[sec:pheno\]. Conclusions are given in Sec. \[sec:summary\]. Elements of the calculation {#sec:calculation} =========================== The structure of the three processes under consideration, $pp\,\to\,\ell^+\nul jjX$, $pp\,\to\,\ell^-\bar\nul jjX$ and $pp\,\to\, \ell^+\ell^-jjX$, is very similar. A discussion of any single one of them is sufficient to clarify our procedures for all, and we use $W^+$ production, i.e., the calculation of the $pp\,\to\,\ell^+\nul jjX$ cross section, for this purpose. [Mutatis mutandis]{}, all the considerations apply to the other processes too. Approximations and general framework {#sec:approximations} ------------------------------------ At tree level, the topological structure of the generic subprocesses contributing to EW $Wjj$ production is depicted in Fig. \[fig:feynBorn\]. Two additional classes of diagrams appear in case of identical quark flavors on two of the fermion lines: - diagrams where both the two virtual vector bosons are time-like. They correspond to diagrams called conversion, abelian and non-abelian annihilation in Ref. [@Boudjema:1996qg], and contain vector-boson pair production with subsequent decay of one of the weak bosons to a pair of jets. [Pars pro toto]{}, we call this class vector-boson pair production in the following. - diagrams obtained by interchange of identical initial- or final-state (anti)quarks, such as in the $uu\,\to\, du\ell^+\nul$ or $d u \, \to \, d d \ell^+ \nul$ subprocesses. These additional diagrams are obtained from the ones shown in Fig. \[fig:feynBorn\] by crossing. In our calculation, we have neglected contributions from vector-boson pair production completely. In addition, any interference effects of the second class with the graphs of Fig. \[fig:feynBorn\] are neglected. This is justified because, in the phase-space region where VBF can be observed experimentally, with widely-separated quark jets of very large invariant mass, the neglected terms are strongly suppressed by large momentum transfer in one or more weak-boson propagators. Color suppression further reduces any interference terms. We have checked with MadEvent [@madgraph] that, at LO, the diagrams that we have not considered and interference effects contribute less than 0.3% to our final results in e.g. Fig. \[fig:scale\_depW\]. Since we expect QCD corrections to the neglected terms to be modest, the above approximations are fully justified within the accuracy of our NLO calculation. Fermion masses are set to zero throughout, because observation of either leptons or (light) quarks in a hadron-collider environment requires large transverse momenta and hence sizable scattering angles and relativistic energies. For the $t$-channel processes which we include, we have used a diagonal form (equal to the identity matrix) for the Cabibbo-Kobayashi-Maskawa matrix, $V_{CKM}$. This approximation is not a limitation of our calculation. As long as no final-state quark flavor is tagged (no $c$ tagging is done, for example), the sum over all flavors, using the exact $V_{CKM}$, is equivalent to our results, due to the unitarity of the $V_{CKM}$ matrix. Tree-level diagrams and real corrections ---------------------------------------- For the $Wjj$ Born amplitude, we need to add the contributions from the 10 Feynman graphs shown in Fig. \[fig:feynBorn\] ($Z$ and $\gamma$ propagators counted as different diagrams), and sum cross sections of all subprocesses producing $W^+$ plus two jets. The same is true for $W^-$ production. For the case of $Zjj$ production, amplitudes which correspond to neutral-current exchange (no change of quark flavors) receive contributions from 24 Feynman graphs at tree level. To obtain the real-emission diagrams, with a final-state gluon, one needs to attach the gluon to the quark lines in all possible ways. For the diagrams in Fig. \[fig:feynBorn\], this gives rise to 45 real-emission graphs. 112 different Feynman graphs contribute to real-emission corrections to $Zjj$ production via neutral-current exchange. The contributions with an initial-state gluon are obtained by crossing the previous diagrams, promoting the final-state gluon as incoming parton, and an initial-state (anti)quark as final-state particle. We again remove all diagrams where two time-like, final-state vector bosons appear such as $g u \,\to \,\ell^+ \nul d Z^$, with $Z^ \,\to\, c \bar c$. Such diagrams, for consistency, must be removed since we have not considered the corresponding Born contributions. Figure \[fig:feyndrop\] clarifies this issue: we drop all initial-gluon contributions in which the gluon couples to the fermion line of the initial quark or antiquark. In fact, these diagrams are strongly suppressed when VBF cuts (see Sec. \[sec:pheno\]) are applied to the final-state jets. Our Monte Carlo program computes all amplitudes numerically, using the formalism of Ref. [@HZ]. The Born amplitudes for $W$ and $Z$ production are taken from Ref. [@Chehime:1992ub]. The real-emission amplitudes for $Z$ production were first given in Ref. [@Rainwater:1996ud]. The corresponding amplitudes for $W$ production were partially programed at the time. We have finalized and tested them for the present application. Virtual corrections {#sec:virtual} ------------------- At NLO, we have to deal with soft and collinear singularities in the virtual and real-emission contributions. Our calculation uses the subtraction method of Catani and Seymour [@CS] to cancel the soft and collinear divergences between virtual and real-emission diagrams. Since these divergences only depend on the color structure of the external partons, the subtraction terms encountered for EW $Vjj$ production are identical in form to those found for Higgs boson production in VBF. Thus, we can use the results described in Ref. [@Figy:2003nv] for the case at hand. The main difference is that the finite parts of the virtual corrections are more complicated than for $Hjj$ production (where only vertex corrections were present). The QCD corrections to EW $Vjj$ production appear as two gauge-invariant subsets, corresponding to corrections to the upper and lower fermion lines in Fig. \[fig:feynBorn\]. Due to the color singlet nature of the exchanged electroweak bosons, any interference terms between subamplitudes with gluons attached to both the upper and the lower quark lines vanish identically at order $\alpha_s$. Hence, it is sufficient to consider radiative corrections to a single quark line only, which we here take as the upper one. Corrections to the lower fermion line are an exact copy. In computing the virtual corrections, we have used the dimensional reduction scheme [@DR_citation]: we have performed the Passarino-Veltman reduction of the tensor integrals in $d=4-2\ep$ dimensions, while the algebra of the Dirac gamma matrices, of the external momenta and of the polarization vectors has been performed in $d=4$ dimensions. We split the virtual corrections into two classes: the virtual corrections along a quark line with only one weak boson attached and the virtual corrections along a quark line with two weak bosons attached. [I.]{} The virtual NLO QCD contribution to any tree level Feynman subamplitude ${\cal M}_B^{(i)}$ which has a single electroweak boson $V$ (of momentum $q$) attached to the upper fermion line, q(k\_1) q(k\_2) + V(q) , appears in the form of a vertex correction, which is factorisable in terms of the original Born subamplitude \[eq:vertexvirt\] [M]{}\_V\^[(i)]{} = [M]{}\_B\^[(i)]{} C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) --+c\_[virt]{} +. Here $\mu_R$ is the renormalization scale, and the boson virtuality $Q^2 = -(k_1-k_2)^2 = - q^2$ is the only relevant scale in the process, since the quarks are assumed to be massless, $k_1^2 = k_2^2=0$. In dimensional reduction, the finite contribution is given by $c_{\rm virt}=\pi^2/3-7$ ($c_{\rm virt}=\pi^2/3-8$ in conventional dimensional regularization). [II.]{} The second class of diagrams are the virtual QCD corrections to the Feynman graphs where two electroweak bosons $V_1$ and $V_2$ (of outgoing momenta $q_1$ and $q_2$) are attached to the same fermion line (see, for example, the upper quark line in Fig. \[fig:feynBorn\] (a, b)). It suffices to consider one of the two possible permutations of $V_1$ and $V_2$, as depicted in Fig. \[fig:boxline\]. The kinematics is given by q(k\_1) q(k\_2)+V\_1(q\_1)+V\_2(q\_2), where $k_1^2=k_2^2=0$ and momentum conservation reads $k_1=k_2+q_1+q_2$. In the following, it is convenient to use the Mandelstam variables for a $2\,\to\, 2$ process which we take as $q\bar q \, \to \, V_1 V_2$. We then define \[eq:kinematics\] s = (k\_1-k\_2)\^2 = (q\_1+q\_2)\^2,t = (k\_1-q\_1)\^2 = (k\_2+q\_2)\^2,u = (k\_1-q\_2)\^2 = (k\_2+q\_1)\^2. In order to use the same notation as in Eq. (\[eq:vertexvirt\]), we define $Q^2=2k_1\cdot k_2\equiv-s$. The two electroweak bosons are always virtual in our calculation, i.e., the effective polarization vectors $\eps_1(q_1)$ and $\eps_2(q_2)$ actually correspond to fermion currents (the charm-quark current and the leptonic-decay currents in the Feynman graphs of Fig. \[fig:feynBorn\] (a, b)). Since fermion masses are neglected, current conservation implies transversity of the effective polarization vectors: $\eps_1\cdot q_1 =\eps_2\cdot q_2 =0$. The expressions that we give in Appendix \[sec:appendix\] exploit this relationship. Our numerical code permits to switch on the missing $\eps_1\cdot q_1$ and $\eps_2\cdot q_2$ terms, allowing us to test gauge invariance. Due to the trivial color structure of the corresponding tree-level diagram, the divergent part (soft and collinear singularities) of the sum of the four diagrams in Fig. \[fig:boxline\] is a multiple of this Born subamplitude, just like for the vertex corrections, \[eq:boxlinefig\] [M]{}\_[boxline]{}\^[(i)]{} &=& [M]{}\_B\^[(i)]{} C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) --+c\_[virt]{}\ &+& C\_F \_(q\_1,q\_2)(-e\^2)g\_\^[V\_1f\_1]{}g\_\^[V\_2f\_2]{} +[O]{}() . Here $\tau$ denotes the quark chirality and the electroweak couplings $g_{\tau}^{Vf}$ follow the notation of Ref. [@HZ], with, e.g., $g_\pm^{\gamma f}=Q_f$, the fermion electric charge in units of $\|e\|$, $g_-^{Wf}=1/(\sqrt{2}\sin\theta_W)$ and $g_-^{Zf}=(T_{3f}-Q_f\sin^2\theta_W)/ (\sin\theta_W\cos\theta_W)$, where $\theta_W$ is the weak mixing angle and $T_{3f}$ is the third component of the isospin of the (left-handed) fermions. A finite contribution of the virtual diagrams, which is proportional to the Born amplitude (the $c_{\rm virt}$ term), is pulled out in correspondence with Eq. (\[eq:vertexvirt\]). The remaining non-universal term, $\tilde{\cal M}_\tau(q_1,q_2)$, is also finite and can be expressed in terms of the finite parts of the Passarino-Veltman $B_0$, $C_0$ and $D_{ij}$ functions, which we denote as $\tilde B_0$, $\tilde C_0$ and $\tilde D_{ij}$. Analytical expressions for these functions, along with the expression for $\tilde{\cal M}_\tau(q_1,q_2)$, are given in Appendix \[sec:appendix\]. An equivalent form for Eq. (\[eq:boxlinefig\]) has been derived where all the $\tilde D_{ij}$ have been reduced to $\tilde B_0$, $\tilde C_0$ and $\tilde D_{0}$ functions. We have checked numerically that the two expressions agree within the numerical precision of the two FORTRAN codes. The factorization of the divergent contributions to the virtual subamplitudes, as multiples of $\MB^{(i)}$, implies that the overall infrared and collinear divergence multiplies the complete Born amplitude (the sum of the Feynman graphs of Fig. \[fig:feynBorn\]). We can summarize this result for the virtual corrections to the upper fermion line by writing the complete virtual amplitude ${\cal M}_V$ as \[eq:box\_tri\_contrib\] [M]{}\_V &=& [M]{}\_B C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) --+c\_[virt]{}\ &+& C\_F(-e\^2)+[O]{}()\ &=& [M]{}\_B C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) --+c\_[virt]{} +\_V , where $\tilde{\cal M}_V$ is finite. The interference contribution in the cross-section calculation is then given by \[eq:virtual\_born\] 2 \_V\^\* = \|\|\^2 C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) --+c\_[virt]{}+2 \_V\^\* . This expression replaces the analogous result for the NLO corrections to $qq\to qqH$, Eq. (2.11) in Ref. [@Figy:2003nv]. The divergent piece appears as the same multiple of the Born amplitude squared as in the $qq\to qqH$ cross section. It cancels explicitly against the phase-space integral of the dipole terms (see Ref. [@CS] and Eq. (2.10) of Ref. [@Figy:2003nv]) \[eq:I\] () = \|\|\^2 C\_F $\frac{4\pi\mu_R^2}{Q^2}$\^(1+) ++9-\^2, which absorbs the real-emission singularities. After this cancellation, all remaining integrals are finite and can, hence, be evaluated in $d=4$ dimensions. This means that the values of $\MB$ and $\tilde{\cal M}_V$ need to be computed in 4 dimensions only and we use the amplitude techniques of Ref. [@HZ] to obtain them numerically. Checks ------ We have verified, both analytically and numerically, the gauge invariance of Eq. (\[eq:box\_tri\_contrib\]): once the extra and $\eps_2\cdot q_2$ terms have been re-inserted in this expression, the individual finite subamplitudes $\tilde{\cal M}_\tau(q_i,q_j)$ vanish upon the replacements $\eps_1\,\to\,q_1$ or $\eps_2\,\to\, q_2$. This is a strong check of the tensor reduction and manipulation of the virtual contributions depicted in Fig. \[fig:boxline\]. We have taken the Born amplitudes for $W$ and $Z$ production from Ref. [@Chehime:1992ub] and use the real-emission amplitudes of Ref. [@Rainwater:1996ud] for $Z$ production. In addition, the $Zjj$ results at the Born level were successfully checked with COMPHEP code [@ilyin]. For $W$ production, the real-emission amplitudes were obtained by modifying the previously tested $Zjjj$ amplitudes [@Rainwater:1996ud]. We have generated equivalent amplitudes with MadGraph [@madgraph], checking their consistency numerically. For the $W^{+}$ case, we have built two totally-independent codes. This has allowed us to check the overall structure of the dipole-formalism terms (common to all the vector-boson fusion processes), and to compare tree-level, real-emission and virtual amplitudes. The two codes agree within the numerical precision of the two FORTRAN programs for the total cross sections and for final-state kinematic distributions. The parton-level Monte Carlo {#sec:MC} ============================ The cross-section contributions discussed above have been implemented in a parton-level Monte Carlo program for $\ell^+\nul jj$, $\ell^-\bar\nul jj$ and $\ell^+\ell^-jj$ production at NLO in QCD, which is very similar to the program for $Hjj$ production by weak-boson fusion described in Ref. [@Figy:2003nv]. As in our previous work, the tree-level and the finite parts of the virtual amplitudes are calculated numerically, using the helicity-amplitude formalism of Ref. [@HZ]. The Monte Carlo integration is performed with a modified version of VEGAS [@vegas]. While many aspects of our present calculation are completely analogous to those described in Ref. [@Figy:2003nv], several new problems appear for the vector-boson production processes which require explanation. In order to deal with $W/Z$ boson decay W/Z(p\_[\_1]{}+p\_[\_2]{}) \_1(p\_[\_1]{}) + \_2(p\_[\_2]{}), we have to introduce a finite $W/Z$ width, $\Gamma_V$, in the resonant poles of the $s$-channel weak-boson propagators. However, in the presence of non-resonant graphs, like those of Figs. \[fig:feynBorn\](e) and (f), this introduces changes in a subclass of Feynman graphs only, which leads to a violation of electroweak gauge invariance, which is guaranteed for the zero-width amplitudes. Such non-gauge-invariant finite-width effects can lead to huge unphysical enhancements at very small photon virtuality and should be avoided [@Argyres:1995ym]. For the case at hand, transverse-momentum cuts on the two final-state tagging jets (see Sec. \[sec:pheno\]) largely eliminate the dangerous phase-space regions with low-virtuality gauge bosons. Nevertheless, a careful handling of the finite-width effects is called for. We have accomplished this using two different schemes.\ [I.]{} In [the overall-factor scheme]{} [@Baur:1991pp], one multiplies all the diagrams shown in Fig. \[fig:feynBorn\], and all virtual and real-emission contributions as well, by an overall factor \[eq:overall\_scheme\] , where $\Gamma_V$ has been assumed to be constant. This way, close to resonance \[($p_{\ell_1} + p_{\ell_2})^2 \sim m_V^2$\], where the sum of the diagrams is dominated by the vector-boson propagator, we recover the result of the resonance approximation. Away from resonance, and, thus, in a subdominant phase-space region, the error that we make, by multiplying all the diagrams by the factor in Eq. (\[eq:overall\_scheme\]), is of the order of $\Gamma_V/m_V\approx 2.7\%$, for both $Z$ and $W$ boson production. The advantage of this scheme is that it preserves full $SU(2)\times U(1)$ gauge invariance, since the gauge-invariant set of zero-width diagrams is multiplied by an overall factor. [II.]{} In [the complex-mass scheme]{} [@Denner:1999gp], one globally replaces $m_V^2 \rightarrow m_V^2 -i m_V\Gamma_V$, also in the definition of the weak mixing angle, $\sin^2\theta_W=1-m_W^2/m_Z^2$. We have implemented a modified complex-mass scheme where we replace $m_V^2 \rightarrow m_V^2 -i m_V\Gamma_V$ in the weak-boson propagators appearing in Fig. \[fig:feynBorn\], but we keep a real value for $\sin^2\theta_W$. With this prescription, the electromagnetic Ward identity relating the tree-level triple-gauge-boson vertex, $-ie\Gamma_{WW\gamma}^{\alpha\beta\mu}$, and the inverse $W$ propagator, $(D_W)^{-1}_{\alpha\beta}(q)$, is preserved [@lopez] $$(q_1-q_2)_\mu \Gamma_{WW\gamma}^{\alpha\beta\mu} = i(D_W)^{-1}_{\alpha\beta}(q_1) - i(D_W)^{-1}_{\alpha\beta}(q_2) \;.$$ This relation removes potential problems with small $q^2$ photon propagators, where gauge-invariance-violating terms, proportional to $\Gamma_W/m_W$, may be enhanced by factors $E_T^2/q^2$, where the hard scale $E_T$ is set by typical transverse momenta of the process. The corresponding enhancement for $Z$-boson propagators is of order $E_T^2/(\|q^2\|+m_Z^2)$ and, hence, small for the energies available at the LHC. Also, we note that the imaginary part of $\sin^2\theta_W=1-(m_W^2-im_W\Gamma_W)/(m_Z^2-im_Z\Gamma_Z)$, in the full complex-mass scheme, is 200 times smaller than the real part and hence well below the naive expectation $\Gamma_V/m_V\approx 2.7\%$ for the size of finite-width corrections. We have used the two different schemes to compute total cross sections with VBF cuts and find agreement at the level of the 0.5% or better. This ambiguity, thus, represents a minor contribution to higher-order electroweak corrections. Inspection of the Feynman graphs of Fig. \[fig:feynBorn\] shows that the non-abelian triple-gauge-boson vertices (TGV) enter via the $WWZ$ and $WW\gamma$ couplings in diagrams like Fig. \[fig:feynBorn\] (d). These graphs receive QCD vertex corrections only and, therefore, factorize according to Eq. (\[eq:vertexvirt\]) in terms of the tree-level TGV graphs, independent of the form of the TGV. In particular, the presence of anomalous $WWZ$ or $WW\gamma$ couplings can easily be taken into account by a simple modification of the Born amplitude. Our program supports anomalous couplings $\kappa_\gamma$, $\kappa_Z$, $\lambda_\gamma$, $\lambda_Z$ etc. [@Hagiwara:1986vm] and thus allows to extend the analysis of anomalous-coupling effects in vector-boson fusion processes [@Baur:1993fv] to NLO QCD accuracy. The requirement of two observable jets, of finite transverse momentum (see Sec. \[sec:pheno\]), is sufficient to render the LO cross section for EW $Wjj$ and $Zjj$ events finite. At NLO, initial-state collinear singularities appear. For $g\to q\bar q$ and $q\to qg$ splitting, these are properly taken into account via the renormalization of quark and gluon distribution functions. An additional collinear divergence exists, however, because of the presence of $t$-channel photons in tree-level graphs, such as in Fig. \[fig:feynBorn\] (a, b, d, e). Real-emission corrections lead to Feynman graphs such as the one shown in Fig. \[fig:feyndrop\] (d): the final-state $d$ and $\bar u$ quarks may lead to observable jets, allowing vanishing momentum transfer for the virtual photon and a corresponding collinear singularity, representing, in the case shown, a QED correction to the LO process $g\gamma\to d\bar u W^+$. This singularity would have to be absorbed into the renormalization of the photon distribution function inside the proton. Alternatively, one may impose a cut, $\|t\|>Q^2_{\gamma,\rm min}$, on the virtuality of the photon and replace the missing piece by the $p\gamma\to VjjX$ cross section, folded with the appropriate photon density in the proton [@Baur:1991pp; @kniehl]. We have chosen this latter approach: all divergent amplitudes are set to zero below $Q^2_{\gamma,\rm min}=4$ GeV$^2$ and $p\gamma\to VjjX$ is considered to be a separate electroweak contribution to $Vjj$ events, which we do not calculate here.\ When imposing typical VBF cuts, with their large-rapidity separation and concomitant invariant mass of the two tagging jets, the $p\gamma\to VjjX$ contribution to the EW $Vjj$ cross section is quite small. For the VBF cuts defined in the next section, with $p_{Tj}>20$ GeV and a rapidity separation of the two tagging jets of $\Delta y_{jj}>4$, the NLO $W^+jj$ cross section, for example, increases by a mere 0.2% when lowering the photon cutoff to $Q^2_{\gamma,\rm min}=0.1$ GeV$^2$ from our 4 GeV$^2$ default value[^2]. This number increases to 0.7% for $\Delta y_{jj}>2$. Because these contributions are negligible, we have not yet implemented the calculation of this small missing piece in our program. In the computation of cross sections and distributions presented below, we have used the CTEQ6M parton distribution functions (PDFs) [@cteq6] with $\alpha_s(m_Z)=0.118$ for all NLO results and CTEQ6L1 parton distributions for all LO cross sections. The CTEQ6 fits include $b$ quarks as an active flavor. For consistency, the $b$ quark is included as an initial- and/or final-state massless parton in all neutral-current processes, i.e., we include only those processes with external $b$ quarks, where no internal top-quark propagator appears via the $btW$ vertex, being forbidden by Feynman rules. Top-quark contributions, obviously, go beyond our massless-fermion approximation and would have to be treated as a separate process. Allowed neutral-current processes with $b$ quarks appear for $Z$ production only. The $b$-quark contributions are quite small, however, affecting the $Z$-boson production cross section at the $1$% level only. We choose $m_Z=91.188$ GeV, $m_W=80.419$ GeV and the measured value of $G_F$ as our electroweak input parameters, from which we obtain $\alpha_{QED}=1/132.51$ and $\sin^2\theta_W=0.2223$, using LO electroweak relations. The decay widths are then calculated as $\Gamma_W=2.099$ GeV and $\Gamma_Z=2.510$ GeV, which agrees with their Particle Data Group [@PDG] values at the level of 0.9% and 0.6% respectively, which is better than the overall theoretical uncertainty we are striving for. In order to reconstruct jets from the final-state partons, the $k_T$ algorithm [@kToriginal], as described in Ref. [@kTrunII], is used, with resolution parameter $D=0.8$. Results for the LHC {#sec:pheno} =================== The parton-level Monte Carlo program described in the previous section has been used to determine the size of the NLO QCD corrections to EW $Vjj$ cross sections at the LHC. Using the $k_T$ algorithm, we calculate the partonic cross sections for events with at least two hard jets, which are required to have \[eq:cuts1\] p\_[Tj]{} 20 [GeV]{} , \|y\_j\| 4.5 . Here $y_j$ denotes the rapidity of the (massive) jet momentum which is reconstructed as the four-vector sum of massless partons of pseudorapidity $\|\eta\|<5$. The two reconstructed jets of highest transverse momentum are called “tagging jets” and are identified with the final-state quarks which are characteristic for vector-boson fusion processes. We consider decays $Z\to \ell^+\ell^-$ and $W\to \ell\nul$ into a single generation of leptons. In order to ensure that the charged leptons are well observable, we impose the lepton cuts \[eq:cuts2\] p\_[T]{} 20 [GeV]{} ,\|\_\| 2.5 , R\_[j]{} 0.4 , where $R_{j\ell}$ denotes the jet-lepton separation in the rapidity-azimuthal angle plane. In addition, the charged leptons are required to fall between the rapidities of the two tagging jets, \[eq:cuts3\] y\_[j,min]{} < \_4 . Cross sections, within the cuts of Eqs. (\[eq:cuts1\])–(\[eq:cuts4\]), are shown in Fig. \[fig:scale\_depW\], for $Wjj$ production, and in Fig. \[fig:scale\_depZ\], for the $Zjj$ case. In both figures, the scale dependence of the LO and NLO cross sections is shown for fixed renormalization and factorization scales, $\mu_R$ and $\mu_F$, which are tied to the masses of the produced vector bosons $m_V$ \[eq:scale.mV\] \_R = \_Rm\_V,\_F = \_Fm\_V . The LO cross sections only depend on $\mu_F=\xi\,m_V$. At NLO we show three cases: (a) $\xi_F=\xi_R=\xi$ (red solid line); (b) $\xi_F=\xi$, $\xi_R=1$ (blue dot-dashed line); and (c) $\xi_R=\xi$, $\xi_F=1$ (green dashed line). While the factorization-scale dependence of the LO result is sizable, the NLO cross sections are quite insensitive to scale variations: allowing a factor 2 variation in either directions, i.e., considering the range $0.5<\xi <2$, the NLO cross sections change by less than 1% in all cases. As a second option, we have considered scales tied to the virtuality of the exchanged electroweak bosons. Specifically, independent scales $Q_i$ are determined as in Eqs. (\[eq:vertexvirt\]) and (\[eq:boxlinefig\]) for radiative corrections on the upper and on the lower quark line, and we set \[eq:scale.Q\] \_[Fi]{}=\_F Q\_i ,\_[Ri]{}=\_R Q\_i. This choice is motivated by the picture of VBF as two independent deep-inelastic scattering type events, with independent radiative corrections on the two electroweak-boson vertices. Resulting $Vjj$ cross sections at NLO are about 1% lower for $\mu_F=\mu_R=Q_i$ than for $\mu_F=\mu_R=m_V$. In the following, we refer to the latter choice as the “$M$ scheme” while the choice $\mu_F=\mu_R=Q_i$ is called the “$Q$ scheme”. As we will see below, a residual NLO scale dependence of about 1%–2% is also typical for distributions, resulting in very stable NLO predictions for $Vjj$ cross sections. In addition to these quite small scale uncertainties, we have estimated the error of the $W^\pm jj$ cross sections due to uncertainties in the determination of the PDFs. This error is determined by calculating the total $Wjj$ cross section, within the cuts of Eqs. (\[eq:cuts1\])–(\[eq:cuts4\]), using two different sets of PDFs with errors, computed by the CTEQ [@cteq6] and MRST [@mrst] Collaborations. Together with the PDF that gives the best fit to the data, the CTEQ6M set provides 40 PDFs, and the MRST2001E 30 PDFs, which correspond to extremal plus-minus variations in the directions of the error eigenvectors of the Hessian, in the space of the fitting parameters. To be on the conservative side, we have added the maximum deviations for each error eigenvector in quadrature, and we have found a total PDF uncertainty of $\pm 4\%$ with the CTEQ PDFs, and of roughly $\pm 2\%$ with the MRST set. For precise comparisons with future LHC data, the residual theoretical error on the jet and lepton distributions must be estimated. As a first example, we show the transverse-momentum distribution of the highest-$p_T$ tagging jet for $W^+jj$ production in Fig. \[fig:pt\_max\_tagj\] (a): the shape of the $p_T$ distribution is fairly similar at LO (red dashed curve) and NLO (black solid line). Both curves were obtained with a scale choice of $\mu_R=\mu_F=m_W$. In the right-hand panel their ratio to the NLO curve with $\mu_R=\mu_F=Q_i$ is shown. The ratio of the two NLO distributions deviates from unity by 2% or less over the entire range, which, again, points to the small QCD dependence of our calculation. In contrast to the stability of the NLO result, the LO curves depend appreciably on the scale choice. The blue dotted line and the red dashed line in Fig. \[fig:pt\_max\_tagj\] (b) give the ratio of the LO curves for $\mu_F=Q_i$ and $\mu_F=m_W$, respectively, to the NLO result. The shape of the LO curves, in particular for a constant scale choice like $\mu_F=m_W$, is quite different from the more reliable NLO result. For transverse-momentum distributions we generally find that the “dynamical” scale choice $\mu_F=Q_i$, at LO, better reproduces the shape of the NLO distributions, and is thus preferable to a fixed scale. At NLO, or higher order, where the definition of the momentum transfer $Q_i$ becomes more problematic, the fixed-scale choice becomes more natural. However, because of the greater stability of the cross-section prediction, the scale selection also becomes less of a phenomenological issue. Rapidity distributions of the two tagging jets are shown in Fig. \[fig:y\_tags\], at LO and NLO, and for two choices of the rapidity-gap requirement, $\Delta y_{jj}>2$ and $\Delta y_{jj}>4$. The shapes of the rapidity distributions for the more central tagging jet, panel (a), and the more forward tagging jet, panel (b), are quite similar at LO and NLO. In fact, the $K$ factors for these distributions are fairly flat, and adequately described by a constant value of about 1.1. The results in Fig. \[fig:y\_tags\] were obtained for a fixed scale $\mu_F=\mu_R=m_W$ and are for $W^-jj$ production. Curves for the $W^+jj$ and $Zjj$ cross sections are very similar in shape and show the preservation of shape between LO and NLO curves. While tagging-jet distributions are quite similar for electroweak $Wjj$ and $Zjj$ events at the LHC, the presence of two charged leptons in the $Zjj$ case results in somewhat more noticeable differences. When considering changes in the lepton $p_T$ cut of Eq. (\[eq:cuts2\]), the transverse momentum of the softer lepton is critical for $Z$ production, while the single charged lepton must be considered for $Wjj$ events. These distributions are shown in Fig. \[fig:pt\_L\_min\] for $W^+$ production (top panels) and $Z$ production (bottom panels). At NLO the scale variations are again very small, at the 1% level, as demonstrated by the ratios of the NLO $p_T$ distributions for $\mu_F=\mu_R=m_V$ and $\mu_F=\mu_R=Q_i$ (solid black lines) in Fig. \[fig:pt\_L\_min\] (b, d). Varying either scales by a factor of 2 leads to the same conclusion of 1%–2% scale uncertainties for the NLO results. Comparing the LO predictions (dashed and dot-dashed curves) with the very precise NLO results shows theoretical errors of the order of 10%. Again, as for the jet $p_T$ distributions discussed earlier, the choice $\mu_F=Q_i$ is better for simulating the shape of the lepton $p_T$ distribution at LO. A fixed scale, $\mu_F=m_V$, predicts too steep a fall-off at large $p_T$. One should note, however, that for the electroweak $Vjj$ processes considered here, these differences are exceptionally small already at LO: the differences between the LO curves in Fig. \[fig:pt\_L\_min\] are of the order of 10% only. In contrast to the lepton transverse-momentum distributions described above, the shape of the lepton-rapidity distributions is virtually unaffected by the NLO corrections: an overall constant $K$ factor is sufficient to describe NLO effects. Larger changes are found when considering angular correlations of the leptons and jets, which we show for $Zjj$ production in Fig. \[fig:delta\_y\_L\_tag\_min\]. The top panels show the minimal rapidity between any of the two leptons and the two tagging jets, $\Delta y_{{\rm tag},l}^{\rm min}$. As before, the tagging jets are taken as the two highest transverse-momentum jets in the event ($p_T$ selection). The two bottom panels show the minimal separation in the rapidity-azimuthal angle plane of the two leptons from any jet (not necessarily the two tagging jets) in the event, $R_{j,l}^{\rm min}$. In both cases, the two scale choices for the NLO result show excellent agreement (black solid lines in Fig. \[fig:delta\_y\_L\_tag\_min\] (b, d)). However, the dynamical $K$ factors K(x) = for $x=\Delta y_{{\rm tag},l}^{\rm min}$ and $x=R_{j,l}^{\rm min}$ show qualitatively different behavior. While $K(\Delta y_{{\rm tag},l}^{\rm min})$ is fairly constant, i.e., the shape of the distribution is well described by the LO approximation, the minimal lepton-jet separation, $d\sigma/dR_{j,l}^{\rm min}$, shifts noticeably to smaller values at NLO. This behavior was to be expected, since additional parton emission in the higher-order calculation reduces lepton isolation. What is remarkable, then, is that the selection of the tagging jets as the two highest-$p_T$ jets does not affect the lepton-tagging jet separation. As for the Higgs boson case [@Figy:2003nv], this selection of the tagging jets provides excellent correspondence of the LO- and NLO-event topology. In order to stress this point we show dijet invariant-mass distributions for the reconstructed jets (not necessarily the two tagging jets) for $W^+jj$ events at LO (red dashed lines) and at NLO (solid black lines) in Fig. \[fig:m\_jj\_min\_Wp\]. The distribution with respect to the minimal dijet invariant mass in the event is shown in Fig. \[fig:m\_jj\_min\_Wp\] (a) while Fig. \[fig:m\_jj\_min\_Wp\] (b) uses the invariant mass of the two tagging jets, $m_{\rm tags}$. At LO, there are only two final-state quarks of $p_T>20$ GeV in each event and, hence, the two curves are identical. At NLO, additional parton emission provides for soft third jets which form low invariant-mass pairs with one of the tagging jets, and this pair shows up as a low-mass peak in $d\sigma/dm_{jj}^{\rm min}$. Generic selections of the two tagging jets in a multijet environment tend to pick up some of these low-mass pairs and lead to substantial differences in the invariant-mass distribution of the two tagging jets at LO and at NLO. The $p_T$ selection of tagging jets, which we have used throughout and for which results are shown in Fig. \[fig:m\_jj\_min\_Wp\] (b), is remarkable in that it preserves the shape of the tagging jet invariant-mass distribution, $d\sigma/dm_{\rm tags}$, when going from LO to NLO. Conclusions {#sec:summary} =========== Vector-boson fusion at the LHC represents a class of electroweak processes which are under excellent control perturbatively. This has been known for some time for the most interesting process in this class: Higgs boson production via VBF has a modest $K$ factor of about 1.05 for the inclusive production cross section [@Han:1992hr] and this result also holds when applying realistic acceptance cuts [@Figy:2003nv]. In the present paper, we have extended this result to the electroweak production of $W$ and $Z$ plus two jets, when the final-state particles are in a kinematic configuration typical of VBF events. More precisely, we have calculated the NLO QCD corrections to electroweak production of $\ell \nul jj$ and $\ell^+\ell^- jj$ at LHC, and we have implemented them in a fully-flexible NLO Monte Carlo program. $K$ factors are of the same size as for the Higgs boson production process, typically ranging between 1.0 and 1.1 for most distributions. What is more important is the stability of the NLO result: residual scale dependence is at the 2% level or below. This is smaller than the present parton-distribution-function uncertainties, which we have calculated for the $W^\pm jj$ cross sections. We estimate 4% PDF errors using CTEQ6M parton distributions and roughly half that size using MRST2001E PDFs. Given the excellent theoretical control which we now have for EW $Vjj$ production, these processes can be used as testing grounds for Higgs boson production in VBF: techniques should be developed to measure hadronic properties, like forward-jet tagging efficiencies or central-jet-veto probabilities, in $Wjj$ or $Zjj$ production at the LHC and to extrapolate these results to Higgs boson production, thus reducing the systematic errors for Higgs boson coupling measurements. We leave such applications for the future. Acknowledgments {#acknowledgments .unnumbered} =============== Part of this work was done at LAPTH in Annecy and D.Z. would like to thank the members of the laboratoire for their hospitality. This research was supported in part by the University of Wisconsin Research Committee with funds granted by the Wisconsin Alumni Research Foundation and in part by the U.S. Department of Energy under Contract No. DE-FG02-95ER40896. C.O. thanks the UK Particle Physics and Astronomy Research Council for supporting his research. Virtual corrections {#sec:appendix} =================== In this appendix, we give the expression for the finite, reduced amplitude $\tilde{\cal M}_\tau(q_1,q_2)$ that appears in Eqs. (\[eq:boxlinefig\]) and (\[eq:box\_tri\_contrib\]), in terms of $\tilde B_0$, $\tilde C_0$ and $\tilde D_{ij}$ functions. Here $\tilde B_0$, $\tilde C_0$ and $\tilde D_{ij}$ are the finite parts of the Passarino-Veltman $B_0$, $C_0$ and $D_{ij}$ functions [@Passarino:1978jh], and are given explicitly below. We have also derived $\tilde{\cal M}_\tau(q_1,q_2)$ in terms of $\tilde B_0$, $\tilde C_0$ and $\tilde D_{0}$ functions, but do not show this expression here, due to its length. We write \[eq:boxlinedef\] \_(q\_1,q\_2) = (k\_2) c\_1 \_1 +c\_2\_2 +c\_q$\sla q_1-\sla q_2$ + c\_b\_2$\sla k_2+\sla q_2$ \_1 (k\_1), where $\eps_1=\eps_1(q_1)$ and $\eps_2=\eps_2(q_2)$ are the effective polarization vectors of the two electroweak gauge bosons. The coefficient function $c_1=c_1(q_1,q_2)$ is given by c\_1 &=& 2T\_$q_2^2,t$ -2 +$q_1^2+q_2^2-3 s-4 t$\ &-& 2 -$q_2^2-t$\ &+& 4 , where T\_$q^2,t$ = B\_0(t)-B\_0(q\^2) +2B\_0(q\^2)+1-2q\^2C\_0(q\^2,t) is defined in terms of the finite parts of the $B_0$ and $C_0$ functions \[eq:b0tilde\] B\_0(q\^2) = 2- and \[eq:c0tilde\] C\_0(q\^2,t) = (\^2- \^2). These expressions are obtained by pulling a common factor $\Gamma(1+\eps)(-s)^{-\eps}\equiv{\Gamma(1+\eps)}/(Q^2)^\eps$ out of all amplitudes and Passarino-Veltman functions, e.g., B\_0(q\^2) &=& = (-q\^2-i0\^+)\^[-]{}\ &=& = . For the other coefficient functions $c_i=c_i(q_1,q_2)$ we find c\_2&=& -2+$q_1^2+q_2^2-s-2 t$ +$q_2^2-s-3 t$\ &+& 4\ &+& 2T\_$q_1^2,t$, c\_q &=& +s + 2 , c\_b &=& -2 { +-2 $q_2^2-t$\ &+& $q_1^2+2 q_2^2$ - $2 q_1^2+q_2^2$ }- 2 s\ &+& 2 {+-2 t - 2\ &+& q\_2\^2 - q\_1\^2- 6 $\xd{312}-\xd{313}$}\ &-& 1t T\_b(q\_1\^2,t)+T\_b(q\_2\^2,t)+B\_0(t) -5+ , with T\_b(q\^2,t) = 2q\^2B\_0(t)-B\_0(q\^2)+tB\_0(t)-q\^2B\_0(q\^2)-2q\^2C\_0(q\^2,t) . For the crossed function $\tilde{\cal M}(q_2,q_1)$, the same expressions as above apply, with the obvious interchange $q_1\leftrightarrow q_2$, $\eps_1\leftrightarrow \eps_2$, and $t\,\to\, u$. The finite part of the $D_0$ function is defined by \[eq:d0tilde\] D\_0(k\_2,q\_2,q\_1) = \^2 +4\_2$1-\frac{t}{q_1^2}$ +4\_2$1-\frac{t}{q_2^2}$- . This expression is well defined when all invariants, $q_1^2$, $q_2^2$ and $t$, are space-like. In our application, we always have one space-like and one time-like weak boson, i.e., exactly one of the two quotients $t/q_i^2$ is positive. In the other quotient simply replace the time-like invariant by $t\,\to \, t+i0^+$ or $q_i^2\,\to\, q_i^2+i0^+$, as in Eqs. (\[eq:b0tilde\]) and (\[eq:c0tilde\]). The remaining finite $\tilde D_{ij}$ functions are obtained from the above expressions for the $\tilde B_0$, $\tilde C_0$, and $\tilde D_0$ functions with the usual Passarino-Veltman recursion relations given in Ref. [@Passarino:1978jh], adapted to the Bjorken-Drell metric, $q_i^2>0$ for a time-like momentum $q_i$. In these recursion relations we need the additional finite $\tilde B_0$ and $\tilde C_0$ functions B\_0(0) &=& 0,\ C\_0(k\_2,q\_1+q\_2) &=& C\_0(s,0,0) = , while C\_0(q\_1,q\_2) = C\_0(q\_1\^2,q\_2\^2,s) is the infrared- and ultraviolet-finite $C_0$ function for massless internal propagators but with nonzero invariants $q_1^2$, $q_2^2$ and $s$. [99]{} ATLAS Collaboration, ATLAS TDR, Report No. CERN/LHCC/99-15 (1999); E. Richter-Was and M. Sapinski, Acta Phys. Pol. B [30]{}, 1001 (1999); B. P. Kersevan and E. Richter-Was, Eur. Phys. J. C [25]{}, 379 (2002) \[arXiv:hep-ph/0203148\]. G. L. Bayatian [et al.]{}, CMS Technical Proposal, Report No. CERN/LHCC/94-38x (1994); R. Kinnunen and D. Denegri, CMS Note No. 1997/057; R. Kinnunen and A. Nikitenko, Report No. CMS TN/97-106; R. Kinnunen and D. Denegri, arXiv:hep-ph/9907291; V. Drollinger, T. Müller and D. Denegri, arXiv:hep-ph/0111312. D. Zeppenfeld, R. Kinnunen, A. Nikitenko and E. Richter-Was, Phys. Rev. D [62]{}, 013009 (2000) \[arXiv:hep-ph/0002036\]; D. Zeppenfeld, in [Proc. of the APS/DPF/DPB Summer Study on the Future of Particle Physics, Snowmass, 2001]{} edited by N. Graf, eConf [C010630]{}, p. 123 (2001) \[arXiv:hep-ph/0203123\]; A. Belyaev and L. Reina, JHEP [0208]{}, 041 (2002) \[arXiv:hep-ph/0205270\]. D. L. Rainwater, arXiv:hep-ph/9908378. J. Campbell and R. K. Ellis, Phys. Rev. D [65]{}, 113007 (2002) \[arXiv:hep-ph/0202176\]; J. Campbell, R. K. Ellis and D. Rainwater, Phys. Rev. D [68]{}, 094021 (2003) \[arXiv:hep-ph/0308195\]. H. Chehime and D. Zeppenfeld, Phys. Rev. D [47]{}, 3898 (1993). D. Rainwater, R. Szalapski and D. Zeppenfeld, Phys. Rev. D [54]{}, 6680 (1996) \[arXiv:hep-ph/9605444\]. V. A. Khoze, M. G. Ryskin, W. J. Stirling and P. H. Williams, Eur. Phys. J. C [26]{}, 429 (2003) \[arXiv:hep-ph/0207365\]. U. Baur and D. Zeppenfeld, arXiv:hep-ph/9309227. D. Rainwater, D. Zeppenfeld and K. Hagiwara, Phys. Rev. D [59]{}, 014037 (1999) \[arXiv:hep-ph/9808468\]; T. Plehn, D. Rainwater and D. Zeppenfeld, Phys. Rev. D [61]{}, 093005 (2000) \[arXiv:hep-ph/9911385\]; S. Asai et al., Report No. ATL-PHYS-2003-005. D. Rainwater and D. Zeppenfeld, Phys. Rev. D [60]{}, 113004 (1999) \[Erratum-ibid. D [61]{}, 099901 (2000)\] \[arXiv:hep-ph/9906218\]; N. Kauer, T. Plehn, D. Rainwater and D. Zeppenfeld, Phys. Lett. B [503]{}, 113 (2001) \[arXiv:hep-ph/0012351\]; C. M. Buttar, R. S. Harper and K. Jakobs, Report No. ATL-PHYS-2002-033; K. Cranmer et al., Report No. ATL-PHYS-2003-002 and Report No. ATL-PHYS-2003-007; S. Asai et al., Report No. ATL-PHYS-2003-005. O. J. Eboli and D. Zeppenfeld, Phys. Lett. B [495]{}, 147 (2000) \[arXiv:hep-ph/0009158\]; B. Di Girolamo, A. Nikitenko, L. Neukermans, K. Mazumdar and D. Zeppenfeld, in arXiv:hep-ph/0203056. D. G. Charlton, arXiv:hep-ex/0110086. The LEP Electroweak Working Group: [http://lepewwg.web.cern.ch/LEPEWWG]{}. T. Han, G. Valencia and S. Willenbrock, Phys. Rev. Lett. [69]{}, 3274 (1992) \[arXiv:hep-ph/9206246\]. T. Figy, C. Oleari and D. Zeppenfeld, Phys. Rev. D [68]{}, 073005 (2003) \[arXiv:hep-ph/0306109\]. F. Boudjema [et al.]{}, arXiv:hep-ph/9601224. T. Stelzer and W. F. Long, Comput. Phys. Commun. [81]{}, 357 (1994) \[arXiv:hep-ph/9401258\]; F. Maltoni and T. Stelzer, JHEP [0302]{}, 027 (2003) \[arXiv:hep-ph/0208156\]. K. Hagiwara and D. Zeppenfeld, Nucl. Phys. [B274]{}, 1 (1986); K. Hagiwara and D. Zeppenfeld, Nucl. Phys. [B313]{}, 560 (1989). S. Catani and M. H. Seymour, Nucl. Phys. [B485]{}, 291 (1997) \[Erratum-ibid. [B510]{}, 503 (1997)\] \[arXiv:hep-ph/9605323\]. Warren Siegel, Phys. Lett. B [84]{}, 193 (1979); Warren Siegel, Phys. Lett. B [94]{}, 37 (1980). V. Ilyin, private communication. G. P. Lepage, J. Comput. Phys. [27]{}, 192 (1978). See, e.g., E. N. Argyres [et al.]{}, Phys. Lett. B [358]{}, 339 (1995) \[arXiv:hep-ph/9507216\]. U. Baur, J. A. Vermaseren and D. Zeppenfeld, Nucl. Phys. [B375]{}, 3 (1992). A. Denner, S. Dittmaier, M. Roth and D. Wackeroth, Nucl. Phys. [B560]{}, 33 (1999) \[arXiv:hep-ph/9904472\]. See, e.g., G. Lopez Castro, J.L.M. Lucio and J. Pestieau, Mod. Phys. Lett. [A6]{}, 3679 (1991); M. Nowakowski and A. Pilaftsis, Z. Phys. [C60]{}, 121 (1993); U. Baur and D. Zeppenfeld, Phys. Rev. Lett. [75]{}, 1002 (1995) \[arXiv:hep-ph/9503344\], and references therein. K. Hagiwara, R. D. Peccei, D. Zeppenfeld and K. Hikasa, Nucl. Phys. [B282]{}, 253 (1987). K. Hagiwara, D. Zeppenfeld and S. Komamiya, Z. Phys. C [29]{}, 115 (1985); B. A. Kniehl, Phys. Lett. B [254]{}, 267 (1991). J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. Nadolsky and W. K. Tung, JHEP [0207]{}, 012 (2002) \[arXiv:hep-ph/0201195\]. K. Hagiwara [et al.]{} \[Particle Data Group Collaboration\], Phys. Rev. D [66]{}, 010001 (2002). S. Catani, Yu. L. Dokshitzer and B. R. Webber, Phys. Lett. B [285]{} 291 (1992); S. Catani, Yu. L. Dokshitzer, M. H. Seymour and B. R. Webber, Nucl. Phys. [B406]{} 187 (1993); S. D. Ellis and D. E. Soper, Phys. Rev. D [48]{} 3160 (1993). G. C. Blazey [et al.]{}, arXiv:hep-ex/0005012. A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C [28]{}, 455 (2003) \[arXiv:hep-ph/0211080\]; A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, arXiv:hep-ph/0308087. G. Passarino and M. J. Veltman, Nucl. Phys. [B160]{}, 151 (1979). [^1]: Another source of $Wjj$ or $Zjj$ events are QCD processes at order $\alpha_s^2\alpha$, sometimes called QCD $Vjj$ production. Within typical VBF cuts, cross sections for these QCD processes are only somewhat larger than those for electroweak production [@Rainwater:1999gg]. One thus needs to calculate NLO QCD corrections for both sources independently, and as a function of phase space. For the QCD processes this was done in Ref. [@Campbell:2002tg]. [^2]: The finite proton mass provides an absolute lower bound on the photon virtuality, $Q^2_\gamma \gsim m_p^2(m_{Vjj}^2/xs)^2$, where $m_{Vjj}$ is the invariant mass of the produced system and $x$ denotes the Feynman $x$ of the colored parton in the subprocesses for $p\gamma\to VjjX$. We have chosen the lower cutoff of $Q^2_{\gamma,\rm min}=0.1$ GeV$^2$ for a very rough simulation of the resulting finite photon flux.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We investigate Non-Hermitian quantum field theoretic model with $\iota g\phi^3$ interaction in 6 dimension. Such a model is PT-symmetric for the pseudo scalar field $\phi$. We analytically calculate the 2-loop $\beta$ function and analyse the system using renormalization group technique. Behavior of the system is studied near the different fixed points. Unlike $g\phi^3$ theory in 6 dimension $\iota g\phi^3$ theory develops a new non trivial fixed point which is energetically stable.' author: - Aditya Dwivedi^1^ - Bhabani Prasad Mandal^2^ title: '2-Loop $\beta$ Function for Non-Hermitian PT Symmetric $\iota g\phi^3$ Theory' --- Introduction ============ Over the past two decades a new field with combined Parity(P)-Time reversal(T) symmetric non-Hermitian systems has emerged and has been one of the most exciting topics in frontier research. It has been shown that such theories can lead to the consistent quantum theories with real spectrum, unitary time evolution and probabilistic interpretation in a different Hilbert space equipped with a positive definite inner product [@real1]-[@real3]. The huge success of such non-Hermitian systems has lead to extension to many other branches of physics and interdisciplinary areas. The novel idea of such theories have been applied in numerous systems leading huge number of application [@pt4]-[@shal2]. Several PT symmetric non-Hermitian models in quantum field theory have also been studied in various context [@pt]-[@qft2]. Deconfinment to confinment transition is realised by PT phase transition in QCD model using natural but unconventional hermitian property of the ghost fields [@pt]. PT symmetric quantum field theory involving non-Hermitian mass term $\alpha\bar{\psi}\gamma^5\psi$ was introduced and further investigated to show the existence of conserved current in such a model to show the consistency of PT symmetry and unitarity of the theory. Theories with non-Hermitian mass term has also been used for alternative description of neutrino mass and dark matter. Neutrino oscillation has been investigated using non-Hermitian PT symmetric model in quantum field theory. Aspects of spontaneous symmetry breaking and Goldston theorem has been studied using non-Hermitian field theoretic model [@x]. Pseudo scalar field theories with $\iota g\phi^3$ interaction has been investigated by several groups [@sha0], [@sha00]. The motivation behind the study of $\iota g\phi^3$ is stated as follows. In ordinary $g\phi^3$ theory ground state is unstable and it decays gradually, whereas in $\iota g\phi^3$ potential is complex. Therefore concept of boundedness of ground state is not going to apply here as explained in Refs.[@re1]-[@re3]. One can calculate the ground state energy density for $\iota g\phi^3$ theory by using perturbation theory and summing all the connected vacuum Feynman diagrams. The Borel summation of this Stieltjes series is real [@re1]-[@re3]. Thus, in view of this one can say that ground state of this theory is stable and $\iota g\phi^3$ theory is physically acceptable. Further in $1+1$ dimension $\iota g\phi^3$ theory provides an important feature in their solutions. The field $\phi$ is pure imaginary soliton wave, which can be useful in the description of hadrons, because they does not change shape after interaction [@sol1]. Due to boundary condition $\phi_{\pm\infty}=0$ and $\phi_{\pm\infty}^\dagger=0$, we can say that these are non topological solitons and these characterise the theory with more than one component. Merging of $\iota g\phi^3$ and $-\iota g\phi^3$ theory is equivalent to Lee-Wick theory which includes both $\phi$ and $\phi^\dagger$ [@sol2]. Such a theory is shown to Hermitian equivalent to Lee-Wick theory which suffers from the existence of the famous ghost state and instability problem [@shal]. Metric operator has been computed for $\iota\phi^3$ theory by making new ansatz [@shal2]. The critical behavior of such theories around the fixed points has been investigated using renormalization group technique at the one loop order [@bend1]. Further renormalization group properties of ordinary $\phi^3$ theory have compares with that of PT symmetric $\iota g\phi^3$ theory in 1-loop $\beta$ function for the later theory [@bend2]. The purpose of the present article is to revisit the $\iota g\phi^3$ theory in 6 dimension in the framework of renormalization group analysis and to examine the theory with higher loop calculation. We calculate 2-loop $\beta$ function explicitly and analyse the behavior of the system near the fixed points. Unlike the usual $ g\phi^3$ theory the PT symmetric $\iota g\phi^3$ theory develops non-trivial fixed point which is energetically stable. The theory is perturbatively renormalizable. Our analysis is consistent with renormalization group study of the same model at the 1-loop level [@bend2].\ Now, we present the plan of our paper. In Section 2 we discuss $g\phi^3$ theory in the framework of RG analysis. earlier. 1-loop calculations of $\beta$ function is carried out in Section 3. 2- loop calculation are presented in Section 4. Section 5 is kept for results and discussions. Theory of $\phi^3$ fields: ========================== We start with the bare Lagrangian density of $\phi^3$ theory in 6 dimension as: $$\L=\frac{1}{2}(\partial\phi_0)^2-\frac{1}{2}{m_0^2\phi_0^2}-\frac{1}{6}{g_0\phi_0^3}$$ where the subscript 0 denotes the bare quantities . This Lagrangian density further can be expressed in terms of renormalized Lagrangian density as $$L=\frac{1}{2}(\partial\phi)^2-\frac{1}{2}m^2\phi^2-\frac{1}{6}{g}{\phi^3}+\frac{1}{2}{C}(\partial\phi)^2-\frac{1}{2}B\phi^2-\frac{1}{6}A\phi^3.$$ Where m, g and $\phi$ are renormalized mass, coupling constant and field respectively and satisfy: $$\sqrt{Z}\phi=\phi_0$$ $${g_0}={Z_0}{Z^{-\frac{3}{2}}}{g}.$$ The coefficient of counter Lagrangian are written as $Z=1+C$, ${m_0^2}Z=m^2+B$, ${Z_0}g=g+A$.\ In the arbitrary n dimension the coupling $g$ is dimensionful, to make it dimensionless a mass scale $\mu$ is introduced. Unrenormalized coupling and mass are divergent in $6$ dimension therefore these quantities can be expanded around $n=6$ in terms of renormalized parameter as $$g_0Z^\frac{3}{2}=\mu^\frac{6-n}{2}\left[g+\sum_{\alpha=1}^{\infty}\frac{e_\alpha(m,g)}{(n-6)^\alpha}\right]$$ $$m_0^2Z=\mu\left[m^2+\sum_{\alpha=1}^\infty\frac{f_\alpha(m,g)}{(n-6)^\alpha}\right]$$ and $$Z=1+\sum_{\alpha=1}^\infty\frac {h_\alpha(m,g)}{(n-6)^\alpha}$$ where $e_\alpha$, $f_\alpha$ and $h_\alpha$ are the different coefficients in the above expansion for coupling, mass and Z terms for the pole at $n=6$ of order $\alpha$. To check the uniqueness of the above series let us write $g$ and $m$ in form of\ $$g\to\left[g+x_1(n-6)+x_2(n-6)^2+......\right]$$ and $$m\to\left[m+y_1(n-6)+y_2(n-6)^2+......\right].$$ On putting these in equations (2) and (3) we obtain $$g_0Z^\frac{3}{2}=\mu^\frac{6-n}{2}\left[g+\sum_{k=1}^\infty x_k(n-6)^k+\sum_{\alpha=1}^\infty\frac{e_\alpha(m',g')}{(n-6)^\alpha}\right]$$ $$m_0^2Z=\mu\left[m^2+\sum_{k=1}^\infty y_k(n-6)^k+\sum_{\alpha=1}^\infty\frac{f_\alpha(m'g')}{(n-6)^\alpha}\right]$$ $$Z=1+\sum_{\alpha=1}^\infty\frac {h_\alpha(m',g')}{(n-6)^\alpha}$$ These equations are different from those which we first considered in equation (2) and (3). For the uniqueness of the series we can put the additional condition that the coefficients of $x_k$ and $y_k$ are 0 in equations (5) and (6). In other words to say that all the transformations of $g$ and $m$ always lead to the series in equations (2),(3) and (4). Since mass parameter $\mu$ is arbitrary we have to check the uniqueness of the series in equations (2) and (3) after changing it too. In order to do that we shift $\mu$ as $\mu'=\mu(1+\epsilon)$ ,where $\epsilon<<1$. Putting $\mu'$ in equations (2) and (3) we obtain - $$g_0Z^\frac{3}{2}=\mu'^{(\frac{6-n}{2})}\left[g+g(\frac{n-6}{2})\frac{\epsilon}{2}+\frac{\epsilon}{2}e_1+\sum_{\alpha=1}^{\infty}\frac{e_\alpha+\frac{\epsilon}{2}e_{\alpha+1}}{(n-6)^\alpha}\right]$$ and $$m_0^2Z=\mu'(1-\epsilon)\left[m^2+\sum_{\alpha=1}^\infty\frac{f_\alpha}{(n-6)^\alpha}\right ].$$ These equations will be put back in the form of equations (2) and (3) when - $$g=\tilde{g}-\frac{\epsilon}{2}(n-6).$$ This transformation makes the above series as $$g_0Z^\frac{3}{2}=(\mu')^\frac{6-n}{2}\left[\tilde{g}+\frac{\epsilon}{2}e_1-\frac{\epsilon}{2}\tilde{g}e_{1,g}+\sum_{\alpha=1}^\infty\frac{e_\alpha(m,\tilde{g})+\frac{\epsilon}{2}e_{\alpha+1}-\frac{\epsilon}{2}\tilde{g}e_{\alpha+1,g}}{(n-6)^\alpha}\right ]$$ $$m_0^2Z=\mu'\left[m^2-\epsilon m^2-\frac{\epsilon}{2} \tilde{g} f_{1,g}+\sum_{\alpha=1}^\infty \frac{f_\alpha(m,\tilde{g})-\frac{\epsilon}{2}f_\alpha-\frac{\epsilon}{2}f_{\alpha+1,g}}{(n-6)^\alpha}\right ]$$ where $e_{\alpha,g}$ $\equiv$ $\frac{\partial{e_\alpha}}{\partial\tilde{g}}$ and similarly for others. We see that changing the mass scale, we have to change the renormalized parameters also. As at n=6, $\tilde{g}=g$, thus new renormalized parameter in terms of old parameters in view of above equations can be written as $$g'=g+\frac{\epsilon}{2}\left[1-g\frac{\partial}{\partial g}\right]e_1(m,g)$$ and $$m'^2=m^2-\epsilon \left[m^2+\frac{g}{2}\frac{\partial}{\partial g}f_1(m,g)\right]$$ It has been shown that $e_\alpha$, $f_\alpha$ and $h_\alpha$ are independent of mass, thus above equation can be written as- $$g'=g+\frac{\epsilon}{2}\beta_g$$ $$m'^2=m^2-\epsilon \gamma_{m^2}$$ where $\gamma_{m^2}=\left[m^2+\frac{g}{2}\frac{\partial}{\partial g}f_1(m,g)\right]$ and, $\beta$ function for the coupling is defined as $$\beta_g=\big[1-g\frac{\partial}{\partial g}]e_1.$$ On deriving the renormalization group equation for n point green function we get the result $$\mu\frac{\partial{m^2}}{\partial\mu}=\beta_{m^2}=-m^2\big[1+\frac{g}{2}\frac{\partial}{\partial g}\big]f_1.$$ From equations (14) and (15) we can see that if we use the ’t Hooft procedure [@thooft1], [@thooft2], the beta function depends only on the one loop pole coefficients. Beta function upto higher order in coupling is found by relating higher order pole coefficients to the single pole coefficients. This shows that higher order pole coefficients are not arbitrary and depends on the one loop pole coefficients. It is well known that if theory is renormalizable in one loop order then it is renormalizable at higher loops also. In view of the above, the series in equation (2), (3) and (4) are changed to the form $$g_0Z^\frac{3}{2}=\mu^\frac{6-n}{2}\left[g+\sum_{\alpha=1}^\infty \frac{e_\alpha(g)}{(n-6)^\alpha}\right]=\mu^\frac{6-n}{2}\left[g+g\sum_{\alpha=1}^\infty \sum_{\gamma=\alpha}^\infty \frac{e_{\alpha\gamma} g^\gamma }{(n-6)^\gamma}\right]$$ $$m_0^2Z=\mu\left[m^2+m^2\sum_{\alpha=1}^\infty \frac{f_\alpha(g)}{(n-6)^\alpha}\right]=\mu\left[m^2+m^2\sum_{\alpha=1}^\infty \sum_{\gamma=\alpha}^\infty \frac{f_{\alpha\gamma} g^\gamma}{(n-6)^\alpha}\right]$$ and$$Z=1+\sum_{\alpha=1}^\infty \frac{h_\alpha(g)}{(n-6)^\alpha}=1+\sum_{\alpha=1}^\infty \sum_{\gamma=\alpha}^\infty \frac{h_{\alpha\gamma}g^\gamma }{(n-6)^\alpha}.$$ Further, by considering one and two loop diagrams, we will calculate the terms in these series expansion taking $\alpha=1$ and 2 and $\gamma=2,4$ which will enable us to calculate the coefficients of the pole terms. This will in term help us to find the $\beta$ function for the theory up to two loops.\ One loop calculation: ===================== We want to use the general formula written in the previous section to calculate the amplitudes explicitly for different graphs in $g\phi^3$ theory in one loop order. Let us consider the following diagrams and counter terms relevant to the calculation of $\beta$ function in one loop order.\ ![(a) One loop self energy diagram, (b) Wave function counter term, (c) Self energy with massive propagator, (d) Mass counter term, (e) Diagram for vertex term, (f) Vertex counter term](IPL1) \ We calculate the amplitudes for these diagrams using the technique given in reference [@woo]. Diagram $1(a)$ is self energy diagram which gives the result\ $$A_{1(a)}=\frac{g^2p^2}{12(4\pi)^3}\left[\frac{2}{(n-6)}-\ln\frac{-p^2}{4\pi\mu^2}-(\gamma-\frac{8}{3})\right].$$ Since the amplitudes of massless two point diagrams are related to the field renormalization coefficient Z. We use equation (18), which suggest that term corresponding to diagram 1(b) is $\frac{1}{2}\left[\frac{h_{12}}{n-6}\right]g^2(\partial\phi)^2$, and amplitude for the diagram is $\left[\frac{h_{12}}{n-6}\right]g^2p^2$. On comparing the pole part of $A_{1(a)}$ with amplitude of diagram $1(b)$ we get, $ h_{12}=\frac{1}{6(4\pi)^3}$. The contribution of 1(c) diagram is\ $$A_{1(c)}=\frac{m^2 g^2}{(4\pi)^\frac{n}{2}}\Gamma(3-\frac{n}{2})\big(\frac{-p^2}{\mu^2}\big) B(\frac{n}{2}-1,\frac{n}{2}-1).$$ Pole part in $A_{1(c)}$ is present due to the pole of gamma function present in the amplitude. Amplitudes of massive two point diagrams are related to the mass renormalization. Using equation (17), we get contribution of diagram 1(d) in lagrangian density is $-\frac{1}{2}f_{12}\left[\frac{m^2g^2\phi^2}{(n-6)}\right]$ and we compare the pole part of $A_{1(c)}$ to the amplitude of $1(d)$ to get $f_{12}=\frac{1}{(4\pi)^3}$. Amplitude of the diagram $1(e)$ is $$A_{1(e)}=\frac{g^3\mu^\frac{\epsilon}{2}}{2(4\pi)^3}\frac{2}{n-6}$$. Three point diagrams are related to the renormalization of the coupling of the theory thus, use of equation (16) suggest that diagram $1(f)$ corresponds to the term $-\frac{1}{6}\left[\frac{e_{12}}{n-6}\right]g^2\phi^2$ and on comparing the coefficient of simple pole of $A_{1{e}}$ with amplitude of diagram $1(f)$ gives $e_{12}=\frac{1}{(4\pi)^3}$. Two Loop calculation ==================== In this section we consider all the two loop diagrams in the $g\phi^3$ theory in 6 dimension which are relevant to our calculation for $\beta$ function. First we consider 2 loop self energy diagrams (Figs. 2(a), 2(b)) and self energy counter terms (Figs. 2(c), 2(d)). Then we consider the diagrams with massive propagator and, finally we consider 2 loop three point diagrams for finding the various pole coefficients. Massless self energy diagrams ------------------------------ We have to consider following diagrams to calculate the pole coefficients of two loop level for the self energy diagram for massless case. ![Two loop diagram for massless case. (a) Self energy with overlapping divergences, (b) Self energy with nested divergences, (c) Self energy with vertex counter term , (d) Self energy with wave function counter term.](10) \ On analysis of these, we can easily see that there are overlapping divergences in these diagrams. After doing the separation of overlapping divergences, we get the contribution from various diagrams. Amplitude for the diagram 2(a) is evaluated as $$A_{2(a)}=\frac{g^4p^2}{4\pi^6}\left[-\frac{1}{6(n-6)^2}-\frac{\ln(\frac{-p^2}{\mu^2})}{6(n-6)}+\frac{1}{6}\frac{(-\gamma+3)}{n-6}\right]+ \mbox {finite part}.$$\ Similarly we find the amplitudes for diagrams 2(b), 2(c) and 2(d) which are written as- $$A_{2(b)}=\frac{2g^4p^2}{(4\pi)^6}\left[\frac{1}{36(n-6)^2}+\frac{ln(\frac{-p^2}{\mu^2})}{36(n-6)}+\frac{12\gamma-43}{432(n-6)}\right]+\mbox {finite part}$$\ $$A_{2(c)}=\frac{g^4p^2}{(4\pi)^6}\left[\frac{1}{3(n-6)^2}+\frac{ln(\frac{-p^2}{\mu^2})}{6(n-6)}+\frac{3\gamma-8}{18(n-6)}\right]+\mbox {finite part}$$\ $$A_{2(d)}=\frac{g^4p^2}{(4\pi)^6}\left[-\frac{1}{18(n-6)^2}-\frac{ln(\frac{-p^2}{\mu^2})}{36(n-6)}-\frac{3\gamma-8}{108(n-6)}\right]+\mbox {finite part}$$\ On adding these amplitudes corresponding to the diagrams 2(a), 2(b), 2(c) and 2(d), we see that logarithmic divergent parts vanish and on comparing the pole parts of the added result using equation (18) and diagrams which give counter terms, as done in the previous section, we get - $$\frac{h_{14}}{n-6}=\frac{13}{432(4\pi)^6(n-6)}$$ and $$\frac{h_{24}}{(n-6)^2}=\frac{5}{36(4\pi)^6(n-6)^2}.$$ Next we proceed to evaluate the massive case. Contribution to mass term ------------------------- Following diagrams contribute to the mass term pole coefficients in the two loop order. ![Two loop diagrams for massive case. 3(a), 3(b), 3(c) and 3(d) are self energy diagrams with massive propagator, (e) self energy vertex counter term with massive propagator, (f) self energy wave function counter term with massive propagator, (g) self energy mass counter term with massive propagator.](IPL2) Contribution for these diagrams are calculated explicitly and given below.\ Amplitude for the diagram 3(a) is $$A_{3(a)}=\frac{g^4m^2}{(4\pi)^6}\left[\frac{1}{(n-6)^2}+\frac{ln(\frac{-p^2}{\mu^2})}{(n-6)}+\frac{4\gamma-9}{4(n-6)}\right]+\mbox {finite part}.$$ Amplitude for the diagrams 3(b), 3(c), 3(d), 3(e), 3(f) and 3(g) are $$A_{3(b)}=\frac{g^4m^2}{(4\pi)^6}\left[-\frac{1}{12(n-6)^2}-\frac{ln(\frac{-p^2}{\mu^2})}{12(n-6)}-\frac{12\gamma-31}{144(n-6)}\right]+\mbox {finite part}$$ $$A_{3(c)}= \frac{g^4m^2}{(4\pi)^6}\left[\frac{1}{2(n-6)^2}+\frac{ln(\frac{-p^2}{\mu^2})}{2(n-6)}+\frac{(4\gamma-9)}{8(n-6)}\right]+\mbox {finite part}$$ $$A_{3(d)}=\frac{g^4m^2}{(4\pi)^6}\left[-\frac{1}{4(n-6)}\right]+\mbox {finite part}$$ $$A_{3(e)}=\frac{g^4m^2}{(4\pi)^6}\left[-\frac{2}{(n-6)^2}-\frac{ln(\frac{-p^2}{\mu^2})}{(n-6)}-\frac{\gamma+2}{(n-6)}\right]+\mbox {finite part}$$ $$A_{3(f)}=-\frac{1}{12}\frac{g^4m^2}{(4\pi)^6}\left[-\frac{2}{(n-6)^2}-\frac{ln(\frac{-p^2}{\mu^2})}{(n-6)}-\frac{\gamma+2}{(n-6)}\right]+\mbox {finite part}$$ $$A_{3(g)}=\frac{g^4m^2}{(4\pi)^6}\left[-\frac{1}{(n-6)^2}-\frac{ln(\frac{-p^2}{\mu^2})}{2(n-6)}-\frac{\gamma+2}{2(n-6)}\right]+\mbox {finite part}$$\ respectively. Adding up all these contributions we see that the logarithmic part in the amplitudes cancels each other and then comparing the added amplitude using equation (17) we get, $$\frac{f_{14}}{(n-6)}=\frac{23}{48(n-6)(4\pi)^6}$$ and $$\frac{f_{24}}{(n-6)^2}=\frac{5}{4(n-6)^2(4\pi)^6}.$$ 2 loop three point diagrams --------------------------- In this section we discuss the 2 loop three point diagrams which are relevant for calculation of $\beta$ function, these diagrams are given in figure 4. ![Two loop three point diagram 4(a), 4(b) and 4(c) represent vertex term, 4(d) vertex term with wave function counter term, 4(e) vertex term with vertex counter term.](30) Amplitudes for 2 loop three point diagrams are given as- $$A_{4(a)}=\frac{g^4}{(4\pi)^6}\left[-\frac{1}{4(n-6)^2}-\frac{ln(\frac{q^2}{\mu^2})}{4(n-6)}-\frac{12\gamma-7}{48(n-6)}-\frac{X}{2(n-6)}\right]+\mbox {finite part}$$ $$A_{4(b)}=\frac{g^4}{(4\pi)^6}\left[\frac{3}{2(n-6)^2}+\frac{3}{2}\frac{ln(\frac{q^2}{\mu^2})}{(n-6)}-\frac{12\gamma-3}{8(n-6)}+\frac{3X}{(n-6)}\right]+\mbox {finite part}$$ $$A_{4(c)}=-\frac{g^4}{4(4\pi)^6}\frac{1}{n-6}+\mbox {finite part}$$ $$A_{4(d)}=\frac{g^4}{(4\pi)^6}\left[\frac{1}{2(n-6)^2}+\frac{ln(\frac{q^2}{\mu^2})}{4(n-6)}-\frac{\gamma}{4(n-6)}+\frac{X}{2(n-6)}\right]+\mbox {finite part}$$ $$A_{4(e)}=\frac{g^4}{(4\pi)^6}\left[-\frac{3}{(n-6)^2}-\frac{3}{2}\frac{ln(\frac{q^2}{\mu^2})}{(n-6)}-\frac{\gamma}{2(n-6)}-\frac{3X}{n-6}\right]+\mbox {finite part}$$ respectively, where $X=\int_0^1{\delta{x}\delta{y}\delta{z}\delta(1-x-y-z)\ln(xy+yz+zx)}$.\ Again adding up all the contribution we see that the logarithmic part is removed and, then on comparing using equation (16) we get the pole term coefficients $$\frac{e_{14}}{n-6}=\frac{23}{48(4\pi)^6(n-6)}$$ and $$\frac{e_{24}}{(n-6)^2}=\frac{5}{4(n-6)^2(4\pi)^6}.$$ Calculation of $\beta$ function for $\iota{g}{\phi^3}$ theory: ============================================================== Now we are ready to calculate the 2 loop $\beta$ function for the theory using equation (14) and other results in the previous section. To calculate the value of $e_1$, we use equations (2) and (16). We expand both series and compare the coefficients of $\frac{1}{n-6}$ which give us the relation- $$e_1=(e_{12}-\frac{3}{2}h_{12})g^3+(e_{14}-\frac{3}{2}h_{14})g^5.$$ Now using equation (14), for $g\phi^3$ theory, we have $$\beta(g)=-\frac{3}{4}\frac{g^3}{(4\pi)^3}-\frac{125}{144}\frac{g^5}{(4\pi)^6}.$$\ For $\iota g\phi^3$ theory, we replace $g$ by $\iota g$ in the above expression of the $\beta$ function, which give $$\beta(\iota{g})=\frac{3}{4}\frac{\iota{g^3}}{(4\pi)^3}-\frac{125}{144}\frac{\iota{g^5}}{(4\pi)^6}.$$\ The fixed points of theory are obtained by putting $\beta(\iota{g})=0$, and given by , $$g=\pm\sqrt{\frac{108(4\pi)^3}{125}}\equiv \tilde{g}(say)\ \mbox{and}, \ g=0.$$ Now to see how the system behaves near the fixed point, we revisit equation(14), $$\mu\frac{\partial{ g}}{\partial{\mu}}=\beta{(g)}$$ Integrating this, using $\beta$ function given in equation (29) $$\int\frac{\partial g}{\alpha{g^3}-\lambda{g^5}}=\int \frac{\partial \mu}{\mu}$$ where $\alpha\equiv\frac{3}{4}\frac{1}{(4\pi)^3}$ and $\lambda\equiv\frac{125}{144}\frac{1}{(4\pi)^6}$. Solution of the above integral is obtained as- $$\mu=\exp (\frac{-1}{\alpha}\frac{1}{2g^2})\big[ \frac{g}{(1-\frac{\lambda}{\alpha}{g^2})^\frac{1}{2}}\big]^{ ( \frac{\lambda}{\alpha^2})}.$$\ This indicates that $g=\sqrt{\frac{\alpha}{\lambda}}=\sqrt{\frac{108(4\pi)^3}{125}}=\tilde{g}$ is a stable fixed point as $ g\to\tilde{g}$, $\mu$ grows. In the high energy range, properties of $\iota{g}\phi^3$ theory in 6 dimension is governed by the RG trajectories near the fixed point $g=\tilde{g}$. We would like to point out that in $\phi^3$ theory in 6 dimension the only fixed point $g=0$ was the trivial one. In case of low energy limits i.e. $\lim{\mu\to 0}$ the coupling ${g\to 0}$ which shows that theory behaves as a free theory in low energy limit and behavior of the theory is governed by the trajectories near the Gaussian fixed point $g=0$. ![Plot of $\beta$ function of coupling $g$ for $\iota g \phi^3$ theory.](11) ![Plot of $\beta$ function of coupling $g$ for $g\phi^3$ theory.](31) The variation of $\beta$ function with coupling $g$ for $\iota g \phi^3$ is shown in fig (5). This behaviour of $\beta$ function is then compared with that of usual $ g \phi^3$ theory, given in fig (6). We find the dependence of mass on coupling using equation (15). To find the value of $f_1$ we use equation (3) and (17). We calculate the $\beta_{m^2}$ for the theory with $\iota g\phi^3$ coupling as, $$\beta_{m^2}=-m^2\big[1-\frac{5}{12}\frac{g^2}{(4\pi^3)}+\frac{97}{216}\frac{g^4}{(4\pi)^6}\big].$$ Equation (31) implies that $m^2=0$ is only fixed point for the $\beta_{m^2}.$\ Using the definition of the beta function, $$\mu\frac{\partial m^2}{\partial\mu}=-m^2\big[1-\frac{5}{12}\frac{g^2}{(4\pi^3)}+\frac{97}{216}\frac{g^4}{(4\pi)^6}\big]$$ we found that $$\frac{1}{m^2}=\mu\exp\big[1-\frac{5}{12}\frac{g^2}{(4\pi^3)}+\frac{97}{216}\frac{g^4}{(4\pi)^6}\big]$$ which shows that on going $\lim{\mu\to\infty}$ we see that ${m^2\to 0}$ which shows that $m^2=0$ is stable fixed point for the beta function for the mass term. Conclusion ========== In this work, we have investigated a PT symmetric non-Hermitian model for scalar fields in 6 dimension to extract some new features of the theory. By considering the appropriate Feynman diagrams in one loop and two loop order and the relevant counter terms for those diagrams, we explicitly evaluate two loop $\beta$ finction for $\iota g \phi^3$ theory in $6$ dimension.Unlike the Hermitian theory, this non-Hermitian theory develops additional non-trivial energetically stable fixed point. The high energy behavior of the theory is governed by this non trivial fixed point (fig 5). On the other hand low energy behaviour is governed by the RG trajectories near the Gaussian fixed point $g=0$. $\beta$ function for the mass term for this non-Hermitian model has been calculated. $m^2=0$ is a stable fixed point for the $\beta$ function for mass term. [99]{} C.M. Bender, S. Boettcher, Phys. Rev. Lett. 80, 5243 (1998). C.M. Bender, Rep. Prog. Phys. 70, 947 (2007), and references therein. A. Mostafazadeh, Int. J. Geom. Meth. Mod. Phys. 7, 1191(2010) and references therein. A. Khare, B.P. Mandal, Phys. Lett. A 272, 53 (2000). M. Znojil, J. Phys. A 36, 7825 (2003). B.P. Mandal, B.K. Mourya, K. Ali, A. Ghatak, Ann. Phys. 363, 185–193 (2015). C.M. Bender, S. Boettcher, P.N. Meisinger, J. Math. Phys. 40, 2201 (1999). B.P. Mandal, Mod. Phys. Lett. A 20, 655 (2005). B.P. Mandal, A. Ghatak, J. Phys. A, Math. Theor. 45, 444022 (2012). C.T. West, T. Kottos, T. Prosen, Phys. Rev. Lett. 104, 054102 (2010). A. Nanayakkara, Phys. Lett. A 304, 67 (2002). C.M. Bender, G.V. Dunne, P.N. Meisinger, M. Simsek, Phys. Lett. A 281, 311–316 (2001). B.P. Mandal, B.K. Mourya, R.K. Yadav, Phys. Lett. A 377, 1043 (2013). G. Levai, J. Phys. A 41, 244015 (2008). C.E. Ruter, K.G. Makris, R. El-Ganainy, D.N. Christodulides, M. Segev, D. Kip, Nat. Phys. 6, 192 (2010). H. Raval, B. P. Mandal, Nuclear phys. B, 946, 114699 (2019). Jean Alexandre, John Ellis, Peter Millington, and Dries Seynaeve, Phys. Rev. D 98, 045001 (2018). Abouzeid M. Shalaby, IJMPA, 34, 1950090 (2019). Abouzeid M. Shalaby, arXiv no. 1811.10998. Abouzeid M. Shalaby, Phys. Rev. D 80, 025006 (2009). Abouzeid M. Shalaby, Phys. Rev. D 79, 107702 (2009). Carl M. Bender, V. Branchina, and Emanuele Messina, Phys. Rev. D 87, 085029 (2013). Carl M. Bender, Vincenzo Branchina, and Emanuele Messina, Phys. Rev. D 85, 085001 (2012). A.J. Macfarlen and G. Woo, Nuclear Physics B 77, 91 (1974). Jean Alexandre, Carl M. Bender, Peter Millington, JHEP 11, 111 (2015). Jean Alexandre, Peter Millington, Dries Seynaeve, Phys. Rev. D 96, 065027 (2017). G. ’t Hooft, Under the spell of gauge principle. G. ’t Hooft, Nuclear Phys. B, 61, 455-468, (1973). E. Caliceti, S. Graffi, M. Maioli, Comm. Math. Phys. 75, 51 (1980). C.M. Bender, G. V. Dunne, J. Math Phys., 40, 4616 (1999). C.M. Bender, E. J. Weniger, J. Math Phys., 42, 2167 (2001). R. Rajaraman, Solitons and Instantons, North-Hollands Pub. (1989). Abouzeid M. Shalaby, Arxiv: 0912.0304V2 (2010)	{ "pile_set_name": "ArXiv" }
--- abstract: 'Automated decision making is used routinely throughout our every-day life. Recommender systems decide which jobs, movies, or other user profiles might be interesting to us. Spell checkers help us to make good use of language. Fraud detection systems decide if a credit card transactions should be verified more closely. Many of these decision making systems use machine learning methods that fit complex models to massive datasets. The successful deployment of machine learning (ML) methods to many (critical) application domains crucially depends on its explainability. Indeed, humans have a strong desire to get explanations that resolve the uncertainty about experienced phenomena like the predictions and decisions obtained from ML methods. Explainable ML is challenging since explanations must be tailored (personalized) to individual users with varying backgrounds. Some users might have received university-level education in ML, while other users might have no formal training in linear algebra. Linear regression with few features might be perfectly interpretable for the first group but might be considered a black-box by the latter. We propose a simple probabilistic model for the predictions and user knowledge. This model allows to study explainable ML using information theory. Explaining is here considered as the task of reducing the “surprise” incurred by a prediction. We quantify the effect of an explanation by the conditional mutual information between the explanation and prediction, given the user background.' author: - 'Alexander Jung and Pedro H. J. Nardelli [^1]' bibliography: - '/Users/alexanderjung/Literature.bib' title: 'An Information-Theoretic Approach to Personalized Explainable Machine Learning' --- Introduction {#sec_intro} ============ Machine learning (ML) methods compute predictions for quantities of interest based on a statistical analysis of large amounts of historical data [@hastie01statisticallearning; @JungCompML; @BishopBook]. These methods are routinely used to power many services within our everyday-life. ML methods power recommendation systems that decide what job ads or which other user profiles could be interesting to us [@Wang2018; @Martinez2009]. Recent breakthroughs in ML, such as in image or text processing [@Goodfellow-et-al-2016], also holds the promise of boosting the level of automation in domains which currently rely mainly on human labour or manual design [@Goodall2016]. A key challenge for the successful and ethically sound deployment of ML methods to critical application domains is the (lack of) explainability of its predictions [@Holzinger2018; @Hagras2018; @Mittelstadt2016; @Wachter2017]. Explanations of predictions, which are used for decisions that affect humans, are increasingly becoming a legal obligation [@Wachter2017]. Beside legal aspects, it also seems that humans have a basic need for understanding decision making processes [@Kagan1972; @Kruglanski1996]. One reason why explainable ML is challenging is that (good) explanations must be tailored to the knowledge of individual users (“explainee”). In general, for a particular prediction, there is no unique explanation that serves equally well a large group of heterogeneous users. Thus, achieving explainable ML would be easier for applications involving a homogenous group of users, like graduate students in a university program. Large-scale applications as, for instance, recommendation systems for video streaming providers typically involve users with very different backgrounds, which can range from graduate studies in ML-related fields to users with no formal training in linear algebra. While linear models involving few hand-crafted features might be viewed as interpretable for the former group it might be considered a “black-box” for the latter group of users. This contribution studies explainable ML within information theory by using a probabilistic model for the data and user background. Loosely speaking, we model the effect of providing an explanation for a prediction as a reduction of the “surprise” incurred by a prediction to the user. This qualitative interpretation of explaining a prediction leads naturally to measuring the quantitative effect of explanations via (conditional) mutual information (MI) between the explanation and the prediction, given the user background (see Section \[sec\_setup\]). Our approach is different from existing work on explainable ML in the sense that we explicitly model the specific knowledge of each individual user. In contrast, most existing methods for explainable ML do not make any assumption about the end-user and her background knowledge. Explainable ML methods can be roughly divided into two groups. The first group of methods uses models that are considered as intrinsically interpretable, like linear regression or small decision trees. The second group of methods, referred to as model-agnostic methods, probe an ML method by perturbing the features of the data point. The most straightforward approach to explainable ML methods is to use models that are considered to be intrinsically interpretable. Such methods include linear models, decision trees and artificial neural networks [@Montavon2018; @Bach2015; @Hagras2018]. Explaining the predictions obtained from such intrinsically interpretable models merely amounts to specifying the model parameters, such as the weights $w_{i}$ of a linear predictor $h(\mathbf{x}) = \sum_{i} w_{i} x_{i}$, or the feature-wise thresholds used in decision trees [@hastie01statisticallearning]. Interpretable models offer an intuitive decomposition of its predictions into a combination of elementary properties of a data point. Defining elementary properties of a data point via the activations of a (deep) neural network renders those models also interpretable (see [@Montavon2018]). Explainable models for sequential decision making have been studied in [@Mcinerney18], where the authors obtain an explainable multi-armed bandit model by using the choice for the action space as the explanation. An explanation can be obtained by notifying the user that only previously purchased items are recommended. In contrast to [@Mcinerney18], our approach uses a probabilistic model for the user background to compute personalized explanations that are optimal in a precise (information-theoretic) sense. A second group of explainable ML, referred to as model agnostic methods, is based on constructing explanations by probing a predictor as a black box [@Ribeiro2018; @Hagras2018]. These methods aim at locally approximating black box models by simpler and interpretable models, such as linear models or shallow decision trees [@Ribeiro2018]. Our approach is also model agnostic as it only requires the statistical distributions of the model prediction. However, in contrast to most model agnostic explainable ML, we do not use local approximations to explain a black box method. Instead, we use a probabilistic model for the predictions and user knowledge. We frame explainable ML within a probabilistic model for ML predictions and user knowledge. This allows to capture the act of explaining a prediction using information-theoretic concepts. The act of explaining provides the user additional information about the prediction delivered by some (arbitrary) ML method. Information theory has already been used for learning optimal explanations [@Chen2018]. In a similar spirit, we also use MI to guide the learning of instance-wise explanations. However, in contrast to [@Chen2018], we also model the effect of the user background on the information provided by an explanation. In a nutshell, while [@Chen2018] uses unconditional MI between explanations and predictions, we use the conditional MI given the user knowledge (see Section \[sec\_optimal\_explanation\]). [Outline and Contribution.]{} In Section \[sec\_setup\], we propose a simple probabilistic model for the features, prediction and user summary of a data point. This probabilistic model allows to quantify the effect of explanations via the conditional between the explanation and the model prediction, given the user background. Our main contribution is the formulation of an information-theoretic concept of optimal personalized explanations. As discussed in Section \[sec\_optimal\_explanation\]), we construct (information-theoretically) optimal personalized explanations by maximizing the conditional MI between explanation and predictions, when conditioning on the user summary of a data point. To the best of our knowledge, we present the first information-theoretic approach to personalized explainable ML. A simple algorithm for computing optimal explanation given the model predictions and user summaries based on i.i.d. samples is presented in Section \[sec\_simple\_XML\]. The proposed algorithm allows to construct personalized explanations that are optimal in an information-theoretic sense. Problem Setup {#sec_setup} ============= We consider a supervised ML problem involving data points with features ${{\mathbf x}}= \big(x_{1},\ldots,x_{{n}}\big)^{T} \in \mathbb{R}^{{n}}$ and label $y \in \mathbb{R}$. Given some labelled training data $$\big(\mathbf{x}^{(1)},y^{(1)}\big),\big(\mathbf{x}^{(2)},y^{(2)}\big),\ldots,\big(\mathbf{x}^{({m})},y^{({m})}\big),$$ ML methods typically learn a predictor (map) $$h(\cdot): \mathbb{R}^{{n}} \rightarrow \mathbb{R}: {{\mathbf x}}\mapsto \hat{y}=h({{\mathbf x}})$$ by requiring $\hat{y}^{(i)} \approx y^{(i)}$ [@hastie01statisticallearning; @BishopBook; @JungCompML]. (OR) at (0.00, 1.50); (data) [user $u$ consumig prediction $\hat{y}$]{}; (dummy) ; (hypothesis) [ML method]{}; (t) [prediction $\hat{y}$]{} ; (g) [explanation $e$]{} ; (hypothesis.east) to \[out=0,in=90\] (t.north); (t.south) to \[out=270,in=0\] (data.east); (data.west) to \[out=180,in=270\] (g.south); (g.north) to \[out=90,in=180\] (hypothesis.west); After learning a predictor $\hat{y}=h({{\mathbf x}})$, it is applied to new data points yielding the prediction $\hat{y} = h({{\mathbf x}})$. In may application, the prediction $\hat{y}$ is then delivered to a human user. The user can be the subscriber of a streaming service [@GomezUribe2016], a dermatologist [@Esteva2017] or a city planner [@Yang2019]. Each user has typically some conception or model for the relation between features ${{\mathbf x}}$ and label $y$ of a data point. Based on the user background, she has some understanding of a data point with features $\mathbf{x}$. Our approach to explainable ML is based on modelling the user understanding of a data point by some summary $u \in \mathbb{R}$. The summary is obtained by a (stochastic) map from the features ${{\mathbf x}}$ of a data point. We will focus on summaries being obtained by a deterministic map $$u(\cdot): \mathbb{R}^{{n}} \rightarrow \mathbb{R}: \mathbf{x} \mapsto u {:=}u(\mathbf{x}).$$ However, our approach also covers stochastic maps characterized by a conditional probability distribution $p(u\| {{\mathbf x}})$. The (user-specific) quantity $u$ represents the understanding of the specific properties of the data point given the user knowledge (modelling assumptions). We interpret $u$ as a “summary” of the data point based on its features ${{\mathbf x}}$ and the intrinsic modelling assumptions of the user. Let us illustrate the concept of the user summary $u$ as a means to represent user knowledge (or background) by two particular choices for $u$. First, the user summary could be the prediction obtained from a simplified model, such as linear regression using few features that the user anticipates as being relevant. Another example for a user summary $u$ could be a higher-level feature, such as eye spacing in facial pictures [@Jeong2015]. We formalize the act of explaining a prediction $\hat{y} = h({{\mathbf x}})$ as presenting some additional quantity $e$ to the user. This “explanation” $e$ can be any quantity that helps the user to understand the prediction $\hat{y}$, given her understanding $u$ of the data point. Loosely speaking, the explanation $e$ contributes to resolving the uncertainty of the user $u$ about the prediction $\hat{y}$ [@Kagan1972]. For the sake of exposition, our focus will be on explanations obtained via a deterministic map $$\label{equ_def_explanation} e(\cdot): \mathbb{R}^{{n}} \rightarrow \mathbb{R}: {{\mathbf x}}\mapsto e {:=}e(\mathbf{x}),$$ from the features ${{\mathbf x}}$ of a data point. However, our approach can be generalized without difficulty to handle explanations obtained by a (stochastic) map. In the end, we only require the specification of the conditional probability distribution $p(e\|{{\mathbf x}})$. Explanations can be constructed in quite different ways. An explanation could be a subset of features of a data point (see [@Ribeiro2016] and Section \[sec\_optimal\_explanation\]). More generally, explanations could be obtained from simple local statistics (averages) of features that are considered related, such as near-by pixels in an image or consecutive samples of an audio signal. Instead of individual features, carefully chosen data points can also serve as an explanation [@Mcinerney18; @Ribeiro2018]. To obtain comprehensible explanations that can be computed efficiently, we must typically restrict the space of possible explanations to a small subset $\mathcal{F}$ of maps . This is conceptually similar to the restriction of the space of possible predictor functions in a ML method to a small subset of maps which is known as the hypothesis space. We consider data points as independent and identically distributed (i.i.d.) realizations of a random variable with fixed underlying probability distribution $p(\mathbf{x},y)$. Modelling the data point as random implies that the user summary $u$, prediction $\hat{y}$ and explanation $e$ are also random variables. The joint distribution $p(u,\hat{y},e,\mathbf{x},y)$ conforms with the Bayesian network [@Pearl1988] (depicted in Figure \[fig\_simple\_prob\_ML\]) since $$\label{equ_joint_prob_factor} p(u,\hat{y},e,\mathbf{x},y) = p(u\|\mathbf{x}) \cdot p(e\|\mathbf{x}) \cdot p(\hat{y}\|\mathbf{x}) \cdot p({{\mathbf x}},y).$$ We measure the amount of additional information provided by an explanation $e$ for a prediction $\hat{y}$ to some user $u$ via the conditional MI [@coverthomas Ch. 2 and 8] $$\label{eq_def_surprise} I(e;\hat{y}\|u) {:=}{{\rm E}}\bigg\{ \log \frac{p( \hat{y},e\|u)}{p(\hat{y}\|u)p(e\|u)} \bigg\}.$$ The conditional MI $I(e;\hat{y}\|u)$ can also be interpreted as a measure for the amount by which the explanation $e$ reduces the uncertainty about the prediction $\hat{y}$ which is delivered to some user $u$. Thus, constructing explanations via solving conforms with the apparent human need to understand observed phenomena, such as the predictions from a ML method [@Kagan1972]. (dp) [data point $(\mathbf{x},y)$]{}; (-3.9,-3.5) rectangle (4,-1.0); at (-3.9,-3) \[above=1mm, right=0mm\] [some user]{}; (exp) [explanation $e$]{}; (user) [summary $u$]{}; (pred) [prediction $\hat{y}$]{}; (dp) edge\[-latex\] (exp) (dp) edge\[-latex\] (user) ; (dp) edge\[-latex\] (pred); Optimal Explanations {#sec_optimal_explanation} ==================== Capturing the effect of an explanation using the probabilistic model offers a principled approach to computing an optimal explanation $e$. We require the optimal explanation $e^{}$ to maximize the conditional MI between the explanation $e$ and the prediction $\hat{y}$ conditioned on the user summary $u$ of the data point. Formally, an optimal explanation $e^{}$ solves $$\label{equ_opt_explanation} I(e^{};\hat{y}\|u) = \sup_{e \in \mathcal{F}} I(e;\hat{y}\|u).$$ The choice for the subset $\mathcal{F}$ of valid explanations offers a trade-off between comprehensibility, informativeness and computational cost incurred by an explanation $e^{}$ (solving ). The maximization problem for obtaining optimal explanations is similar to the approach in [@Chen2018]. However, while [@Chen2018] uses the unconditional MI between explanation and prediction, involves the conditional MI given the user summary $u$. Let us illustrate the concept of optimal explanations using a linear regression method. We model the features ${{\mathbf x}}$ as a realization of a multivariate normal random vector with zero mean and covariance matrix ${\mathbf{C}}_{x}$, $$\label{equ_feature_vector_Gaussian} {{\mathbf x}}\sim \mathcal{N}(\mathbf{0},{\mathbf{C}}_{x}).$$ The predictor and the user summary are linear functions of the features, $$\label{equ_pred_summary} \hat{y} {:=}\mathbf{w}^{T} \mathbf{x} \mbox{, and } u {:=}\mathbf{v}^{T} \mathbf{x}.$$ We construct explanations via subsets of individual features $x_{i}$ that are considered most relevant for a user to understand the prediction $\hat{y}$ (see [@Montavon2018 Definition 2] and [@Molnar2019]). Thus, we consider explanations of the form $$\label{equ_def_explanation} e {:=}\{ x_{i} \}_{i \in \mathcal{E}} \mbox{ with some subset } \mathcal{E} \subseteq \{1,\ldots,{n}\}.$$ The complexity of an explanation $e$ is measured by the number $\|\mathcal{E}\|$ of features that contribute to it. We limit the number of features contributing to an explanation by a fixed (small) sparsity level, $$\| \mathcal{E} \| \leq {s}(\ll {n}).$$ Modelling the feature vector ${{\mathbf x}}$ as Gaussian implies that the prediction $\hat{y}$ and user summary $u$ obtained from is jointly Gaussian for a given $\mathcal{E}$ . Basic properties of multivariate normal distributions [@coverthomas Ch. 8], allow to develop as $$\begin{aligned} \label{equ_sup_mi_Gauss} \max_{\substack{\mathcal{E} \subseteq \{1,\ldots,{n}\} \\ \|\mathcal{E}\| \leq s }} & I(e;\hat{y}\|u) \nonumber \\ &= h(\hat{y}\|u) - h(\hat{y}\|u,\mathcal{E}) \nonumber \\[3mm] &= (1/2) \log {\mathbf{C}}_{\hat{y}\|u} - (1/2) \log \det {\mathbf{C}}_{\hat{y}\|u,\mathcal{E}} \nonumber \\[3mm] & = (1/2) \log \sigma^2_{\hat{y}\|u} - (1/2) \log \sigma^{2}_{\hat{y}\|u,\mathcal{E}}. \end{aligned}$$ Here, $\sigma^2_{\hat{y}\|u}$ denotes the conditional variance of the prediction $\hat{y}$, conditioned on the user summary $u$. Similarly, $\sigma^{2}_{\hat{y}\|u,\mathcal{E}}$ denotes the conditional variance of $\hat{y}$, conditioned on the user summary $u$ and the subset $\{x_r\}_{r \in \mathcal{E}}$ of features. The last step in follows from the fact that $\hat{y}$ is a scalar random variable. The first component of the last expression in does not depend on the choice $\mathcal{E}$ for the explanation $e$ (see ). Therefore, the optimal choice $\mathcal{E}$ solves $$\label{equ_sup_m_sigma} \sup_{\|\mathcal{E}\| \leq {s}} - (1/2) \log \sigma^{2}_{\hat{y}\|u,\mathcal{E}}.$$ The maximization is equivalent to $$\label{equ_min_variance} \inf_{\|\mathcal{E}\| \leq {s}} \sigma^{2}_{\hat{y}\|u,\mathcal{E}}.$$ In order to solve , we relate the conditional variance $\sigma^{2}_{\hat{y}\|u,\mathcal{E}}$ to a particular decomposition $$\label{equ_def_linear_model} \hat{y} = \alpha u + \sum_{i \in \mathcal{E}} \beta_{i} x_{i} + \varepsilon.$$ For an optimal choice of the coefficients $\alpha$ and $\beta_{i}$, the variance of the error term in is given by $\sigma^{2}_{\hat{y}\|u,\mathcal{E}}$. Indeed, $$\label{equ_def_optimal_coef_linmodel} \min_{\alpha,\beta_{i} \in \mathbb{R}} {{\rm E}}\big\{ \big(\hat{y} - \alpha u - \sum_{i \in \mathcal{E}} \beta_{i} x_{i} \big)^{2}\big\} = \sigma^{2}_{\hat{y}\|u,e}.$$ Inserting into , an optimal choice $\mathcal{E}$ (of feature) for the explanation of prediction $\hat{y}$ to user $u$ is obtained from $$\begin{aligned} & \inf_{\|\mathcal{E}\| \leq {s}} \min_{\alpha,\beta_{i} \in \mathbb{R}} {{\rm E}}\big\{ \big(\hat{y} - \alpha u - \sum_{i \in \mathcal{E}} \beta_{i} x_{i} \big)^{2}\big\} \label{equ_final_opt_E} \\ &= \min_{ \\| {\bm \beta} \\|_{0} \leq s } {{\rm E}}\big\{ \big(\hat{y} - \alpha u - {\bm \beta}^{T} {{\mathbf x}}\big)^{2}\big\} \label{equ_final_opt_beta}.\end{aligned}$$ An optimal subset $\mathcal{E}_{\rm opt}$ of features defining the explanation $e$ is obtained from any solution ${\bm \beta}_{\rm opt}$ of via $$\label{equ_opt_expl_support} \mathcal{E}_{\rm opt} = \operatorname{supp}{\bm \beta}_{\rm opt}.$$ A Simple XML Algorithm {#sec_simple_XML} ====================== Under a Gaussian model for the features of data points, Section \[sec\_optimal\_explanation\] shows how to construct optimal explanations via the (support of the) solutions ${\bm \beta}_{\rm opt}$ of the sparse linear regression problem . In order to obtain a practical algorithm for computing (approximately) optimal explanations , we need to approximate the expectation in with an empirical average over i.i.d. samples $\big({{\mathbf x}}^{(i)},\hat{y}^{(i)},u^{(i)}\big)$ of features, predictions and user summaries. This results in Algorithm \[alg:xml\]. explanation sparsity ${s}$, training samples $\big({{\mathbf x}}^{(i)},\hat{y}^{(i)},u^{(i)}\big)$ for $i=1,\ldots,{m}$ compute $\widehat{\bm \beta}$ by solving $$\label{equ_P0} \widehat{\bm \beta} \in \operatorname{arg\;min}_{ \\| {\bm \beta} \\|_{0} \leq s } \sum_{i=1}^{{m}} \big(\hat{y}^{(i)} - \alpha u^{(i)} - {\bm \beta}^{T} {{\mathbf x}}^{(i)} \big)^{2}$$ feature set $\widehat{\mathcal{E}} {:=}{\rm supp} \widehat{\bm \beta}$ Note that Algorithm \[alg:xml\] is interactive since the user has to provide samples $u^{(i)}$ of its summary for the data points with features ${{\mathbf x}}^{(i)}$. Based on the user input $u^{(i)}$, for $i=1,\ldots,{m}$, Algorithm \[alg:xml\] learns an optimal subset $\mathcal{E}$ of features that are used for the explanation of predictions. The sparse regression problem becomes intractable for large feature length ${n}$. However, if the features are weakly correlated with each other and the user summary $u$, the solutions of can be found by convex optimization. Indeed, for a wide range of settings, sparse regression can be solved via a convex relaxation, known as the least absolute shrinkage and selection operator (Lasso) [@HastieWainwrightBook], $$\label{equ_Lasso} \widehat{\bm \beta} \!\in\!\operatorname{arg\;min}_{ {\bm \beta} \in \mathbb{R}^{{n}}} \sum_{i=1}^{{m}} \big(\hat{y}^{(i)} - \alpha u^{(i)} - {\bm \beta}^{T} {{\mathbf x}}^{(i)} \big)^{2} + \lambda \\| {\bm \beta} \\|_{1}.$$ We have already a good understanding of choosing the Lasso parameter $\lambda$ in such that its solutions coincide with the solutions of (see, e.g., [@HastieWainwrightBook]). Numerical Experiments ===================== We verify the ability of Algorithm \[alg:xml\] to provide explainable ML using a computer vision application. In particular, we consider data points representing square patches of a greyscale aerial photograph of Helsinki city area.[^2] The goal is to predict the greyscale value $y$ of the center (“target”) pixel. In order to predict the greyscale value of the $i$th pixel $y^{(i)}$ we use the greyscale values $x_{j}^{(i)}$ of close-by pixels $j \in \mathcal{P}^{(i)}$. As depicted in Figure \[fig:patch\_target\], the neighbourhood $j \in \mathcal{P}^{(i)}$ is constituted by two rectangular areas that are adjacent to pixel $i$. ![The greyscale level of a particular target pixel within an aerial photograph can be predicted based on the greyscale values of nearby pixels within adjacent rectangles (indicated).[]{data-label="fig:patch_target"}](patchtarget.png){width="\columnwidth"} A user having some prior experience in processing natural images might consider the average greyscale value $$u^{(i)} = (1/ \| \mathcal{P}^{(i)}\|)\sum_{j \in \mathcal{P}^{(i)}} x_{j}^{(i)}$$ as a reasonable summary of the features ${{\mathbf x}}^{(i)} {:=}\{ x_{j}^{(i)}\}_{j \in \mathcal{P}^{(i)}}$. We refer to the Python notebook <https://github.com/alexjungaalto/ResearchPublic/blob/master/itxml.ipynb> for the results of the experiments. Conclusion ========== We have introduced a simple probabilistic model for the predictions of a ML method and the user background. The user background is represented by a summary of the features of a data point. The effect of an explanation is measured by the conditional MI between prediction and explanation, given the user summary of a data point. [^1]: AJ is with the Department of Computer Science, Aalto University, Finland. PHJN is with Lappeenranta-Lahti University of Technology, Finland. This work is partly supported by Academy of Finland via: (a) ee-IoT project n.319009, (b) FIREMAN consortium CHIST-ERA/n.326270, and (c) EnergyNet Research Fellowship n.321265/n.328869. [^2]: The data is freely available via the online map service <https://kartta.hel.fi/>	{ "pile_set_name": "ArXiv" }
--- abstract: 'Evidence is emerging that the luminous X-ray sources in the cores of globular clusters may often consist of, or perhaps even as a class be dominated by, ultracompact (P1 hr) binary stars. To the two such systems already known, in NGC6624 and NGC6712, we now add evidence for two more. We detect large amplitude variability in the candidate optical counterpart for the X-ray source in the core of NGC6652. Although the available observations are relatively brief, the existing [Hubble Space Telescope]{} data indicate a strong 43.6 min periodic modulation of the visible flux of semi-amplitude 30%. Further, although the orbital period of the source in NGC1851 is not yet explicitly measured, we demonstrate that previous correlations of optical luminosity with X-ray luminosity and accretion disk size, strengthened by recent data, strongly imply that the period of that system is also less than 1 hr. Thus currently there is evidence that 4 of the 7 globular cluster X-ray sources with constrained periods are ultracompact, a fraction far greater than that found in X-ray binaries the field.' author: - 'Eric W. Deutsch, Bruce Margon, and Scott F. Anderson' title: 'Ultracompact X-ray Binaries in Globular Clusters: Variability of the Optical Counterpart of X1832–330 in NGC6652[^1] ' --- -0.5in 9.20in 0.216in Accepted for publication in The Astrophysical Journal Letters\ [received 1999 November 5; accepted 1999 December 13]{} INTRODUCTION ============ It has long been recognized that highly luminous ($L_X\sim10^{36-37}$ erg s$^{-1}$) X-ray sources in the cores of globular clusters are grossly overrepresented with respect to the general galactic population (Katz 1975; Clark 1975): clusters contain $10^{-4}$ of the Galaxy’s mass, but $10^{-1}$ of the low-mass X-ray binary (LMXB) sources. A variety of lines of evidence tell us that these objects are neutron stars in very close binary systems, but the precise mechanisms which enhance their formation in clusters, and protect them from disruption thereafter, are obscure. Essentially all of the bright cluster sources are also X-ray bursters, frequently emitting $L_X\sim10^4~L_\odot$ in just a few seconds. Close binaries also have a profound effect on cluster dynamics: just a few such objects in a cluster have a store of orbital kinetic energy which can equal or exceed the orbital energy of all $10^5$ single cluster stars. Thus the study of these rare and odd objects also has important macroscopic implications for the dynamical evolution of clusters (Elson et al. 1987, Hut et al. 1992). Considerable recent progress in the understanding of the intense bursting X-ray sources in globular cluster cores is in large part due to [Hubble Space Telescope]{} ([HST]{}) identifications and follow-on studies of optical/UV counterparts, and to the realization that at least two of the cluster sources are exotic, ultra-short period double-degenerate binary systems: $P=11$ min for X1820–303 in NGC6624 (Stella et al. 1987; Anderson et al. 1997) and $P=21$ min for X1850–087 in NGC6712 (Homer et al. 1996). The optical/UV studies with [HST]{} have in one sense proven highly successful: a plausible optical counterpart has been identified and/or studied in detail in each of the clusters carefully scrutinized thus far. However, the diversity in properties of the six counterparts now identified is enormous, with optical luminosities ranging from $M_B$=6 to $M_B$=1, and confirmed binary orbital periods ranging from 11 min to 17 hr. The only optical counterpart candidate thus far for X1832–330 in NGC6652 was advanced by Deutsch et al. (1998b; hereafter Paper I). The object, denoted Star 49, exhibits a UV excess in the [HST]{} data similar to other known LMXB optical counterparts, and similar absolute magnitude to the optical counterpart of the LMXB in NGC1851. However, the region surveyed in Paper I does not completely cover the [ROSAT]{} X-ray error circle derived in that work, and the images are not very deep. Furthermore, the position of Star 49 is discrepant at the $2.3\sigma$ level with the X-ray coordinates. For these reasons, Paper I suggests that while Star 49 is the best candidate for the optical counterpart, its identification remains tentative. ANALYSIS ======== Since the initial search for the optical counterpart and discovery of Star 49 in Paper I, additional WFPC2 observations have become available in the [Hubble Data Archive]{}. Here we discuss three orbits of F555W (V) and F814W (I) imaging data obtained on 1997 September 5, as well as one orbit of F555W, F439W (B), and F218W imaging data obtained on 1995 September 13. In the former observation, seventeen $\sim$20 s F555W and F814W exposures were taken on the first orbit, twelve 160 s F555W exposures on the second orbit, and twelve 160 s F814W exposures on the third. For the latter observation, fine lock was not achieved and the stellar images are elongated, the cluster is miscentered on the CCDs, and the F218W exposures failed completely. The retake data for this failed observation were successful, and are discussed in Paper I, but did not include the proposed optical counterpart. The early, poor quality data, however, do actually include Star 49, are usable, and will be briefly discussed here as they were overlooked in Paper I. All these data were acquired during unrelated programs to study the cluster NGC6652 itself. Despite the suboptimal sampling of the WFPC2 [Wide Field]{} CCDs, on which the optical counterpart falls in these observations, the cluster is sufficiently sparse and Star 49 sufficiently unblended that aperture photometry is entirely adequate to measure magnitudes for this object and a set of nearby comparison stars. Aperture corrections are taken from Table 2(a) in Holtzman et al. (1995b). The photometric measurements have not been corrected for geometric distortions, nor is any correction for charge transfer efficiency losses (Holtzman et al. 1995b) applied; for most of the images, these effects should contribute errors of only a few percent. We use the photometric zero points for the STMAG system from Table 9 ($Z_{STMAG}$) in Holtzman et al. (1995a). Systematic errors for all magnitudes due to uncertainties in detector performance and absolute calibration are [$\sim$]{}5%. Five nearby reference stars also photometered show no variability to the limit of the derived uncertainties; three of the five are of comparable brightness of Star 49. For Star 49 itself, variability is suggested in the 1995 epoch observations, and large amplitude variability is clearly evident in the 1997 epoch observations. This large-amplitude variability, when coupled with the UV excess demonstrated in Paper I, lends considerable confidence that this object is the correct identification of the optical counterpart to X1832–330. From the three orbits of F555W and F814W observations, we find $<m_{555}>=19.48$ and $<m_{814}>=19.90$. To create a single light curve of all the data, we subtract 0.4 mag from $m_{814}$ to create approximately filter-independent magnitudes. This result is searched for periodic components using algorithms described in Horne & Baliunas (1986). In Fig. 1a we show the Fourier transform with the CLEAN algorithm (Roberts et al. 1987) applied to remove the effects of the window function. We find the strong peak seen at $43.7\pm0.7$ min to have 99.5% significance, based on the original (i.e. prior to CLEANing) periodogram, using methods in Horne & Baliunas (1986). In Fig. 1b we show the entire light curve ($m_{555}$, $m_{814}-0.4$) for the three orbit observation. Uncertainty bars are provided for each datapoint, although they are sometimes smaller than the symbols themselves. A non-linear least squares fitting algorithm is used to determine the best fit sinusoid, which is overplotted on the light curve. The result is a best-fit period $43.6 \pm 0.6$ min, semi-amplitude $0.30 \pm 0.05$ mag, and mean magnitude $19.49 \pm 0.03$. The sinusoid does describe the broad trends in the data quite well, but clearly a large amount of aperiodic flickering is also evident. In Fig. 1c we show the light curve averaged into 10 phase bins. Photometric uncertainties are smaller than the symbols. A few points deviate significantly from the sinusoid fit, most likely due to the strong flickering and small amount of data. During 1998 November 28–30, we obtained $\sim35$ ks of X-ray observation on X1832–330 with [Rossi X-ray Timing Explorer]{}. The data are processed through the standard [Ftools]{} package to obtain a calibrated light curve. Two Type I X-ray burst events are evident, confirming the bursting nature of this source first reported by in ’t Zand et al. (1998) with [BeppoSAX]{} observations. After background subtraction, we measure a persistent countrate $\sim100$ s$^{-1}$, which is $\sim6$ mCrab, similar to fluxes reported previously for this object. A search of the background-subtracted light curve reveals no significant periodicities, except for some power at half the [RXTE]{} orbital period, apparently induced by the calculated background model. In particular, there does not appear to be any significant power at the 43.6 min optical period. The rms scatter in the X-ray light curve is consistent with Poisson noise; we find no evidence for flickering, which might be expected based on our optical observations. However, as the X-ray and optical observations were made over a year apart, no firm conclusions can be drawn from the lack of X-ray flickering. A further analysis of the light curve, spectra, and bursts in these X-ray data will be discussed elsewhere. Mukai & Smale (2000) present an X-ray observation of X1832–330 from a 1996 [ASCA]{} observation. They also find no periodic X-ray modulation, but they do provide evidence for X-ray variability with a similar timescale as the flickering seen in our optical light curve. DISCUSSION ========== The peak in the periodogram of the optical data has 99.5% significance, indicating that it is quite improbable that uncorrelated, Gaussian noise would randomly generate such a strong periodic signal. However, the evident scatter in the photometry is not due to measurement uncertainties, but rather an inherent flickering in the source, and this behavior is likely to produce significantly correlated “noise”. Thus the formal significance calculation for the periodicity may overestimate the actual confidence. Our derived period is close to, but statistically different from, half the [HST]{} orbital period. Two further tests increase our confidence that this behavior is not an artifact of the [HST]{} orbit. As noted in §2.1, several stars near to and of similar magnitude to Star 49 have been measured from the same data, and show no variability at this or other periods. We have also randomized the association of observation times versus magnitudes for Star 49 and rereduced the data. Although periods due to incomplete removal of the window function should then remain, as the observing times are identical in these randomized data and in the original observations, the resulting periodograms show no significant power at 43 min. We find that only $\sim5\times10^{-4}$ of 10000 trials show a peak at any frequency as high as the 43 min one. Although the marked variability we report here adds confidence to the identification of Star 49 with the X-ray source, the mediocre agreement of the X-ray position with the object still leaves some uncertainty. This issue will almost surely be settled by a scheduled observation by the [Chandra X-ray Observatory]{}, which should yield a highly accurate position. However, if we accept that our observed optical variations in Star 49 are indeed periodic, there are few alternatives to identifying this object as a LMXB, given its measured characteristics, irrespective of the issues of the X-ray position. The most extreme SX Phe stars, for example, have periods less than 1 hr, but do not display the marked flickering we observe, so stellar pulsation seems implausible. If the period is instead orbital, the flickering implies a mass-transfer system. However, no classical cataclysmic variables (CVs) are known with periods less than 1 hr, and although CVs share the colors and flickering of Star 49, they are typically 3–4 mag less luminous than our object in any case. The He-rich AM CVn stars have the appropriate colors, flickering and period range, but are thought to have $M_V\sim10$ (Warner 1995), so would be $\sim10^2$ fainter than Star 49. Thus, given the measured period, luminosity, colors, and flickering of this object, one would likely conclude it is an LMXB even without knowledge of the fact that there is indeed a bright X-ray source observed in the region. We summarize a variety of parameters for all of the globular cluster LMXB sources and their host clusters in Table 1. Cluster data are primarily compiled from Djorgovski (1993) and other references in the same volume. Column 8 lists ${\rm F_X}$ values, which we derive by taking a mean of the [RXTE]{} ASM flux measurements since 1996, and applying a simple correction to convert approximately to $\mu$Jy. In column 9, we apply the distance correction and give an approximate X-ray luminosity in $10^{36}$ erg s$^{-1}$ for the 2–10 keV ASM band. The absolute calibration is only an estimate and should be treated with caution, but the relative values are likely reliable. Finally, in column 10 we list $\xi=B_0+2.5{\rm\,log\,F_X(\mu Jy})$, the parameter used by van Paradijs & McClintock (1995) to characterize the ratio of X-ray to optical flux. That the optical luminosity should depend upon the X-ray luminosity and the size of the accretion disk has been quantified by van Paradijs & McClintock (1994; hereafter vPM94). They define the parameter $\Sigma=(L_X/L_{\rm Edd})^{1/2} (P/1\,{\rm hr})^{2/3}$ and find a strong correlation, such that M$_V=1.57(\pm0.24)-2.27(\pm0.32)\log \Sigma$. In Fig. \[periodrel\] we show a similar figure as in vPM94, but we use M$_B$ instead of M$_V$, which is likely to be reasonable as vPM94 find an average $(B-V)_0=-0.09\pm0.14$ for field LMXBs. The data for globular cluster LMXBs are derived here and from Deutsch et al. (1998a), and are plotted with large diamonds, approximately indicating the entire known range of optical and X-ray luminosity. The solid line indicates the best fit to all LMXBs by vPM94. The dotted lines denote the apparent full range of possible values (using vPM94’s best fit slope). For NGC1851, no orbital period is known, but the optical and X-ray luminosities are measured. We therefore draw a dashed line which is likely to encompass the probable range of orbital periods, $0.2-0.85$ hr, where the lower bound is taken to be the shortest orbital period known and the upper value is the maximum period implied by our dotted line bounds. We thus predict that the orbital period of X0512–401 will prove to be less than 1 hr. Based on model accretion disks, Deutsch et al. (2000) also infer that X0512–401 must have orbital period less than one hour. As listed in Table 1, an eclipse period of 12.4 hr was recently reported by in ’t Zand et al. (2000) for the LMXB in the globular cluster Terzan 6. Using behavior exhibited by the GC LMXB sources in Fig. 2, we can now infer that the optical counterpart of that source (for which there has not yet been a search) will have $M_B\sim2\pm1$. The high inclination ($i>74^\circ$) inferred from the eclipse behavior by in ’t Zand et al. (2000) suggests that the luminosity may well be at the fainter end of the above range, and thus similar to the optical counterpart in NGC6441. However, the high reddening to Terzan 6 will making discovery of the optical counterpart by conventional means ([HST]{} observations of UV-excess) extremely difficult (Deutsch et al. 1998b). A search for eclipses with infrared imaging of the X-ray error circle may be the easiest method of isolating the optical/IR counterpart of this source. CONCLUSIONS =========== The ultraviolet-excess candidate for the optical counterpart of the intense X-ray source in NGC6652 suggested by Deutsch et al. (1998b) is found to be highly photometrically variable. Although the data are of limited length, the evidence is strong that a significant component of the variability is periodic, with $P=43.6$ min, most likely the orbital period of the system. Regardless of whether or not the variability is periodic, the marked amplitude of the variations significantly strengthens the case for the identification of the object with the bursting X-ray source. Somewhat tempering this conclusion is the poor positional agreement of the object with the only extant X-ray data, although [Chandra X-ray Observatory]{} observations will almost surely settle this issue. The lack of similar X-ray variability is inconclusive. Star 49 is clearly unusual regardless of its association with the X-ray source, but the probability of two such unrelated objects falling within a few arcseconds of each other is presumably modest, so we favor the identification of the star and the X-ray burster, pending the [Chandra]{} data. We examine a homogeneous set of [HST]{} data on globular cluster X-ray source counterparts, including Star 49 in NGC6652, and find that they fit well with the correlation of optical luminosity, X-ray luminosity, and accretion disk size previously discussed by vPM94. Even if the 43 min period is not confirmed, the orbital period of X1832–330 must still be less than $\sim2$ hr if it is to follow the relation of this diagram. Using a somewhat different argument, Mukai & Smale (2000) also infer that X1832–330 in NGC6652 is a short period system. The correlation in this diagram also strongly implies that the X-ray source in NGC1851 must have orbital period $P<1$ hr as well. A similar conclusion is reached by Deutsch et al. (2000) via an independent argument, through examination of the spectral energy distribution of that object. Thus four of the seven central globular cluster X-ray sources where orbital periods are constrained or known are inferred to be ultracompact, a fraction considerably in excess of that for field low mass X-ray binaries; only $\sim7$% of field LMXBs with known periods have $P<1$ hr in the compilation of van Paradijs (1995). In fact, if ultracompact systems in GC LMXBs were as rare as in the field, then the binomial probability that at least four out of seven systems are by chance found to be ultracompact is only $7\times10^{-4}$. One can readily imagine multiple explanations for this significant overabundance of compact systems, although the unique dynamic environment of cluster cores is certainly a most tempting factor to invoke. It is not clear whether the same explanation applies even to all members of this small sample (the double-degenerate systems in NGC6624 and NGC6712 may be unique), or that observational selection may be at work. [RXTE]{} ASM data products were provided by the ASM/RXTE teams at MIT and at the [RXTE]{} SOF and GOF at NASA’s GSFC. Support for this work was provided by NASA through grants NAG5-7330 and NAG5-7932, as well as grant AR-07990.01 from the STScI, which is operated by AURA, Inc. Anderson, S. F., Margon, B., Deutsch, E. W., Downes, R. A., & Allen, R. G. 1997, ApJ, 482, L69 Clark, G. W. 1975, ApJ, 199, L143 Deutsch, E. W., Anderson, S. F., Margon, B., & Downes, R. A. 1998a, , 493, 775 Deutsch, E. W., Margon, B., & Anderson, S. F. 1998b, AJ, 116, 1301 (Paper I) Deutsch, E. W., Margon, B., & Anderson, S. F. 2000, MNRAS, submitted Djorgovski, S. 1993, in ASP Conf. Ser. 50, Structure and Dynamics of Globular Clusters, ed. S. G. Djorgovski & G. Meylan (San Francisco: ASP), 373 Elson, R., Hut, P., & Inagaki, S. 1987, ARAA, 25, 565 Holtzman, J. A., Burrows, C. J., Casertano, S., Hester, J. J., Trauger, J. T., Watson, A. M., & Worthey, G. 1995a, PASP, 107, 1065 Holtzman, J. A., et al. 1995b, PASP, 107, 156 Homer, L., Charles, P. A., Naylor, T., van Paradijs, J., Aurière, M., & Koch-Miramond, L. 1996, MNRAS, 282, L37 Horne, J. H., & Baliunas, S. L. 1986, ApJ, 302, 757 Hut, P., McMillan, S. L. W., Goodman, J., Mateo, M., Phinney, E. S., Pryor, C., Richer, H. B., Verbunt, F., & Weinberg, M. 1992, PASP, 104, 981 in ’t Zand, J. J. M., et al. 2000, A&A, submitted (preprint astro-ph/9910107) in ’t Zand, J. J. M., Verbunt, F., Heise, J., Muller, J. M., Bazzano, A., Cocchi, M., Natalucci, L., & Ubertini, P. 1998, A&A, 329, L37 Katz, J. I. 1975, Nature, 253, 698 Mukai, K., & Smale, A. P. 2000, ApJ, in press (preprint astro-ph/9911392) Roberts, D. H., Lehár, J., & Dreher, J. W. 1987, AJ, 93, 968 Stella, L., Priedhorsky, W., & White, N. E. 1987, ApJ, 312, L17 van Paradijs, J. 1995, in X-Ray Binaries, ed. W. H. G. Lewin, J. van Paradijs, & E. P. J. van den Heuvel (Cambridge: Cambridge Univ.), 536 van Paradijs, J., & McClintock, J. E. 1994, A&A, 290, 133 (vPM94) van Paradijs, J., & McClintock, J. E. 1995, in X-Ray Binaries, ed. W. H. G. Lewin, J. van Paradijs, & E. P. J. van den Heuvel (Cambridge: Cambridge Univ.), 58 Warner, B. 1995, Cataclysmic Variable Stars (Cambridge: Cambridge Univ. Press) [llccrrrrrr]{} 6.8in X0512$-$401 & NGC 1851 & 0.02 & 15.43 & $ -1.29$ & $<1$ & 5.60 & 5.5 & 1.9 & 22.88X1724$-$307 & Terzan 2 & 1.42 & 14.37 & $ -0.25$ & & & 33.6 & 4.3 & X1730$-$335 & Liller 1 & 3.00 & 14.68 & $ 0.20$ & & & 12.8 & 2.2 & X1732$-$304 & Terzan 1 & 1.64 & 13.85 & $ -0.71$ & & & 6.4 & 0.5 & X1745$-$203 & NGC 6440 & 1.00 & 14.64 & $ -0.34$ & & & 5.5 & 0.9 & X1745$-$248 & Terzan 5 & 1.87 & 14.50 & $ -0.28$ & & & & & X1746$-$370 & NGC 6441 & 0.45 & 15.15 & $ -0.53$ & 5.70 & 2.43 & 28.8 & 7.6 & 21.22X1747$-$313 & Terzan 6 & 2.04 & 14.16 & $ -0.61$ & 12.36 & & 31.8 & 3.4 & X1820$-$303 & NGC 6624 & 0.27 & 14.54 & $ -0.37$ & 0.19 & 2.99 & 269.5 & 40.6 & 23.61X1832$-$330 & NGC 6652 & 0.10 & 14.85 & $ -0.99$ & 0.73 & 5.59 & 10.9 & 2.2 & 23.03X1850$-$087 & NGC 6712 & 0.46 & 14.16 & $ -1.01$ & 0.33 & 4.48 & 7.8 & 0.8 & 20.87X2127+119 & NGC 7078 & 0.09 & 15.11 & $ -2.17$ & 17.10 & 0.66 & 13.9 & 3.5 & 18.63 [^1]: Based on observations with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555.	{ "pile_set_name": "ArXiv" }
--- author: - 'A. R. Ridden-Harper' - 'I. A. G. Snellen' - 'C. U. Keller' - 'R. J. de Kok' - 'E. Di Gloria' - 'H. J. Hoeijmakers' - 'M. Brogi' - 'M. Fridlund' - 'B. L. A. Vermeersen' - 'W. van Westrenen' bibliography: - '55Cnce.bib' title: 'Search for an exosphere in sodium and calcium in the transmission spectrum of exoplanet 55 Cancri e [^1]' --- [The atmospheric and surface characterization of rocky planets is a key goal of exoplanet science. Unfortunately, the measurements required for this are generally out of reach of present-day instrumentation. However, the planet Mercury in our own solar system exhibits a large exosphere composed of atomic species that have been ejected from the planetary surface by the process of sputtering. Since the hottest rocky exoplanets known so far are more than an order of magnitude closer to their parent star than Mercury is to the Sun, the sputtering process and the resulting exospheres could be orders of magnitude larger and potentially detectable using transmission spectroscopy, indirectly probing their surface compositions.]{} [The aim of this work is to search for an absorption signal from exospheric sodium (Na) and singly ionized calcium (Ca$^+$) in the optical transmission spectrum of the hot rocky super-Earth 55 Cancri e. Although the current best-fitting models to the planet mass and radius require a possible atmospheric component, uncertainties in the radius exist, making it possible that 55 Cancri e could be a hot rocky planet without an atmosphere.]{} [High resolution (R$\sim$110000) time-series spectra of five transits of 55 Cancri e, obtained with three different telescopes (UVES/VLT, HARPS/ESO 3.6m & HARPS-N/TNG) were analysed. Targeting the sodium D lines and the calcium H and K lines, the potential planet exospheric signal was filtered out from the much stronger stellar and telluric signals, making use of the change of the radial component of the orbital velocity of the planet over the transit from $-$57 to +57 km sec$^{-1}$.]{} [Combining all five transit data sets, we detect a signal potentially associated with sodium in the planet exosphere at a statistical significance level of 3$\sigma$. Combining the four HARPS transits that cover the calcium H and K lines, we also find a potential signal from ionized calcium (4.1 $\sigma$). Interestingly, this latter signal originates from just one of the transit measurements - with a 4.9$\sigma$ detection at this epoch. Unfortunately, due to the low significance of the measured sodium signal and the potentially variable Ca$^+$ signal, we estimate the p-values of these signals to be too high (corresponding to $<$4$\sigma$) to claim unambiguous exospheric detections. By comparing the observed signals with artificial signals injected early in the analysis, the absorption by Na and Ca$^+$ are estimated to be at a level of $\sim 2.3\times 10^{-3}$ and $\sim 7.0\times 10^{-2}$ respectively, relative to the stellar spectrum.]{} [If confirmed, the $3\sigma$ signal would correspond to an optically thick sodium exosphere with a radius of 5 $ \mathrm{R_\oplus}$, which is comparable to the Roche lobe radius of the planet. The 4.9$\sigma$ detection of Ca$^+$ in a single HARPS data set would correspond to an optically thick Ca$^+$ exosphere approximately five times larger than the Roche lobe radius. If this were a real detection, it would imply that the exosphere exhibits extreme variability. Although no formal detection has been made, we advocate that probing the exospheres of hot super-Earths in this way has great potential, also knowing that Mercury’s exosphere varies significantly over time. It may be a fast route towards the first characterization of the surface properties of this enigmatic class of planets.]{} Introduction ============ Transit and radial velocity surveys have revealed a new class of rocky planets orbiting their parent stars at very short distances (0.014 - 0.017 au[^2]). Their evolutionary history is unknown. They may be rocky planets formed at significantly larger distances that subsequently migrated inwards, or could originally be gas-rich planets which lost their gaseous envelopes during migration through tidal heating and/or direct stellar irradiation [@Raymond2008]. Insights into the composition of these hot rocky planets would help to distinguish between the different scenarios. Although the first secondary eclipse measurements have been presented in the literature [@Demory2012], showing them to be indeed very hot, with observed surface or atmospheric temperatures of 1300 - 3000K, detailed observations that could reveal atmospheric or surface compositions are beyond the reach of current instrumentation. One physical process that could reveal information of a planet’s surface composition, potentially also with current instruments, is that of sputtering. Atomic species are ejected from the planet surface by the intense stellar wind of charged particles, creating an extended exosphere around the planet. This process is well known from planet Mercury in our own solar system. It has an exosphere composed of atomic species including sodium (Na), calcium (Ca) and magnesium (Mg), which are thought to be the results of sputtering, thermal vaporisation, photon-stimulated desorption, and meteoroid impact vaporisation. Since the discovery of an emission spectrum of sodium in the exosphere of Mercury by [@Potter1985], it has been subsequently detected many times [see @Killen2007 for a review] in emission, and less commonly in absorption during the transit of Mercury in front of the Sun [@Potter2013]. These decades of observations have revealed that sodium in the exosphere of Mercury shows a great deal of spatial and temporal variability. Rapid variations at a 50% level on timescales of a day in the ion-sputtering component of the sodium in Mercury’s exosphere due to variability in the magnetosphere have been observed, as well as latitudinal and/or longitudinal variations [@Potter1990; @Killen2007]. In addition, long-term variations on timescales of months to years in photon-stimulated desorption [@Lammer2003; @Killen2007] and radiation pressure acceleration [@Smyth1995; @Killen2007] have been observed, as well as variations in meteoritic vaporisation [@Morgan1988; @Killen2007]. [@Mura2011] argued for the first time that such exospheres resulting from the sputtering process may be observable for hot rocky exoplanets. Their simulations for CoRoT-7b suggest that it may have a high escape rate of species such as Na, Ca$^+$, and Mg$^+$ which likely form a tail tens of planetary radii long. [@Guenther2011] observed a transit of CoRoT-7b with the UVES instrument on the VLT with a focus on Na, Ca, and Ca$^+$. While [@Guenther2011] express their derived upper limits in units of stellar luminosity ($2-6\times10^{-6}L_{}$), these limits appear to correspond to an absorption level on the order of approximately 3$\times$10$^{-3}$ smeared out over a 55 km sec$^{-1}$ velocity bin due to the change in the radial component of the orbital velocity of the planet during their long exposures. In this paper we target the exoplanet 55 Cancri e, whose host star has an apparent magnitude of $V=5.95$, 200 times brighter than CoRoT-7. In addition, [@Schaefer2009] argue that a tidally locked hot rocky super-Earth could have a magma ocean that releases vapours to produce a silicate based atmosphere. Their models show that Na is likely the most abundant constituent of such an atmosphere, which they believe could form a large cloud of Na through interaction with the stellar wind. Considerable progress regarding the detection and study of exospheres of hot gas giant exoplanets has already been made. Hydrogen exospheres extending beyond the Roche lobe have been repeatedly detected around HD 209458b [@Vidal-Madjar2003; @Vidal-Madjar2004] and HD 189733b [@Lecavelier_des_Etangs2010], where the hydrogen signal from HD 189733b has been claimed to show temporal variation [@Lecavelier_des_Etangs2012]. Heavier atoms and ions have been detected in the exosphere of HD 209458b, including C$^+$ [@Vidal-Madjar2004; @Linsky2010] and, more tentatively, O [@Vidal-Madjar2004], Mg [@Vidal-Madjar2013] and Si$^{2+}$ [@Linsky2010]. Exospheric studies have recently also been extended to smaller planets with the detection of hydrogen around the warm Neptune GJ 436b [@Ehrenreich2015]. We note that no hydrogen exosphere was detected around 55 Cancri e [@Ehrenreich2012], which is the object of this study. The hot, rocky super-Earth type planet, 55 Cancri e (or $\mathbf{\rho^1}$ Cancri e, 55 Cnc e) orbits a bright (V=5.95) 0.905 $M_\sun$ star. It has a very short orbital period of 17.7 hours (see Table \[table:55Cncparameters\] for uncertainties), a radius of $2.173$ $R_\oplus$ [@Gillon2012], a mass of $8.09$ $M_\oplus$, and an inferred average density of $5.51$ $\mathrm{g cm^{-3}}$ [@Nelson2014]. Transits of 55 Cnc e have been detected with broadband photometry from space in the visible [@Winn2011] and infra-red [@Demory2011], and recently also from the ground [@deMooij2014]. There is significant debate in the literature about the chemical composition and interior structure of 55 Cnc e. Using the internal structure model by [@Valencia2006; @Valencia2010], [@Gillon2012] argue that 55 Cnc e is likely a rocky, oxygen-rich planet composed of silicates with a gaseous envelope consisting of either a mixture of hydrogen and helium of approximately 0.1% by mass or a water atmosphere of approximately 20% by mass. However, because such a low mass H-He atmosphere would escape over a timescale of millions of years, while a water-vapour atmosphere could survive over billions of years, the water-vapour atmosphere interpretation is favoured, where the water-vapour is in a super-critical form due to its high temperature. Furthermore, [@Ehrenreich2012] found that 55 Cnc e lacks a H exosphere which could be the result of complete H loss from the atmosphere in the past. In contrast, [@Madhusudhan2012] claim that if 55 Cnc e were to be a carbon-rich planet, a different structure is possible where Fe, C (in the form of graphite and diamond), SiC, and silicates of a wide range of mass fractions could explain its density without the need for a gaseous envelope. While the C/O ratio of 55 Cnc was previously thought to be $>$1, a subsequent analysis by [@Teske2013] found that it more likely has a C/O ratio of 0.8. This value is lower than the value adopted by [@Madhusudhan2012] of 1.12$\pm$0.19; however, it still corresponds to the predicted minimum value of 0.8 necessary for the formation of a carbon-rich condensate under the assumption of equilibrium [@Larimer1975]. Furthermore, [@Demory2015] report a 4$\sigma$ detection of variability in the day-side thermal emission of 55 Cnc e, with the emissions varying by a factor of 3.7 between 2012 to 2013. They also tentatively suggest variations in the transit depth and calculate the planetary radii to range from 1.75$\pm$0.13 $R_\oplus$ to 2.25$\pm$0.17 $R_\oplus$ with a mean value of 1.92$\pm$0.08 $R_\oplus$, which is approximately 2$\sigma$ smaller than the value published by [@Gillon2012] of 2.17$\pm$0.10$R_\oplus$ based on Spitzer+MOST data. We believe that this smaller radius implies that the need for a significant atmosphere to explain the planet’s radius is significantly reduced. If 55 Cnc e does not have an atmosphere, its surface would be directly exposed to stellar radiation, making it analogous to Mercury. It is likely that the processes which produce the exosphere of Mercury would be much more pronounced on 55 Cnc e because it receives a bolometric flux from its host star that is approximately 500 times greater than Mercury receives from the Sun. This corresponds to an equilibrium temperature of 55 Cnc e of almost 2000 K. [@Demory2015] claim to have detected brightness temperatures which vary from 1300 K to 3000 K; however, the mechanism which causes this variability is not understood. [@Demory2016] report the observation of a complete phase curve of 55 Cnc e in the 4.5$\mu$m channel of the Spitzer Space Telescope Infrared Array Camera which allowed them to construct a longitudinal thermal brightness map due to 55 Cnc e being tidally locked to its host star. This map revealed that 55 Cnc e has a strong day-night temperature contrast with temperatures of 2700 K and 1380 K on the day and night sides respectively. Furthermore, they found that the day side exhibits highly asymmetric thermal emissions, with a hot spot located 41 degrees east of the substellar point. These observations were interpreted as being either due to an atmosphere with heat recirculation confined to the day side only, or a planet without an atmosphere with low-viscosity magma flows on the surface. Atmospheric escape rate arguments indicate that it is unlikely that 55 Cnc e has a thick atmosphere, so the magma ocean interpretation is favoured. In this paper, we aim to search for an absorption signal from exospheric sodium (Na) and singly ionized calcium (Ca$^+$) in the optical transmission spectrum of the hot rocky super-Earth 55 Cnc e. This paper is structured as follows. Section \[sec:obdata\] describes the data and Section \[sec:analysis\] explains the methods used in this analysis. Section \[sec:results\] presents the results, and Section \[sec:discussion\_and\_conclusion\] discusses the interpretation of the results and concludes. [@l@c@c@]{} Parameter & Value & Source\ \ \ Distance (pc) & 12.34 $\pm$ 0.11 & [@vanLeeuwen2007]\ Radius (R$_\sun$) & 0.943$\pm$0.010 & [@vonBraun2011]\ Luminosity (L$_\sun$) & 0.582 $\pm$ 0.014 & [@vonBraun2011]\ T$_{EFF}$ (K) & 5196$\pm$24 & [@vonBraun2011]\ Mass (M$_\sun$) & 0.905$\pm$0.015 & [@vonBraun2011]\ log $g$ & 4.45$\pm$0.01 & [@vonBraun2011]\ Radial velocity & 27.58 $\pm$ 0.07 & [@Nidever2002]\ \ \ \ Period (days) & $0.7365449 \pm 0.000005$ & [@Gillon2012]\ Orb. radius (AU) & 0.0154$\pm$0.0001 & \\ Radius $(R_\oplus)$ & $2.173 \pm 0.098$ & [@Gillon2012]\ Mass $(M_\oplus)$ & $8.09\pm0.26$ & [@Nelson2014]\ Density $(\mathrm{g cm^{-3}})$ & $5.51\pm^{1.32}_{1.00}$ & [@Nelson2014]\ \ \ Observational data \[sec:obdata\] ================================= High-dispersion spectral time series of five transits of 55 Cnc e taken with three different telescopes were used for our analysis. The five data sets each cover one transit including observations just before and after the transit. We observed one transit using the Ultraviolet and Visual Echelle Spectrograph (UVES) [@Dekker2000] installed on the Nasmyth B focus of the Very Large Telescope (VLT) at the Paranal Observatory. Furthermore, we retrieved additional data sets from observatory archives. Two of these were observed with the High Accuracy Radial Velocity Planet Searcher (HARPS) [@Mayor2003] located at the ESO 3.6m Telescope at the La Silla Observatory, and two from its northern-hemisphere copy - HARPS-N [@Cosentino2012] located at the 3.6m Telescopio Nazionale Galileo at the Roque de los Muchachos Observatory. An overview of all observations is shown in Table \[table:timing\]. UVES data --------- 138 UVES spectra were obtained of 55 Cnc. The transit timing, dates, exposure times, observational cadence and phase coverage are presented in Table \[table:timing\]. The observations were made using the red arm of UVES, utilizing grating CD\#3 with a central wavelength of 580.0 nm, resulting in a wavelength range of $4726.5 - 6835.1$ $\AA$. A resolving power of $R\approx 110000$ was achieved using a slit width of 0.3$''$ and image slicer \#3 to minimize the slit losses. Using no charge-coupled device (CCD) binning, the sampling is two pixels per spectral element [@D'Odorico2000]. Unfortunately, cirrus clouds were present during our observations which considerably decreased the signal-to-noise ratio (S/N) in the spectra, ranging from S/N=180 during relatively good spells down to S/N=60 per pixel. HARPS data ---------- The HARPS data used for our analysis cover two transits and were originally taken for ESO programme 288.C-5010 (PI: A. Triaud) which was used by [@Lopez-Morales2014] to investigate the Rossiter-Mclaughlin effect in 55 Cnc e. We retrieved the pipeline-reduced data from the ESO Science Archive Facility[^3]. HARPS has a resolving power of R$\approx$115000 and a wavelength range of $\mathrm{3800 - 6910}$ $\AA$. It is enclosed in a vacuum vessel, pressure and temperature controlled to a precision of $\pm$0.01 mbar and $\pm$0.01 K respectively, resulting in a wavelength precision of $\lesssim$ 0.5 m s$^{-1}$ night$^{-1}$ [@Bonfils2013]. It has two fibres which feed the spectrograph with light from the telescope and calibration lamp. The fibre aperture on the sky is 1$''$. The CCD has a pixel size of 15 $\mathrm{\mu m}$ and a sampling of 3.2 pixels per spectral element [@Mayor2003]. The transit timing, dates, exposure times, observational cadence, and phase coverage are presented in Table \[table:timing\]. HARPS-N data ------------ The HARPS-N observations also cover two transits, and were originally taken in TNG Observing programme CAT13B\_33 (PI: F. Rodler), also to investigate the Rossiter McLaughlin effect by the same team [@Lopez-Morales2014]. The pipeline-reduced data was retrieved by us from the TNG data archive [^4]. HARPS-N is a copy of HARPS so its properties are all identical or very similar to HARPS at ESO. It has a slightly different wavelength range of $3830 - 6900$ $\AA$ and a sampling of 3.3 pixels per FWHM. It also has a greater temperature stability than HARPS of 0.001 K, giving a short-term precision of 0.3 m s$^{-1}$ and a global long-term precision of better than 0.6 m$^{-1}$ [^5]. The fibre aperture on sky and spectral resolution are identical to those of HARPS. The transit timing, dates, exposure times, observational cadence, and phase coverage are presented in Table \[table:timing\]. We note that an additional five publicly available data sets[^6] of the transit of 55 Cnc e were obtained with HARPS-N by [@Bourrier2014] to investigate the Rossiter-McLaughlin effect. The individual spectra of these data sets have exposure times of 360 seconds, which is twice the average exposure time of the data used in this study. Therefore, we chose to not use these data sets because due to the very rapid change in the radial component of the orbital velocity of the planet (114 km s$^{-1}$ over the transit), any planet signature would be smeared out by ten pixels, significantly decreasing its peak. Data analysis \[sec:analysis\] ============================== In our analysis we concentrate on the H and K lines of ionized calcium (at 3968.47 Å and 3933.66 Å respectively) and the two sodium D lines (5889.95 Å and 5895.92 Å). While the sodium lines are covered by all data sets, the ionized calcium lines are only present in the HARPS and HARPS-N data and not in the UVES data. Processing of UVES spectra \[sec:UVESprocessing\] ------------------------------------------------- The observed spectra are dominated by stellar and possible telluric absorption lines which are orders of magnitude stronger than the expected planet features. Since the stellar and telluric absorption lines are quasi-fixed in wavelength (the stellar lines change in radial velocity by approximately 1.4 m sec$^{-1}$ during the four hour observations [@McArthur2004]) and the radial component of the orbital velocity of the planet changes by tens of km sec$^{-1}$ during the transit, the change in the Doppler shift of the planet lines can be used to separate the planet signal from that of the star and the Earth’s atmosphere. The procedure we used to carry out this processing is very similar to that used in previous work [eg. @Snellen2010; @Hoeijmakers2015] and the individual steps are summarised below. 1. Extraction of wavelength calibrated 1D spectra. The UVES data were reduced using the standard ESO UVES reduction pipeline [@Ballester2000] which was executed with Gasgano[^7] and EsoRex [^8]. The pipeline produced a one-dimensional wavelength calibrated spectrum for each order for each exposure. 2. Normalization of the spectra to a common flux level. Variation in instrumental throughput (for example, due to slit losses) and atmospheric absorption result in the spectra having different baseline fluxes. To normalise the individual spectra to a common flux level, every spectrum was divided through its median value. The median value of a spectrum was used to minimize the influence of cosmic ray hits. This scaling can be performed because this analysis does not depend on the absolute flux, but instead only on the relative changes in flux as a function of wavelength. 3. Alignment of spectra. It is important for our analysis that all of the individual stellar spectra are in the same intrinsic wavelength frame. Since the radial component of the barycentric velocity changes during an observing night, and the absolute wavelength solution of UVES is unstable at the subpixel level, the spectra need to be re-aligned to a common wavelength frame. To do this, Gaussians were fitted to narrow stellar lines close to the sodium D lines in each spectrum to determine the offset relative to a Kurucz model stellar spectrum with atmospheric parameters of $\mathrm{T_{eff}}$ = 5000 K, log(g) = 4.5 [@Castelli2004] that was Doppler shifted to account for the system velocity of 55 Cnc of $27.58 \pm 0.07$ $\mathrm{km s^{-1}}$. These offsets were then used to update the intrinsic wavelength solution for the star. The normalized and aligned spectra are shown in the top panel of Fig. \[fig:pcaremoved\]. 4. [Removal of cosmic rays. The standard UVES data reduction recipes do not remove cosmic ray hits for observations made with the image slicer. Therefore, after the UVES spectra were normalized and aligned, cosmic rays were removed by fitting a linear function at each wavelength position through all spectra, so that cosmic rays could be identified as being outliers from the fit. They were then replaced with the interpolated value from the fit. This process was iterated twice with different threshold values so it only identified very strong cosmic rays on the first iteration. This was necessary because the presence of very strong cosmic rays could skew the linear fit and cause weaker cosmic rays to be missed.]{} 5. [Removal of the stellar absorption features. The stellar spectrum of 55 Cnc was assumed to be constant during a night of observations. This allowed the stellar features to be removed by dividing every observed spectrum from a single night by the median of all of the observed spectra from that night. This only slightly weakened the strength of potential planet absorption lines because they changed wavelength significantly ($>$100 pixels) during the transit. The resulting spectra are shown in the second panel of Fig.\[fig:pcaremoved\].]{} 6. [Removal of large systematic trends. Significant systematic trends in the residual spectra in the wavelength direction became apparent after the stellar features had been removed. These trends, which differed for different spectra, were fitted with a linear slope, that was subsequently removed at the beginning of Step 2 (see above). Steps 2 to 5 were redone, after which we proceeded with Step 7.]{} 7. Removal of telluric lines with principal component analysis (PCA). Telluric absorption lines change in strength, mainly due to the change in airmass during observations, but also possibly due to variations in the water vapour content of the Earth’s atmosphere. We removed the telluric absorptions in the sodium D region using PCA (also know as singular value decomposition) over the time domain. This method relies on the assumption that all of the telluric lines vary in the same way and is discussed in Section 2 of [@dekok2013]. Since Step 3 of our data analysis aligned the spectra to the stellar rest frame, the telluric lines show a small shift in position during the night since they are in the rest frame of the observer. However, the PCA algorithm was able to mostly remove the misaligned telluric lines, as shown in the third panel of Fig. \[fig:pcaremoved\], while the components that were removed are shown in Fig. \[fig:pca\_components\]. Some weak residual features from the telluric lines remain after the PCA. These are probably caused by the line width of the telluric lines changing slightly during the night. The PCA algorithm is a blind process so if it is allowed to remove a large number of components, it will eventually remove all variation in the data, including the planet signal. However, only four PCA components were required to remove the telluric lines. By injecting artificial planet signals (see Section \[subsect:injected\] and the lower panel of Fig. \[fig:pcaremoved\]) we show that the planet signal is left intact by this procedure. ![image](ProcessingExampleUvesPCA4Injected.png){width="15cm"} ![image](PCA_components.png){width="15cm"} 8. [Weighting by noise as a function of wavelength. The values at each wavelength position were subsequently weighted down by the noise, derived from the standard deviation of the residuals as function of time at that position. Practically, this only influences the region directly around the cores of the two stellar sodium D lines, as can be seen by comparing panels 3 and 4 of Fig. \[fig:pcaremoved\]. This naturally weighs down the contribution from the spectra during which the planet absorption overlaps with the strong stellar sodium lines. The effect this has on the planet signal is illustrated in panel 5 of Fig. \[fig:pcaremoved\]. An artificial planet signal is injected into the observed spectra (see section \[subsect:injected\]) as a 3% absorption relative to the stellar spectrum. This signal is weighted with a box-shaped transit profile which is reasonable for a small planet such as 55 Cnc e. At the mid-transit point when the planet signal is at the same wavelength as the stellar signal (and thus falls in the cores of the Na D lines), the retrieved planet signal is reduced by a factor of approximately]{} 10. Processing of HARPS and HARPS-N data ------------------------------------ Except for small differences, the processing of the HARPS and HARPS-N data was performed in a similar way to that of the UVES data explained above. Since the data retrieved from the data archives of both telescopes is completely reduced and wavelength-calibrated, Step 1 was not needed. In addition, the wavelength calibration of both HARPS and HARPS-N is stable at the 1 m sec$^{-1}$ level, and delivered to the user in the restframe of the star. Therefore also step 3 was not needed. In addition to the sodium D lines (5889.95Å and 5895.92Å), the HARPS and HARPS-N spectra also cover the Ca H and K lines (at 3968.47Å and 3933.66Å respectively). While the sodium lines are covered by all data sets, the ionized calcium H and K lines are only present in the HARPS and HARPS-N data and not in the UVES data. Combining the different data sets \[subsection:combining\_Na\_signal\] ---------------------------------------------------------------------- The final step in the analysis is to merge the signal from the two sodium D lines (and calcium H and K lines) and combine the signal from all the in-transit spectra. In addition, we also combine the data sets from the different telescopes. Two regions of 16$\AA$ centred on the positions of the Na D$_2$ and D$_1$ lines in the residual spectra were averaged with weights proportional to the relative theoretical line strengths. Each spectrum was subsequently shifted to the planet rest frame and added in time over the transit. The signal from the different data sets was subsequently combined using weights proportional to the total in-transit S/N of the data set. The unfavourable observing conditions during the UVES observations caused the UVES observations to have a comparable total S/N to the HARPS and HARPS-N observations (see Table \[table:SNR\]). The final data from the calcium H and K lines were produced in the same way using a weighting ratio of Ca K/Ca H = 2 for the two lines. ----------- ------- ---- ------ UVES 116.3 47 977 HARPS-N B 161.2 18 838 HARPS-N A 223.3 19 1192 HARPS B 109.6 24 658 HARPS A 143.9 24 863 ----------- ------- ---- ------ : Signal-to-noise-ratios (S/N) of the data sets.[]{data-label="table:SNR"} Injection of artificial planet signals \[subsect:injected\] ----------------------------------------------------------- A useful technique to determine the magnitude of the absorption signal of the planet relative to the stellar spectrum is to inject artificial planet signals early on in the data processing so that the artificial signals are processed in the same way as a real signal would be. This also allows us to check to what level our analysis removes any planet signal. The injection of artificial signals was carried out in a similar way to [@Snellen2010] and [@Hoeijmakers2015]. The artificial planet signals of the sodium $\mathrm{D}_1$ (5895.92 Å) and $\mathrm{D}_2$ (5889.95 Å) lines were generated using Gaussian line profiles of equal width, and with amplitudes with a ratio of $\mathrm{D}_2/\mathrm{D}_1 = 2$. These relative line strengths were calculated according to equation 1 in [@Sharp2007]. We do note that these equations in principle only hold for local thermodynamic equilibrium, while planet exospheres may be better described by radiative transfer algorithms which use a non-Maxwellian velocity distribution function such as in [@Chaufray2013]. The quantum parameters which describe the Na D line transitions were obtained from the Vienna Atomic Line Database (VALD) [@Kupka2000]. The relative strengths of the Na D lines are practically independent of temperature across the range of 1000 to 3000 K. We assumed $\mathrm{T} = 2000$ $\mathrm{K}$. Using the orbital parameters from [@Gillon2012] and assuming a circular orbit [@Demory2012], the radial velocity of the planet can be calculated at the time of each exposure to determine the central wavelengths of the Doppler-shifted sodium lines. The planet signal was injected into the in-transit data by multiplying the observed spectra with the artificial absorption model according to $$\rm F(\lambda)_{\rm injected} = {[1 - A \times F_{\rm model}(\lambda,\rm v_{rad})]F_{\rm obs}(\lambda)},$$ where $\rm F_{obs}(\lambda)$ is the observed spectrum, $\rm F_{model}(\lambda,v_{rad})$ is the Doppler shifted sodium model spectrum, with $\rm A$ as a scaling parameter that sets the amplitude of the sodium D$_2$ line, and $\rm F(\lambda)_{injected}$ is the output spectrum. During an exposure, the radial component of the orbital velocity of the planet changes significantly. For example, the observations taken with HARPS-N have an exposure time of 240 seconds, during which the planet radial velocity changes by approximately 5 km sec$^{-1}$, corresponding to six resolution elements. Thus, even for an intrinsically narrow planet absorption, the observed signal cannot be narrower than five or six pixels. Therefore, the injected artificial sodium lines have a width equal to this observational broadening, which is different for each data set, as shown in Table \[tab:widths\]. The same procedure was used to inject an artificial absorption signal of ionized calcium, using a relative line ratio of Ca K/Ca H = 2 as calculated from [@Sharp2007]. Dataset width $\mathrm{(kms^{-1})}$ width (Å) width (pixels) ----------- ----------------------------- ----------- ---------------- UVES 1.3 0.026 1.0 HARPS A 4.0 0.078 4.9 HARPS B 4.0 0.078 4.9 HARPS-N A 5.3 0.105 6.5 HARPS-N B 5.3 0.105 6.5 : An estimation of the width of the absorption signal from 55 Cnc e based on the average change of its radial velocity during the exposure of each spectrum in each data set.[]{data-label="tab:widths"} [l l l l l l l]{} & & & & &\ & & & & &\ & & & & &\ & 0.871 & 0.850 & 0.850 & 0.944 & 0.939\ & 0.106 & 0.108 & 0.074 & 0.077 & 0.093\ & 109.0 & 265.8 & 264.3 & 211.4 & 211.4\ & 60 & 240 & 240 & 180 & 180$^\dagger$\ & 04:43:03.805 & 02:10:18.576 & 02:18:54.502 & 01:01:25.593 & 03:56:16.478\ & 06:16:30.913 & 04:05:54.424 & 04:14:02.316 & 01:17:26.886 & 04:16:50.727\ & 07:00:13.153 & 04:49:36.664 & 04:57:44.556 & 02:01:09.126 & 05:00:32.967\ & 07:43:55.393 & 05:33:18.904 & 05:41:26.797 & 02:44:51.366 & 05:44:15.207\ & 08:52:37.259 & 06:44:16.990 & 06:15:53.853 & 03:22:45.983 & 06:39:14.017\ & 52 & 26 & 25 & 5 & 6\ & 47 & 18 & 19 & 24 & 24\ & 39 & 17 & 9 & 12 & 17\ & 138 & 61 & 53 & 41 & 47\ Results \[sec:results\] ======================= Sodium ------ The results for sodium are shown in Fig. \[fig:all\_datasets\_no\_injected\]. The left and right panels show the unbinned data and data binned by five pixels (or 0.05 $\AA$) respectively. From top to bottom the panels show the two HARPS data sets, the two from HARPS-N, the UVES data set, and the signal combined from all telescopes. In these panels, the stellar and telluric signals have been removed so all that remains is residual noise and a possible absorption signal from 55 Cnc e. The noise has an approximately Gaussian distribution, so the statistical significance of the detection can be estimated by comparing the depth of the absorption signal to the standard deviation of the noise. Since any planet signal is expected to be broadened due to the long exposure times, the S/N in the unbinned data, calculated as described above, may be underestimated. While there is a hint of planet absorption in the individual UVES data set, this is somewhat more pronounced in the combined (binned) data. This signal has a statistical significance of 3.2$\sigma$ and $3.3\sigma$ in the unbinned and binned data respectively. The binned and unbinned versions of the combined data are also overlayed in Fig. \[fig:binned7perbinNoInjected\] for clarity. By injecting artificial signals at various levels relative to the stellar spectrum we can estimate the strength of the retrieved signal. If real, the planet sodium lines in the combined data are at a level of 2.3$\times 10^{-3}$ with respect to the star. Ionized calcium --------------- The results for ionized calcium are shown in Fig. \[fig:CaIIAllDatasets\]. The individual panels are the same as in Fig. \[fig:all\_datasets\_no\_injected\], except that UVES is not included because the wavelength range of the UVES data does not cover the calcium H and K lines. In contrast to the sodium data, an interesting signal can be seen in the first HARPS-A data set. It shows a feature that has a statistical significance of 4.9$\sigma$, although it is blueshifted with respect to the planet rest frame by approximately 4 km sec$^{-1}$. An overlay of the binned and unbinned data of the HARPS-A data set is shown in Fig. \[fig:CaOverlay\]. The signal does not appear in the other data sets, resulting in a S/N of less than 4 in the combined data. The contribution to the 4.9$\sigma$ Ca$^+$ signal from each individual spectrum of the HARPS-A data set is shown in Fig. \[fig:Camatrix\]. This figure presents the data in the rest-frame of 55 Cnc e so that the features that contribute to the signal lie on a vertical line that is blueshifted by approximately 4 km sec$^{-1}$. The transit duration is indicated in the figure. It can be seen that there are contributions from multiple spectra during transit, indicating that the signal is not caused by a random spurious feature in a single spectrum. If the exosphere is extended beyond the Roche lobe, one would expect it to be distorted and hence possess different velocities relative to the planet and possibly be detectable just before or after transit. However, the S/N in the data is not sufficient to see such distortions or extended absorption signatures. If real, the planet calcium H and K lines in the HARPS-A data set are at a level of $7.0\times 10^{-2}$ with respect to the star. To assess whether the Ca$^+$ signal could originate from variability in the stellar Ca$^+$ H and K lines, we investigated the emission from the cores of the H and K lines in all data sets (Fig. \[fig:CaHandK\_cores\]). We found no evidence for variations in excess of that expected from Poisson noise within each transit data set. Although one dataset (HARPS-N A, hence not corresponding to that showing Ca$^+$) exhibits stronger Ca H+K emission (55 Cnc has a known 39 day periodicity in its Ca$^+$ H and K stellar emission lines; [@Fischer2008]), it shows no variability during the transit. Also, the radial velocity of 55 Cnc e changes by $\pm$57 km sec$^{-1}$ which causes a Doppler shift of $\pm$0.75Å relative to the core of the lines. Since the signals across spectra are combined in the planet rest frame, only the spectra taken close to the mid-transit point (where the planet signal is at the same wavelength as the stellar lines) could be influenced by variability in the Ca$^+$ H and K emission. Therefore, even if there was some variability in emission during a night of observations, its impact on the results would still be limited. ![image](Na_all_plots_5points_per_bin.png){width="15cm"} ![image](Na_single_weighted_avg.png){width="12cm"} ![image](Ca_all_plots_5points_per_bin.png){width="15cm"} ![image](HARPSA_CaII_5points_per_bin.png){width="12cm"} ![image](CaIIMatrixBlue.png){width="15cm"} ![image](CaHandK_cores_medians.png){width="15cm"} Discussion and conclusions \[sec:discussion\_and\_conclusion\] ============================================================== In this paper we carried out a search for neutral sodium (Na) and singly ionized calcium (Ca$^+$) in the exosphere of the exoplanet 55 Cnc e with transmission spectroscopy. This search yielded a $3.3\sigma$ detection of Na after combining five transit data sets and a 4.9$\sigma$ detection of Ca$^+$ in only one transit data set. We estimated the p-value of the Ca$^+$ detection in one of our four HARPS(-N) data sets. The probability of observing a spurious 4.9 $\sigma$ signal is very low at 4.8$\times 10^{-7}$. However, we would have observed such a signal at any velocity between $-$50 and +50 km sec$^{-1}$ in the planet rest frame, corresponding to about approximately 20 positions. Multiplying this by the number of transits observed means that we had approximately 80 opportunities to observe a spurious signal, meaning that we can estimate that the false alarm probability is $\sim80 \times 4.8\times 10^{-7} \approx 4 \times 10^{-5}$, corresponding to $<4 \sigma$. For this estimate we do not take into account our freedom to use a certain width for the probed signal. In addition, the possible impact of unquantified correlated noise in the data may also increase the p-values. We therefore think these data are as yet insufficient to claim definite detections of the planet exosphere. As discussed above, the spectral resolution of any potential planet signal is broadened due to the change of the radial component of the planet orbital velocity during an exposure. As shown in Table \[tab:widths\], this ‘instrumental’ broadening is five to six pixels for the HARPS and HARPS-N data, and below the intrinsic spectral resolution of the spectrograph for the UVES data, due to the significantly shorter exposure times. In addition, the absorption from sodium and ionized calcium could be intrinsically broadened due to a strong velocity field in the planet exosphere [@Mura2011]. If the observed signals from either sodium or ionized calcium are real, they are indeed broad at the five to six pixel (4 km sec$^{-1}$) level, which is much broader than the intrinsic width of the Na D$_2$ line in the exosphere of Mercury, previously observed by [@Potter2013], of approximately 20 m$\AA$ or approximately 1 km sec$^{-1}$. It is not clear whether to expect the blueshift of approximately 4 km sec$^{-1}$ as measured for the potential ionized calcium signal. On the one hand, the photon-ionization lifetime of singly-ionized calcium is estimated to be significantly longer than that of neutral sodium, which could allow it to accumulate a significant acceleration as it is picked up by the stellar wind and dragged in the anti-stellar direction. On the other hand, if the planet has a significant dipole magnetic field, the ionized calcium may be trapped in the planet’s magnetic field [@Mura2011]. If the ionized calcium signal is real, it would imply that this signal is highly variable, since it is only visible in one of the four data sets. As discussed in the introduction, we do know that Mercury’s exosphere is highly variable, on a range of timescales from days, to months, to years [@Killen2007]. It is not clear at this stage whether we would expect similar behaviour for the exospheres of hot rocky super-Earths. If 55 Cnc e were to have a significant atmosphere, a confirmed detection would be evidence of atmospheric blow-off. On the other hand, if 55 Cnc e does not have a thick atmosphere, as is suggested by [@Demory2016] as being a likely interpretation of its longitudinal thermal brightness map, the exosphere would likely be produced by sputtering of the surface. We modelled the exosphere of 55 Cnc e to first order as an optically thick ring around the planet. Ignoring subtle effects like stellar limb darkening, the fraction of starlight absorbed by sodium and ionized calcium in the exosphere of 55 Cnc e would correspond to an outer radius of the exosphere of 5 $\mathrm{R_\oplus}$ and 25 $\mathrm{R_\oplus}$ respectively - 2.3 and 12 times the radius of the planet. We compare this to the Roche radius of 55 Cnc e, calculated using $$R_R = \frac{0.49q^{2/3}}{0.6q^{2/3} + \mathrm{ln}(1+q^{1/3})},$$ where $q = M_{planet}/M_{star}$ [@Eggleton1983], which is found to be $R_R = 5.35$ $R_\oplus$. Hence, the possible sodium signal, if optically thick, would come from a region as large as the planet’s Roche lobe, while that of ionized calcium would be significantly larger. If the sodium exosphere were not optically thick, it would also need to be significantly larger than the planet’s Roche lobe. If the Ca$^+$ exosphere really were to have a radius of 25 $\mathrm{R_\oplus}$, it would have an earlier ingress and a delayed egress compared to what would be expected from the radius of the planet as determined by broadband photometry. Using [@Mandel2002], the transit duration for 55 Cnc e with a 25 $\mathrm{R_\oplus}$ exosphere was found to last 26 minutes longer than the broadband transit duration. This corresponds to a range of orbital phases of -0.054 to 0.054. There may be a hint of this early ingress and delayed egress, as can be seen in Fig. \[fig:Camatrix\], however, the S/N is not sufficient to make definite claims. Although no formal detection has been made, we advocate that probing the exospheres of hot super-Earths in this way has great potential, also knowing that Mercury’s exosphere varies significantly over time. It may be a fast route towards the first characterization of the surface properties of this enigmatic class of planets. Our team is pursuing a transit monitoring programme with UVES to further investigate the possible variable signal from ionized calcium. A. R. R.-H. is grateful to the Planetary and Exoplanetary Science (PEPSci) programme of the Netherlands Organisation for Scientific Research (NWO) for support. I. A. G. S. acknowledges support from an NWO VICI grant (639.043.107). M. B. acknowledges support by NASA, through Hubble Fellowship grant HST-HF2-51336 awarded by the Space Telescope Science Institute. We thank the anonymous referee for their constructive comments. [^1]: Based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programmes 092.C-0178 & 288.C-5010 and the Telescopio Nazionale Galileo under programme CAT13B\_33. [^2]: 0.014 is the semi-major axis of GJ 1214b, identified from exoplanets.org as the shortest semi-major axis with filter MINSI\[mjupiter\] < 10M$_\oplus$ and 0.017 comes from the definition of an ultra-short period planet (USP) from [@Demory2015] of P < 0.75 days. [^3]: <http://archive.eso.org/cms.html> [^4]: <http://ia2.oats.inaf.it/archives/tng> [^5]: <http://www.tng.iac.es/instruments/harps/> [^6]: TNG programme IDs: OPT12B\_13, OPT13B\_30, OPT14A\_34 [^7]: <https://www.eso.org/sci/software/gasgano.html> [^8]: <http://www.eso.org/sci/software/cpl/esorex.html>	{ "pile_set_name": "ArXiv" }
--- author: - 'V. Bosch-Ramon' title: ' Multifrequency Behavior of Microquasars in the GeV–TeV era: A review ' --- Introduction ============ Microquasars are X-ray binaries with non-thermal jets (e.g. @mir99 [@rib05]), being called high-mass microquasars when hosting a massive star, and low-mass microquasars otherwise. The energy powering the non-thermal emission in microquasars can be either of accretion or black-hole rotation origin. The magnetic and kinetic power is channelled through a jet launched from the inner regions of the accretion disk (e.g. @bla77 [@bla82]), and part of this power is eventually converted into relativistic particles and radiation. For several decades, microquasars were considered strong candidates to gamma-ray sources (e.g. @cha85; see also @cha89 [@lev96; @par00]), but they have not become fully recognized as powerful gamma-ray emitters until recent years, after the most recent generation of ground-based Cherenkov (HESS, MAGIC, VERITAS) and satellite-borne instruments ([Fermi]{}, [AGILE]{}) arrived. The most relevant cases are the microquasars Cygnus X-1[^1] and Cygnus X-3 [@alb07; @sab10; @tav09; @abd09a; @sab11]. Other sources, like for instance SS 433, Scorpius X-1 or GRS 1915$-$105, have been also observed in GeV and TeV energies but only upper limits have been obtained [@sai09; @ace09; @ale10; @bor10]. It is noteworthy that there are four other binary systems that may be also microquasars: LS I+61+303 (e.g. @alb06 [@abd09b; @pit09]), LS 5039 (e.g. @aha05 [@abd09c; @pit09]), HESS J0632$+$057 (e.g. @hin09 [@fal11; @mol11]), and 1FGL J1018.6$-$5856 (e.g. @cor11), although they could as well host a non-accreting pulsar. Regarding the microquasar and pulsar scenarios, LS I+61+303 and LS 5039 have been extensively discussed in the literature (see, e.g., @bos09 and references therein). Although the detected radio emission is already evidence of particle acceleration in microquasar jets, the finding of microquasar gamma-ray emission proves that these sources can very efficiently channel accretion or black-hole rotational energy into radiation. In addition, together with this high efficiency, the temporal characteristics of the detected radiation may favor leptonic models, although hadronic mechanisms cannot be discarded. Also, the extreme conditions under which gamma-rays are produced can put restrictions in the emitter structure. Morphological studies can be also of help, since non-thermal processes can take place not only at the binary scales, but also far away (e.g. the jet termination region). Although the complexity of microquasar phenomenology can make the characterization of the ongoing processes difficult, high quality data together with semi-analytical modeling can provide sensible information on the non-thermal physics of the sources. Numerical calculations are also important, since they can inform about the conditions of the background plasma in which emission takes place. In this paper, we briefly review relevant aspects of the non-thermal emission in microquasars. We will focus mainly in the GeV and TeV energy bands, for which photon production requires extreme conditions in these sources. In Figure \[mq\], a sketch of the microquasar scenario is presented. ![image](vbosch_2011_01_fig1.eps){width="6cm"} Non-thermal emission in microquasars ==================================== Microquasar jets can produce non-thermal populations of relativistic particles via diffusive shock acceleration or other mechanisms at different spatial scales (e.g. @rie07). These particles, electrons, protons or even heavy nuclei, can interact with the background matter, radiation and magnetic fields to produce non-thermal emission from radio to gamma-rays. In the GeV and TeV bands, the most efficient process is inverse Compton (IC), but in systems with very high density regions, like SS 433, Cygnus X-3 and possibly Cygnus X-1, proton-proton interactions may be also relevant. Also, in systems with very dense fields of energetic target photons, photomeson production and even photodisintegration of nuclei may be efficient. The gamma-ray emission can take place at different scales, although certain regions suffer from strong absorption via pair creation (e.g. deep inside the system or at the jet base), and some others may (or may not) lack enough targets (e.g. the jet largest scales). Below, we discuss farther the non-thermal phenomena at different scales in high- and low-mass microquasars. For a general review on the efficiency of leptonic and hadronic processes under typical microquasar conditions, see [@bos09] and references therein. Microquasar emitting sites -------------------------- Different emitting regions can be considered when understanding the non-thermal emission from microquasars. Jets are the best acceleration sites given the large amount of energy that they transport. Different forms of dissipation can take place in them through shocks, velocity gradients and turbulence (e.g. @rie07), as well as magnetic reconnection (e.g. @zen01), which can lead to generate non-thermal particle populations. The jet formation itself, interaction with an accretion disc wind, or recollimation and internal shocks can accelerate particles at the jet base. The presence of non-thermal electrons in the region can lead to the production of gamma rays through IC with accretion photons, or with photons produced by the same electrons via synchrotron emission (e.g. @bos06b). The base of the jet is possibly the region in which hadronic processes may be the most efficient, given the high density of matter and photons in there, and the hardness of the latter (e.g. @lev01 [@rom08]). However, the local radiation fields could also strongly suppress the GeV emission via gamma-ray absorption and pair creation (see, e.g., @rom08 [@cer11]). For low ambient magnetic fields, electromagnetic cascades can increase the effective transparency of the source to gamma-rays [@akh85]. An example of a (leptonic) low-mass microquasar spectral energy distribution, with its high-energy radiation mainly coming from the base of the jet, is shown in Fig. \[lmqs\]. ![ Computed spectral energy distribution of the non-thermal emission from 1E 1740.7$-$2942 for two situations. In one case, the hard X-rays come from a corona, whereas in the other, they are of synchrotron origin and come from the jet. Gamma-ray absorption in the accretion disk and corona photon fields has been taken into account (see the two dips below and above $\sim$ GeV energies). For details, see [@bos06b]. []{data-label="lmqs"}](vbosch_2011_01_fig2.ps){width="6.5cm"} In high-mass microquasars, the strong radiation and mass-loss from the star can render significant non-thermal radiation, in particular at high energies, whereas radio may be at least partially free-free absorbed. The considered most efficient high-energy channel is typically IC with stellar photons (e.g. @bos06a), interaction that is anisotropic and has specific lightcurve and spectral features (e.g. @kha08). Anisotropic IC may be behind the orbital modulation of the GeV lightcurve of Cygnus X-3 seen by [Fermi]{} [@abd09a; @dub10]. Absorption of TeV emission in the stellar photon field is likely to be significant for compact high-mass systems, like Cygnus X-3 and Cygnus X-1. That may be the reason why the former has not been detected in the TeV range (see @ale10 and references therein), and why the evidence of detection of Cygnus X-1 by MAGIC may imply an emitter outside the binary system (see @bos08a). As before, for low enough magnetic fields (see @kha08), electromagnetic cascades can increase the effective transparency of these two sources (see, e.g., @bed07 [@ore07; @bed10]). As discussed below, the role of pair creation cannot be neglected in the context of broadband non-thermal emission. At the binary scales, absorption of GeV photons is not expected since this band is below the gamma-ray energy threshold for pair creation, around $\sim 10-100$ GeV for stellar photons peaking in the UV. Proton-proton, photomeson production and photodisintegration have also been proposed as possible mechanisms of gamma-ray emission at these scales (e.g. @rom03 [@aha06; @bed05]). An example of a (leptonic) spectral energy distribution of a high-mass microquasar is shown in Fig. \[hm\]. ![image](vbosch_2011_01_fig3.ps){width="6.5cm"} The interaction of the jets with the stellar wind cannot be neglected in microquasars with a massive companion. The impact of the wind on the jet triggers strong shocks, good candidates for particle acceleration, jet bending, and potentially jet disruption (e.g. @per08 [@per10]). This interaction can generate high-energy emission [@per08], but the specific properties can depend on the level of inhomogeneity of the stellar wind (e.g. @ara09 [@ara11]; also Perucho & Bosch-Ramon, in preparation). Figure \[jetint\] shows the density map resulting from a 3-dimensional simulation of a microquasar jet interacting with the wind of the companion. ![image](vbosch_2011_01_fig4.ps){width="11cm"} Far from the binary system, say at milliarcsecond to second scales, the jet propagates unaffected by significant external disturbances. However, there are different mechanisms that may lead to energy dissipation, particle heating/acceleration and subsequent radiation, like velocity gradients and Kelvin-Helmholzt instabilities in the jet walls. Shear acceleration has been proposed for instance to explain extended emission from large scale jets in microquasars and AGNs [@rie07]. All this could generate fresh relativistic particles that could emit in radio by synchrotron. Very powerful ejections could also be bright enough to be detectable, from radio to gamma-rays, far away from the binary (e.g. @ato99). It is noteworthy that, unless there is not significant previous jet activity, the wall of a continuous jet, or a transient ejection, are always to encounter diluted and hot jet material. This material was reprocessed in the jet reverse shock, where jet and ISM pressures balance, and was swept backwards filling the so-called cocoon. Only the presence of a strong wind, either from the accretion disc or the star, can clean this material out up to a certain distance from the microquasar. However, the jet material will unavoidably end up embedded in the cocoon plasma before the reverse shock is reached. The pressure of the cocoon can trigger a recollimation shock in the jet, which becomes collimated and suffers pinching. The jet fed cocoon drives a slow forward shock in the ISM, much denser and cooler than the jet. This complex dynamical behavior has associated the production of non-thermal emission, which possibly may reach gamma-ray energies. An interesting situation arises when the microquasar has a high-mass companion and the proper motion velocity is $\gtrsim 10^7$ cm s$^{-1}$, in which case the jet can be completely disrupted before reaching the ISM, as illustrated in Fig. \[ls\]. Farther discussion of jets interacting with the ISM can be found in [@bor09], [@bos11b], and references therein. ![Density map resulting from a 2-dimensional slab simulation, in which a jet propagates in an environment characterized by the microquasar motion in the ISM. The shocked stellar wind, deflected by the microquasar proper motion, comes from the top and impact the jet from a side (see @bos11b).[]{data-label="ls"}](vbosch_2011_01_fig5.ps){width="6.5cm"} Jets, or their termination region, are not the only possible emitting sites in microquasars. The inner regions of the accretion structures (e.g. disc, corona/ADAF and the like) may also contain non-thermal populations of particles (e.g. @bis76 [@pin82; @spr88; @gie99; @rom10]). At the binary system scales, and in particular with high-mass companions, the stellar wind is dense, carries magnetic field, and is embedded in a dense photon bath by the star. Therefore, for those systems with very efficient particle acceleration in the jet, electrons and protons could diffuse out of it and radiate their energy in the environment. Also, gamma-ray absorption due to pair creation in the stellar photon field can inject electrons and positrons in the wind, also leading to broadband non-thermal emission, as shown for instance in [@bos08b]. This emission may be actually behind a substantial fraction of the milliarcsecond radiation in a TeV emitting microquasar [@bos11a]. An example of this is shown in Fig. \[raw\], in which 5 GHz maps are presented for different orbital phases in a TeV emitting binary. ![image](vbosch_2011_01_fig6.eps){width="35.00000%"}![image](vbosch_2011_01_fig7.eps){width="35.00000%"}\ ![image](vbosch_2011_01_fig8.eps){width="35.00000%"}![image](vbosch_2011_01_fig9.eps){width="35.00000%"}\ Final remarks ============= Microquasar can efficiently accelerate particles up to very high energies and produce gamma-rays within and outside the binary system. It is however unclear currently why some sources emit gamma-rays and others do not. A key point may be the presence of a massive star, which as discussed here can affect the jet significantly, with the formation of particle acceleration sites, and also offering dense target photon and matter fields, suitable for gamma-ray production at the binary scales. Low-mass microquasars could in principle produce gamma-rays, and a reason for remaining undetected yet may be that their GeV emission is too dim for the present instrumentation. At TeV energies, in the context of leptonic models, strong synchrotron cooling and IC scattering deep in the Klein-Nishina regime (as expected in the hard accretion photon fields of low-mass microquasars) may also prevent their detection in TeV. The lack of a dense stellar wind in low-mass systems makes also a difference with high-mass ones. It is worth noting that the four mentioned gamma-ray microquasar candidates harbor massive stars, which seems to be a common feature in most of the known gamma-ray binaries (not only microquasars). At the jet base, the situation seems to be basically the same in high- and low-mass systems. In both object types, the compactness of the region could imply that gamma rays are absorbed. Presently, gamma rays have not been detected at the largest scales, and thus it is not clear whether there is an intrinsic difference between high- and low-mass microquasars at these scales, which may be the case accounting for their different environments. I want to thank the organizers for their kind invitation. The research leading to these results has received funding from the European Union Seventh Framework Program (FP7/2007-2013) under grant agreement PIEF-GA-2009-252463. V.B.-R. acknowledges support by the Spanish Ministerio de Ciencia e Innovación (MICINN) under grants AYA2010-21782-C03-01 and FPA2010-22056-C06-02. Abdo, A. A., et al. 2009a, Science, 326, 1512 Abdo, A. A., et al. 2009b, ApJ, 701, L123 Abdo, A. A., et al. 2009c, ApJ, 706, L56 Acero, F. et al. 2009, A&A, 508, 1135 Albert, J. et al. 2006, Science, 312, 1771 Albert, J., et al. 2007, ApJ, 665, L51 Aleksic, J. 2010, ApJ, 721, 843 Aleksic, J. et al. 2011, ApJL, in press \[astro-ph/1103.5677\] Akharonian, F. A. & Vardanian, V. V. 1985, Ap&SS, 115, 31 Aharonian, F. A. et al. 2005, Science, 309, 746 Aharonian, F., Anchordoqui, L., Khangulyan, D., Montaruli, T. 2006, J. Phys. Conf. Ser., 39, 408 \[astro-ph/0508658\] Araudo, A. T., Bosch-Ramon, V., Romero, G. E. 2009, A&A, 503, 673 Araudo, A. T., Bosch-Ramon, V., Romero, G. E. Proceedings of the 25th Texas Symposium on Relativistic Astrophysics - TEXAS 2010, Heidelberg, Germany \[astro-ph/1104.1730\] Atoyan, A. M. & Aharonian, F. A. 1999, MNRAS, 302, 253 Bednarek, W. 2005, MNRAS, 631, 466 Bednarek, W. & Giovannelli, F. 2007, A&A, 464, 437 Bednarek, W. 2010, MNRAS, 406, 689 Bisnovatyi-Kogan, G. S., Blinnikov, S. I. 1976, SvAL, 2 191 Blandford, R. D. & Znajek, R. L. 1977, MNRAS, 179, 433 Blandford, R. D. & Payne, D. G. 1982, MNRAS, 199, 883 Bordas, P., Bosch-Ramon, V., Paredes, J. M., Perucho, M. 2009, A&A, 497, 325 Bordas, P., Bosch-Ramon, V., Paredes, J. M., 2010, IJMPD, 19, 749 \[astro-ph/1011.2130\] Bosch-Ramon, V. & Khangulyan, D. 2009, IJMPD, 18, 347 \[astro-ph/0805.4123\] Bosch-Ramon, V. & Khangulyan, D. 2011, PASJ, in press \[astro-ph/1105.2172\] Bosch-Ramon, V. et al. 2006a, A&A, 457, 1011 Bosch-Ramon, V., Romero, G E., Paredes, J. M. 2006b, A&A, 447, 263 Bosch-Ramon, V., Khangulyan D., Aharonian F. A. 2008a, A&A, 489, L21 Bosch-Ramon, V., Khangulyan D., Aharonian F. A. 2008b, A&A, 482, 397 Bosch-Ramon, V., Perucho, M., Bordas, P. 2011, A&A, 528, 89 Cerutti, B. et al. 2011, A&A, 529, 120 Chadwick, P. M., et al. 1985, Nature, 318, 642 Chardin, G. & Gerbier, G. 1989, A&A, 210, 52 Corbet, R. H. D., Cheung, C. C., Kerr, M., 2011, ATel, 3221, 1 Dubus, G., Cerutti, B., Henri, G. 2010, MNRAS, 404, L55 Falcone, A., Bongiorno, S., Stroh, M., Holder, J. 2011, ATel, 3152, 1 Gierlinski, M. et al. 1999, MNRAS, 309, 496 Hinton, J. A., Skilton, J. L., Funk, S., et al. 2009, ApJ, 690, L101 Khangulyan, D., Aharonian, F., & Bosch-Ramon, V. 2008, MNRAS, 383, 467 Levinson, A. & Blandford, R. D. 1996, ApJ, 456, L29 Levinson, A. & Waxman, E. 2001, PhRvL, 87, 1101 Mirabel, I. F. & Rodríguez, L. F. 1999, ARA&A 37, 409 Moldón, J., Ribó, M., Paredes, J. M. 2011, ATel, 3180, 1 Orellana, M. et al. 2007, A&A, 476, 9 Paredes, J.M., Martí, J., Ribó, M., Massi, M. 2000 Science, 288, 2340 Perucho, M. & Bosch-Ramon, V. 2008, A&A, 482, 917 Perucho, M., Bosch-Ramon, V., & Khangulyan, D. 2010, A&A, 512, L4 Pineault, S. 1982, A&A, 109, 294 Pittori, C., et al. 2009, A&A, 506, 1563 Ribó, M. in Future Directions in High Resolution Astronomy: The 10th Anniversary of the VLBA, (ASPC, 2005) 340, 421 \[astro-ph/0402134\] Rieger, F. M., Bosch-Ramon, V., Duffy, P. 2007, Ap&SS, 309, 119 Romero, G. E., Torres, D. F., Kaufman Bernadó, M. M., Mirabel, I. F. 2003, A&A, 410, L1 Romero, G. E., Vila, G. S. 2008, A&A, 485, 623 Romero, G. E., Vieyro, F. L., Vila G. S. 2010, A&A, 519, 109 Sabatini, S. et al. 2010, ApJ, 712, L10 Sabatini, S., 2011, these proceedings Saito, T. et al. 2009, Proc. 31st ICRC \[astro-ph/0907.1017\] Spruit, H. C. 1988, A&A, 194, 319 Tavani, M., et al. 2009, Nature, 462, 620 Zenitani, S. & Hoshino, M. 2001, ApJ, 562, L63 [DISCUSSION]{} [WOLFGANG KUNDT’s Comment:]{} In your thoughtful review on VHE radiation from microquasars you also addressed the expected VHE emission region. How realistic was your cartoon? In my understanding of all the jet sources, their high-energy leptons are created by magnetic reconnection in the distorted magnetosphere of its central rotator, and post-accelerated by its outgoing frequency waves, on scales $\la$ that of the Blandford & Rees de-Laval nozzle, some $10^{15\pm 1}$ cm. This scale was also found by Martin Kluczykont for M 87. It should exceed that of the accretion disc. [VALENTÍ BOSCH-RAMON:]{} In standard models of jet formation, the jet launching region is $\sim 10^2-10^3$ $R_{\rm Sch}$, so $\sim 10^9$ cm for a stellar mass black hole. In that framework, within that region the jet would be magnetically dominated and magnetic reconnection may be important, but farther, particle acceleration is likely to take place through kinetic energy dissipation, via some diffusive acceleration process of the Fermi type. Of course, the issue is still open. [IMMACOLATA DONNARUMMA:]{} How could the GeV detection of Cygnus X-1 challenge the theoretical interpretation of microquasar activity? May you compare the case of Cygnus X-1 with the one of Cygnus X-3? [VALENTÍ BOSCH-RAMON:]{} The detection of GeV emission from Gygnus X-1 shows that the emitter is not close from the accretion disc nor the jet base. It also shows that the luminosity budget is very high. All this also applies to Cygnus X-3, but in this object this GeV lightcurve is modulated along the orbit, very likely of leptinic IC origin. Such evidence for leptonic emission still lacks in the case of Cygnus X-1. The fast variability or orbital modulation imply for both sources that the GeV emission can hardly take place far from the binary. [^1]: This source has been detected in GeV and TeV energies with significances close, but slightly below, 5 $\sigma$, and thus these detections are still to be firmly established.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We compute the general expression of the one-loop vertex correction in an arbitrary plane-wave background field for the case of two on-shell external electrons and an off-shell external photon. The properties of the vertex corrections under gauge transformations of the plane-wave background field and of the radiation field are studied. Concerning the divergences of the vertex correction, the infrared one is cured by assigning a finite mass to the photon, whereas the ultraviolet one is shown to be renormalized exactly as in vacuum. Finally, the corresponding expression of the vertex correction within the locally-constant crossed field is also derived and the high-field asymptotic is shown to scale according to the Ritus-Narozhny conjecture.' author: - 'A.' - 'M. A.' title: 'One-loop vertex correction in a plane wave' --- Introduction ============ The predictions of QED agree with experiments with impressive accuracy (see, e.g., Refs. [@Hanneke_2008; @Sturm_2011]). The great success of QED has called for testing this theory under more extreme conditions as, for example, those provided by intense background electromagnetic fields. An electromagnetic field is denoted as “intense” in the realm of QED if it is of the order of the so-called “critical” field of QED: $F_{cr}=m^2/\|e\|=1.3\times 10^{16}\;\text{V/cm}=4.4\times 10^{13}\;\text{G}$ (from now on we employ units with $\epsilon_0=\hbar=c=1$ and $m$ and $e<0$ denote the electron mass and charge, respectively) [@Landau_b_4_1982; @Fradkin_b_1991; @Dittrich_b_1985]. Importantly, the presence of intense background electromagnetic fields allows for testing QED on a sector where nonlinear effects with respect to the background field strongly affect physical processes and the dynamics of charged particles. This sector is somewhat alternative to the high-energy one successfully investigated via conventional accelerators and it thus can serve as an independent ground test of QED. High-power optical lasers are becoming a suitable tool to test QED at critical field strengths, which correspond to laser intensities of the order of $10^{29}\;\text{W/cm$^2$}$. In fact, although available lasers have reached peak intensities $I_0$ of the order of $5.5\times 10^{22}\;\text{W/cm$^2$}$ [@Yoon_2019] and upcoming facilities aim at $I_0\sim 10^{23}\text{-}10^{24}\;\text{W/cm$^2$}$ [@APOLLON_10P; @ELI; @CoReLS; @XCELS], the Lorentz invariance of the theory implies that the effective laser field strength at which a process occurs is the one experienced by the charges in their rest frame [@Mitter_1975; @Ritus_1985; @Ehlotzky_2009; @Reiss_2009; @Di_Piazza_2012; @Dunne_2014]. Since the amplitude of the laser field is boosted by a factor of the order of the relativistic Lorentz factor of the charge, an electron for definiteness, one can see that the strong-field QED regime, in which the background strength is effectively of the order of $F_{cr}$, can be entered already at intensities of the order of $10^{23}\;\text{W/cm$^2$}$, if the laser field counterpropagates with respect to an electron/positron with energy of the order of $500\;\text{MeV}$. In order to test QED in the strong-field regime by means of intense optical fields, it is essential that both experiments and theoretical predictions are correspondingly accurate. However, as it is understandable, first experiments in this regime have so far been designed especially to show the occurrence of phenomena like nonlinear Compton scattering [@Bula_1996], nonlinear Breit-Wheeler pair production [@Burke_1997; @Bamber_1999], and radiation reaction [@Cole_2018; @Poder_2018], without aiming at obtaining high-accuracy results. Correspondingly, on the theory side, the basic strong-field QED processes like nonlinear Compton scattering [@Goldman_1964; @Nikishov_1964; @Ritus_1985; @Baier_b_1998; @Ivanov_2004; @Boca_2009; @Harvey_2009; @Mackenroth_2010; @Boca_2011; @Mackenroth_2011; @Seipt_2011; @Seipt_2011b; @Dinu_2012; @Krajewska_2012; @Dinu_2013; @Seipt_2013; @Krajewska_2014; @Wistisen_2014; @Harvey_2015; @Seipt_2016; @Seipt_2016b; @Angioi_2016; @Harvey_2016b; @Angioi_2018; @Di_Piazza_2018_c; @Dinu_2018; @Alexandrov_2019; @Di_Piazza_2019; @Ilderton_2019_b] and nonlinear Breit-Wheeler pair production [@Reiss_1962; @Nikishov_1964; @Narozhny_2000; @Roshchupkin_2001; @Reiss_2009; @Heinzl_2010b; @Mueller_2011b; @Titov_2012; @Nousch_2012; @Krajewska_2013b; @Jansen_2013; @Augustin_2014; @Meuren_2015; @Meuren_2016; @Di_Piazza_2019; @King_2020] have been studied in detail at tree level by approximating the laser field as a plane wave (see also the reviews [@Mitter_1975; @Ritus_1985; @Ehlotzky_2009; @Reiss_2009; @Di_Piazza_2012]). However, even under the plane-wave approximation, the radiative corrections of these processes have never been computed. The reason is that calculations including the effects of the external laser field exactly are significantly more complex than the corresponding calculations in vacuum. The standard technique, in fact, is to work within the so-called Furry picture [@Furry_1951], where the electron-positron field is quantized in the presence of the background field [@Fradkin_b_1991; @Landau_b_4_1982]. This requires that the Dirac equation can be solved analytically in the presence of the background field, which has been achieved in Ref. [@Volkov_1935] in the case of a plane wave (see also Ref. [@Landau_b_4_1982]), the corresponding states being known as Volkov states. An alternative, equivalent technique is the so-called operator technique, first proposed by Schwinger [@Schwinger_1951] and then developed for the case of a background plane wave [@Baier_1976_a; @Baier_1976_b; @Di_Piazza_2007; @Di_Piazza_2008_b; @Di_Piazza_2013; @Di_Piazza_2018_d], which does not require the explicit solution of the Dirac equation in the plane-wave field. Going back to the radiative corrections, a systematic study has been only carried out in the special case of a zero-frequency plane wave or a constant crossed field, i.e., a constant and uniform electromagnetic field with electric and magnetic field having the same amplitude and being perpendicular to each other, from the early works of Ritus and Narozhny [@Ritus_1970; @Ritus_1972; @Narozhny_1979; @Narozhny_1980; @Morozov_1981] to the more recent one [@Mironov_2020] (see also Ref. [@Akhmedov_1983] and the reviews in Refs. [@Akhmedov_2011; @Fedotov_2017]), where higher-loop Feynman diagrams have been evaluated. However, so far, in the case of a general plane wave with an arbitrary polarization and shape, only the one-loop mass operator (see Fig. \[FD\_MO\]) and the one-loop polarization operator (see Fig. \[FD\_PO\]) have been computed in Ref. [@Baier_1976_a] and in Refs. [@Becker_1975; @Baier_1976_b], respectively (see also Ref. [@Meuren_2013] for an alternative derivation of the polarization operator). ![The one-loop mass operator in an intense plane wave. The double lines represent exact electron states and propagator in a plane wave (Volkov states and propagator, respectively) [@Landau_b_4_1982].[]{data-label="FD_MO"}](Figure_1.pdf){width="0.6\columnwidth"} ![The one-loop polarization operator in an intense plane wave. The double lines represent exact electron propagators in a plane wave (Volkov propagators) [@Landau_b_4_1982].[]{data-label="FD_PO"}](Figure_2.pdf){width="0.6\columnwidth"} The one-loop vertex correction in a general plane wave (see Fig. \[FD\_VC\]) has never been evaluated, whereas the corresponding quantity in a constant crossed field was computed in Ref. [@Morozov_1981]. ![The one-loop Feynman diagram corresponding to the vertex correction. The double lines represent exact electron states and propagator in a plane wave (Volkov states and propagator, respectively) [@Landau_b_4_1982].[]{data-label="FD_VC"}](Figure_3.pdf){width="0.4\columnwidth"} The purpose of the present paper is to fill this gap and, indeed, to compute the one-loop vertex correction in an arbitrary plane wave for the case of two on-shell external electrons and an off-shell external photon. It is worth mentioning here that the computation of the vertex-correction function is not only important to evaluate the leading-order radiative corrections of strong-field QED processes. There is also a more fundamental reason related to the so-called Ritus-Narozhny conjecture [@Ritus_1970; @Narozhny_1979; @Narozhny_1980; @Morozov_1981] about the high-energy behavior of radiative corrections in strong-field QED in a constant crossed field. As we have mentioned, a constant crossed field is a constant and uniform electromagnetic field $F_0^{\mu\nu}=(\bm{E}_0,\bm{B}_0)$ such that the two field Lorentz-invariants $\bm{E}_0^2-\bm{B}_0^2$ and $\bm{E}_0\cdot\bm{B}_0$ vanish. Now, in a constant crossed field radiative corrections depend only on the Lorentz- and gauge-invariant quantum nonlinearity parameter $\chi_0=\sqrt{-(p_{\mu}F_0^{\mu\nu})^2}/mF_{cr}$ [@Mitter_1975; @Ritus_1985; @Ehlotzky_2009; @Reiss_2009; @Di_Piazza_2012], where $p^{\mu}$ is the four-momentum of the particle at hand and where the metric tensor $\eta^{\mu\nu}=\text{diag}(+1,-1,-1,-1)$ is employed. The Ritus-Narozhny conjecture states that at $\chi_0\gg 1$ the effective coupling of QED in a constant crossed field scales as $\alpha\chi_0^{2/3}$. Since, apart from irrelevant prefactors, the energy of the particle enters radiative corrections only through $\chi_0$ at $\chi_0\gg 1$, the Ritus-Narozhny conjecture implies an asymptotic high-energy behavior of strong-field QED in a constant crossed field qualitatively different from the logarithmic one of QED in vacuum [@Jauch_b_1976; @Itzykson_b_1980; @Landau_b_4_1982; @Schwartz_b_2014]. The physical relevance of the Ritus-Narozhny conjecture is broadened by the so-called locally-constant field approximation (LCFA), stating that in the limit of low-frequency plane waves the probabilities of QED processes reduce to the corresponding probabilities in a constant crossed field averaged over the phase-dependent plane-wave profile [@Ritus_1985]. In Ref. [@Podszus_2019] we have investigated the one-loop mass and polarization operator to show that, if one first performs in the general expression of these quantities the high-energy limit, one indeed recovers the typical logarithmic behavior of QED as in vacuum (see also Ref. [@Ilderton_2019]). Below, we will also investigate the vertex correction within the LCFA, whereas the high-energy asymptotic will be presented elsewhere. The paper is organized as follows: In Sec. \[Notation\] we introduce the basic notation of the paper. In Sec. \[VC\_General\] the general form of the vertex-correction function is derived by means of the operator technique. In Sec. \[VC\_GI\] the properties of the vertex-correction function under gauge transformations of the radiation field and of the plane-wave background field are studied. In Sec. \[VC\_CP\] we show how to regularize and renormalize the vertex-correction function in the ultraviolet. The expression of the vertex-correction function within the LCFA is derived in Sec. \[VC\_LCFA\] and, finally, the main conclusions of the paper are reported in Sec. \[VC\_Conclusions\]. An appendix contains some technical considerations on a component of the vertex-correction function. Notation {#Notation} ======== The notation employed below is the same as in Ref. [@Di_Piazza_2018_d] but it is convenient to report here the main definitions. As we have mentioned in the Introduction, the present paper focuses on studying radiative corrections in a general plane-wave field. The latter is described by the four-vector potential $A^{\mu}(\phi)$, which only depends on the light-cone time $\phi=t-\bm{n}\cdot \bm{x}$. Here, the unit vector $\bm{n}$ defines the propagation direction of the plane wave, which can be used to introduce two useful four-dimensional quantities: $n^{\mu}=(1,\bm{n})$ and $\tilde{n}^{\mu}=(1,-\bm{n})/2$ (note that $\phi=(nx)$). Assuming obvious differential properties of the four-vector potential $A^{\mu}(\phi)$ and its derivatives, it is clear that it is a solution of the free wave equation $\square A^{\mu}=0$, where $\square=\partial_{\nu}\partial^{\nu}$, and it is assumed to fulfill the Lorenz-gauge condition $\partial_{\mu}A^{\mu}=0$, with the additional constraint $A^0(\phi)=0$. Thus, if we represent $A^{\mu}(\phi)$ in the form $A^{\mu}(\phi)=(0,\bm{A}(\phi))$, then the Lorenz-gauge condition implies $\bm{n}\cdot\bm{A}'(\phi)=0$, with the prime in a function of $\phi$ indicating its derivative with respect to $\phi$. If we make the additional assumption that $\bm{A}(\phi)$ vanishes for $\phi\to\pm\infty$, the equality $\bm{n}\cdot\bm{A}'(\phi)=0$ implies that $\bm{n}\cdot\bm{A}(\phi)=0$. By introducing two four-vectors $a_j^{\mu}=(0,\bm{a}_j)$, with $j=1,2$, such that $(na_j)=-\bm{n}\cdot\bm{a}_j=0$ and $(a_ia_j)=-\bm{a}_i\cdot\bm{a}_j=-\delta_{ij}$, the most general form of the vector potential $\bm{A}(\phi)$ reads $\bm{A}(\phi)=\psi_1(\phi)\bm{a}_1+\psi_2(\phi)\bm{a}_2$, where the two functions $\psi_j(\phi)$ are arbitrary provided that they vanish for $\phi\to\pm\infty$ and they feature the differential properties mentioned above when the four-vector potential $A^{\mu}(\phi)$ was introduced. The field tensor $F^{\mu\nu}(\phi)=\partial^{\mu}A^{\nu}(\phi)-\partial^{\nu}A^{\mu}(\phi)$ of the plane wave is given by $F^{\mu\nu}(\phi)=n^{\mu}A^{\prime\,\nu}(\phi)-n^{\nu}A^{\prime\,\mu}(\phi)$ and below we will also use its integral $\mathscr{F}^{\mu\nu}(\phi)=\int_{-\infty}^{\phi}d\phi'F^{\mu\nu}(\phi')=n^{\mu}A^{\nu}(\phi)-n^{\nu}A^{\mu}(\phi)$ (note that the tensor $\mathscr{F}^{\mu\nu}(\phi)$ is gauge invariant). The four-dimensional quantities $n^{\mu}$, $\tilde{n}^{\mu}$, and $a^{\mu}_j$ fulfill the completeness relation: $\eta^{\mu\nu}=n^{\mu}\tilde{n}^{\nu}+\tilde{n}^{\mu}n^{\nu}-a_1^{\mu}a_1^{\nu}-a_2^{\mu}a_2^{\nu}$ (note that $(n\tilde{n})=1$ and $(\tilde{n}a_j)=0$). Below, we will refer to the longitudinal ($n$) direction as the direction along $\bm{n}$ and to the transverse ($\perp$) plane as the plane spanned by the two perpendicular unit vectors $\bm{a}_j$. In this respect, together with the light-cone time $\phi=t-x_n$, with $x_n=\bm{n}\cdot \bm{x}$, we also introduce the remaining three light-cone coordinates $T=(\tilde{n}x)=(t+x_n)/2$, and $\bm{x}_{\perp}=(x_{\perp,1},x_{\perp,2})=-((xa_1),(xa_2))=(\bm{x}\cdot\bm{a}_1,\bm{x}\cdot\bm{a}_2)$. Analogously, the light-cone coordinates of an arbitrary four-vector $v^{\mu}=(v^0,\bm{v})$ will be indicated as $v_-=(nv)=v^0-v_n$, with $v_n=\bm{n}\cdot \bm{v}$, $v_+=(\tilde{n}v)=(v^0+v_n)/2$, and $\bm{v}_{\perp}=(v_{\perp,1},v_{\perp,2})=-((va_1),(va_2))=(\bm{v}\cdot\bm{a}_1,\bm{v}\cdot\bm{a}_2)$. Since we will employ the operator technique, it is convenient to also introduce the momenta operators $P_{\phi}=-i\partial_{\phi}=-(\tilde{n}P)=-(i\partial_t-i\partial_{x_n})/2$, $P_T=-i\partial_T=-(nP)=-(i\partial_t+i\partial_{x_n})$, and $\bm{P}_{\perp}=(P_{\perp,1},P_{\perp,2})=-i(\bm{a}_1\cdot\bm{\nabla},\bm{a}_2\cdot\bm{\nabla})$. These operators are the momenta conjugated to the light-cone coordinates in the sense that the commutator between the operator corresponding to each light-cone coordinate and the associated momentum operator is equal to the imaginary unit (all other possible commutators vanish): $[\phi,P_{\phi}]=[T,P_T]=i$ and $[X_{\perp,j},P_{\perp,k}]=i\delta_{jk}$, which are equivalent to the commutation relations $[X^{\mu},P^{\nu}]=-i\eta^{\mu\nu}$, with $P^{\mu}=i\partial^{\mu}$. The commutation relations $[X^{\mu},P^{\nu}]=-i\eta^{\mu\nu}$ imply that $[P^{\mu},f(X)]=i\partial_X^{\mu}f(X)$, where $f(X)$ is an arbitrary function of the four-position operator that can be expanded in Taylor series and $\partial_X^{\mu}=\partial/\partial X_{\mu}$. Analogously, it can easily be shown that $\exp[if(X)]P^{\mu}\exp[-if(X)]=P^{\mu}+\partial^{\mu}f(X)$ and then formally that $\exp[if(X)]g(P)\exp[-if(X)]=g(P+\partial f(X))$, where $g(P)$ is a function of the four-momentum that can be expanded in Taylor series \[this identity has to be intended to apply to the Taylor series expansion of the function $g(P)$\]. The same commutation relations imply that $\exp[ig(P)]X^{\mu}\exp[-ig(P)]=X^{\mu}-\partial_P^{\mu}g(P)$ and that $\exp[ig(P)]f(X)\exp[-ig(P)]=f(X-\partial_Pg(P))$, where $\partial_P^{\mu}=\partial/\partial P_{\mu}$ \[as above, this identity has to be intended to apply to the Taylor series expansion of the function $f(X)$\]. In particular, we will consider the case where the functions in the exponents are linear either in $X^{\mu}$ or in $P^{\mu}$: $$\begin{aligned} \label{Trans_X} \exp(i(Xq))g(P)\exp(-i(Xq))&=g(P+q),\\ \label{Trans_P} \exp(i(Py))f(X)\exp(-i(Py))&=f(X-y),\end{aligned}$$ where $q^{\mu}$ and $y^{\mu}$ are constant four-vectors. In addition, the commutation relations $[\phi,P_{\phi}]=[T,P_T]=i$ imply in particular the identities $$\begin{aligned} \label{Trans_phi} \exp(ia\phi)\tilde{g}(P_{\phi})\exp(-ia\phi)&=\tilde{g}(P_{\phi}-a),\\ \label{Trans_PT} \exp(ibP_T)\tilde{f}(T)\exp(-ibP_T)&=\tilde{f}(T+b),\end{aligned}$$ with $a$ and $b$ being two constants and $\tilde{f}(T)$ and $\tilde{g}(P_{\phi})$ being two arbitrary functions, which we will use below. Note that if $\|x\rangle$ ($\|p\rangle$) is the eigenstate of the four-position (four-momentum) operator $X^{\mu}$ ($P^{\mu}=i\partial^{\mu}$) with eigenvalue $x^{\mu}$ ($p^{\mu}$), i.e., $X^{\mu}\|x\rangle=x^{\mu}\|x\rangle$ ($P^{\mu}\|p\rangle=p^{\mu}\|p\rangle$), then, by normalizing the eigenstates $\|x\rangle$ ($\|p\rangle$) such that $\langle x\|y\rangle=\delta^{(4)}(x-y)$ \[$\langle p\|q\rangle=(2\pi)^4\delta^{(4)}(p-q)$\], it is $\langle x\|p\rangle=\exp(-i(px))=\exp[-i(p_+\phi+p_-T-\bm{p}_{\perp}\cdot\bm{x}_{\perp})]$ and $P_{\phi}\|p\rangle=-p_+\|p\rangle$, $P_T\|p\rangle=-p_-\|p\rangle$, and $\bm{P}_{\perp}\|p\rangle=\bm{p}_{\perp}\|p\rangle$. Also, the operator completeness relations hold $$\begin{aligned} \label{C_x} \int d^4x\,\|x\rangle\langle x\|&=1,\\ \int \frac{d^4p}{(2\pi)^4}\,\|p\rangle\langle p\|&=1.\end{aligned}$$ The Volkov states are the exact, analytical solutions of the Dirac equation in a plane wave [@Volkov_1935; @Landau_b_4_1982]. The positive-energy Volkov states $U_s(p,x)$ can be classified by means of the asymptotic momentum quantum numbers $\bm{p}$ (and then the energy $\varepsilon =\sqrt{m^2+\bm{p}^2}$) and of the asymptotic spin quantum number $s=1,2$ in the remote past, i.e. for $t\to-\infty$ (for notational simplicity, we have indicated the functional dependence on the four components of the electron four-momentum $p^{\mu}=(\varepsilon ,\bm{p})$, although the energy is a function of the linear momentum). Following the general notation in Ref. [@Landau_b_4_1982], these states can be written as $U_s(p,x)=E(p,x)u_s(p)$, where $$\label{E_p} E(p,x)=\bigg[1+\frac{e\hat{n}\hat{A}(\phi)}{2p_-}\bigg]\text{e}^{i\left\{-(px)-\int_{-\infty}^{\phi}d\varphi\left[\frac{e(pA(\varphi))}{p_-}-\frac{e^2A^2(\varphi)}{2p_-}\right]\right\}},$$ and where $u_s(p)$ are the free, positive-energy spinors normalized as $u^{\dag}_s(p)u_{s'}(p)=2\varepsilon \delta_{ss'}$ [@Landau_b_4_1982]. In Eq. (\[E\_p\]) we have introduced the notation $\hat{v}=\gamma^{\mu}v_{\mu}$ for a generic four-vector $v^{\mu}$, with $\gamma^{\mu}$ being the Dirac matrices, which satisfy the anti-commutation relations $\{\gamma^{\mu},\gamma^{\nu}\}=2\eta^{\mu\nu}$ [@Landau_b_4_1982]. The electron Green’s function $G(x,x')$ in the general plane-wave background electromagnetic field described by the four-vector potential $A^{\mu}(\phi)$ is defined by the equation $$\{\gamma^{\mu}[i\partial_{\mu}-eA_{\mu}(\phi)]-m\}G(x,x')=\delta^{(4)}(x-x').$$ In order to uniquely identify the Green’s function, boundary conditions have also to be specified. Here, we always assume the Feynman prescription corresponding to the shift $m\to m-i0$ [@Landau_b_4_1982]. Within the operator technique the operator $G$ corresponding to the Green’s function $G(x,x')$ is defined via the equation $G(x,x')=\langle x\|G\|x'\rangle$, i.e., as $$G=\frac{1}{\hat{\Pi}-m+i0},$$ where $\Pi^{\mu}=P^{\mu}-eA^{\mu}(\Phi)$. Now, we have explicitly shown in Ref. [@Di_Piazza_2018_d] (see also Refs. [@Baier_1976_a; @Baier_1976_b; @Di_Piazza_2007]) that the operator $G$ can be written in the form $$\label{G_1} \begin{split} G&=(\hat{\Pi}+m)\frac{1}{\hat{\Pi}^2-m^2+i0}=(\hat{\Pi}+m)(-i)\int_0^{\infty}ds\, e^{-im^2s}e^{2isP_TP_{\phi}}\\ &\times e^{-i\int_0^sds'[\bm{P}_{\perp}-e\bm{A}_{\perp}(\Phi-2s'P_T)]^2}\Big\{1-\frac{e}{2P_T}\hat{n}[\hat{A}(\Phi-2sP_T)-\hat{A}(\Phi)]\Big\}, \end{split}$$ where the prescription $m^2\to m^2-i0$ is understood. Below, we will also need the equivalent expression $$\label{G_2} \begin{split} G=&\frac{1}{\hat{\Pi}^2-m^2+i0}(\hat{\Pi}+m)=(-i)\int_0^{\infty}ds\, e^{-im^2s}\Big\{1+\frac{e}{2P_T}\hat{n}[\hat{A}(\Phi+2sP_T)-\hat{A}(\Phi)]\Big\}\\ &\times e^{-i\int_0^sds'[\bm{P}_{\perp}-e\bm{A}_{\perp}(\Phi+2s'P_T)]^2}e^{2isP_TP_{\phi}}(\hat{\Pi}+m). \end{split}$$ General expression of the one-loop vertex correction {#VC_General} ==================================================== The one-loop vertex correction corresponds to the Feynman diagram in Fig. \[FD\_VC\], where we have implicitly assumed that the photon four-momentum $q^{\mu}$ is outgoing. Note that the two external electron lines correspond to real electrons, i.e., the four-momenta $p^{\mu}$ and $p^{\prime\,\mu}$ are on-shell ($p^2=p^{\prime\,2}=m^2$), whereas at the moment we make no assumptions about the outgoing photon, i.e., in particular, $q^2\neq 0$. If we denote by $s$ ($s'$) the spin quantum number of the incoming (outgoing) electron and by $l$ the polarization quantum number of the outgoing photon, the amplitude $-ie\Gamma_{s,s',l}(p,p',q)$ corresponding to the Feynman diagram in Fig. \[FD\_VC\] can be written as $$-ie\Gamma_{s,s',l}(p,p',q)=-e^3\int d^4x\, d^4y\, d^4z\,\bar{U}_{s'}(p',y)\gamma^{\lambda}G(y,z)\hat{e}^_l(q)e^{i(qz)}G(z,x)\gamma^{\nu}U_s(p,x)D_{\lambda\nu}(x-y),$$ where $e^{\mu}_l(q)$ is the polarization four-vector of the outgoing photon. Here, we have introduced the photon propagator $D^{\lambda\nu}(x)$ and we work in the Lorenz gauge such that $$D^{\lambda\nu}(x)=\int\frac{d^4k}{(2\pi)^4}\frac{\eta^{\lambda\nu}}{k^2-\kappa^2+i0}e^{-i(kx)},$$ where $\kappa^2$ is the square of a fictitious photon mass, which has been introduced to avoid infrared divergences. By using the completeness relation in Eq. (\[C\_x\]) and the translation properties in Eq. (\[Trans\_X\]), the amplitude can be written in the semi-operator form as $$\label{Gamma_2} \begin{split} -ie\Gamma_{s,s',l}(p,p',q)&=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\bar{U}_{s'}(p',x)e^{i(kx)}\gamma^{\lambda}G e^{i(qx)}\hat{e}^_l(q)Ge^{-i(kx)}\gamma_{\lambda}U_s(p,x)\\ &=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\\ &\quad\times\bar{U}_{s'}(p',x)\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0} e^{i(qx)}\hat{e}^_l(q)\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}U_s(p,x)\\ &=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\\ &\quad\times\bar{U}_{s'}(p',x)\gamma^{\lambda}[\hat{\Pi}(\phi)+\hat{k}+m]\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0} e^{i(qx)}\hat{e}^_l(q)\\ &\qquad\times\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0}[\hat{\Pi}(\phi)+\hat{k}+m]\gamma_{\lambda}U_s(p,x), \end{split}$$ where $\Pi^{\mu}(\phi)=i\partial^{\mu}-eA^{\mu}(\phi)$. By using the fact that $[\hat{\Pi}(\phi)-m]U_s(p,x)=[\hat{\Pi}(\phi)-m]U_{s'}(p',x)=0$, we obtain $$\begin{split} -ie\Gamma_{s,s',l}(p,p',q)&=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\\ &\quad\times\bar{U}_{s'}(p',x)[2\Pi^{\lambda}(\phi)+\gamma^{\lambda}\hat{k}]\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0} e^{i(qx)}\hat{e}^_l(q)\\ &\qquad\times\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0}[2\Pi_{\lambda}(\phi)+\hat{k}\gamma_{\lambda}]U_s(p,x). \end{split}$$ Now, we notice that \[see Eq. (\[E\_p\])\] $$\Pi^{\lambda}(\phi)U_s(p,x)=\left[\pi^{\lambda}_p(\phi)+i\frac{e\hat{n}\hat{A}'(\phi)}{2p_-}n^{\lambda}\right]U_s(p,x),$$ where $$\pi^{\lambda}_p(\phi)=p^{\lambda}-eA^{\lambda}(\phi)+\frac{e(pA(\phi))}{p_-}n^{\lambda}-\frac{e^2A^2(\phi)}{2p_-}n^{\lambda}$$ is the classical kinetic four-momentum of an electron in the plane wave $A^{\mu}(\phi)$, with $\lim_{\phi\to\pm\infty}\pi^{\lambda}_p(\phi)=p^{\lambda}$. The kinetic four-momentum $\pi^{\lambda}_p(\phi)$ is clearly a gauge-invariant four-vector and, by using the tensor $\mathscr{F}^{\mu\nu}(\phi)$ (see Sec. \[Notation\]), it can be written in the manifestly gauge-invariant form as $$\label{pi} \pi^{\lambda}_p(\phi)=p^{\lambda}-\frac{ep_{\mu}\mathscr{F}^{\mu\lambda}(\phi)}{p_-}+\frac{e^2p_{\mu}\mathscr{F}^{\mu\rho}(\phi)\mathscr{F}_{\rho\nu}(\phi)p^{\nu}}{2p^3_-}n^{\lambda}.$$ In this way, the quantity $-ie\Gamma_{s,s',l}(p,p',q)$ can be written as $$\begin{split} -ie\Gamma_{s,s',l}(p,p',q)&=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\\ &\quad\times\bar{U}_{s'}(p',x)\left[2\pi_{p'}^{\lambda}(\phi)+i\frac{e\hat{n}\hat{A}'(\phi)}{p'_-}n^{\lambda}+\gamma^{\lambda}\hat{k}\right]\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0} e^{i(qx)}\hat{e}^_l(q)\\ &\qquad\times\frac{1}{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0}\left[2\pi_{p,\lambda}(\phi)+i\frac{e\hat{n}\hat{A}'(\phi)}{p_-}n_{\lambda}+\hat{k}\gamma_{\lambda}\right]U_s(p,x). \end{split}$$ At this point, it is convenient to use the representations in Eq. (\[G\_1\]) and in Eq. (\[G\_2\]) for the second and the first square Volkov propagator $1/\{[\hat{\Pi}(\phi)+\hat{k}]^2-m^2+i0\}$, respectively: $$\begin{split} &-ie\Gamma_{s,s',l}(p,p',q)=e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\int_0^{\infty}ds\int_0^{\infty}du\,\frac{e^{i(qx)}}{k^2-\kappa^2+i0}\\ &\quad\times\bar{U}_{s'}(p',x)\left[2\pi_{p'}^{\lambda}(\phi)+i\frac{e\hat{n}\hat{A}'(\phi)}{p'_-}n^{\lambda}+\gamma^{\lambda}\hat{k}\right]e^{-im^2s}e^{-2is(p_--q_-+k_-)(P_{\phi}-k_++q_+)}\\ &\quad\times e^{-i\int_0^sds'[\bm{p}_{\perp}-\bm{q}_{\perp}+\bm{k}_{\perp}-e\bm{A}_{\perp}(\phi+2s'(p_--q_-+k_-))]^2}\left\{1+\frac{e\hat{n}[\hat{A}(\phi+2s(p_--q_-+k_-))-\hat{A}(\phi)]}{2(p_--q_-+k_-)}\right\}\\ &\quad\times \hat{e}^_l(q)e^{-im^2u}\left\{1-\frac{e\hat{n}[\hat{A}(\phi-2u(p_-+k_-))-\hat{A}(\phi)]}{2(p_-+k_-)}\right\}e^{-i\int_0^udu'[\bm{p}_{\perp}+\bm{k}_{\perp}-e\bm{A}_{\perp}(\phi-2u'(p_-+k_-))]^2}\\ &\quad\times e^{-2iu(p_-+k_-)(P_{\phi}-k_+)}\left[2\pi_{p,\lambda}(\phi)+i\frac{e\hat{n}\hat{A}'(\phi)}{p_-}n_{\lambda}+\hat{k}\gamma_{\lambda}\right]U_s(p,x), \end{split}$$ where we have exploited the fact that Volkov states are eigenstates of the operators $P_T$ and $\bm{P}_{\perp}$. Indeed, the only operator remaining in this equation is $P_{\phi}$. Now, we use the translation property in Eq. (\[Trans\_P\]) and, analogously to the vacuum case, we write the amplitude $-ie\Gamma_{s,s',l}(p,p',q)$ in the form $$\label{Gamma^mu} -ie\Gamma_{s,s',l}(p,p',q)=-ie\int d^4x\,e^{i(qx)}\bar{U}_{s'}(p',x)\Gamma^{\mu}(p,p',q;\phi)U_s(p,x)e^_{l,\mu}(q),$$ where $$\label{Gamma_3} \begin{split} -ie\Gamma^{\mu}(p,p',q;\phi)&=e^3\int\frac{d^4k}{(2\pi)^4}\int_0^{\infty}ds\int_0^{\infty}du\,\frac{e^{-im^2(s+u)}}{k^2-\kappa^2+i0}\\ &\quad\times e^{2ik_+[s(p'_-+k_-)+u(p_-+k_-)]}e^{i\left\{p'_+(\phi_s-\phi)+\int_{\phi}^{\phi_s}d\phi'\left[-\frac{e\bm{p}'_{\perp}\cdot\bm{A}_{\perp}(\phi')}{p'_-}+\frac{e^2\bm{A}^2_{\perp}(\phi')}{2p'_-}\right]\right\}}\\ &\quad\times e^{-i\int_0^sds'[\bm{p}'_{\perp}+\bm{k}_{\perp}-e\bm{A}_{\perp}(\phi_{s'})]^2}e^{-i\int_0^udu'[\bm{p}_{\perp}+\bm{k}_{\perp}-e\bm{A}_{\perp}(\phi_{u'})]^2}\\ &\quad\times e^{i\left\{-p_+(\phi_u-\phi)-\int_{\phi}^{\phi_u}d\phi'\left[-\frac{e\bm{p}_{\perp}\cdot\bm{A}_{\perp}(\phi')}{p_-}+\frac{e^2\bm{A}^2_{\perp}(\phi')}{2p_-}\right]\right\}}M^{\mu}(\phi,k,s,u). \end{split}$$ Here, we have introduced the quantities $$\begin{aligned} \label{phi_s} \phi_s&=\phi+2s(p'_-+k_-),\\ \label{phi_u} \phi_u&=\phi-2u(p_-+k_-),\end{aligned}$$ and the matrix $$\begin{split} M^{\mu}(k,s,u;\phi)&=\left\{1-\frac{e\hat{n}[\hat{A}(\phi_s)-\hat{A}(\phi)]}{2p'_-}\right\}\left[2\pi_{p'}^{\lambda}(\phi_s)+i\frac{e\hat{n}\hat{A}'(\phi_s)}{p'_-}n^{\lambda}+\gamma^{\lambda}\hat{k}\right]\\ &\quad\times\left\{1+\frac{e\hat{n}[\hat{A}(\phi_s)-\hat{A}(\phi)]}{2(p'_-+k_-)}\right\}\gamma^{\mu}\left\{1-\frac{e\hat{n}[\hat{A}(\phi_u)-\hat{A}(\phi)]}{2(p_-+k_-)}\right\}\\ &\quad\times \left[2\pi_{p,\lambda}(\phi_u)+i\frac{e\hat{n}\hat{A}'(\phi_u)}{p_-}n_{\lambda}+\hat{k}\gamma_{\lambda}\right]\left\{1+\frac{e\hat{n}[\hat{A}(\phi_u)-\hat{A}(\phi)]}{2p_-}\right\}. \end{split}$$ Note that the integrals in $T$ and $\bm{x}_{\perp}$ can be easily taken and enforce the conservation laws $p_-=p'_-+q_-$ and $\bm{p}_{\perp}=\bm{p}'_{\perp}+\bm{q}_{\perp}$, typical of problems in a plane-wave background field. As a related remark, it is clear that the quantity $\Gamma^{\mu}(p,p',q;\phi)$, unlike the corresponding vacuum expression, depends also on the plane-wave phase. Finally, the definition in Eq. (\[Gamma\^mu\]) is consistent with the idea that computing the amplitude of the vertex in a plane wave up to one loop, one can use the substitution rule $\gamma^{\mu}\to \gamma^{\mu}+\Gamma^{\mu}(p,p',q;\phi)$, as for the vertex correction in vacuum. The phase in Eq. (\[Gamma\_3\]) can be written in a compact form by turning the integral from $\phi$ to $\phi_s$ (from $\phi$ to $\phi_u$) into an integral in $s'$ ($u'$) like that in the third line of Eq. (\[Gamma\_3\]). By exponentiating also the denominator $k^2-\kappa^2+i0$ in the photon propagator, the quantity $\Gamma^{\mu}(p,p',q;\phi)$ can be written as $$\Gamma^{\mu}(p,p',q;\phi)=e^2\int\frac{d^4k}{(2\pi)^4}\int_0^{\infty}ds\int_0^{\infty}du\int_0^{\infty}dt\,e^{iSk^2-i\kappa^2t+2i(k\tilde{F})} M^{\mu}(k,s,u;\phi),$$ where $S=u+s+t$ and $$\label{F} \tilde{F}^{\mu}=\int_0^sds'\pi_{p'}^{\mu}(\phi_{s'})+\int_0^udu'\pi_p^{\mu}(\phi_{u'}).$$ As next step, we can perform the integrals in $d^4k$ analytically by shifting the four-momentum $k^{\mu}$ by setting $k^{\prime\mu}=k^{\mu}+\tilde{F}^{\mu}/S$, which, since all components of $\tilde{F}^{\mu}$ except $\tilde{F}_-$ depend on $k_-$, implies that $k^{\mu}=k^{\prime\mu}-\tilde{G}^{\mu}/S$, where $$\label{tG} \tilde{G}^{\mu}=\int_0^sds'\pi_{p'}^{\mu}(\tilde{\psi}_{s'})+\int_0^udu'\pi_p^{\mu}(\tilde{\psi}_{u'}),$$ such that $\tilde{G}_-=\tilde{F}_-=sp'_-+up_-$. Here, we have introduced the two shifted phases $$\begin{aligned} \label{tpsi_p} \tilde{\psi}_s&=\phi+2s\tau'_-+2sk_-,\\ \label{tpsi_pp} \tilde{\psi}_u&=\phi-2u\tau_--2uk_-,\end{aligned}$$ where $$\begin{aligned} \label{tau_p_m} \tau'_-&=p'_--\frac{\tilde{G}_-}{S}=\frac{tp'_--uq_-}{S}, \\ \label{tau_m} \tau_-&=p_--\frac{\tilde{G}_-}{S}=\frac{tp_-+sq_-}{S}.\end{aligned}$$ After the shift of the four-momentum $k^{\mu}$, we can write $\Gamma^{\mu}(p,p',q;\phi)$ in the form $$\Gamma^{\mu}(p,p',q;\phi)=e^2\int_0^{\infty}dsdudt\int\frac{d^4k}{(2\pi)^4}e^{-i\kappa^2t-i\frac{\tilde{G}^2}{S}+iSk^2}\tilde{L}(\tilde{Q}^{\prime\lambda}+\gamma^{\lambda}\hat{k})\tilde{C}^{\mu}(\tilde{Q}_{\lambda}+\hat{k}\gamma_{\lambda})\tilde{R},$$ where $$\begin{aligned} \label{tL} \tilde{L}&=1-\frac{e\hat{n}[\hat{A}(\tilde{\psi}_s)-\hat{A}(\phi)]}{2p'_-},\\ \tilde{Q}^{\lambda}&=2\pi_p^{\lambda}(\tilde{\psi}_u)+i\frac{e\hat{n}\hat{A}'(\tilde{\psi}_u)}{p_-}n^{\lambda}-\frac{\hat{\tilde{G}}}{S}\gamma^{\lambda},\\ \tilde{C}^{\mu}&=\left\{1+\frac{e\hat{n}[\hat{A}(\tilde{\psi}_s)-\hat{A}(\phi)]}{2(\tau'_-+k_-)}\right\}\gamma^{\mu}\left\{1-\frac{e\hat{n}[\hat{A}(\tilde{\psi}_u)-\hat{A}(\phi)]}{2(\tau_-+k_-)}\right\},\\ \tilde{Q}^{\prime\lambda}&=2\pi_{p'}^{\lambda}(\tilde{\psi}_s)+i\frac{e\hat{n}\hat{A}'(\tilde{\psi}_s)}{p'_-}n^{\lambda}-\gamma^{\lambda}\frac{\hat{\tilde{G}}}{S},\\ \label{tR} \tilde{R}&=1+\frac{e\hat{n}[\hat{A}(\tilde{\psi}_u)-\hat{A}(\phi)]}{2p_-}.\end{aligned}$$ The integral in $d^4k$ in $\Gamma^{\mu}(p,p',q;\phi)$ is complicated by the fact that the variable $k_-$ is contained in the argument of the four-vector potential of the plane wave. Thus, we first take the integral in $d^2\bm{k}_{\perp}$, which is Gaussian: $$\Gamma^{\mu}(p,p',q;\phi)=-i\alpha\int_0^{\infty}\frac{dsdudt}{S}\int\frac{dk_-dk_+}{(2\pi)^2}e^{-i\kappa^2t-i\frac{\tilde{G}^2}{S}+2iSk_-k_+}\tilde{M}^{\mu}(k_-,k_+,s,u,t;\phi),$$ where $\alpha=e^2/4\pi\approx 1/137$ is the fine-structure constant and where $$\begin{split} \tilde{M}^{\mu}(k_-,k_+,s,u,t;\phi)=&\tilde{L}\left[(\tilde{Q}^{\prime\lambda}+k_-\gamma^{\lambda}\hat{\tilde{n}})\tilde{C}^{\mu}(\tilde{Q}_{\lambda}+k_-\hat{\tilde{n}}\gamma_{\lambda})-\frac{i}{2S}\gamma^{\lambda}\gamma_{\perp,i}\tilde{C}^{\mu}\gamma_{\perp,i}\gamma_{\lambda}\right]\tilde{R}\\ &+k_+\tilde{L}[\gamma^{\lambda}\hat{n}\tilde{C}^{\mu}(\tilde{Q}_{\lambda}+k_-\hat{\tilde{n}}\gamma_{\lambda})+(\tilde{Q}^{\prime\lambda}+k_-\gamma^{\lambda}\hat{\tilde{n}})\tilde{C}^{\mu}\hat{n}\gamma_{\lambda}]\tilde{R}\\ &+k_+^2\tilde{L}\gamma^{\lambda}\hat{n}\gamma^{\mu}\hat{n}\gamma_{\lambda}\tilde{R}. \end{split}$$ Finally, the integral in $dk_+$ results in a delta function and its first and second derivatives all evaluated at $2Sk_-$. This allows then also to take the integral in $dk_-$ and, after straightforward manipulations, the resulting expression of $\Gamma^{\mu}(p,p',q;\phi)$ can be written as $$\label{Gamma_f} \begin{split} \Gamma^{\mu}(p,p',q;\phi)&=-\frac{i\alpha}{4\pi}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t}\bigg\{e^{-i\frac{\tilde{G}^2}{S}}\tilde{L}(S\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+2i\tilde{C}^{\mu})\tilde{R}\\ &\quad\left.\left.+\frac{i}{2}\frac{d}{dk_-}\left[e^{-i\frac{\tilde{G}^2}{S}}\tilde{L}(\gamma^{\lambda}\hat{n}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\hat{n}\gamma_{\lambda})\tilde{R}\right]+\frac{\hat{n}}{S}n^{\mu}\frac{d^2}{dk^2_-}\left(e^{-i\frac{\tilde{G}^2}{S}}\right)\right\}\right\vert_{k_-=0}\\ &=-\frac{i\alpha}{4\pi}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t-i\frac{\tilde{G}^2}{S}}\\ &\quad\times\Bigg\{\tilde{L}\left[S\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+2i\tilde{C}^{\mu}+\frac{1}{2S}\frac{d\tilde{G}^2}{dk_-}(\gamma^{\lambda}\hat{n}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\hat{n}\gamma_{\lambda})\right]\tilde{R}\\ &\quad\left.\left.+\frac{i}{2}\frac{d}{dk_-}\left[\tilde{L}(\gamma^{\lambda}\hat{n}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\hat{n}\gamma_{\lambda})\tilde{R}\right]-\frac{\hat{n}}{S^2}n^{\mu}\left[\frac{1}{S}\left(\frac{d\tilde{G}^2}{dk_-}\right)^2+i\frac{d^2\tilde{G}^2}{dk^2_-}\right]\right\}\right\vert_{k_-=0}. \end{split}$$ This expression can be further manipulated especially to simplify its matrix structure. However, it is first convenient to make the following considerations related to the Ward identity to be fulfilled by $\Gamma^{\mu}(p,p',q;\phi)$ [@Itzykson_b_1980]. From now on we assume that $q_->0$. Thus, by using the three four-vectors $$\begin{aligned} N^{\mu}&=q^{\mu}-\frac{q^2n^{\mu}}{2q_-},\\ \label{Lambda_i} \Lambda^{\mu}_i&=a_i^{\mu}+\frac{q_{\perp,i}n^{\mu}}{q_-},\end{aligned}$$ with $i=1,2$ together with $n^{\mu}$, one can build a light-cone basis such that $$\eta^{\mu\nu}=\frac{N^{\mu}n^{\nu}+n^{\mu}N^{\nu}}{q_-}-\Lambda^{\mu}_1\Lambda^{\nu}_1-\Lambda^{\mu}_2\Lambda^{\nu}_2.$$ Then, the quantity $\Gamma^{\mu}(p,p',q;\phi)e^_{l,\mu}(q)$, which is the one finally required here, can be written as \[recall that we work in the Lorenz gauge where $(qe^_l(q))=0$\] $$\label{Gamma_exp} \begin{split} \Gamma^{\mu}(p,p',q;\phi)e^_{l,\mu}(q)&=\frac{e^_{l,-}(q)}{q_-}\Gamma_q(p,p',q;\phi)-\frac{q^2e^_{l,-}(q)}{q_-^2}\Gamma_-(p,p',q;\phi)\\ &\quad-(\Gamma(p,p',q;\phi)\Lambda_1)(\Lambda_1e^_l(q))-(\Gamma(p,p',q;\phi)\Lambda_2)(\Lambda_2e^_l(q))\\ &=\frac{e^_{l,-}(q)}{q_-}\Gamma_q(p,p',q;\phi)\\ &\quad-\left[\frac{q^2e^_{l,-}(q)}{q_-^2}+\frac{q_{\perp,1}}{q_-}(\Lambda_1e^_l(q))+\frac{q_{\perp,2}}{q_-}(\Lambda_2e^_l(q))\right]\Gamma_-(p,p',q;\phi)\\ &\quad+\Gamma_{\perp,1}(p,p',q;\phi)(\Lambda_1e^_l(q))+\Gamma_{\perp,2}(p,p',q;\phi)(\Lambda_2e^_l(q)) \end{split}$$ where $\Gamma_q(p,p',q;\phi)=(\Gamma^{\mu}(p,p',q;\phi)q_{\mu})$ and all other symbols are defined in analogy to the definitions given in the introduction. Now, since the structure of the function $\Gamma^{\mu}(p,p',q;\phi)$ is complicated because of the presence of the plane wave, it is clear that the components $\Gamma_-(p,p',q;\phi)=(\Gamma^{\mu}(p,p',q;\phi)n_{\mu})$ and $\Gamma_{\perp,j}(p,p',q;\phi)=-(\Gamma^{\mu}(p,p',q;\phi)a_{j,\mu})$ are relatively easy to work out because the quantities $n^{\mu}$ and $a_i^{\mu}$ characterize the plane wave. For example, we observe that all the terms proportional to $n^{\mu}$ in $\Gamma^{\mu}(p,p',q;\phi)$ can be ignored in the computation of the components $\Gamma_-(p,p',q;\phi)$ and $\Gamma_{\perp,j}(p,p',q;\phi)$. The apparently most complicated term is, therefore, $\Gamma_q(p,p',q;\phi)$, which is related to the fact that by itself the vertex-correction function is not gauge invariant. However, the gauge invariance of QED guarantees that the component $\Gamma_q(p,p',q;\phi)$ of the vertex-correction function does not to contribute to any transition amplitude involving on-shell external electrons/positrons. We explicitly prove this statement in the case under consideration with an incoming and an outgoing electron, the other possible cases being proved in an analogous way. First, we start back from Eq. (\[Gamma\_2\]) and we apply the same procedure to prove the Ward identity [@Mitter_1975; @Morozov_1981]. From the second equality in Eq. (\[Gamma\_2\]) and from the definition of $\Gamma^{\mu}(p,p',q;\phi)$ in Eq. (\[Gamma\^mu\]), we obtain $$\Gamma_q(p,p',q;\phi)=-ie^2\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}\hat{q}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}.$$ now, by writing $\hat{q}=\hat{\Pi}(\phi)+\hat{k}-m-[\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m]$ it is clear that we can express $\Gamma_q(p,p',q;\phi)$ as the difference of two terms containing only one propagator in the plane wave: $$\label{Gamma_q_0} \Gamma_q(p,p',q;\phi)=-ie^2\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\gamma^{\lambda}\left[\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}-\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\right]\gamma_{\lambda}.$$ At this point, we observe that in computing, for example, the one-loop radiative corrections to nonlinear Compton scattering (see the leading-order diagram in Fig. \[FD\_NCS\_LO\], for the kinematical situation corresponding to the case under study) ![The leading-order Feynman diagram corresponding to nonlinear Compton scattering of an off-shell photon. The double lines represent exact electron states in a plane wave (Volkov states) [@Landau_b_4_1982].[]{data-label="FD_NCS_LO"}](Figure_4.pdf){width="0.4\columnwidth"} we also have to include the remaining diagrams listed in Fig. \[FD\_NCS\_MO\_PO\]. ![The one-loop Feynman diagrams corresponding, together with the one-loop vertex correction in Fig. \[FD\_VC\], to the leading-order radiative corrections of nonlinear Compton scattering of an off-shell photon. The double lines represent exact electron states and propagator in a plane wave (Volkov states and propagator, respectively) [@Landau_b_4_1982].[]{data-label="FD_NCS_MO_PO"}](Figure_5.pdf){width="\columnwidth"} If we indicate as $i\mathcal{M}^{(1)}_{s,s',\mu}(p,p',q)e_l^{\,\mu}(q)$ the amplitude of the one-loop radiative corrections to nonlinear Compton scattering represented by the diagrams in Figs. \[FD\_VC\] and \[FD\_NCS\_MO\_PO\], the gauge invariance of QED implies that $\mathcal{M}^{(1)}_{s,s',\mu}(p,p',q)q^{\mu}=0$ [@Peskin_b_1995]. Since the contribution corresponding to Fig. \[FD\_NCS\_MO\_PO\].c is by itself gauge invariant [@Baier_1976_b; @Meuren_2013], by summing the contributions from Fig. \[FD\_VC\] and from Figs. \[FD\_NCS\_MO\_PO\].b and \[FD\_NCS\_MO\_PO\].c, we obtain \[see also Eq. (\[Gamma\_q\_0\])\] $$\begin{split} i\mathcal{M}^{(1)}_{s,s',\mu}(p,p',q)q^{\mu}&=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\\ &\times\bar{U}_{s'}(p',x)\left\{\gamma^{\lambda}\left[\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}e^{i(qx)}-e^{i(qx)}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\right]\gamma_{\lambda}\right.\\ &\quad+\hat{q}e^{i(qx)}\frac{1}{\hat{\Pi}(\phi)-m+i0}\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}\\ &\quad\left.+\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}\frac{1}{\hat{\Pi}(\phi)-m+i0}\hat{q}e^{i(qx)}\right\}U_s(p,x). \end{split}$$ Now, by using the fact that $[\hat{\Pi}(\phi)-m]U_s(p,x)=[\hat{\Pi}(\phi)-m]U_{s'}(p',x)=0$, we first replace $\hat{q}$ with $\hat{\Pi}(\phi)+\hat{q}-m$ ($m-\hat{\Pi}(\phi)+\hat{q}$) in the third (fourth) line of this equation and then we move the exponential $\exp[i(qx)]$ to the left of all other operators by exploiting the identity in Eq. (\[Trans\_X\]). The result is $$\begin{split} i\mathcal{M}^{(1)}_{s,s',\mu}(p,p',q)q^{\mu}&=-e^3\int d^4x\int\frac{d^4k}{(2\pi)^4}\,\frac{e^{i(qx)}}{k^2-\kappa^2+i0}\\ &\times\bar{U}_{s'}(p',x)\left\{\gamma^{\lambda}\left[\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}-\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\right]\gamma_{\lambda}\right.\\ &\quad\left.+\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}-\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}\gamma_{\lambda}\right\}U_s(p,x), \end{split}$$ which indeed vanishes identically. This result indicates that gauge invariance implies that the component $\Gamma_q(p,p',q;\phi)$ can be ignored as it will always be compensated by the corresponding contributions arising from the mass operators (some properties of $\Gamma_q(p,p',q;\phi)$ are discussed in the appendix). At this point, we can consider the other component $\Gamma_-(p,p',q;\phi)$, whose structure is particularly easy. In fact, starting from Eq. (\[Gamma\_f\]), we have that $$\label{Gamma_-} \begin{split} \Gamma_-(p,p',q;\phi)&=-\frac{i\alpha}{4\pi}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t-i\frac{G^2}{S}}\left(SLQ^{\prime\lambda}\hat{n}Q_{\lambda}R+2i\hat{n}\right)\\ &=-\frac{i\alpha}{2\pi}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t-i\frac{G^2}{S}}\left[\left(2S(\pi_s\pi_u)+\frac{G^2}{S}+i\right)\hat{n}-2G_-(\hat{\pi}_sR+L\hat{\pi}_u)\right.\\ &\left.\quad-\hat{G}\hat{\pi}_s\hat{n}-\hat{n}\hat{\pi}_u\hat{G}+2\tau_-L\hat{G}+2\tau'_-\hat{G}R+2\frac{G_-}{S}\hat{G}-\frac{G_-^2}{S}\frac{\hat{\Delta}_s\hat{n}\hat{\Delta}_u}{p_-p'_-}\right], \end{split}$$ where $$\begin{aligned} \pi_s^{\mu}&=\pi^{\mu}_{p'}(\psi_s),\quad \Delta^{\mu}_s=e[A^{\mu}(\psi_s)-A^{\mu}(\phi)], & L&=1-\frac{\hat{n}\hat{\Delta}_s}{2p'_-}, \quad \psi_s=\phi+2s\tau'_-,\\ \pi_u^{\mu}&=\pi^{\mu}_p(\psi_u),\quad \Delta^{\mu}_u=e[A^{\mu}(\psi_u)-A^{\mu}(\phi)], & R&=1+\frac{\hat{n}\hat{\Delta}_u}{2p_-},\quad \psi_u=\phi-2u\tau_-,\\ C^{\mu}&=\left(1+\frac{\hat{n}\hat{\Delta}_s}{2\tau'_-}\right)\gamma^{\mu}\left(1-\frac{\hat{n}\hat{\Delta}_u}{2\tau_-}\right), & G^{\mu}&=\int_0^sds'\pi^{\mu}_{p'}(\psi_{s'})+\int_0^udu'\pi^{\mu}_p(\psi_{u'}), \\ Q^{\lambda}&=2\pi_u^{\lambda}+i\frac{e\hat{n}\hat{A}'(\psi_u)}{p_-}n^{\lambda}-\frac{\hat{G}}{S}\gamma^{\lambda},& Q^{\prime\lambda}&=2\pi_s^{\lambda}+i\frac{e\hat{n}\hat{A}'(\psi_s)}{p'_-}n^{\lambda}-\gamma^{\lambda}\frac{\hat{G}}{S}.\end{aligned}$$ As we have mentioned, in order to compute the components $\Gamma_{\perp,j}(p,p',q;\phi)=-(\Gamma^{\mu}(p,p',q;\phi)a_{j,\mu})$ we can effectively assume that the matrix $\hat{n}$ anticommutes with $\gamma^{\mu}$. In the following four equations, with an abuse of notation, we use the equal symbol also for two matrices that are equal to each other up to terms proportional to $n^{\mu}$, which can anyway be ignored in the computation of $\Gamma_{\perp,j}(p,p',q;\phi)$. Going through the terms in Eq. (\[Gamma\_f\]) in order of complexity, one can easily show that $$\begin{aligned} &LC^{\mu}R=\gamma^{\mu}+\frac{G_-}{2p'_-\tau'_-S}\hat{n}\hat{\Delta}_s\gamma^{\mu}-\frac{G_-}{2p_-\tau_-S}\gamma^{\mu}\hat{n}\hat{\Delta}_u,\\ \begin{split} &L(\gamma^{\lambda}\hat{n}C^{\mu}Q_{\lambda}+Q^{\prime\lambda}C^{\mu}\hat{n}\gamma_{\lambda})R=-\frac{2\tau_-}{p'_-}\hat{n}\hat{\Delta}_s\gamma^{\mu}+\frac{2\tau'_-}{p_-}\gamma^{\mu}\hat{n}\hat{\Delta}_u-\frac{4G_-}{S}\gamma^{\mu}+\frac{4G^{\mu}}{S}\hat{n}\\ &\qquad-2\hat{n}\gamma^{\mu}\hat{\pi}_s-2\hat{\pi}_u\gamma^{\mu}\hat{n}, \end{split}\\ \begin{split} &\frac{d}{dk_-}\left[\tilde{L}(\gamma^{\lambda}\hat{n}\tilde{C}^{\mu}\tilde{Q}_{\lambda}+\tilde{Q}^{\prime\lambda}\tilde{C}^{\mu}\hat{n}\gamma_{\lambda})\tilde{R}\right]_{k_-=0}=8\left(\frac{G_1^{\mu}}{S}-\frac{s\tau_-}{p'_-}\mathcal{A}_s^{\prime\,\mu}+\frac{u\tau'_-}{p_-}\mathcal{A}_u^{\prime\,\mu}\right)\hat{n}\\ &\qquad+4s\left(1+\frac{\tau_-}{p'_-}\right)\hat{n}\gamma^{\mu}\hat{\mathcal{A}}_s'-4u\left(1+\frac{\tau'_-}{p_-}\right)\hat{\mathcal{A}}_u'\gamma^{\mu}\hat{n}, \end{split}\\ \begin{split} &LQ^{\prime\,\lambda}C^{\mu}Q_{\lambda}R=4(\pi_s\pi_u)\left(\gamma^{\mu}+\frac{G_-}{2p'_-\tau'_-S}\hat{n}\hat{\Delta}_s\gamma^{\mu}-\frac{G_-}{2p_-\tau_-S}\gamma^{\mu}\hat{n}\hat{\Delta}_u\right)\\ &\qquad+2i\frac{\tau_-}{p'_-}\hat{n}\hat{\mathcal{A}}'_s\gamma^{\mu}+2i\frac{\tau'_-}{p_-}\gamma^{\mu}\hat{n}\hat{\mathcal{A}}'_u-\frac{2}{S}LC^{\mu}\hat{G}\hat{\pi}_sR-\frac{2}{S}L\hat{\pi}_u\hat{G}C^{\mu}R\\ &\qquad-\frac{2}{S^2}L\hat{G}\left(\gamma^{\mu}+\frac{\gamma^{\mu}\hat{\Delta}_s\hat{n}}{2\tau'_-}-\frac{\hat{\Delta}_u\hat{n}\gamma^{\mu}}{2\tau_-}\right)\hat{G}R \end{split}\end{aligned}$$ where $$\begin{split} G_1^{\mu}&=\frac{d}{d\phi}\left[\int_0^sds'\,s'\pi^{\mu}_{p'}(\psi_{s'})-\int_0^udu'\,u'\pi^{\mu}_p(\psi_{u'})\right]\\ &=\frac{1}{2\tau'_-}\left[s\pi_{p'}^{\mu}(\psi_s)-\int_0^sds'\pi_{p'}^{\mu}(\psi_{s'})\right]+\frac{1}{2\tau_-}\left[u\pi_p^{\mu}(\psi_u)-\int_0^udu'\pi_p^{\mu}(\psi_{u'})\right] \end{split}$$ and $$\mathcal{A}^{\mu}_{s/u}=eA^{\mu}(\psi_{s/u})$$ (the prime on these quantities indicates the derivative with respect to $\phi$). In this way, we obtain the following expressions of the transverse components $\Gamma_{\perp,i}(p,p',q;\phi)$: $$\label{Gamma_j} \begin{split} \Gamma_{\perp,j}(p,p',q;\phi)&=\frac{i\alpha}{2\pi}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t-i\frac{G^2}{S}}\\ &\quad\times\left\{(2S(\pi_s\pi_u)+i)\left(\hat{a}_j+\frac{G_-}{2p'_-\tau'_-S}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{G_-}{2p_-\tau_-S}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\right.\\ &\qquad-L(Ca_j)\hat{G}\hat{\pi}_sR-L\hat{\pi}_u\hat{G}(Ca_j)R-\frac{1}{S}L\hat{G}\left(\hat{a}_j+\frac{\hat{a}_j\hat{\Delta}_s\hat{n}}{2\tau'_-}-\frac{\hat{\Delta}_u\hat{n}\hat{a}_j}{2\tau_-}\right)\hat{G}R\\ &\qquad-\frac{2(GG_1)}{S}\left(\frac{\tau_-}{p'_-}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{\tau'_-}{p_-}\hat{a}_j\hat{n}\hat{\Delta}_u+\frac{2G_-}{S}\hat{a}_j-\frac{2(Ga_j)}{S}\hat{n}+\hat{n}\hat{a}_j\hat{\pi}_s+\hat{\pi}_u\hat{a}_j\hat{n}\right)\\ &\qquad+2i\left(\frac{(G_1a_j)}{S}+s(\mathcal{A}'_sa_j)-u(\mathcal{A}'_ua_j)\right)\hat{n}\\ &\qquad\left.+i\left[s-(u+t)\frac{\tau_-}{p'_-}\right]\hat{\mathcal{A}}_s'\hat{n}\hat{a}_j-i\left[u-(s+t)\frac{\tau'_-}{p_-}\right]\hat{a}_j\hat{n}\hat{\mathcal{A}}_u'\right\} \end{split}$$ Equations (\[Gamma\_exp\]), (\[Gamma\_-\]), and (\[Gamma\_j\]) are the main results of the paper and, as it can easily be shown, they reduce to the result in vacuum as, e.g., on page 339 of Ref. [@Itzykson_b_1980] \[as it is shown in the appendix, the component $\Gamma_q(p,p',q;\phi)$ vanishes in vacuum\]. We notice that the pre-exponential matrices in all terms feature symmetry properties such that they can all be written as the sum of two classes of terms with the second one, being obtained from the first one by: 1) taking the Dirac conjugate, 2) swapping all indexes $s$ and $u$ in each quantity. We have exploited this symmetry in the computations presented below. Finally, we observe that all the terms in Eqs. (\[Gamma\_exp\]), (\[Gamma\_-\]), and (\[Gamma\_j\]) have at most three gamma matrices except the three terms on the third line of Eq. (\[Gamma\_j\]) [^1]. The three terms in the third line of Eq. (\[Gamma\_j\]) can be easily reduced to expressions containing at most five gamma matrices: $$\label{5_gamma_1} \begin{split} &L(Ca_j)\hat{G}\hat{\pi}_sR=\left(\hat{a}_j+\frac{G_-}{2Sp'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j -\frac{G_-}{2Sp_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\hat{G}\hat{\pi}_s\\ &+\left(\hat{a}_j+\frac{G_-}{2Sp'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j\right)\left(p'_-\hat{G}-G_-\hat{\pi}_s\right)\frac{\hat{\Delta}_u}{p_-}\\ &+\frac{\hat{a}_j\hat{n}}{2p_-\tau_-}[2(G_-(\Delta_u\pi_s)-p'_-(\Delta_uG))\hat{\Delta}_u-(G_-\Delta_u^2+2\tau_-(\Delta_uG))\hat{\pi}_s+(p'_-\Delta_u^2+2\tau_-(\Delta_u\pi_s))\hat{G}], \end{split}$$ $$\label{5_gamma_2} \begin{split} &L\hat{\pi}_u\hat{G}(Ca_j)R=\hat{\pi}_u\hat{G}\left(\hat{a}_j+\frac{G_-}{2Sp'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j -\frac{G_-}{2Sp_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\\ &+\frac{\hat{\Delta}_s}{p'_-}\left(p_-\hat{G}-G_-\hat{\pi}_u\right)\left(\hat{a}_j-\frac{G_-}{2Sp_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\\ &+[2(G_-(\Delta_s\pi_u)-p_-(\Delta_sG))\hat{\Delta}_s-(G_-\Delta_s^2+2\tau'_-(\Delta_sG))\hat{\pi}_u+(p_-\Delta_s^2+2\tau'_-(\Delta_s\pi_u))\hat{G}]\frac{\hat{n}\hat{a}_j}{2p'_-\tau'_-}, \end{split}$$ $$\label{5_gamma_3} \begin{split} &L\hat{G}\left(\hat{a}_j+\frac{\hat{a}_j\hat{\Delta}_s\hat{n}}{2\tau'_-}-\frac{\hat{\Delta}_u\hat{n}\hat{a}_j}{2\tau_-}\right)\hat{G}R=\frac{G_-}{2}\left(\frac{\hat{G}\hat{a}_j\hat{\Delta}_s\hat{n}\hat{\Delta}_u}{p_-\tau'_-}+\frac{\hat{\Delta}_s\hat{n}\hat{\Delta}_u\hat{a}_j\hat{G}}{p'_-\tau_-}\right)+G_-\left(\frac{\hat{a}_j\hat{\Delta}_s\hat{G}}{\tau'_-}+\frac{\hat{G}\hat{\Delta}_u\hat{a}_j}{\tau_-}\right)\\ &+\left[(p'_-+\tau'_-)(Ga_j)+G_-(\Delta_sa_j)\right]\frac{\hat{\Delta}_s\hat{n}\hat{G}}{p'_-\tau'_-}+\left[(p_-+\tau_-)(Ga_j)+G_-(\Delta_ua_j)\right]\frac{\hat{G}\hat{n}\hat{\Delta}_u}{p_-\tau_-}-G^2\hat{a}_j\\ &-\left[(p'_-+\tau'_-)G^2+\frac{\tau'_-G^2_-}{p_-\tau_-}\Delta^2_u\right]\frac{\hat{\Delta}_s\hat{n}\hat{a}_j}{2p'_-\tau'_-}-\left[(p_-+\tau_-)G^2+\frac{\tau_-G^2_-}{p'_-\tau'_-}\Delta^2_s\right]\frac{\hat{a}_j\hat{n}\hat{\Delta}_u}{2p_-\tau_-}+2(Ga_j)\hat{G}\\ &-[G_-\Delta_s^2+2p'_-(G\Delta_s)]\frac{\hat{a}_j\hat{n}\hat{G}}{2p'_-\tau'_-}-[G_-\Delta_u^2+2p_-(G\Delta_u)]\frac{\hat{G}\hat{n}\hat{a}_j}{2p_-\tau_-}\\ &+\frac{G_-}{p_-p'_-}\left[(Ga_j)+\frac{G_-}{\tau'_-}(\Delta_sa_j)+\frac{G_-}{\tau_-}(\Delta_ua_j)\right]\hat{\Delta}_s\hat{n}\hat{\Delta}_u-G^2\left[\frac{(\Delta_sa_j)}{\tau'_-}+\frac{(\Delta_ua_j)}{\tau_-}\right]\hat{n}, \end{split}$$ Below, we will further investigate the structure of the vertex correction and discuss its divergences. Gauge-invariance properties of the vertex-correction function {#VC_GI} ============================================================= The first aspect we would like to discuss is about the gauge invariance of the expression of $\Gamma_{s,s',l}(p,p',q)$ obtained above. On the one hand, it is clear that $\Gamma_{s,s',l}(p,p',q)$ is invariant under a gauge transformation of the plane wave four-vector potential, as it can be proved by replacing $A^{\mu}(\phi)$ with $A^{\mu}(\phi)+\partial^{\mu}f(\phi)=A^{\mu}(\phi)+n^{\mu}f'(\phi)$, with $f(\phi)$ being an arbitrary function of $\phi$ \[we recall, in particular, that $\pi^{\mu}_p(\phi)$ is the kinetic four-momentum of an electron in a plane wave and it is therefore gauge invariant, see Eq. (\[pi\])\]. Now, concerning a gauge transformation of the radiation field and, in particular, of the external photon, we have already discussed that $\Gamma^{\mu}(p,p',q;\phi)$ is written in a form that automatically fulfills the Ward identity, in such a way that one-loop radiative corrections are gauge invariant. In addition, we study here the effect of the additional term $\delta\Gamma^{(\xi)}_{s,s',l}(p,p',q)$ brought about by considering the photon propagator $D^{(\xi)\,\lambda\nu}(x)$ [@Itzykson_b_1980] $$D^{(\xi)\,\lambda\nu}(x)=\int\frac{d^4k}{(2\pi)^4}\frac{e^{-i(kx)}}{k^2-\kappa^2+i0}\left[\eta^{\lambda\nu}+\left(1-\frac{1}{\xi}\right)\frac{k^{\lambda}k^{\nu}}{k^2-\kappa^2+i0}\right],$$ in an arbitrary gauge parametrized by the constant $\xi$ (the Lorenz gauge corresponds to $\xi=1$). It is clear from Eq. (\[Gamma\_2\]) that $$\begin{split} \delta\Gamma^{(\xi)}_{s,s',l}(p,p',q)=&-ie^2\left(1-\frac{1}{\xi}\right)\int d^4x \int\frac{d^4k}{(2\pi)^4}\,\frac{1}{(k^2-\kappa^2+i0)^2}\\ &\times\bar{U}_{s'}(p',x)\hat{k}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}e^{i(qx)}\hat{e}^_l(q)\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\hat{k}U_s(p,x). \end{split}$$ Now, since $[\hat{\Pi}(\phi)-m]U_s(p,x)=[\hat{\Pi}(\phi)-m]U_{s'}(p',x)=0$, we have that $\bar{U}_{s'}(p',x)\hat{k}=\bar{U}_{s'}(p',x)[\hat{\Pi}(\phi)+\hat{k}-m]$ and analogously $\hat{k}U_s(p,x)=[\hat{\Pi}(\phi)+\hat{k}-m]U_s(p,x)$. Thus, the two electron propagators in $\delta\Gamma^{(\xi)}_{s,s',l}(p,p',q)$ simplify and this quantity can be written in the form $$\delta\Gamma^{(\xi)}_{s,s',l}(p,p',q)=Z^{(\xi)}\int d^4x\,e^{i(qx)}\bar{U}_{s'}(p',x)\hat{e}^_l(q)U_s(p,x),$$ with $$Z^{(\xi)}=-ie^2\left(1-\frac{1}{\xi}\right)\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{(k^2-\kappa^2+i0)^2}$$ being a logarithmically divergent, gauge-dependent constant. However, since $\delta\Gamma^{(\xi)}_{s,s',l}(p,p',q)$ has exactly the same structure of the tree-level matrix element of (virtual) nonlinear Compton scattering, the constant $Z^{(\xi)}$ can be absorbed in the renormalization of the electric charge exactly as in vacuum [@Itzykson_b_1980]. Thus, we conclude that the gauge-dependent part of the vertex correction can be absorbed in the renormalization of the electric charge and below we will continue to work in the Lorenz gauge. Convergence properties of the vertex-correction function {#VC_CP} ======================================================== Analogously as the corresponding quantity in vacuum, the quantity $\Gamma^{\mu}(p,p',q;\phi)$ is logarithmically divergent in the ultraviolet, as it can be ascertained from the integral in $d^4k$ in Eq. (\[Gamma\_2\]) (see, e.g., the book [@Itzykson_b_1980] for the analysis of the vacuum case). Now, if we imagine to expand the exact Volkov propagators in powers of the external field (see also Fig. \[FD\_VC\]), it is clear that, since the divergence of the corresponding vacuum amplitude is logarithmic, all resulting terms depending on the field are ultraviolet convergent because the loop contains at least three vacuum electron propagators apart from the photon propagator. It is important to stress here that this does not imply that the whole field-dependent part of $\Gamma_{s,s',l}(p,p',q)$ is ultraviolet convergent because the terms dependent on the field exclusively through the external electron states are still logarithmically divergent. For this reason, the correct way of regularizing the vertex correction in the plane wave is to regularize the quantity $\Gamma^{\mu}(p,p',q;\phi)$ (see also Ref. [@Morozov_1981]). Since the divergence at hand is only logarithmic one first writes $\Gamma^{\mu}(p,p',q;\phi)=\Gamma^{\mu}(p,p',q;\phi)-\Gamma_0^{\mu}(p,p',q)+\Gamma_0^{\mu}(p,p',q)$, where $\Gamma_0^{\mu}(p,p',q)=\Gamma^{\mu}(p,p',q;\phi)\|_{A^{\mu}(\phi)=0}$, and notices that $\Gamma^{\mu}(p,p',q;\phi)-\Gamma_0^{\mu}(p,p',q)$ is ultraviolet convergent. Then, one can regularize the vacuum expression $\Gamma_0^{\mu}(p,p',q)$ exactly as in the vacuum, i.e., by subtracting the same expression evaluated for $q^{\mu}=0$ and for $\hat{p}=\hat{p}'=m$ [@Itzykson_b_1980] (notice that the conservation laws in a plane wave already imply that $p^{\prime\mu}=p^{\mu}$ because these four-momenta are on-shell). In conclusion, by assuming that $e$ indicates the physical electron charge, we continue by investigating the regularized vertex function $\Gamma_R^{\mu}(p,p',q;\phi)=\Gamma^{\mu}(p,p',q;\phi)-\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}$. By using the master integrals $$\begin{aligned} \int\frac{d^4k}{(2\pi)^4}e^{iSk^2}&=-\frac{i}{16\pi^2S^2},\\ \int\frac{d^4k}{(2\pi)^4}k^{\mu}k^{\nu}e^{iSk^2}&=\frac{\eta^{\mu\nu}}{4}\int\frac{d^4k}{(2\pi)^4}k^2e^{iSk^2}=\frac{\eta^{\mu\nu}}{32\pi^2S^3},\end{aligned}$$ it is straightforward to take the integral in $d^4k$ in $\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}$ and to obtain the result $$\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}=-i\frac{\alpha}{2\pi}\gamma^{\mu}\int_0^{\infty}\frac{dsdudt}{S^3}\,e^{-i\kappa^2t-i\frac{(s+u)^2}{S}m^2}\left\{m^2\left[2t-\frac{(s+u)^2}{S}\right]+i\right\}.$$ From the derivations, it is clear that $\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}$ has only components $\Gamma_{0,-}(p,p,0)\|_{\hat{p}=m}$ and $\Gamma_{0,\perp,j}(p,p,0)\|_{\hat{p}=m}$, and then that $\Gamma_{R,q}(p,p',q;\phi)=\Gamma_q(p,p',q;\phi)$, which, as we have mentioned, can be shown to vanish for $A^{\mu}(\phi)=0$ (see the appendix). Now, we would like to investigate the convergence properties of the proper time integrals in $\Gamma_{R,-}(p,p',q;\phi)$ and $\Gamma_{R,\perp,j}(p,p',q;\phi)$. It is first convenient to use the following identity [@Schubert_2001] $$\begin{split} \int_0^{\infty}ds\int_0^{\infty}du\int_0^{\infty}dt&=\int_0^{\infty}ds\int_0^{\infty}du\int_0^{\infty}dt\int_0^{\infty}dS\,\delta(S-s-u-t)\\ &=\int_0^{\infty}dS\int_0^Sds\int_0^Sdu\int_0^Sdt\,\delta(S-s-u-t)\\ &=\int_0^{\infty}dS\,S^2\int_0^1dx\int_0^1dy\int_0^1dz\,\delta(1-x-y-z), \end{split}$$ where in the last line we performed the changes of variables $s=xS$, $u=yS$, and $t=zS$. By setting $$\label{int_3d} \int_{\delta} dxdydz=\int_0^1dx\int_0^1dy\int_0^1dz\,\delta(1-x-y-z),$$ it is instructive to report the expression of $\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}$ in terms of the new variables: $$\Gamma_0^{\mu}(p,p,0)\|_{\hat{p}=m}=-i\frac{\alpha}{2\pi}\gamma^{\mu}\int_0^{\infty}dS\int_{\delta}dxdydz\,e^{-i\kappa^2zS-im^2(x+y)^2S}\left\{m^2[2z-(x+y)^2]+\frac{i}{S}\right\},$$ because it clearly shows that only the term whose integrand is proportional to $i$ is (logarithmically) divergent (in the limit $S\to 0$). This divergence is related with the ultraviolet logarithmic divergence of the vertex-correction function. Keeping in mind that $z=1-x-y$ \[see Eq. (\[int\_3d\])\], another divergence for $x+y\to 0$ arises for a massless photon ($\kappa^2=0$), which corresponds to the infrared divergence of the vertex-correction function. By means of the above change of variables, we obtain $$\label{Gamma_R_-} \begin{split} &\Gamma_{R,-}(p,p',q;\phi)=\frac{\alpha}{2\pi}\hat{n}\int_0^{\infty}\frac{dS}{S}\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left[e^{-ig^2S}-e^{-im^2(x+y)^2S}\right]\\ &\quad-\frac{i\alpha}{2\pi}\int_0^{\infty}dS\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\Bigg\{e^{-ig^2S}\Bigg[\left(2(\pi_s\pi_u)+g^2\right)\hat{n}-2g_-(\hat{\pi}_sR+L\hat{\pi}_u)-\hat{g}\hat{\pi}_s\hat{n}-\hat{n}\hat{\pi}_u\hat{g}\\ &\left.\quad+2\tau_-L\hat{g}+2\tau'_-\hat{g}R+2g_-\hat{g}-g_-^2\frac{\hat{\Delta}_s\hat{n}\hat{\Delta}_u}{p_-p'_-}\right]-m^2\hat{n}e^{-im^2(x+y)^2S}[2z-(x+y)^2]\Bigg\}, \end{split}$$ and $$\label{Gamma_R_j} \begin{split} &\Gamma_{R,\perp,j}(p,p',q;\phi)=-\frac{\alpha}{2\pi}\hat{a}_j\int_0^{\infty}\frac{dS}{S}\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left[e^{-ig^2S}-e^{-im^2(x+y)^2S}\right]\\ &\quad+\frac{i\alpha}{2\pi}\int_0^{\infty}dS\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left\langle e^{-ig^2S}\left\{2(\pi_s\pi_u)\left(\hat{a}_j+\frac{g_-}{2p'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{g_-}{2p_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\right.\right.\\ &\quad+\frac{i}{S}\left(\frac{g_-}{2p'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{g_-}{2p_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)-L(Ca_j)\hat{g}\hat{\pi}_sR-L\hat{\pi}_u\hat{g}(Ca_j)R\\ &\quad-L\hat{g}\left(\hat{a}_j+\frac{\hat{a}_j\hat{\Delta}_s\hat{n}}{2\tau'_-}-\frac{\hat{\Delta}_u\hat{n}\hat{a}_j}{2\tau_-}\right)\hat{g}R-2S(gg_1)\left(\frac{\tau_-}{p'_-}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{\tau'_-}{p_-}\hat{a}_j\hat{n}\hat{\Delta}_u+2g_-\hat{a}_j-2(ga_j)\hat{n}\right.\\ &\quad+\hat{n}\hat{a}_j\hat{\pi}_s+\hat{\pi}_u\hat{a}_j\hat{n}\bigg)+2i\left((g_1a_j)+x(\mathcal{A}'_sa_j)-y(\mathcal{A}'_ua_j)\right)\hat{n}+i\left[x-(y+z)\frac{\tau_-}{p'_-}\right]\hat{\mathcal{A}}_s'\hat{n}\hat{a}_j\\ &\quad-i\left[y-(x+z)\frac{\tau'_-}{p_-}\right]\hat{a}_j\hat{n}\hat{\mathcal{A}}_u'\Bigg\}-m^2\hat{a}_je^{-im^2(x+y)^2S}[2z-(x+y)^2]\Bigg\rangle, \end{split}$$ where $$\begin{split} &L(Ca_j)\hat{g}\hat{\pi}_sR=\left(\hat{a}_j+\frac{g_-}{2p'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j -\frac{g_-}{2p_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\hat{g}\hat{\pi}_s\\ &+\left(\hat{a}_j+\frac{g_-}{2p'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j\right)\left(p'_-\hat{g}-g_-\hat{\pi}_s\right)\frac{\hat{\Delta}_u}{p_-}\\ &+\frac{\hat{a}_j\hat{n}}{2p_-\tau_-}[2(g_-(\Delta_u\pi_s)-p'_-(\Delta_ug))\hat{\Delta}_u-(g_-\Delta_u^2+2\tau_-(\Delta_ug))\hat{\pi}_s+(p'_-\Delta_u^2+2\tau_-(\Delta_u\pi_s))\hat{g}], \end{split}$$ $$\begin{split} &L\hat{\pi}_u\hat{g}(Ca_j)R=\hat{\pi}_u\hat{g}\left(\hat{a}_j+\frac{g_-}{2p'_-\tau'_-}\hat{n}\hat{\Delta}_s\hat{a}_j-\frac{g_-}{2p_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right) \\ &+\frac{\hat{\Delta}_s}{p'_-}\left(p_-\hat{g}-g_-\hat{\pi}_u\right)\left(\hat{a}_j-\frac{g_-}{2p_-\tau_-}\hat{a}_j\hat{n}\hat{\Delta}_u\right)\\ &+[2(g_-(\Delta_s\pi_u)-p_-(\Delta_sg))\hat{\Delta}_s-(g_-\Delta_s^2+2\tau'_-(\Delta_sg))\hat{\pi}_u+(p_-\Delta_s^2+2\tau'_-(\Delta_s\pi_u))\hat{g}]\frac{\hat{n}\hat{a}_j}{2p'_-\tau'_-}, \end{split}$$ $$\begin{split} &L\hat{g}\left(\hat{a}_j+\frac{\hat{a}_j\hat{\Delta}_s\hat{n}}{2\tau'_-}-\frac{\hat{\Delta}_u\hat{n}\hat{a}_j}{2\tau_-}\right)\hat{g}R=\frac{g_-}{2}\left(\frac{\hat{g}\hat{a}_j\hat{\Delta}_s\hat{n}\hat{\Delta}_u}{p_-\tau'_-}+\frac{\hat{\Delta}_s\hat{n}\hat{\Delta}_u\hat{a}_j\hat{g}}{p'_-\tau_-}\right)+g_-\left(\frac{\hat{a}_j\hat{\Delta}_s\hat{g}}{\tau'_-}+\frac{\hat{g}\hat{\Delta}_u\hat{a}_j}{\tau_-}\right)\\ &+\left[(p'_-+\tau'_-)(ga_j)+g_-(\Delta_sa_j)\right]\frac{\hat{\Delta}_s\hat{n}\hat{g}}{p'_-\tau'_-}+\left[(p_-+\tau_-)(ga_j)+g_-(\Delta_ua_j)\right]\frac{\hat{g}\hat{n}\hat{\Delta}_u}{p_-\tau_-}-g^2\hat{a}_j\\ &-\left[(p'_-+\tau'_-)g^2+\frac{\tau'_-g^2_-}{p_-\tau_-}\Delta^2_u\right]\frac{\hat{\Delta}_s\hat{n}\hat{a}_j}{2p'_-\tau'_-}-\left[(p_-+\tau_-)g^2+\frac{\tau_-g^2_-}{p'_-\tau'_-}\Delta^2_s\right]\frac{\hat{a}_j\hat{n}\hat{\Delta}_u}{2p_-\tau_-}+2(ga_j)\hat{g}\\ &-[g_-\Delta_s^2+2p'_-(g\Delta_s)]\frac{\hat{a}_j\hat{n}\hat{g}}{2p'_-\tau'_-}-[g_-\Delta_u^2+2p_-(g\Delta_u)]\frac{\hat{g}\hat{n}\hat{a}_j}{2p_-\tau_-}\\ &+\frac{g_-}{p_-p'_-}\left[(ga_j)+\frac{g_-}{\tau'_-}(\Delta_sa_j)+\frac{g_-}{\tau_-}(\Delta_ua_j)\right]\hat{\Delta}_s\hat{n}\hat{\Delta}_u-g^2\left[\frac{(\Delta_sa_j)}{\tau'_-}+\frac{(\Delta_ua_j)}{\tau_-}\right]\hat{n}, \end{split}$$ and where it is clear that also in the case of $\Gamma_{R,\perp,j}(p,p',q;\phi)$ the only term requiring regularization is the one analogous to that in the first line of Eq. (\[Gamma\_R\_-\]). Due to the above change of variables, the various quantities appearing in $\Gamma_{R,-}(p,p',q;\phi)$, and $\Gamma_{R,\perp,j}(p,p',q;\phi)$ have to be interpreted as $$\begin{aligned} \tau'_-&=zp'_--yq_-=(1-x-y)p'_--yq_-, & \pi_s^{\mu}&=\pi^{\mu}_{p'}(\theta'_S), & \Delta^{\mu}_s &=\mathcal{A}^{\mu}(\theta'_S)-\mathcal{A}^{\mu}(\phi), \\ \tau_-&=zp_-+xq_-=(1-x-y)p_-+xq_-, & \pi_u^{\mu}&=\pi^{\mu}_p(\theta_S), & \Delta^{\mu}_u&=\mathcal{A}^{\mu}(\theta_S)-\mathcal{A}^{\mu}(\phi),\end{aligned}$$ where $$\begin{aligned} \theta'_S&=\phi+2x\tau'_-S=\phi+2x[(1-x-y)p'_--yq_-]S,\\ \theta_S&=\phi-2y\tau_-S=\phi-2y[(1-x-y)p_-+xq_-]S.\end{aligned}$$ The formal definitions of the other quantities like $L$, $R$, $C^{\mu}$, $Q^{\lambda}$, and $Q^{\prime\,\lambda}$ remain unchanged and the additional quantities $$g^{\mu}=\frac{G^{\mu}}{S}=x\int_0^1d\eta\,\pi^{\mu}_{p'}(\theta'_{\eta S})+y\int_0^1d\eta\,\pi^{\mu}_p(\theta_{\eta S})$$ and $$\begin{split} g_1^{\mu}&=\frac{G_1^{\mu}}{S^2}=\frac{d}{d\phi}\left[x^2\int_0^1d\eta\,\eta\pi^{\mu}_{p'}(\theta'_{\eta S})-y^2\int_0^1d\eta\,\eta\pi^{\mu}_p(\theta'_{\eta S})\right]\\ &=\frac{x}{2\tau'_-S}\left[\pi^{\mu}_{p'}(\theta'_S)-\int_0^1d\eta\,\pi^{\mu}_{p'}(\theta'_{\eta S})\right]+\frac{y}{2\tau_-S}\left[\pi^{\mu}_p(\theta_S)-\int_0^1d\eta\,\pi^{\mu}_p(\theta_{\eta S})\right], \end{split}$$ which is regular in the limit $S\to 0$ (and also in the limits $\tau_-\to 0$ and $\tau'_-\to 0$), have been also introduced. The locally-constant field approximation {#VC_LCFA} ======================================== In this section, we would like to investigate the regularized vertex-correction function $\Gamma_R^{\mu}(p,p',q;\phi)$ in the so-called locally-constant field approximation (LCFA) [@Reiss_1962; @Ritus_1985; @Baier_b_1998; @Di_Piazza_2012]. Under this approximation quantum processes in an external field are assumed to form over a length much shorter than the typical length where the external field significantly varies [@Reiss_1962; @Ritus_1985; @Baier_b_1998; @Di_Piazza_2012]. As a general condition of validity of the LCFA in a plane wave, one assumes that the strength of the vector potential of the plane wave times the elementary charge is much larger than the electron mass. This condition is based on the idea that the strength of the vector potential scales as the strength of the electric field of the wave times the typical field wavelength, and that the LCFA applies for larger and larger wavelengths (see Refs. [@Baier_1989; @Khokonov_2002; @Di_Piazza_2007; @Wistisen_2015; @Harvey_2015; @Dinu_2016; @Di_Piazza_2018_c; @Alexandrov_2019; @Di_Piazza_2019; @Ilderton_2019_b; @Podszus_2019; @Ilderton_2019; @Raicher_2020] for more refined results and investigations about the validity of the LCFA). In order to study the structure of the vertex-correction function $\Gamma_R^{\mu}(p,p',q;\phi)$, it is first useful to exploit the general structure of the external plane wave, in particular, to rewrite the phase $G^2/S$ in a convenient form \[see Eqs. (\[Gamma\_-\]) and (\[Gamma\_j\])\]. By starting from the identity $v^2=2v_+v_--\bm{v}^2_{\perp}$, valid for a generic four-vector $v^{\mu}$, it can easily be shown that $$\label{G^2} \begin{split} G^2&=\frac{s(p'_-s+p_-u)}{p'_-}(m^2+\delta m_s^2)+\frac{u(p'_-s+p_-u)}{p_-}(m^2+\delta m_u^2)\\ &\quad+usp_-p'_-\left\{\frac{1}{p'_-}\left[\bm{\pi}_{p',\perp}(\phi)-\frac{1}{s}\int_0^sds'\bm{\Delta}_{s',\perp}\right]-\frac{1}{p_-}\left[\bm{\pi}_{p,\perp}(\phi)-\frac{1}{u}\int_0^udu'\bm{\Delta}_{u',\perp}\right]\right\}^2, \end{split}$$ where we have introduced the laser-induced square mass corrections $$\begin{aligned} \label{m_corr_s} \delta m_s^2&=\frac{1}{s}\int_0^sds'\bm{\mathcal{A}}^2_{\perp}(\psi_{s'})-\frac{1}{s^2}\left[\int_0^sds'\bm{\mathcal{A}}_{\perp}(\psi_{s'})\right]^2=\frac{1}{s}\int_0^sds'\bm{\Delta}^2_{s',\perp}-\frac{1}{s^2}\left(\int_0^sds'\bm{\Delta}_{s',\perp}\right)^2,\\ \label{m_corr_u} \delta m_u^2&=\frac{1}{u}\int_0^udu'\bm{\mathcal{A}}^2_{\perp}(\psi_{u'})-\frac{1}{u^2}\left[\int_0^udu'\bm{\mathcal{A}}_{\perp}(\psi_{u'})\right]^2=\frac{1}{u}\int_0^udu'\bm{\Delta}^2_{u',\perp}-\frac{1}{u^2}\left(\int_0^udu'\bm{\Delta}_{u',\perp}\right)^2.\end{aligned}$$ We notice that for the present case of on-shell electrons, the three quantity $G^2$ is non-negative, a property which will be used below. Also, we observe that for the evaluation of $\Gamma_R^{\mu}(p,p',q;\phi)$ the phase $\phi$ is fixed but the vector potential depends on the integration variables $s$, $u$, and $t$ (or $x$, $y$, and $S$). Thus, within the integration region, the terms in the phases depending on the vector potential become larger and larger, leading in turn to highly-oscillating integrands. Thus, the largest contributions to the integrals in the proper times come from the regions where these variables are sufficiently small that the squares of the mass corrections $\delta m_s^2$, and $\delta m_u^2$ are of the order of $m^2$ [@Di_Piazza_2013] (see also Ref. [@Meuren_2015b] for a study of the subleading contributions arising from the saddle points of the phases). In order to implement this idea, we assume that the variables $s$ and $u$ in Eqs. (\[m\_corr\_s\])-(\[m\_corr\_u\]) in the regions mainly contributing to the corresponding integrals are sufficiently small to expand the integrands in those equations for $\psi_{s'}$, and $\psi_{u'}$ around $\phi$ (the validity of this assumption is checked a posteriori). It is appropriate to perform the expansions up to terms proportional to the second derivative of $\bm{\mathcal{A}}_{\perp}(\phi)$ because the leading-order contributions to $\delta m_s^2$, and to $\delta m_u^2$ turn out to be proportional to $\bm{\mathcal{A}}_{\perp}^{\prime\,2}(\phi)$, i.e., to the square of the first-order correction. Indeed, all contributions proportional to $\bm{\mathcal{A}}''_{\perp}(\phi)$ cancel out and one obtains $$\begin{aligned} \label{m_corr_LCFA} \delta m_s^2&\approx \frac{1}{3}m^2\left[\frac{t\chi_{p'}(\phi)-u\chi_q(\phi)}{S}\right]^2m^4s^2,&& \delta m_u^2\approx \frac{1}{3}m^2\left[\frac{t\chi_p(\phi)+s\chi_q(\phi)}{S}\right]^2m^4u^2,\end{aligned}$$ where $\chi_p(\phi)=p_-\|\bm{\mathcal{E}}_{\perp}(\phi)\|/m^3$, $\chi_{p'}(\phi)=p'_-\|\bm{\mathcal{E}_{\perp}}(\phi)\|/m^3$, and $\chi_q(\phi)=q_-\|\bm{\mathcal{E}}_{\perp}(\phi)\|/m^3=\chi_{p'}(\phi)-\chi_{p'}(\phi)$, with $\bm{\mathcal{E}}_{\perp}(\phi)=-\bm{\mathcal{A}}'_{\perp}(\phi)$ (recall that we have assumed that $q_-> 0$, such that $\chi_q(\phi)\ge 0$). Now, in order to obtain the range of validity of the above approximations, it is easier to consider the typical situation in which $p_-\sim p'_-\sim q_-$ and to indicate as $p_{0,-}$ this common light-cone energy scale. Correspondingly, by indicating as $E_0$ and $\omega_0$, the amplitude and the typical angular frequency of the background plane wave ($\omega_0$ can also be thought as the inverse of the typical time interval over which the background field varies significantly), we construct the well-known Lorentz- and gauge-invariant parameters $\xi_0=\mathcal{E}_0/m\omega_0$, $\chi_0=p_{0,-}\mathcal{E}_0/m^3$ and $\eta_0=\chi_0/\xi_0=\omega_0p_{0,-}/m^2$ [@Ritus_1985; @Baier_b_1998; @Di_Piazza_2012], with $\mathcal{E}_0=\|e\|E_0$. The above approximations are all valid if the integrals are formed over regions of $s$ and $u$ such that $\omega_0sp_{0,-}=m^2s\eta_0\ll 1$ and $\omega_0up_{0,-}=m^2u\eta_0\ll 1$ (note that the additional proper time variable $t$ appears in the equations in a way that the relevant conditions and estimates do not involve it). Now, from the expressions of the mass corrections within the LFCA, it is easily seen that if $\chi_0\sim 1$ ($\chi_0\gg 1$), then the integrals are formed over the region where $s,u\lesssim 1/m^2$ ($s,u\lesssim 1/\chi_0^{2/3}m^2$). Since here we are interested in situations where $\chi_0\gtrsim 1$, we can for simplicity use the single expression $s,u\lesssim 1/\chi_0^{2/3}m^2$, such that the LCFA is valid if $\eta_0/\chi_0^{2/3}=\chi_0^{1/3}/\xi_0\ll 1$ [@Baier_1989; @Khokonov_2002; @Di_Piazza_2007; @Dinu_2016; @Di_Piazza_2018_c; @Di_Piazza_2019; @Ilderton_2019_b; @Podszus_2019; @Ilderton_2019]. By expanding also the terms in the second line of Eq. (\[G\^2\]) up to the second derivative of $\bm{A}_{\perp}(\phi)$, we obtain that at the leading order in the LCFA the phase $G^2/S$ reads $$\label{Phase_LCFA_1} \begin{split} \frac{G^2}{S}&=g^2S\approx\frac{p'_-s+p_-u}{p'_-S}m^2s\left\{1+\frac{1}{3}\left[\frac{t\chi_{p'}(\phi)-u\chi_q(\phi)}{S}\right]^2m^4s^2\right\}\\ &\quad+\frac{p'_-s+p_-u}{p_-S}m^2u\left\{1+\frac{1}{3}\left[\frac{t\chi_p(\phi)+s\chi_q(\phi)}{S}\right]^2m^4u^2\right\}\\ &\quad+\frac{usp_-p'_-}{S}\left\{\frac{1}{p'_-}\left[\bm{p}'_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)+m^3s\frac{t\bm{\chi}_{\perp,p'}(\phi)-u\bm{\chi}_{\perp,q}(\phi)}{S}\right]\right.\\ &\left.\qquad-\frac{1}{p_-}\left[\bm{p}_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)-m^3u\frac{t\bm{\chi}_{\perp,p}(\phi)+s\bm{\chi}_{\perp,q}(\phi)}{S}\right]\right\}^2\\ &\quad+\frac{4}{3}\frac{usp_-p'_-}{S}\left(\frac{s^2\tau_-^{\prime\,2}}{p'_-}-\frac{u^2\tau_-^2}{p_-}\right)\bm{\mathcal{E}}_{\perp}'(\phi)\cdot\bm{\mathcal{V}}_{\perp}(\phi), \end{split}$$ where $\bm{\chi}_{\perp,p}(\phi)=p_-\bm{\mathcal{E}}_{\perp}(\phi)/m^3$ ($\chi_p(\phi)=\|\bm{\chi}_{\perp,p}(\phi)\|$), $\bm{\chi}_{\perp,p'}(\phi)=p'_-\bm{\mathcal{E}}_{\perp}(\phi)/m^3$ ($\chi_{p'}(\phi)=\|\bm{\chi}_{\perp,p'}(\phi)\|$), $\bm{\chi}_{\perp,q}(\phi)=q_-\bm{\mathcal{E}}_{\perp}(\phi)/m^3=\bm{\chi}_{\perp,p}(\phi)-\bm{\chi}_{\perp,p'}(\phi)$ ($\chi_q(\phi)=\|\bm{\chi}_{\perp,q}(\phi)\|$), and where $$\bm{\mathcal{V}}_{\perp}(\phi)=\frac{1}{p'_-}\left[\bm{p}'_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)\right]-\frac{1}{p_-}\left[\bm{p}_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)\right].$$ The appearance of $\bm{\mathcal{A}}_{\perp}(\phi)$ in the vector $\bm{\mathcal{V}}_{\perp}(\phi)$ seems to indicate that indeed the last line of Eq. (\[Phase\_LCFA\_1\]) is leading order in the LCFA because $\|\bm{\mathcal{V}}_{\perp}(\phi)\|$ scales as $1/\omega_0$. We should however recall that the final object to be computed is $\Gamma_{s,s',l}(p,p',q)$ in Eq. (\[Gamma\^mu\]). Now, if we compute the phase $\Phi(p,p',q;\phi)$ resulting from the functions in Eq. (\[Gamma\^mu\]) other than $\Gamma^{\mu}(p,p',q;\phi)$, after taking the integrals in $\bm{x}_{\perp}$ and $T$, we obtain (apart from an inessential constant) $$\label{Phi_NCS} \Phi(p,p',q;\phi)=\frac{q_-}{2p_-p'_-}\int_0^{\phi}d\phi'\left\{m^2+\frac{p_-p'_-}{q_-^2}q^2+\left[\bm{p}_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi')-\frac{p_-}{q_-}\bm{q}_{\perp}\right]^2\right\}$$ together with the conservation laws $\bm{p}_{\perp}=\bm{p}'_{\perp}+\bm{q}_{\perp}$ and $p_-=p'_-+q_-$. It is clear that, apart from the term proportional to $q^2$, this is the phase of nonlinear Compton scattering, as given, e.g., in Ref. [@Di_Piazza_2018_c]. By using the conservation laws $\bm{p}_{\perp}=\bm{p}'_{\perp}+\bm{q}_{\perp}$ and $p_-=p'_-+q_-$, it is easy to show that $$\bm{\mathcal{V}}_{\perp}(\phi)=\frac{q_-}{p_-p'_-}\left[\bm{p}_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)-\frac{p_-}{q_-}\bm{q}_{\perp}\right].$$ Since the LCFA corresponds to evaluate the remaining integral in $\phi$ in Eq. (\[Gamma\^mu\]), it is clear that in the region where most of the photons are emitted (and tacitly assuming that the virtuality $q^2$ is less or of the order of $m^2$), it is $\|\bm{\mathcal{V}}_{\perp}(\phi)\|\sim m$ and the last line in Eq. (\[Phase\_LCFA\_1\]) can be neglected. Finally, by applying the changes of variables discussed above, we obtain the final expression of $g^2S$ within the LCFA in the form $$\label{Phase_LCFA} \begin{split} g^2S&\approx\frac{p'_-x+p_-y}{p'_-}xm^2S\left\{1+\frac{1}{3}[z\chi_{p'}(\phi)-y\chi_q(\phi)]^2x^2m^4S^2\right\}\\ &\quad+\frac{p'_-x+p_-y}{p_-}ym^2S\left\{1+\frac{1}{3}[z\chi_p(\phi)+x\chi_q(\phi)]^2y^2m^4S^2\right\}\\ &\quad+xyp_-p'_-S\left\langle\frac{1}{p'_-}\left\{\bm{p}'_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)+xm^3S[z\bm{\chi}_{\perp,p'}(\phi)-y\bm{\chi}_{\perp,q}(\phi)]\right\}\right.\\ &\left.\qquad-\frac{1}{p_-}\left\{\bm{p}_{\perp}-\bm{\mathcal{A}}_{\perp}(\phi)-ym^3S[z\bm{\chi}_{\perp,p}(\phi)+x\bm{\chi}_{\perp,q}(\phi)]\right\}\right\rangle^2, \end{split}$$ where $z=1-x-y$. Finally, we point out that have explicitly proved that in the constant-crossed field case $\bm{\mathcal{A}}_{\perp}(\phi)=-\bm{\mathcal{E}}_0\phi$, the phase $g^2S$ reduces to the phase computed in Ref. [@Morozov_1981]. The same can be verified starting from Eq. (\[Phase\_LCFA\]) by setting $\bm{\mathcal{E}}_{\perp}(\phi)=\bm{\mathcal{E}}_0$ and we note that our expression of the phase of the vertex-correction function is not only more general but also much more compact than that presented in Ref. [@Morozov_1981]. A comparison of the final expression of the pre-exponent was not carried out as the form presented in Ref. [@Morozov_1981] has a very different structure from ours, due to employed transformations there, which are appropriate only to the constant-crossed field case. Passing now to the pre-exponents of the components $\Gamma_{R,-}(p,p',q;\phi)$ \[see Eq. (\[Gamma\_R\_-\])\] and $\Gamma_{R,\perp,j}(p,p',q;\phi)$ \[see Eq. (\[Gamma\_R\_j\])\], one has to expand the field-dependent terms in Eqs. (\[Gamma\_R\_-\])-(\[Gamma\_R\_j\]) around the phase $\phi$. Taking into account that the final quantity to be evaluated is $-ie\Gamma_{s,s',l}(p,p',q)$ in Eq. (\[Gamma\^mu\]), a lengthy by straightforward calculation shows that $$\label{Gamma_R_-_LCFA} \begin{split} &\Gamma_{R,-}(p,p',q;\phi)\approx\frac{\alpha}{2\pi}\hat{n}\int_0^{\infty}\frac{dS}{S}\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left[e^{-ig^2S}-e^{-im^2(x+y)^2S}\right]\\ &\quad-\frac{i\alpha}{2\pi}\int_0^{\infty}dS\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\Bigg\langle e^{-ig^2S}\Bigg\{(p_-+p'_--g_-)\left(\frac{1-x}{p'_-}+\frac{1-y}{p_-}\right)m^2\hat{n}-2m^2\hat{n}\\ &\quad+2m[(p_-+p'_--g_-)(x+y)-g_-]+[(2-x)(p_-b'+p'_-b)-g_-b][(2-y)(p_-b'+p'_-b)-g_-b']\frac{\bm{\mathcal{A}}^{\prime\,2}_{\perp}}{p_-p'_-}\hat{n}\\ &\quad+(1-x)(1-y)p_-p'_-\bm{\mathcal{V}}_{\perp}^2\hat{n}-2(1-x)(1-y)p_-p'_-\left(\frac{2-x}{1-x}\frac{b'}{p'_-}+\frac{2-y}{1-y}\frac{b}{p_-}\right)\bm{\mathcal{V}}_{\perp}\cdot\bm{\mathcal{A}}'_{\perp}\hat{n}\\ &\quad+2m\left[g_--(p_-+p'_--g_-)(x+y)\right]\left(\frac{b'}{p'_-}+\frac{b}{p_-}\right)\hat{n}\hat{\mathcal{A}}'\\ &\left.\quad-(p_-+p'_--g_-)\left[x\frac{b'}{p'_-}\hat{\mathcal{A}}'\hat{n}\left(\frac{q_-}{p_-}m+p'_-\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\right)+y\frac{b}{p_-}\left(\frac{q_-}{p'_-}m+p_-\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\right)\hat{n}\hat{\mathcal{A}}'\right]\right\}\\ &\quad-m^2\hat{n}e^{-im^2(x+y)^2S}[2z-(x+y)^2]\Bigg\rangle, \end{split}$$ where $g^2$ is obtained from Eq. (\[Phase\_LCFA\]), $b=y\tau_-S$, $b'=x\tau'_-S$, and where all fields and derivatives are evaluated at $\phi$. Analogously, one obtains the following expression for $\Gamma_{R,\perp,j}(p,p',q;\phi)$ within the LCFA: $$\label{Gamma_R_j_LCFA} \begin{split} &\Gamma_{R,\perp,j}(p,p',q;\phi)\approx-\frac{\alpha}{2\pi}\hat{a}_j\int_0^{\infty}\frac{dS}{S}\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left[e^{-ig^2S}-e^{-im^2(x+y)^2S}\right]\\ &\quad+\frac{i\alpha}{2\pi}\int_0^{\infty}dS\int_{\delta} dxdydz\,e^{-i\kappa^2zS}\left\langle e^{-ig^2S}\left\{2[(\pi_s\pi_u)-((\pi_s+\pi_u)g)]\left(\hat{a}_j+\frac{g_-b'}{p'_-\tau'_-}\hat{n}\hat{\mathcal{A}}'\hat{a}_j+\frac{g_-b}{p_-\tau_-}\hat{a}_j\hat{n}\hat{\mathcal{A}}'\right)\right.\right.\\ &\quad+\frac{i}{S}\left(\frac{g_-b'}{p'_-\tau'_-}\hat{n}\hat{\mathcal{A}}'\hat{a}_j+\frac{g_-b}{p_-\tau_-}\hat{a}_j\hat{n}\hat{\mathcal{A}}'\right)+L(Ca_j)\hat{\pi}_s\hat{g}R+L\hat{g}\hat{\pi}_u(Ca_j)R\\ &\quad-L\hat{g}\left(\hat{a}_j+\frac{b'}{\tau'_-}\hat{a}_j\hat{\mathcal{A}}'\hat{n}+\frac{b}{\tau_-}\hat{\mathcal{A}}'\hat{n}\hat{a}_j\right)\hat{g}R-2S(gg_1)\left\{\left[\frac{2b'\tau_-}{p'_-}+(2-y)b+xb'\right]\hat{n}\hat{\mathcal{A}}'\hat{a}_j\right.\\ &\quad+\left[\frac{2b\tau'_-}{p_-}+(2-x)b'+yb\right]\hat{a}_j\hat{n}\hat{\mathcal{A}}'+m\left[\frac{p'_-}{p_-}-\frac{p_-}{p'_-}+x\left(1-\frac{p'_-}{p_-}\right)-y\left(1-\frac{p_-}{p'_-}\right)\right]\hat{n}\hat{a}_j\\ &\quad-p_-(1-y)\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\hat{n}\hat{a}_j-p'_-(1-x)\hat{n}\hat{a}_j\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\bigg\}+i(x-y)(x+y-2)\bm{\mathcal{A}}'_{\perp}\cdot\bm{a}_j\hat{n}\\ &\quad+i\left[x-(y+z)\frac{\tau_-}{p'_-}\right]\hat{\mathcal{A}}'\hat{n}\hat{a}_j-i\left[y-(x+z)\frac{\tau'_-}{p_-}\right]\hat{a}_j\hat{n}\hat{\mathcal{A}}'\Bigg\}-m^2\hat{a}_je^{-im^2(x+y)^2S}[2z-(x+y)^2]\Bigg\rangle, \end{split}$$ where $$\begin{split} &(\pi_s\pi_u)-((\pi_s+\pi_u)g)\approx-\frac{p_-+p'_-}{2g_-}g^2-\frac{g_-}{p_-+p'_-}m^2\\ &\quad+\frac{1}{2}\left(1-\frac{g_-}{p_-+p'_-}\right)\left\{\left(\frac{p_-}{p'_-} +\frac{p'_-}{p_-}\right)m^2+\left[\bm{\mathcal{V}}_{\perp}-2\left(\frac{b'}{p'_-}+\frac{b}{p_-}\right)\bm{\mathcal{A}}'_{\perp}\right]^2p_-p'_-\right\}\\ &\quad-\frac{p_-^2p_-^{\prime\,2}}{2g_-(p_-+p'_-)}\left[(x-y)\bm{\mathcal{V}}_{\perp}-\frac{(xb'+yb)q_-+2xbp'_--2yb'p_-}{p_-p'_-}\bm{\mathcal{A}}'_{\perp}\right]^2, \end{split}$$ $$\begin{split} &L(Ca_j)\hat{\pi}_s\hat{g}R+L\hat{g}\hat{\pi}_u(Ca_j)R-L\hat{g}\left(\hat{a}_j+\frac{b'}{\tau'_-}\hat{a}_j\hat{\mathcal{A}}'\hat{n}+\frac{b}{\tau_-}\hat{\mathcal{A}}'\hat{n}\hat{a}_j\right)\hat{g}R\\ &\quad\approx g^2\left(\hat{a}_j+\frac{g_-b'}{p'_-\tau'_-}\hat{n}\hat{\mathcal{A}}'\hat{a}_j+\frac{g_-b}{p_-\tau_-}\hat{a}_j\hat{n}\hat{\mathcal{A}}'\right)+2\bigg[\left(\left(\pi_s+\pi_u-g\right)a_j\right)\\ &\quad-S(p_-+p'_--g_-)(x-y)\bm{\mathcal{A}}'_{\perp}\cdot\bm{a}_j\bigg]L\hat{g}R-2m(ga_j)\left[1-\left(\frac{b'}{p'_-}+\frac{b}{p_-}\right)S\hat{n}\hat{\mathcal{A}}'\right]\\ &\quad-mS(x\hat{a}_j\hat{\mathcal{A}}'-y\hat{\mathcal{A}}'\hat{a}_j)\hat{n}\left[\left(by-b'x-2\frac{bg_-}{p_-}\right)\hat{\mathcal{A}}'+m\left(y+x\frac{p'_-}{p_-}\right)-xp'_-\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\right]\\ &\quad-mS\left[\left(by-b'x+2\frac{b'g_-}{p'_-}\right)\hat{\mathcal{A}}'+m\left(x+y\frac{p_-}{p'_-}\right)+yp_-\bm{\gamma}_{\perp}\cdot\bm{\mathcal{V}}_{\perp}\right]\hat{n}(x\hat{a}_j\hat{\mathcal{A}}'-y\hat{\mathcal{A}}'\hat{a}_j)\\ &\quad-2i(p_-+p'_--g_-)(x+y)S\varepsilon^{\lambda\mu\nu\rho}n_{\lambda}\mathcal{A}'_{\mu}a_{j,\nu}\tilde{n}_{\rho}\gamma^5L\hat{g}R, \end{split}$$ $$\begin{split} L\hat{g}R&\approx m(x+y)\left[1-\left(\frac{b}{p_-} +\frac{b'}{p'_-}\right)\hat{n}\hat{\mathcal{A}}'\right]-\frac{xb'}{2p'_-}\hat{\mathcal{A}}'\hat{n}\left(m\frac{q_-}{p_-}+p'_-\bm{\mathcal{V}}_{\perp}\cdot\bm{\gamma}_{\perp}\right)\\ &\quad-\frac{yb}{2p_-}\left(m\frac{q_-}{p'_-}+p_-\bm{\mathcal{V}}_{\perp}\cdot\bm{\gamma}_{\perp}\right)\hat{n}\hat{\mathcal{A}}'-\left(\frac{1}{3}\frac{xb^{\prime\,2}}{p'_-}+\frac{1}{3}\frac{yb^2}{p_-}+\frac{g_-bb'}{p_-p'_-}\right)\hat{n}\bm{\mathcal{A}}_{\perp}^{\prime\,2}, \end{split}$$ $$\begin{split} \gamma^5L\hat{g}R&\approx \gamma^5\left[m(y-x)+m\left(\frac{xb}{p_-} -\frac{yb'}{p'_-}\right)\hat{n}\hat{\mathcal{A}}'-\frac{xb'}{2p'_-}\hat{\mathcal{A}}'\hat{n}\left(m\frac{q_-}{p_-}+p'_-\bm{\mathcal{V}}_{\perp}\cdot\bm{\gamma}_{\perp}\right)\right.\\ &\quad\left.+\frac{yb}{2p_-}\left(m\frac{q_-}{p'_-}-p_-\bm{\mathcal{V}}_{\perp}\cdot\bm{\gamma}_{\perp}\right)\hat{n}\hat{\mathcal{A}}'-\left(\frac{1}{3}\frac{xb^{\prime\,2}}{p'_-}+\frac{1}{3}\frac{yb^2}{p_-}+\frac{g_-bb'}{p_-p'_-}\right)\hat{n}\bm{\mathcal{A}}_{\perp}^{\prime\,2}\right], \end{split}$$ $$\label{pi_pi_a} ((\pi_s+\pi_u)a_j)\approx-\left[2(b-b')\bm{\mathcal{A}}'_{\perp}+\frac{p_-+p'_-}{q_-}\bm{q}_{\perp}+2\frac{p_-p'_-}{q_-}\bm{\mathcal{V}}_{\perp}\right]\cdot\bm{a}_j,$$ $$\label{g_a} (ga_j)\approx\left[(xb'-yb)\bm{\mathcal{A}}'_{\perp}-(yp_-+xp'_-)\frac{\bm{q}_{\perp}}{q_-}-(x+y)\frac{p_-p'_-}{q_-}\bm{\mathcal{V}}_{\perp}\right]\cdot\bm{a}_j,$$ $$\begin{split} (gg_1)&\approx\left[\frac{x^3b'}{6}+\frac{y^3b}{6}+xy^2p'_-\left(\frac{2b}{3p_-}+\frac{b'}{2p'_-}\right)+yx^2p_-\left(\frac{2b'}{3p'_-}+\frac{b}{2p_-}\right)\right]\bm{\mathcal{A}}_{\perp}^{\prime\,2}\\ &\quad-\frac{xy}{2}(yp'_-+xp_-)\bm{\mathcal{V}}_{\perp}\cdot\bm{\mathcal{A}}'_{\perp}, \end{split}$$ where $\gamma^5=i\gamma^0\gamma^1\gamma^2\gamma^3$ and $\varepsilon^{\mu\nu\lambda\rho}$ is the completely antisymmetric tensor with $\varepsilon^{0123}=+1$. In the above equations, the quantity $g^2$ is given in Eq. (\[Phase\_LCFA\]) and we have taken into account that finally we need the matrix elements of these matrices between $\bar{U}_{s'}(p',x)$ and $U_s(p,x)$. The appearance of $\bm{q}_{\perp}$ in Eqs. (\[pi\_pi\_a\])-(\[g\_a\]) may suggest that large terms (i.e., of the order of $1/\omega_0$) could appear in total probabilities, if one imagines to carry out the integral over the transverse photon momentum, as one has to shift the variable $\bm{q}_{\perp}$ by a vector containing $\bm{\mathcal{A}}_{\perp}(\phi)$ \[see Eq. (\[Phi\_NCS\])\] in order to perform the resulting Gaussian integral. However, this does not represent a problem because the vector $\bm{q}_{\perp}$ is always multiplied by $\bm{a}_j$ \[see Eqs. (\[pi\_pi\_a\])-(\[g\_a\])\]. Indeed, the first equality in Eq. (\[Gamma\_exp\]) shows that the components $\Gamma_{R,\perp,j}(p,p',q;\phi)$ are only auxiliary quantities, as one finally needs to compute the components $(\Gamma_R(p,p',q;\phi)\Lambda_j)$. Since by definition the vector $\bm{\Lambda}_{\perp,j}$ is perpendicular to $\bm{q}_{\perp}$ \[see Eq. (\[Lambda\_i\])\], by replacing $a^{\mu}_j$ with $\Lambda_j^{\mu}$ in $\Gamma_{R,\perp,j}(p,p',q;\phi)$, it is easy to see that the $(\Gamma_R(p,p',q;\phi)\Lambda_j)$ depends on the vector $\bm{q}_{\perp}$ only through $\bm{\Lambda}_{\perp,j}$, a quantity which then, due to completeness, drops out once one computes total probabilities. Finally, we comment on the scaling of the radiative corrections due to the vertex correction at $\chi_0\gg 1$. This is more easily done in the case of $\Gamma_{R,-}(p,p',q;\phi)$ because one knows that the corresponding amplitude in nonlinear Compton scattering is simply proportional to $\bar{u}_{s'}(p')\hat{n}u_s(p)$ (before one regularizes the amplitude by integrating by parts, see, e.g., [@Mackenroth_2011]). Now, the structure of the phase in Eq. (\[Phase\_LCFA\]) shows that at large values of $\chi_0$ the main contribution to the integral in $S$ comes from the region $S\lesssim 1/\chi^{2/3}_0$. Thus, the terms in the preexponent in Eq. (\[Gamma\_R\_-\_LCFA\]) proportional to $p_-^2\bm{A}_{\perp}^{\prime\,2}(\phi)S^2=e^2p_-^2\bm{\mathcal{E}}_{\perp}^2(\phi)S^2$ give rise to the scaling $\alpha\chi_0^{2/3}$ of the radiative corrections in agreement with the results in Ref. [@Morozov_1981]. Conclusions {#VC_Conclusions} =========== We have computed the general expression of the one-loop vertex correction in an arbitrary plane-wave background field for the case of two on-shell external electrons and an off-shell external photon. By employing the operator technique within the Furry picture, we have obtained a relatively compact expression, which takes into account exactly the background plane-wave field. By showing explicitly that the vertex correction fulfills a generalized Ward identity, we have singled out the corresponding terms invariant under a gauge transformation of the external photon. As expected, the vertex-correction function features an infrared divergence, which is cured by assigning a small, finite mass to the photon. The ultraviolet divergence of the vertex correction has, instead, been shown to be renormalized as in vacuum. The important special case of the locally-constant field approximation has been investigated in detail. We have shown that in the high-field regime $\chi_0\gg 1$ the vertex-correction function induces radiative corrections which scale according to the Ritus-Narozhny conjecture as $\alpha\chi^{2/3}_0$, where $\chi_0$ is the amplitude of the quantum nonlinearity parameter. The component $\Gamma_q(p,p',q;\phi)$ of the vertex-correction function ======================================================================= In this appendix, we evaluate more explicitly the component $\Gamma_q(p,p',q;\phi)$ of the vertex-correction function although we have seen that it does not contribute to any transition matrix element. General structure of $\Gamma_q(p,p',q;\phi)$ -------------------------------------------- As we have mentioned in the main text, from the second equality in Eq. (\[Gamma\_2\]) and from the definition of $\Gamma^{\mu}(p,p',q;\phi)$ in Eq. (\[Gamma\^mu\]), we obtain $$\Gamma_q(p,p',q;\phi)=-ie^2\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\gamma^{\lambda}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}\hat{q}\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\gamma_{\lambda}.$$ Now, by writing $\hat{q}=\hat{\Pi}(\phi)+\hat{k}-m-[\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m]$ it is clear that we can express $\Gamma_q(p,p',q;\phi)$ as the difference of two terms containing only one propagator in the plane wave, which significantly simplifies its expression: $$\Gamma_q(p,p',q;\phi)=-ie^2\int\frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2-\kappa^2+i0}\gamma^{\lambda}\left[\frac{1}{\hat{\Pi}(\phi)+\hat{k}-\hat{q}-m+i0}-\frac{1}{\hat{\Pi}(\phi)+\hat{k}-m+i0}\right]\gamma_{\lambda}.$$ At this point, by following exactly the same steps as in the main text \[see Eq. (\[Gamma\_f\])\], it is easy to obtain the resulting expression $$\begin{split} \Gamma_q(p,p',q;\phi)&=\frac{\alpha}{4\pi}\int_0^{\infty}\frac{dsdt}{(s+t)^2}e^{-i\kappa^2t-i\frac{\tilde{G}^2_s}{s+t}}\left\{ 2\left[1-\frac{e\hat{n}[\hat{A}(\tilde{\psi}_{0,s})-\hat{A}(\phi)]}{2p'_-}\right]\left[\hat{\pi}_{p'}(\tilde{\psi}_{0,s})+\frac{\hat{\tilde{G}}_s}{s+t}\right]\right.\\ &\left.\left.\qquad\qquad-\frac{\hat{n}}{(s+t)^2}\frac{d\tilde{G}^2_s}{dk_-}+\frac{e\hat{n}[\hat{A}(\tilde{\psi}_{0,s})-\hat{A}(\phi)]}{\tau'_{0,-}+k_-}\left[\hat{\pi}_{p'}(\tilde{\psi}_{0,s})-\frac{\hat{\tilde{G}}_s}{s+t}\right]\right\}\right\vert_{k_-=0}\\ &\quad-\frac{\alpha}{4\pi}\int_0^{\infty}\frac{dudt}{(u+t)^2}e^{-i\kappa^2t-i\frac{\tilde{G}^2_u}{u+t}}\left\{ 2\left[\hat{\pi}_p(\tilde{\psi}_{0,u})+\frac{\hat{\tilde{G}}_u}{u+t}\right]\left[1+\frac{e\hat{n}[\hat{A}(\tilde{\psi}_{0,u})-\hat{A}(\phi)]}{2p_-}\right]\right.\\ &\left.\left.\qquad\qquad-\frac{\hat{n}}{(u+t)^2}\frac{d\tilde{G}^2_u}{dk_-}-\left[\hat{\pi}_p(\tilde{\psi}_{0,u})-\frac{\hat{\tilde{G}}_u}{u+t}\right]\frac{e\hat{n}[\hat{A}(\tilde{\psi}_{0,u})-\hat{A}(\phi)]}{\tau_{0,-}+k_-}\right\}\right\vert_{k_-=0}, \end{split}$$ where $$\begin{aligned} \tilde{G}^{\mu}_s&=\int_0^sds'\pi^{\mu}_{p'}(\tilde{\psi}_{0,s'}), && \tilde{\psi}_{0,s}=\phi+2s\tau'_{0,-}+2sk_-,&&\tau'_{0,-}=\frac{t}{s+t}p'_-,\\ \tilde{G}^{\mu}_u&=\int_0^udu'\pi^{\mu}_p(\tilde{\psi}_{0,u'}), && \tilde{\psi}_{0,u}=\phi-2u\tau_{0,-}-2uk_-,&&\tau_{0,-}=\frac{t}{u+t}p_-.\end{aligned}$$ Finally, by evaluating the remaining derivatives with respect to $k_-$ as $$\begin{aligned} \frac{d\tilde{G}^2_s}{dk_-}&=4(\tilde{G}_s\tilde{G}_{1,s}),\\ \frac{d\tilde{G}^2_u}{dk_-}&=4(\tilde{G}_u\tilde{G}_{1,u}),\end{aligned}$$ where $$\begin{aligned} \tilde{G}^{\mu}_{1,s}&=\int_0^sds'\,s'\pi^{\prime\,\mu}_{p'}(\tilde{\psi}_{0,s'})=\frac{1}{2(\tau'_{0,-}+k_-)}\left[s\pi^{\mu}_{p'}(\tilde{\psi}_{0,s})-\int_0^sds'\pi^{\mu}_{p'}(\tilde{\psi}_{0,s'})\right],\\ \tilde{G}^{\mu}_{1,u}&=-\int_0^udu'\,u'\pi^{\prime\,\mu}_p(\tilde{\psi}_{0,u'})=\frac{1}{2(\tau_{0,-}+k_-)}\left[u\pi^{\mu}_p(\tilde{\psi}_{0,u})-\int_0^udu'\pi^{\mu}_p(\tilde{\psi}_{0,u'})\right],\end{aligned}$$ we obtain $$\label{Gamma_q} \begin{split} \Gamma_q(p,p',q;\phi)&=\frac{\alpha}{2\pi}\int_0^{\infty}\frac{dsdt}{(s+t)^2}e^{-i\kappa^2t-i\frac{G^2_s}{s+t}}\left\{\left(1-\frac{\hat{n}\hat{\Delta}_{0,s}}{2p'_-}\right)\left[\hat{\pi}_{p'}(\psi_{0,s})+\frac{\hat{G}_s}{s+t}\right]\right.\\ &\left.\qquad\qquad-\frac{2\hat{n}}{(s+t)^2}(G_sG_{1,s})+\frac{\hat{n}\hat{\Delta}_{0,s}}{2\tau'_{0,-}}\left[\hat{\pi}_{p'}(\psi_{0,s})-\frac{\hat{G}_s}{s+t}\right]\right\}\\ &\quad-\frac{\alpha}{2\pi}\int_0^{\infty}\frac{dudt}{(u+t)^2}e^{-i\kappa^2t-i\frac{G^2_u}{u+t}}\left\{\left[\hat{\pi}_p(\psi_{0,u})+\frac{\hat{G}_u}{u+t}\right]\left(1+\frac{\hat{n}\hat{\Delta}_{0,u}}{2p_-}\right)\right.\\ &\left.\qquad\qquad-\frac{2\hat{n}}{(u+t)^2}(G_uG_{1,u})-\left[\hat{\pi}_p(\psi_{0,u})-\frac{\hat{G}_u}{u+t}\right]\frac{\hat{n}\hat{\Delta}_{0,u}}{2\tau_{0,-}}\right\}, \end{split}$$ where $$\Delta^{\mu}_{0,s/u}=e[A^{\mu}(\psi_{0,s/u})-A^{\mu}(\phi)],$$ and where all the quantities without the tilde are the same as those with the tilde but with $k_-=0$: $$\begin{aligned} G^{\mu}_s&=\int_0^sds'\pi^{\mu}_{p'}(\psi_{0,s'}), && G^{\mu}_{1,s}=\frac{1}{2\tau'_{0,-}}\left[s\pi_{p'}^{\mu}(\psi_{0,s})-\int_0^sds'\pi_{p'}^{\mu}(\psi_{0,s'})\right], && \psi_{0,s}=\phi+2s\tau'_{0,-},\\ G^{\mu}_u&=\int_0^udu'\pi^{\mu}_p(\psi_{0,u'}), && G^{\mu}_{1,u}=\frac{1}{2\tau_{0,-}}\left[u\pi_p^{\mu}(\psi_{0,u})-\int_0^udu'\pi_p^{\mu}(\psi_{0,u'})\right], && \psi_{0,u}=\phi-2u\tau_{0,-}.\end{aligned}$$ Regularization of $\Gamma_q(p,p',q;\phi)$ ----------------------------------------- Analogously to the other components of the vertex-correction function, the component $\Gamma_q(p,p',q;\phi)$ has in principle to be regularized as it is apparently logarithmically divergent in the ultraviolet. However, in the case $\Gamma_{R,q}(p,p',q;\phi)$, actually, it is not necessary to perform any subtraction of vacuum terms because, as we will show now, it vanishes for $A^{\mu}(\phi)=0$. It is convenient to perform the change of variable $s=x\sigma$ and $t=(1-x)\sigma$ ($u=x\sigma$ and $t=(1-x)\sigma$) in the first (second) integral in Eq. (\[Gamma\_q\]) [^2] and we obtain $$\label{Gamma_R_q} \begin{split} &\Gamma_{R,q}(p,p',q;\phi)=\frac{\alpha}{2\pi}\int_0^{\infty}\frac{d\sigma}{\sigma}\int_0^1dx\,e^{-i\kappa^2(1-x)\sigma}\\ &\quad\times\left\langle e^{-ix^2\sigma g_s^2}\left\{\left(1-\frac{\hat{n}\hat{\Delta}_{0,s}}{2p'_-}\right)[\hat{\pi}_{p'}(\theta'_{0,\sigma})+x\hat{g}_s]-2x^3\sigma(g_sg_{1,s})\hat{n}+\frac{\hat{n}\hat{\Delta}_{0,s}}{2(1-x)p'_-}[\hat{\pi}_{p'}(\theta'_{0,\sigma})-x\hat{g}_s]\right\}\right.\\ &\quad-\left.e^{-ix^2\sigma g_u^2}\left\{[\hat{\pi}_p(\theta_{0,\sigma})+x\hat{g}_u]\left(1+\frac{\hat{n}\hat{\Delta}_{0,u}}{2p_-}\right)-2x^3\sigma(g_ug_{1,u})\hat{n}-[\hat{\pi}_p(\theta_{0,\sigma})-x\hat{g}_u]\frac{\hat{n}\hat{\Delta}_{0,u}}{2(1-x)p_-}\right\}\right\rangle, \end{split}$$ where $$\begin{aligned} \theta'_{0,\sigma}&=\phi+2x(1-x)p'_-\sigma,\\ \theta_{0,\sigma}&=\phi-2x(1-x)p_-\sigma,\end{aligned}$$ and $$\begin{aligned} g^{\mu}_s=\frac{G^{\mu}_s}{s}=\int_0^1d\eta\,\pi_{p'}^{\mu}(\theta'_{0,\eta\sigma}),&& g^{\mu}_{1,s}=\frac{G^{\mu}_{1,s}}{s^2}=\frac{1}{2x(1-x)\sigma p'_-}\left[\pi_{p'}^{\mu}(\theta'_{0,\sigma})-\int_0^1d\eta\,\pi_{p'}^{\mu}(\theta'_{0,\eta\sigma})\right],\\ g^{\mu}_u=\frac{G^{\mu}_u}{u}=\int_0^1d\eta\,\pi_p^{\mu}(\theta_{0,\eta \sigma}),&& g^{\mu}_{1,u}=\frac{G^{\mu}_{1,u}}{u^2}=\frac{1}{2x(1-x)\sigma p_-}\left[\pi_p^{\mu}(\theta_{0,\sigma})-\int_0^1d\eta\,\pi_p^{\mu}(\theta_{0,\eta\sigma})\right].\end{aligned}$$ Note that $g_{1,s}^{\mu}$ and $g_{1,u}^{\mu}$ tend to constant values for $\sigma\to 0$. Since $\Delta^{\mu}_{0,s}$ and $\Delta^{\mu}_{0,u}$ vanish at $\sigma=0$, the only problematic terms are those in the square brackets in Eq. (\[Gamma\_R\_q\]) containing exclusively $\hat{\pi}_{p'}(\theta'_{0,\sigma})+x\hat{g}_s$ and $\hat{\pi}_p(\theta_{0,\sigma})+x\hat{g}_u$. However, since $\Gamma_q(p,p',q;\phi)$ will finally be multiplied by Volkov states both from the left and on the right, we can replace $\hat{\pi}_{p'}(\theta'_{0,\sigma})+x\hat{g}_s$ and $\hat{\pi}_p(\theta_{0,\sigma})+x\hat{g}_u$ in the terms which do not contain other gamma matrices as $$\begin{aligned} \hat{\pi}_{p'}(\theta'_{0,\sigma})+x\hat{g}_s&\rightarrow m(1+x)+\hat{\pi}_{p'}(\theta'_{0,\sigma})-\hat{\pi}_{p'}(\phi)+x\int_0^1d\eta\,[\hat{\pi}_{p'}(\theta'_{0,\eta\sigma})-\hat{\pi}_{p'}(\phi)],\\ \hat{\pi}_p(\theta_{0,\sigma})+x\hat{g}_u&\rightarrow m(1+x)+\hat{\pi}_p(\theta_{0,\sigma})-\hat{\pi}_p(\phi)+x\int_0^1d\eta\,[\hat{\pi}_p(\theta_{0,\eta\sigma})-\hat{\pi}_p(\phi)].\end{aligned}$$ In this way, the only remaining divergent terms are those proportional to $m(1+x)$ but these divergences cancel each other in Eq. (\[Gamma\_R\_q\]), which can be conveniently written in the manifestly convergent form $$\begin{split} \Gamma_{R,q}(p,p',q;\phi)&=\frac{\alpha}{2\pi}\int_0^{\infty}\frac{d\sigma}{\sigma}\int_0^1dx\,e^{-i\kappa^2(1-x)\sigma}\Bigg\langle m(1+x)\left(e^{-ix^2\sigma g_s^2}-e^{-ix^2\sigma g_u^2}\right)\\ &+e^{-ix^2\sigma g_s^2}\left\{\hat{\pi}_{p'}(\theta'_{0,\sigma})-\hat{\pi}_{p'}(\phi)+x\int_0^1d\eta\,[\hat{\pi}_{p'}(\theta'_{0,\eta\sigma})-\hat{\pi}_{p'}(\phi)]\right.\\ &\quad\left.+\frac{x}{1-x}\frac{\hat{n}\hat{\Delta}_{0,s}\hat{\pi}_{p'}(\theta'_{0,\sigma})}{2p'_-}-\frac{x(2-x)}{1-x}\frac{\hat{n}\hat{\Delta}_{0,s}\hat{g}_s}{2p'_-}-2x^3\sigma(g_sg_{1,s})\hat{n}\right\}\\ &-e^{-ix^2\sigma g_u^2}\left\{\hat{\pi}_p(\theta_{0,\sigma})-\hat{\pi}_p(\phi)+x\int_0^1d\eta\,[\hat{\pi}_p(\theta_{0,\eta\sigma})-\hat{\pi}_p(\phi)]\right.\\ &\quad\left.\left.-\frac{x}{1-x}\frac{\hat{\pi}_p(\theta_{0,\sigma})\hat{n}\hat{\Delta}_{0,u}}{2p_-}+\frac{x(2-x)}{1-x}\frac{\hat{g}_u\hat{n}\hat{\Delta}_{0,u}}{2p_-}-2x^3\sigma(g_ug_{1,u})\hat{n}\right\}\right\rangle. \end{split}$$ This expression also shows that $\Gamma_{R,q}(p,p',q;\phi)$ tends to zero for $A^{\mu}(\phi)\to 0$. Some considerations about the LCFA ---------------------------------- As we have mentioned in the main text, in order to study the structure component $\Gamma_{R,q}^{\mu}(p,p',q;\phi)$ of the vertex-correction function, it is useful to rewrite the phases $G_s^2/(s+t)$ and $G_u^2/(u+t)$ in a convenient form \[see Eq. (\[Gamma\_q\])\]. Indeed, since, as we have seen in the main text, the component $\Gamma_{R,q}(p,p',q;\phi)$ does not contribute to any transition matrix element, we only report here some considerations about these phases. By starting from the identity $v^2=2v_+v_--\bm{v}^2_{\perp}$, valid for a generic four-vector $v^{\mu}$, it can easily be shown that $$\begin{aligned} \label{G_s^2} G_s^2&=s^2(m^2+\delta m_{0,s}^2),\\ \label{G_u^2} G_u^2&=u^2(m^2+\delta m_{0,u}^2),\end{aligned}$$ where we have introduced the laser-induced square mass corrections $$\begin{aligned} \label{m_corr_0s} \delta m_{0,s}^2&=\frac{1}{s}\int_0^sds'\bm{\mathcal{A}}^2_{\perp}(\psi_{0,s'})-\frac{1}{s^2}\left[\int_0^sds'\bm{\mathcal{A}}_{\perp}(\psi_{0,s'})\right]^2=\frac{1}{s}\int_0^sds'\bm{\Delta}^2_{0,s',\perp}-\frac{1}{s^2}\left(\int_0^sds'\bm{\Delta}_{0,s',\perp}\right)^2,\\ \label{m_corr_0u} \delta m_{0,u}^2&=\frac{1}{u}\int_0^udu'\bm{\mathcal{A}}^2_{\perp}(\psi_{0,u'})-\frac{1}{u^2}\left[\int_0^udu'\bm{\mathcal{A}}_{\perp}(\psi_{0,u'})\right]^2=\frac{1}{u}\int_0^udu'\bm{\Delta}^2_{0,u',\perp}-\frac{1}{u^2}\left(\int_0^udu'\bm{\Delta}_{0,u',\perp}\right)^2.\end{aligned}$$ We notice that for the present case of on-shell electrons, the quantities $G_s^2$ and $G_u^2$ are non-negative. By proceeding as in the main text we arrive to the approximated expressions $$\begin{aligned} \label{m_corr_0_LCFA} \delta m_{0,s}^2&\approx \frac{1}{3}m^2\left(\frac{t}{s+t}\right)^2m^4s^2\chi_{p'}^2(\phi),&& \delta m_{0,u}^2\approx \frac{1}{3}m^2\left(\frac{t}{u+t}\right)^2m^4u^2\chi_p^2(\phi),\end{aligned}$$ where $\chi_p(\phi)=p_-\|\bm{\mathcal{E}}_{\perp}(\phi)\|/m^3$ and $\chi_{p'}(\phi)=p'_-\|\bm{\mathcal{E}_{\perp}}(\phi)\|/m^3$, with $\bm{\mathcal{E}}_{\perp}(\phi)=-\bm{\mathcal{A}}'_{\perp}(\phi)$. These approximated expressions are valid if $\eta_0/\chi_0^{2/3}=\chi_0^{1/3}/\xi_0\ll 1$ [@Baier_1989; @Khokonov_2002; @Di_Piazza_2007; @Dinu_2016; @Di_Piazza_2018_c; @Di_Piazza_2019; @Ilderton_2019_b; @Podszus_2019; @Ilderton_2019]. The final expressions of the phases $G_s^2/(s+t)$ and $G_u^2/(u+t)$ within the LCFA are $$\begin{aligned} \frac{G_s^2}{s+t}&=\frac{s}{s+t}m^2s\left[1+\frac{1}{3}\frac{t^2}{(s+t)^2}m^4s^2\chi_{p'}^2(\phi)\right],\\ \frac{G_u^2}{u+t}&=\frac{u}{u+t}m^2u\left[1+\frac{1}{3}\frac{t^2}{(u+t)^2}m^4u^2\chi_p^2(\phi)\right].\end{aligned}$$ The above expressions simplify by means of the mentioned changes of variables: $$\begin{aligned} \label{Phase_0s_LCFA} \frac{G_s^2}{s+t}&=x^2m^2\sigma\left[1+\frac{1}{3}x^2(1-x)^2m^4\sigma^2\chi_{p'}^2(\phi)\right],\\ \label{Phase_0u_LCFA} \frac{G_u^2}{u+t}&=x^2m^2\sigma\left[1+\frac{1}{3}x^2(1-x)^2m^4\sigma^2\chi_p^2(\phi)\right].\end{aligned}$$ [97]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**, ()]{} @noop [, ()]{} @noop []{} (, ) @noop []{} (, ) @noop []{} (, ) @noop [**, ()]{} @noop [, ()]{} @noop , @noop , @noop , @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop []{} (, ) @noop [**, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [ ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop []{} (, ) @noop []{} (, ) @noop []{} (, ) @noop [**, ()]{} @noop [, ()]{} @noop []{} (, ) @noop [**, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [ ()]{} @noop [**, ()]{} [^1]: The terms with three gamma matrices can be further reduced according to the identity $$\begin{split} \hat{A}\hat{B}\hat{C}&=\frac{1}{4}\text{tr}(\gamma_{\mu}\hat{A}\hat{B}\hat{C})\gamma^{\mu}-\frac{1}{4}\text{tr}(\gamma^5\gamma_{\mu}\hat{A}\hat{B}\hat{C})\gamma^5\gamma^{\mu}\\ &=\hat{A}(BC)-\hat{B}(AC)+\hat{C}(AB)+i\varepsilon_{\mu\nu\lambda\rho}\gamma^5\gamma^{\mu}A^{\nu}B^{\lambda}C^{\rho}, \end{split}$$ where $\gamma^5=i\gamma^0\gamma^1\gamma^2\gamma^3$ and $\varepsilon^{\mu\nu\lambda\rho}$ is the completely antisymmetric tensor with $\varepsilon^{0123}=+1$, which is valid for three arbitrary four-vectors $A^{\mu}$, $B^{\mu}$, and $C^{\mu}$. [^2]: Note that in this section and afterwards the symbol $x$ should not be confused with a spacetime point as it indicates a single, real variable.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove that a nonempty, proper subset $S$ of the complex plane can be approximated in a strong sense by polynomial filled Julia sets if and only if $S$ is bounded and $\hat{\mathbb{C}} \setminus \textrm{int}(S)$ is connected. The proof that such a set is approximable by filled Julia sets is constructive and relies on Fekete polynomials. Illustrative examples are presented. We also prove an estimate for the rate of approximation in terms of geometric and potential theoretic quantities.' address: - 'Department of Mathematics, University of Chicago, Chicago, IL 60637, United States.' - 'Department of Mathematics, Stony Brook University, Stony Brook, NY 11794-3651, United States.' author: - 'Kathryn A. Lindsey' - Malik Younsi date: May 19 2017 title: Fekete polynomials and shapes of Julia sets --- [^1] [^2] Introduction ============ The study of possible shapes of polynomial Julia sets was instigated by the first author in [@LIN2], who proved that any Jordan curve in the complex plane $\mathbb{C}$ can be approximated by polynomial Julia sets arbitrarily well in the Hausdorff distance, thereby answering a question of W.P. Thurston. \[thmKL\] Let $E \subset \mathbb{C}$ be any closed Jordan domain. Then for any $\epsilon>0$, there exists a polynomial $P$ such that $$d(E,\mathcal{K}(P))<\epsilon, \quad d(\partial E, \mathcal{J}(P))<\epsilon.$$ Here $\mathcal{K}(P):=\{z \in \mathbb{C} : P^m(z) \nrightarrow \infty \,\, \mbox{as}\,\, m \to \infty\}$ is the filled Julia set of $P$, $\mathcal{J}(P):=\partial \mathcal{K}(P)$ is the Julia set of $P$ and $d$ is the Hausdorff distance. Note that Theorem \[thmKL\] remains valid if $E$ is any nonempty connected compact set with connected complement, by a simple approximation process. Approximating compact sets by fractals has proven over the years to be a fruitful technique in the study of important problems in complex analysis, such as the universal dimension spectrum for harmonic measure (cf. the work of Carelson–Jones [@CAR2], Binder–Makarov–Smirnov [@BIN], and the references therein). Other related works include [@BIS], where it was shown that any connected compact set in the plane can be approximated by dendrite Julia sets, and [@KIM], containing applications to computer graphics. An important feature of the approach in [@LIN2] which seems to be absent from other works is that it is constructive and can easily be implemented to obtain explicit images of Julia sets representing various shapes. In this paper, we prove the following generalization of Theorem \[thmKL\]. \[mainthm1\] Let $E \subset \mathbb{C}$ be any nonempty compact set with connected complement. Then for any $\epsilon>0$, there exists a polynomial $P$ such that $$d(E,\mathcal{K}(P))<\epsilon, \quad d(\partial E, \mathcal{J}(P))<\epsilon.$$ Note that here $E$ is not assumed to be connected. This supersedes [@LIN2 Theorem 4.3], which deals with rational maps instead of polynomials. The proof of Theorem \[mainthm1\] generalizes the method introduced in [@LIN2] by putting it in the more natural framework of potential theory. More precisely, our approach is based on the observation that in Theorem \[thmKL\], the roots of the polynomial $P$ are equidistributed with respect to harmonic measure on $\partial E$. In the connected case, such points can be obtained as images of equally spaced points on the unit circle under the Riemann map. This, however, does not extend to disconnected sets and in this case, the explicit construction of equidistributed sequences of points is more difficult. Nevertheless, this is a classical problem in potential theory which has been extensively studied in the past. For instance, Fekete described in [@FEK] the following construction : Let $E$ be a nonempty compact set with connected complement, and let $n \geq 2$. A Fekete $n$-tuple for $E$ is any $n$-tuple $w_1^n, \dots, w_n^n \in E$ which maximizes the product $$\prod_{j<k}\|w_j^n-w_k^n\|.$$ Note that by the maximum principle, any Fekete $n$-tuple lies in $\partial E$. Fekete points are classical objects of potential theory which have proven over the years to be of fundamental importance to a variety of problems related to polynomial interpolation. For instance, they can be used to give a proof of Hilbert’s Lemniscate Theorem, which states that any compact plane set with connected complement can be approximated by polynomial lemniscates. In fact, as we will see, Hilbert’s theorem is closely related to Theorem \[mainthm1\]. For more information on Fekete points and their applications, we refer the reader to [@BOS] and [@SAF Chapter III, Section 1]. It is not difficult to show that Fekete points are indeed equidistributed with respect to harmonic measure on $\partial E$, in the sense that the counting measures $$\mu_n:= \frac{1}{n} \sum_{j=1}^n \delta_{w_j^n}$$ converge $\operatorname{weak}^$ to the harmonic measure (see Proposition \[weakconvergence\]). Combining this with the method put forward in [@LIN2] gives a constructive proof of Theorem \[mainthm1\]. Other equidistributed sequences of points include Leja points, which share many nice properties with Fekete points but are easier to compute numerically. Theorem \[thmExplicitLeja\] proves that our approximation technique works if Leja points, instead of Fekete points, are used. The technique in [@LIN2] for constructing polynomials whose filled Julia sets approximate a given Jordan domain required first obtaining an approximation of the Riemann map on the complement of that domain. Using Leja points or Fekete points obviates this step, and thus constitutes an advance in the constructive process. Note that Theorem \[mainthm1\] not only yields approximation of the set $E$ by the filled Julia set of a polynomial, but also approximation of its boundary by the corresponding Julia set. Consequently, we introduce the following definition. \[DefTotApprox\] A nonempty proper subset $S$ of the plane is totally approximable* by a collection $\mathcal{C}$ of nonempty proper subsets of the plane if for any $\epsilon>0$, there exists $C \in \mathcal{C}$ such that $$d(S,C)<\epsilon, \quad d(\partial S, \partial C)<\epsilon.$$ Note that closeness of two sets in the Hausdorff distance neither implies nor is implied by closeness of their boundaries, see Figure \[f:HausdorffDistance\]. On the other hand, it is easy to see that closeness of both the sets and their boundaries implies closeness of the complements, c.f. Lemma \[lem2\]. \[f:HausdorffDistance\] ![A unit disk, a unit disk with a small disk removed, and a ring of small disks. The first two are close in the Hausdorff distance, but their boundaries are not. The boundaries of the first and third are close in the Hausdorff distance, but the sets themselves are not. ](HausdorffDistance-eps-converted-to.pdf "fig:"){width=".6\linewidth"} With this definition, Theorem \[mainthm1\] states that any nonempty compact plane set $E$ with connected complement is totally approximable by polynomial filled Julia sets. More generally, using an approximation process, one can prove that this remains true if $E$ is replaced by any nonempty bounded set whose interior has connected complement. This condition also turns out to be necessary. \[mainthm3\] A non-empty proper subset $S$ of the complex plane $\mathbb{C}$ is totally approximable by polynomial filled Julia sets if and only if $S$ is bounded and $\mathbb{C} \setminus \operatorname{int}(S)$ is connected. Our method also gives a precise estimate for the rate of approximation in Theorem \[mainthm1\]. More precisely, let $E \subset \mathbb{C}$ be a compact set with connected complement $\Omega$ in the Riemann sphere ${\widehat{\mathbb{C}}}$. We also assume that the interior of $E$ is not empty. In particular, the Green’s function $g_{\Omega}(\cdot,\infty)$ for $\Omega$ with pole at $\infty$ exists. For $n \in \mathbb{N}$, define $s_n(E)$ to be the infimum of $s>0$ for which there exists a polynomial $p_n$ of degree $n$ such that $$E \subset \mathcal{K}(p_n) \subset E_s,$$ where $E_s:= E \cup \{z \in \Omega : g_\Omega(z,\infty) \leq s\}$. \[mainthm2\] Let $E \subset \mathbb{C}$ be a uniformly perfect compact set with nonempty interior and connected complement. Then there exists a real number $c=c(E)$ depending only on $E$ such that $$s_n(E) \leq c \, \frac{\log{n}}{\sqrt{n}} \qquad (n \geq 1).$$ Recall that a compact subset $E$ of the plane is uniformly perfect if there exists a real number $a>0$ such that for any $z \in E$ and for any $0<r<\operatorname{diam}(E)$, there is a point $w \in E$ with $ar \leq \|z-w\| \leq r$. This condition on $E$ is not too restrictive, since, for example, uniformly perfect sets include compact sets consisting of finitely many nontrivial connected components. A direct consequence of Theorem \[mainthm2\] is that $s_n(E) \to 0$ as $n \to \infty$. In particular, Theorem \[mainthm2\] implies Theorem \[mainthm1\] for uniformly perfect sets, since the compact sets $E_s$ shrink to $E$ as $s$ decreases to $0$. We also mention that results similar to Theorem \[mainthm2\] were obtained in [@AND] and [@LAI], but for the rate of approximation by polynomial lemniscates in Hilbert’s Lemniscate Theorem. Finally, we address the natural question of whether a nonempty connected compact set with connected complement can be approximated by Jordan domain Julia sets. We prove that, in this case, the Julia sets of our approximating polynomials are not only Jordan curves but quasicircles. \[t:connected\] Let $E$ be a nonempty connected compact set with connected complement. Then there are polynomials $(P_n)_{n \geq 1}$ whose filled Julia sets are closed Jordan domains totally approximating $E$ in the sense of Definition \[DefTotApprox\]. Moreover, the polynomials can be constructed such that each $P_n$ is a hyperbolic polynomial whose Julia set $\mathcal{J}(P_n)$ is a quasicircle. The remainder of the paper is organized as follows. Section \[sec2\] contains various preliminaries from potential theory. In Section \[sec3\], we prove an explicit version of Theorem \[mainthm1\] using polynomials with zeros at points of any equidistributed sequence with respect to harmonic measure. Section \[sec4\] is devoted to the special case of Fekete polynomials, which yields a proof of Theorem \[mainthm2\]. Then, in Section \[sec5\], we give an alternate proof of Theorem \[mainthm1\] using Hilbert’s Lemniscate Theorem. Section \[sec6\] contains the proof of Theorem \[mainthm3\]. In Section \[sec7\], we prove Theorem \[t:connected\]. Lastly, Section \[sec8\] is devoted to numerical examples. We first discuss Leja points and prove that our approximation scheme remains valid if Fekete points are replaced by Leja points. This allows us to compute images of Julia sets representing various disconnected shapes. Preliminaries from potential theory {#sec2} =================================== This section contains various preliminaries from potential theory, including the notions of potential, logarithmic capacity and Green’s function. The proofs are quite standard and are found in [@RAN] for example, but we include them for the reader’s convenience. We begin with the notions of potential and energy of a measure. Let $\mu$ be a finite positive Borel measure on $\mathbb{C}$ with compact support. 1. The potential of $\mu$ is the function $p_\mu : \mathbb{C} \to [-\infty,\infty)$ defined by $$p_\mu(z):=\int \log{\|z-w\|} \, d\mu(w) \qquad (z \in \mathbb{C}).$$ 2. The energy of $\mu$, noted $I(\mu)$, is given by $$I(\mu):= \int \int \log{\|z-w\|} \, d\mu(z) \, d\mu(w) = \int p_\mu(z) \, d\mu(z).$$ One can check that $p_\mu$ is subharmonic on $\mathbb{C}$, harmonic on $\mathbb{C} \setminus (\operatorname{supp}\mu)$ and that $$\label{eqpotential} p_{\mu}(z) = \mu(\mathbb{C}) \log{\|z\|} + O(\|z\|^{-1})$$ as $z \to \infty$. Now, let $E$ be a nonempty compact subset of $\mathbb{C}$. We assume that the complement of $E$ is connected. Let $\mathcal{P}(E)$ denote the collection of all Borel probability measures supported on $E$. A measure $\nu \in \mathcal{P}(E)$ is an equilibrium measure for $E$ if $$I(\nu) = \sup_{\mu \in \mathcal{P}(E)} I(\mu).$$ A standard $\operatorname{weak}^$-convergence argument shows that $E$ always has an equilibrium measure, see [@RAN Theorem 3.3.2]. Moreover, if the supremum in the definition is not $-\infty$, then the equilibrium measure $\nu$ for $E$ is always unique and supported on $\partial E$ ([@RAN Theorem 3.7.6]). It also coincides with $\omega_{\Omega}(\cdot, \infty)$, the harmonic measure for $\Omega:={\widehat{\mathbb{C}}}\setminus E$ and $\infty$. The energy of the equilibrium measure is used to define the logarithmic capacity of $E$. The logarithmic capacity* of $E$ is defined by $$\operatorname{cap}(E):=e^{I(\nu)},$$ where $\nu$ is the equilibrium measure for $E$. The quantity $-I(\nu)=-\log{\operatorname{cap(E)}}$ is usually called Robin’s constant. For example, the capacity of a closed disk is equal to its radius and the capacity of a segment is equal to a quarter of its length. For the rest of the paper, we shall suppose that $\operatorname{cap}(E)>0$. This can always be assumed without loss of generality in Theorem \[mainthm1\], adjoining a small disk in $E$ if necessary. The condition of positive logarithmic capacity is enough to ensure that the Green’s function exists. Let $\Omega:={\widehat{\mathbb{C}}}\setminus E$ be the complement of $E$ in the Riemann sphere. The Green’s function for $\Omega$ with pole at $\infty$ is the unique function $g_{\Omega}(\cdot,\infty):\Omega \to (0,\infty]$ such that 1. $g_{\Omega}(\cdot,\infty)$ is harmonic on $\Omega \setminus \{\infty\}$; 2. $g_{\Omega}(\infty,\infty)=\infty$ and as $z \to \infty$, $$g_{\Omega}(z,\infty) = \log{\|z\|} + O(1);$$ 3. $g_{\Omega}(z,\infty) \to 0$ as $z \to \zeta$, for all $\zeta \in \partial \Omega$ except possibly a set of zero logarithmic capacity. The Green’s function for $\Omega$ with pole at $\infty$ can be recovered from the potential of the equilibrium measure $\nu$. \[TheoFrostman\] If $\nu$ is the equilibrium measure for $E$, then 1. $p_\nu \geq I(\nu)$ on $\mathbb{C}$; 2. $p_\nu(z) = I(\nu)$ for all $z \in E$ except possibly a set of zero logarithmic capacity; 3. $g_{\Omega}(z,\infty) = p_{\nu}(z)-I(\nu) \qquad (z \in \Omega \setminus \{\infty\}).$ See ([@RAN Theorem 3.3.4]). Note that combining (iii) with Equation (\[eqpotential\]), we get $$I(\nu)=\log \operatorname{cap}(E) = \lim_{z \to \infty} (\log{\|z\|} - g_{\Omega}(z,\infty)).$$ For sufficiently nice compact sets $E$, the potential $p_\nu$ is continuous on $\mathbb{C}$ and the exceptional set of zero capacity in (ii) is in fact empty. More precisely, again under the assumption that $\operatorname{cap}(E)>0$, the equality $p_\nu(z) = I(\nu)$ holds for all $z \in E$ if and only if every point of $\partial E$ is regular for the Dirichlet Problem. A sufficient condition for this to hold is that $E$ is uniformly perfect. See [@RAN Theorem 3.1.3] and [@RAN Theorem 4.2.4] Constructive approximation by Julia sets {#sec3} ======================================== In this section, we prove an explicit version of Theorem \[mainthm1\]. As before, let $E \subset \mathbb{C}$ be a compact set with connected complement $\Omega$. We will construct polynomials whose filled Julia sets totally approximate $E$, in the sense of Definition \[DefTotApprox\]. First, we can assume that the interior of $E$ is not empty, adjoining a small disk in $E$ if necessary. We can further suppose without loss of generality that $0$ is an interior point of $E$, since otherwise it suffices to conjugate the polynomial by a suitable translation. Note that the above ensures that $\operatorname{cap}(E)>0$. For reasons that will soon become clear, we shall assume in addition that $E$ is uniformly perfect. Now, let $(q_n)_{n \geq 1}$ be any sequence of monic polynomials of degree $n$ having all their zeros in $\partial E$ such that the counting measures $$\mu_n:= \frac{1}{n} \sum_{\zeta \in q_n^{-1}(\{0\})} \delta_{\zeta}$$ converge $\operatorname{weak}^$ to $\nu$, the equilibrium measure for $E$. We will need the following lemma on the behavior of $\|q_n\|^{1/n}$. \[lemmaTom\] We have 1. $\|q_n\|^{1/n} \to e^{p_{\nu}}$ locally uniformly on $\Omega = {\widehat{\mathbb{C}}}\setminus E$; 2. $\lim_{n \to \infty} \\|q_n\\|_E^{1/n} = \operatorname{cap}(E)$. Here $\\|q_n\\|_E = \sup_{z \in E} \|q_n(z)\|.$ The following proof is due to Thomas Ransford. To prove (i), fix $z \in \mathbb{C} \setminus E$. Since the function $w \mapsto \log{\|z-w\|}$ is continuous on $\partial E$, we have, by $\operatorname{weak}^$ convergence, $$\int_{\partial E} \log{\|z-w\|} \, d\mu_n(w) \to \int_{\partial E} \log{\|z-w\|} \, d\nu(w),$$ i.e. $$\label{simpleconv} \frac{1}{n} \log{\|q_n(z)\|} \to p_\nu(z),$$ and this holds for all $z \in \mathbb{C} \setminus E$. Now, let $D$ be a disk whose closure is contained in $\mathbb{C} \setminus E$. Note that the sequence $(q_n^{1/n})_{n \geq 1}$ is uniformly bounded on $D$, so by Montel’s theorem and (\[simpleconv\]), every subsequence has a subsequence converging uniformly to some analytic function $h$ on $D$ with $\|h\|=e^{p_{\nu}}$. This shows that $\|q_n\|^{1/n} \to e^{p_{\nu}}$ locally uniformly on $\mathbb{C} \setminus E$. The fact that $\|q_n\|^{1/n} \to e^{p_{\nu}}$ uniformly near $\infty$ follows from Equation (\[eqpotential\]). To prove (ii), first note that $$\int_{\partial E} \frac{1}{n} \log{\|q_n(z)\|} \, d\nu(z) = \int_{\partial E} \int_{\partial E} \log{\|z-w\|} \, d\nu(z) d\mu_n(w) = \int_{\partial E} p_{\nu}(w) \, d\mu_n(w) \geq I(\nu),$$ the last inequality because $p_{\nu} \geq I(\nu)$ on $\mathbb{C}$, by Theorem \[TheoFrostman\]. It follows that $$\label{equa1} \\|q_n\\|_E ^{1/n} \geq e^{I(\nu)}=\operatorname{cap}(E) \qquad (n \geq 1).$$ For the other direction, recall that since $E$ is uniformly perfect, the potential $p_{\nu}$ is continuous on $\mathbb{C}$ and satisfies $p_{\nu}=I(\nu)$ everywhere on $E$ (see the remark following Theorem \[TheoFrostman\]). Thus, given $\epsilon>0$, there exists a bounded neighborhood $V$ of $E$ such that $p_{\nu} \leq I(\nu) + \epsilon$ on $V$. By (i), the functions $\|q_n\|^{1/n}$ converge uniformly to $e^{p_\nu}$ on $\partial V$, so there exists $N$ such that, for all $n \geq N$ and $z \in \partial V$, we have $$\|q_n(z)\|^{1/n} \leq e^{p_\nu(z)}+\epsilon \leq e^{\epsilon}\operatorname{cap}(E) + \epsilon.$$ By the maximum principle, the same inequality holds for all $n \geq N$ and all $z \in E$, from which it follows that $$\limsup_{n \to \infty} \\|q_n\\|_E^{1/n} \leq e^{\epsilon}\operatorname{cap}(E) + \epsilon.$$ Letting $\epsilon \to 0$ and combining with (\[equa1\]), we get (ii). We can now prove the following explicit version of Theorem \[mainthm1\]. \[thmExplicit\] Let $E$ and $(q_n)_{n \geq 1}$ be as above. For $s>0$ and $n \in \mathbb{N}$, define the polynomial $$P_{n,s}(z):=z\frac{e^{-ns/2}}{\operatorname{cap}(E)^n} q_n(z).$$ Then for any bounded neighborhood $U$ of $E$, there exist $s$ and $n$ such that $$E \subset \operatorname{int}(\mathcal{K}(P_{n,s})) \subset U.$$ In particular, the set $E$ is totally approximable by the filled Julia sets of the polynomials $P_{n,s}$, since $U$ can be made arbitrarily close to $E$. Fix $s>0$ sufficiently small so that $$g_{\Omega}(z,\infty)>s \qquad (z \notin U).$$ By Lemma \[lemmaTom\], we can choose $n \in \mathbb{N}$ sufficiently large so that the following conditions hold : 1. $\displaystyle \left(\frac{R}{r}\right)^{1/n} \frac{\\|q_n\\|_E^{1/n}}{\operatorname{cap}(E)} < e^{s/2} \qquad (z \in E)$ 2. $\displaystyle \left\| \frac{1}{n} \log{\|q_n(z)\|} - p_\nu(z) \right\| \leq \frac{s}{4} \qquad (z \notin U)$ 3. $\displaystyle r e^{ns/4} > R,$ where $r, R>0$ are such that $\overline{\mathbb{D}}(0,r) \subset E$ and $U \subset \mathbb{D}(0,R)$. First, note that condition (i) implies that for $z \in E$, $$\|P_{n,s}(z)\| < R \frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \\|q_n\\|_E \leq R e^{-ns/2} e^{ns/2}\frac{r}{R} = r,$$ so that in particular, $P_{n,s}(E) \subset \mathbb{D}(0,r) \subset E$, which clearly implies that $E \subset \operatorname{int}(\mathcal{K}(P_{n,s}))$. On the other hand, if $z \notin U$, then by (ii) and Theorem \[TheoFrostman\], we get $$g_\Omega(z,\infty) -\frac{s}{4} \leq \frac{1}{n} \log{\|q_n(z)\|} - I(\nu)$$ and thus $$\exp \left(n \left(g_\Omega(z,\infty) - \frac{s}{2} - \frac{s}{4} \right) \right) \leq \frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \|q_n(z)\| = \left\|\frac{P_{n,s}(z)}{z}\right\|.$$ Since $g_\Omega(z,\infty) > s$, we obtain $$\label{eqb} e^{ns/4}\|z\| < \|P_{n,s}(z)\|.$$ Now, by (iii), we get $$\|P_{n,s}(z)\| > re^{ns/4} > R,$$ so that $P_{n,s}(z) \notin U$ whenever $z \notin U$. Equation (\[eqb\]) then shows that the iterates of $P_{n,s}$ converge to $\infty$ on ${\widehat{\mathbb{C}}}\setminus U$, so that ${\widehat{\mathbb{C}}}\setminus U \subset {\widehat{\mathbb{C}}}\setminus \mathcal{K}(P_{n,s})$ or, equivalently, $ \mathcal{K}(P_{n,s}) \subset U$. Fekete points {#sec4} ============= In this section, we will see how using Fekete points for the zeros of the polynomials $q_n$ of Theorem \[thmExplicit\] gives a more explicit version of that result which has Theorem \[mainthm2\] as a corollary. First, we need some definitions and the basic properties of Fekete points. As before, let $E$ be a nonempty compact set with connected complement, and assume that $\operatorname{cap}(E)>0$. For $n \geq 2$, the $n$-th diameter of $E$ is given by $$\delta_n(E):=\sup_{w_1,\dots,w_n\in E} \prod_{j<k}\|w_j-w_k\|^{\frac{2}{n(n-1)}}.$$ Recall from the introduction that an $n$-tuple $w_1^n,\dots,w_n^n \in E$ for which the supremum is attained is called a Fekete $n$-tuple for $E$. Since $E$ is compact and nonempty, a Fekete $n$-tuple always exists, though it need not be unique. Moreover, the maximum principle shows that any Fekete $n$-tuple lies in $\partial E$. Also, the quantity $\delta_2(E)$ is the usual diameter of $E$, and $\delta_n(E) \leq \delta_2(E)$ for all $n$. In fact, the sequence $(\delta_n(E))_{n \geq 2}$ is decreasing and has a limit, which is nothing other than the logarithmic capacity of $E$. \[thmFekSze\] The sequence $(\delta_n(E))_{n \geq 2}$ is decreasing and $$\lim_{n \to \infty} \delta_n(E) = \operatorname{cap}(E).$$ See [@RAN Theorem 5.5.2]. A Fekete polynomial for $E$ of degree $n$ is a polynomial of the form $$q(z):=\prod_{j=1}^n(z-w_j^n),$$ where $w_1^n,\dots,w_n^n$ is a Fekete $n$-tuple for $E$. As before, we denote the sup-norm of the polynomial $q$ on $E$ by $\\|q\\|_E$, i.e. $$\\|q\\|_E= \sup_{z \in E} \|q(z)\|.$$ \[thmFek\] Let $q$ be a monic polynomial of degree $n$. Then $$\\|q\\|_E^{1/n} \geq \operatorname{cap}(E).$$ If in addition $q$ is a Fekete polynomial for $E$, then $$\\|q\\|_E^{1/n} \leq \delta_n(E).$$ To prove the first statement, it suffices to observe that since $E \subset q^{-1}\left(\overline{\mathbb{D}}(0,\\|q\\|_E)\right)$, we have $$\operatorname{cap}(E) \leq \operatorname{cap}\left(q^{-1}\left(\overline{\mathbb{D}}(0,\\|q\\|_E)\right) \right),$$ by the monotonicity of logarithmic capacity. The quantity on the right-hand side is easily seen to be equal to $\operatorname{cap}\left(\overline{\mathbb{D}}(0,\\|q\\|_E)\right)^{1/n} = \\|q\\|_E^{1/n}$, since in this case the Green’s function can be identified explicitly. This proves the first statement. To prove the second statement, write $q(z):=\prod_{j=1}^n(z-w_j^n)$ where $w_1^n,\dots,w_n^n$ is a Fekete $n$-tuple for $E$. If $z \in E$, then $z,w_1^n,\dots,w_n^n$ is an $(n+1)$-tuple in $E$, so $$\prod_{i=1}^n \|z-w_i^n\| \prod_{j<k} \|w_j^n-w_k^n\| \leq \delta_{n+1}(E)^{n(n+1)/2},$$ and hence $$\|q(z)\| \leq \frac{\delta_{n+1}(E)^{n(n+1)/2}}{\delta_n(E)^{n(n-1)/2}} \leq \frac{\delta_n(E)^{n(n+1)/2}}{\delta_n(E)^{n(n-1)/2}} = \delta_n(E)^n,$$ as required. Before stating the next result, we need the notion of Harnack distance. Let $D$ be a domain in ${\widehat{\mathbb{C}}}$. Given $z,w \in D$, the Harnack distance between $z$ and $w$ is the smallest number $\tau_D(z,w)$ such that, for every positive harmonic function $u$ on $D$, $$\tau_D(z,w)^{-1} u(w) \leq u(z) \leq \tau_D(z,w)u(w).$$ The existence of $\tau_D$ is a simple consequence of Harnack’s inequality. It is not difficult to show that $\log \tau_D(z,w)$ is a continuous semimetric on $D$ ([@RAN Theorem 1.3.8]). \[thmBernstein\] Let $E$ be as before, and let $\Omega={\widehat{\mathbb{C}}}\setminus E$. If $q$ is a polynomial of degree $n$, then $$\left( \frac{\|q(z)\|}{\\|q\\|_E}\right)^{1/n} \leq e^{g_\Omega(z,\infty)} \qquad (z \in \Omega).$$ If in addition $q$ is a Fekete polynomial for $E$, then $$\left( \frac{\|q(z)\|}{\\|q\\|_E}\right)^{1/n} \geq e^{g_\Omega(z,\infty)} \left( \frac{\operatorname{cap}(E)}{\delta_n(E)} \right)^{\tau_\Omega(z,\infty)} \qquad (z \in \Omega).$$ To prove the first statement, assume without loss of generality that $q$ is monic, and define $$u(z):= \frac{1}{n} \log{\|q(z)\|} - \frac{1}{n} \log{\\|q\\|_E} - g_\Omega(z,\infty) \qquad (z \in \Omega \setminus \{\infty\}).$$ Then $u$ is subharmonic on $\Omega \setminus \{\infty\}$. Also, $$u(z)=\log{\|z\|} - \frac{1}{n} \log{\\|q\\|_E} - \log{\|z\|} + \log{\operatorname{cap}(E)} + o(1)$$ as $z \to \infty$, and therefore setting $$u(\infty) := \log{\operatorname{cap}(E)} - \frac{1}{n} \log{\\|q\\|_E}$$ makes $u$ subharmonic on $\Omega$. Now, since $\partial \Omega \subset E$, we have $$\limsup_{z \to \zeta} u(z) \leq 0$$ for all $\zeta \in \partial \Omega$, and thus $u \leq 0$ on $\Omega$ by the maximum principle for subharmonic functions. Note that $u(\infty) \leq 0$ by Theorem \[thmFek\]. This implies the result. To prove the second statement, note that if $q$ is a Fekete polynomial for $E$, then in particular all its zeros lie in $E$, so that $u$ is actually harmonic on $\Omega$. Also, from the first part of the proof, we have $u \leq 0$ on $\Omega$. Therefore, applying the definition of Harnack distance to $-u$, we get $$u(z) \geq \tau_\Omega(z,\infty)u(\infty) \qquad (z \in \Omega).$$ Now, by Theorem \[thmFek\], $$u(\infty) = \log{\operatorname{cap}(E)} - \frac{1}{n} \log{\\|q\\|_E} \geq \log{\operatorname{cap}(E)} - \log{\delta_n(E)}.$$ Combining the last two inequalities yields the desired conclusion. \[CorBern\] For $z \in \Omega$, we have $$\|z\| \leq R(E) e^{g_\Omega(z,\infty)},$$ where $R(E):= \sup_{w \in E} \|w\|$. This follows from the first part of Theorem \[thmBernstein\] applied to $q(z):=z^n$. The following result shows that Fekete points are equidistributed with respect to the harmonic measure and therefore can be used as zeros of the approximating polynomials $q_n$ of Theorem \[thmExplicit\]. \[weakconvergence\] Let $E$ be as before and for each $n \geq 2$, let $$\mu_n:= \frac{1}{n} \sum_{j=1}^n \delta_{w_j^n},$$ where $w_1^n, w_2^n, \dots, w_n^n$ is a Fekete $n$-tuple for $E$ and $\delta_{w_j^n}$ is the unit point mass at the point $w_j^n$. Then $\mu_n \to \nu$ $\operatorname{weak}^$, where $\nu$ is the equilibrium measure for $E$. First, recall that $\nu$ and $\mu_n$ for $n \geq 2$ are all supported on $\partial E$. Now, let $\mu$ be any $\operatorname{weak}^$-limit of the sequence of measures $(\mu_n)$. Then by the monotone convergence theorem, we have $$I(\mu) = \lim_{m \to \infty} \int_{\partial E} \int_{\partial E} \max(\log{\|z-w\|},-m) d\mu(z) d\mu(w).$$ Now, by the Stone-Weierstrass theorem, we have that $\mu_n \times \mu_n \to \mu \times \mu$ $\operatorname{weak}^$ on $\partial E \times \partial E$, so the above equality becomes $$\begin{aligned} I(\mu) &=& \lim_{m \to \infty} \lim_{n \to \infty} \int_{\partial E} \int_{\partial E} \max(\log{\|z-w\|},-m) d\mu_n(z) d\mu_n(w)\\ &=& \lim_{m \to \infty} \lim_{n \to \infty} \frac{1}{n^2} \sum_{j=1}^n \sum_{k=1}^n \max(\log{\|w_j^n-w_k^n\|},-m)\\ &\geq& \lim_{m \to \infty} \lim_{n \to \infty} \left(\frac{-m}{n} + \frac{2}{n^2} \sum_{j < k} \log{\|w_j^n-w_k^n\|}\right)\\ &=& \lim_{n \to \infty} \frac{n-1}{n} \log{\delta_n(E)}.\end{aligned}$$ This last limit is equal to $\log \operatorname{cap}(E) = I(\nu)$, by Theorem \[thmFekSze\]. The maximality and uniqueness of the equilibrium measure $\nu$ then implies that $\mu=\nu$, as required. Combining this with Lemma \[lemmaTom\], we get that if $w_1^n, w_2^n, \dots, w_n^n$ is a Fekete $n$-tuple for $E$, then $$\frac{1}{n} \sum_{j=1}^n \log{\|z-w_j^n\|} \to p_\nu(z)$$ as $n \to \infty$ for $z \in \Omega$, and the convergence is uniform on compact subsets. The following result due to Pritsker yields a precise rate of convergence, under the additional assumption that $E$ is uniformly perfect. \[thmPRI\] Let $E$, $w_1^n, w_2^n,\dots,w_n^n$, $n \geq 2$, be as above, and assume in addition that $E$ is uniformly perfect. Then $$\label{eqprit1} \left\| \frac{1}{n} \sum_{j=1}^n \log{\|z-w_j^n\|} - p_\nu(z) \right\| \leq C \frac{\log n}{\sqrt{n}} \qquad \left(z \in \Omega, \, g_\Omega(z,\infty)>\frac{1}{n}\right)$$ where $C=C(E)$ is a constant depending only on $E$. \[defPRI\] We define the Pritsker constant* of $E$ to be the smallest constant $C(E)$ for which (\[eqprit1\]) holds. We shall also need the following result, also due to Pritsker, on the rate of convergence of the $n$-th diameter to the logarithmic capacity. \[thmPRI2\] Under the assumptions of Theorem \[thmPRI\], we have $$\log{\frac{\delta_n(E)}{\operatorname{cap}(E)}} \leq C_2 \frac{\log{n}}{\sqrt{n}},$$ where $C_2=C_2(E)$ is a constant depending only on $E$. See [@PRI Theorem 2.2] and its corollary. We now have everything that we need in order to prove a more explicit version of Theorem \[thmExplicit\] based on Fekete points. As a consequence, we shall obtain a proof of Theorem \[mainthm2\] on the rate of approximation by Julia sets. As before, let $E \subset \mathbb{C}$ be a compact set with connected complement $\Omega$. As in Section \[sec3\], assume without loss of generality that $0$ is an interior point of $E$, so that in particular $\operatorname{cap}(E)>0$ and the Green’s function $g_\Omega(\cdot,\infty)$ is well-defined. We now introduce two positive geometric quantities associated with the set $E$, which we call the inner radius and outer radius of $E$. These are defined by $$r(E):= \operatorname{dist}(0,\partial E)$$ and $$R(E):= \sup_{w \in E} \|w\|$$ respectively. Note that $\overline{\mathbb{D}}(0,r(E)) \subset E \subset \overline{\mathbb{D}}(0,R(E)) $. Finally, let us assume as well that $E$ is uniformly perfect. Recall that in this case, the Pritsker constant of $E$ is defined to be the smallest constant $C(E)$ such that $$\left\| \frac{1}{n} \sum_{j=1}^n \log{\|z-w_j^n\|} - p_\nu(z) \right\| \leq C(E) \frac{\log n}{\sqrt{n}} \qquad \left(z \in \Omega, \, g_\Omega(z,\infty)>\frac{1}{n}\right)$$ where $w_1^n,w_2^n,\dots,w_n^n$ are the points of a Fekete $n$-tuple for $E$ and $\nu$ is the equilibrium measure for $E$. \[thmExplicit2\] Let $s>0$, and suppose that $n$ is sufficiently large so that the following conditions hold : 1. $\displaystyle \frac{1}{n} \leq s$ 2. $\displaystyle C(E) \frac{\log n}{\sqrt{n}} \leq \frac{s}{4}$ 3. $\displaystyle e^{ns/4} \geq \frac{R(E)e^s}{r(E)}$ 4. $\displaystyle \left(\frac{R(E)e^s}{r(E)}\right)^{1/n} \frac{\delta_ n(E)}{\operatorname{cap}(E)} \leq e^{s/2}$. Then the polynomial $$P_n(z)=P_{n,s}(z):=z\frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \prod_{j=1}^n (z-w_j^n)$$ satisfies $$E \subset \operatorname{int} (\mathcal{K}(P_n)) \subset E_s,$$ where $E_s:= E \cup \{z \in \Omega : g_\Omega(z,\infty) \leq s\}$. In particular, the set $E$ is totally approximable by the filled Julia sets of the polynomials $P_n$’s, since the compact sets $E_s$ shrink to $E$ as $s$ decreases to $0$. The idea is very similar to the proof of Theorem \[thmExplicit\]. Let $\Omega_s := {\widehat{\mathbb{C}}}\setminus E_s = \{z \in \Omega : g_{\Omega}(z,\infty)>s\}$. Note that $${\widehat{\mathbb{C}}}\setminus \overline{\mathbb{D}}(0,R(E)e^s) \subset \Omega_s,$$ by Corollary \[CorBern\]. Now, let $z \in \Omega_s$. Then $g(z,\infty) > 1/n$ by (i), so that by Theorem \[thmPRI\], $$\left\| \frac{1}{n} \sum_{j=1}^n \log{\|z-w_j^n\|} - p_\nu(z) \right\| \leq C(E) \frac{\log n}{\sqrt{n}} \leq \frac{s}{4},$$ where we used condition (ii). By Theorem \[TheoFrostman\], we get $$g_\Omega(z,\infty) -\frac{s}{4} \leq \frac{1}{n} \sum_{j=1}^n \log{\|z-w_j^n\|} - I(\nu)$$ and thus $$\exp \left(n \left(g_\Omega(z,\infty) - \frac{s}{2} - \frac{s}{4} \right) \right) \leq \|Q_n(z)\|,$$ where $$Q_n(z):= \frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \prod_{j=1}^n (z-w_j^n) = \frac{P_n(z)}{z}.$$ Since $g_\Omega(z,\infty) > s$, we obtain $$\label{eq} \frac{R(E)e^s}{r(E)} \leq e^{ns/4} < \|Q_n(z)\|,$$ by (iii), and this holds for all $z \in \Omega_s$. On the other hand, if $z \in E$, then by Theorem \[thmFek\] and (iv), $$\label{eqq} \|Q_n(z)\| \leq e^{-ns/2}\left(\frac{\delta_n(E)}{\operatorname{cap}(E)}\right)^n \leq e^{-ns/2} e^{ns/2} \frac{r(E)}{R(E)e^s} = \frac{r(E)}{R(E)e^s}.$$ Now, for $z \in \Omega_s$, we have, by (\[eq\]), $$\|P_n(z)\| = \|z\| \|Q_n(z)\| > r(E) \frac{R(E)e^s}{r(E)} = R(E)e^s$$ so that $P_n(z) \in {\widehat{\mathbb{C}}}\setminus \overline{\mathbb{D}}(0,R(E)e^s) \subset \Omega_s$. Moreover, again by (\[eq\]), $$\|P_n(z)\| = \|z\| \|Q_n(z)\| > \|z\| \frac{R(E)e^s}{r(E)},$$ where $$\frac{R(E)e^s}{r(E)}>1.$$ It follows that the iterates of $P_n$ converge to $\infty$ on $\Omega_s$, so that $\Omega_s \subset {\widehat{\mathbb{C}}}\setminus \mathcal{K}(P_n)$ and thus $\mathcal{K}(P_n) \subset E_s$. Finally, for $z \in E$, we have, by (\[eqq\]), $$\|P_n(z)\| = \|z\| \|Q_n(z)\| < R(E)e^s \frac{r(E)}{R(E)e^s} = r(E)$$ and thus $P_n(E) \subset \mathbb{D}(0,r(E)) \subset E$. Clearly, this implies that $E \subset \operatorname{int} (\mathcal{K}(P_n))$, which completes the proof of the theorem. We can now use Theorem \[thmExplicit2\] to obtain a precise estimate for the rate of approximation. Recall from the introduction that $s_n(E)$, $n \in \mathbb{N}$, is defined to be the infimum of $s>0$ for which there exists a polynomial $p_n$ of degree $n$ such that $$E \subset \mathcal{K}(p_n) \subset E_s,$$ where $E_s:= E \cup \{z \in \Omega : g_\Omega(z,\infty) \leq s\}$. Solving for $s$ in Theorem \[thmExplicit2\], one easily obtains the existence of a constant $c'=c'(E)$ such that $$s_n(E) \leq c' \left( \frac{\log{n}}{\sqrt{n}} + \log{\frac{\delta_n(E)}{\operatorname{cap}(E)}} \right) \qquad (n \geq 1).$$ It follows from this and Theorem \[thmPRI2\] that there is a constant $c=c(E)$ depending only on $E$ such that $$s_n(E) \leq c \, \frac{\log{n}}{\sqrt{n}} \qquad (n \geq 1),$$ thereby proving Theorem \[mainthm2\]. Hilbert’s Lemniscate Theorem {#sec5} ============================ In this section, we give a proof of Theorem \[mainthm1\] using a classical result of Hilbert on the approximation of planar sets by polynomial lemniscates. We first present a proof of this latter result based on Fekete polynomials. \[Hilbert\] Let $E \subset \mathbb{C}$ be a compact set with connected complement $\Omega$, and let $U$ be an open neighborhood of $E$. Then there exists a polynomial $Q$ such that $$\|Q(z)\| \leq 1 \qquad (z \in E)$$ and $$\|Q(z)\|>1 \qquad (z \in \mathbb{C} \setminus U).$$ Again, we can assume that $\operatorname{cap}(E)>0$. Define $$A:= \inf_{z \in {\widehat{\mathbb{C}}}\setminus U} g_\Omega(z,\infty)$$ and $$B:= \sup_{z \in {\widehat{\mathbb{C}}}\setminus U} \tau_\Omega(z,\infty),$$ so that $A>0$ and $B<\infty$. By Theorem \[thmBernstein\], if $q$ is a Fekete polynomial for $E$ of degree $n$, then $$\left( \frac{\|q(z)\|}{\\|q\\|_E}\right)^{1/n} \geq e^{A} \left( \frac{\operatorname{cap}(E)}{\delta_n(E)} \right)^{B} \qquad (z \in \mathbb{C} \setminus U).$$ Here we used the fact that $\delta_n(E) \geq \operatorname{cap}(E)$ for all $n$, cf. Theorem \[thmFekSze\]. Finally, since $\delta_n(E) \to \operatorname{cap}(E)$ as $n \to \infty$, the right-hand side will exceed $1$ for all sufficiently large $n$, and setting $Q:=q/\\|q\\|_E$ then gives the desired polynomial $Q$. We now use Theorem \[Hilbert\] to give another proof of Theorem \[mainthm1\]. Again, we can assume without loss of generality that $0$ belongs to $E$ and that it is an interior point. Let $U$ be a bounded neighborhood of $E$, and let $Q$ be the polynomial as in Hilbert’s Theorem \[Hilbert\]. Dividing $Q$ by a constant slightly larger than one if necessary, we can assume that $\\|Q\\|_E < 1$. Let $R,r>0$ such that $\mathbb{D}(0,r) \subset E$ and $U \subset \mathbb{D}(0,R)$. For $k \geq 1$, define the polynomial $P_k(z):=zQ^k(z)$. Here $Q^k$ denotes the $k$-th power of the polynomial $Q$. Then for $k$ sufficiently large, we have $$P_k(E) \subset \mathbb{D}(0,r) \subset E$$ and $$P_k\left({\widehat{\mathbb{C}}}\setminus U \right) \subset {\widehat{\mathbb{C}}}\setminus \mathbb{D}(0,R) \subset {\widehat{\mathbb{C}}}\setminus U,$$ since $P_k \to 0$ uniformly on $E$ and $P_k \to \infty$ uniformly on ${\widehat{\mathbb{C}}}\setminus U$. This implies that $E \subset \mathcal{K}(P_k)$. Also, the inequality $$\|P_k(z)\| \geq \|z\| \min_{{\widehat{\mathbb{C}}}\setminus U}\|Q\| \qquad (z \in {\widehat{\mathbb{C}}}\setminus U),$$ where $\displaystyle{\min_{{\widehat{\mathbb{C}}}\setminus U}\|Q\|>1}$, shows that the iterates of $P_k$ tend to $\infty$ on ${\widehat{\mathbb{C}}}\setminus U$, so that ${\widehat{\mathbb{C}}}\setminus U \subset {\widehat{\mathbb{C}}}\setminus \mathcal{K}(P_k)$, i.e. $\mathcal{K}(P_k) \subset U$. Since $U$ can be made arbitrarily close to $E$, the result follows. Totally approximable sets {#sec6} ========================= In this section, we prove Theorem \[mainthm3\], which states that a nonempty proper subset $S$ of the complex plane $\mathbb{C}$ is totally approximable by polynomial filled Julia sets if and only if $S$ is bounded and $\mathbb{C} \setminus \operatorname{int}(S)$ is connected. First, we need an approximation result. \[TotApproxProp\] Let $S$ be a nonempty proper subset of the complex plane $\mathbb{C}$. If $S$ is bounded and $\mathbb{C} \setminus \operatorname{int}(S)$ is connected, then $S$ is totally approximable by a collection $\{K^\epsilon\}$ of nonempty compact sets with connected complement. The proof of Proposition \[TotApproxProp\] uses the following classical result from plane topology, known as Janiszewski’s Theorem. \[LemmaJan\] Let $A$ and $B$ be closed sets in ${\widehat{\mathbb{C}}}$ such that $A \cap B$ is connected. If $A$ and $B$ both have connected complement, then $A \cup B$ also has connected complement. See [@POM Theorem 1.9]. We can now proceed with the proof of Proposition \[TotApproxProp\]. For a set $A$ and $\epsilon>0$, we denote by $U_\epsilon(A)$ the open $\epsilon$-neighborhood of the set $A$. For $\epsilon>0$, let $G^\epsilon$ be a grid of squares in $\mathbb{C}$ of sidelength $\epsilon/2$. Let $H^\epsilon$ be the union of the closed squares in $G^\epsilon$ which intersect $\partial S$. Note that since $S$ is bounded, the set $H^\epsilon$ consists of finitely many closed squares and is therefore compact. Now, let $F^\epsilon$ be defined by taking each square $Q$ in $H^\epsilon$ and replacing it by a smaller square ${\widetilde}{Q}$ with the same center of sidelength $\epsilon/4$. Then it is easy to see that the complement of $F^\epsilon$ is connected. Now, define a collection of compact sets $\{K^\epsilon\}$ by $$K^\epsilon:= E^\epsilon \cup F^\epsilon,$$ where $E^\epsilon:= \mathbb{C} \setminus U_\epsilon(\mathbb{C} \setminus \operatorname{int}(S))$. Note that the sets $E^\epsilon$ and $F^\epsilon$ are disjoint, since $F^\epsilon$ is contained in the $(\epsilon/\sqrt{2})$-neighborhood of $\partial S$. See Figure 2. \[fig1\] ![On the left, a nonempty bounded subset $S$ of the plane whose interior has connected complement. On the right, the corresponding approximating set $K^\epsilon$. ](fig1-eps-converted-to.pdf "fig:"){width=".6\linewidth"} \[f:fig1\] We claim that for every $\epsilon>0$, the complement of $K^\epsilon$ is connected. Indeed, first note that the complement of $E^\epsilon$ is connected. To see this, note that $$\mathbb{C} \setminus E^\epsilon = \bigcup_{z \notin \operatorname{int}(S)} \mathbb{D}(z,\epsilon),$$ hence each component of $\mathbb{C} \setminus E^\epsilon$ contains a disk $\mathbb{D}(z,\epsilon)$ for some $z \notin \operatorname{int}(S)$ and therefore must intersect $\mathbb{C} \setminus \operatorname{int}(S)$. Since $\mathbb{C} \setminus \operatorname{int}(S)$ is connected and $\mathbb{C} \setminus E^\epsilon \supset \mathbb{C} \setminus \operatorname{int}(S)$, it follows that each component of $\mathbb{C} \setminus E^\epsilon$ has to contain $\mathbb{C} \setminus \operatorname{int}(S)$. Thus there can only be one component; in other words, the set $\mathbb{C} \setminus E^\epsilon$ is connected. Finally, since $F^\epsilon$ also has connected complement and $E^\epsilon, F^\epsilon$ are disjoint, we can apply Lemma \[LemmaJan\] to deduce that $K^\epsilon$ also has connected complement. Next, we claim that $S$ is totally approximable by the sets $K^\epsilon$. It suffices to prove that $d(\partial S, \partial K^\epsilon) \to 0$ and $d(S,K^\epsilon) \to 0$ as $\epsilon \to 0$. First, assume for a contradiction that $d(\partial S, \partial K^{\epsilon_n})> \delta$ for all $n$, for some $\delta>0$ and some sequence $\epsilon_n$ strictly decreasing to $0$. Then for each $n$, either $\partial S \nsubseteq U_\delta(\partial K^{\epsilon_n})$ or $\partial K^{\epsilon_n} \nsubseteq U_\delta(\partial S)$. However, note that if $z \in \partial S$, then $z$ belongs to a square $Q$ in $H^\epsilon$, so that the distance between $z$ and the boundary of the corresponding square ${\widetilde}{Q}$ in $F^\epsilon$ is less than $\epsilon/(4\sqrt{2})$, hence $\partial S \subset U_\epsilon(\partial K^\epsilon)$, for all $\epsilon>0$. We can therefore assume that the second case, $\partial K^{\epsilon_n} \nsubseteq U_\delta(\partial S)$, holds for all $n$. Now, recall that $F^\epsilon$ is contained in the $(\epsilon/\sqrt{2})$-neighborhood of $\partial S$. It follows that for all sufficiently large $n$, we have $\partial E^{\epsilon_n} \nsubseteq U_\delta(\partial S)$. Then for each such $n$, there is a point $z_n \in \partial E^{\epsilon_n} \subset \operatorname{int}(S)$ such that the disk $\mathbb{D}(z_n,\delta)$ does not intersect the boundary of $S$, hence must be contained in $\operatorname{int}(S)$. Passing to a subsequence if necessary, we can assume that $z_n \to z_0 \in \overline{S}$. It is easy to see that $z_0$ necessarily belongs to $\partial S$, since the sets $E^{\epsilon_n}$ form a compact exhaustion of $\operatorname{int}(S)$ with $E^{\epsilon_n} \subset \operatorname{int}(E^{\epsilon_{n+1}})$ for all $n$. This contradicts the fact that for sufficiently large $n$, the disk $\mathbb{D}(z_0,\delta/2)$ is contained in the disk $\mathbb{D}(z_n,\delta)$, a subset of $\operatorname{int}(S)$. Therefore $d(\partial S, \partial K^\epsilon) \to 0$ as $\epsilon \to 0$. It remains to prove that $d(S,K^\epsilon) \to 0$ as $\epsilon \to 0$. Again, assume for a contradiction that $d(S,K^{\epsilon_n})>\delta$ for all $n$, for some $\delta>0$ and some sequence $\epsilon_n$ strictly decreasing to $0$. Since $E^\epsilon \subset \operatorname{int}(S)$ and $F^\epsilon$ is contained in the $(\epsilon/\sqrt{2})$-neighborhood of $\partial S$, we have $K^\epsilon \subset U_\epsilon(S)$, for all $\epsilon>0$. We can therefore assume that for each $n$, $S \nsubseteq U_\delta(K^{\epsilon_n})$. Then for each $n$, there exists $z_n \in S$ such that the disk $\mathbb{D}(z_n,\delta)$ does not intersect $K^{\epsilon_n}$. Assume without loss of generality that $z_n \to z_0 \in \overline{S}$. Then for all sufficiently large $n$, the disk $\mathbb{D}(z_0,\delta/2)$ does not intersect $K^{\epsilon_n}$. If $z_0 \in \partial S$, then for sufficiently large $n$, the disk $\mathbb{D}(z_0,\delta/2)$ contains a square in $F^{\epsilon_n}$, and we get a contradiction. If $z_0 \in \operatorname{int}(S)$, we get a contradiction with the fact that the sets $E^{\epsilon_n}$ form a compact exhaustion of $\operatorname{int}(S)$. This completes the proof of the proposition. Now, recall from Theorem \[mainthm1\] that any nonempty compact set with connected complement is totally approximable by polynomial filled Julia sets. Combining this with Proposition \[TotApproxProp\], we get that any nonempty bounded set whose interior has connected complement is also totally approximable by polynomial filled Julia sets. This proves the converse implication in Theorem \[mainthm3\]. For the proof of the other implication, we need the following elementary lemmas. \[lem1\] For any set $A \subset \mathbb{C}$, we have $$U_\epsilon(A) = A \cup U_\epsilon(\partial A).$$ Clearly, we have $A \subset U_\epsilon(A)$. Also, if $z \in U_\epsilon(\partial A)$, then $\mathbb{D}(z,\epsilon) \cap \partial A \neq \emptyset$, so that $\mathbb{D}(z,\epsilon) \cap A \neq \emptyset$, from which we deduce that $z \in U_\epsilon(A)$. This shows that $A \cup U_\epsilon(\partial A) \subset U_\epsilon(A)$. For the other inclusion, if $z \in U_\epsilon(A)$ but $z \notin A$, then the disk $\mathbb{D}(z,\epsilon)$ intersects both $A$ and $\mathbb{C} \setminus A$. By connectedness, it must intersect $\partial A$, so that $z \in U_\epsilon(\partial A)$. \[lem2\] Let $A,B$ be two nonempty proper subsets of the plane with $d(A,B)<\epsilon$ and $d(\partial A, \partial B) < \epsilon$. Then $d(\mathbb{C} \setminus A, \mathbb{C} \setminus B) < 2\epsilon$. We first show that $\mathbb{C} \setminus A \subset U_{2\epsilon}(\mathbb{C} \setminus B)$. Let $z \in \mathbb{C} \setminus A$. If $z \notin U_\epsilon(A)$, then $z \in \mathbb{C} \setminus B$, since $B \subset U_\epsilon(A)$. On the other hand, if $z \in U_\epsilon(A)$, then by Lemma \[lem1\], we have that $z \in U_\epsilon(\partial A) \subset U_{2\epsilon}(\partial B)$. It follows that $\mathbb{D}(z,2\epsilon) \cap \partial B \neq \emptyset$, so that $\mathbb{D}(z,2\epsilon) \cap (\mathbb{C} \setminus B) \neq \emptyset$, from which we deduce that $z \in U_{2\epsilon}(\mathbb{C} \setminus B)$. In both cases, we get that $z \in U_{2 \epsilon}(\mathbb{C} \setminus B)$. This proves that $\mathbb{C} \setminus A \subset U_{2\epsilon}(\mathbb{C} \setminus B)$. The same argument with $A$ and $B$ interchanged yields the inclusion $\mathbb{C} \setminus B \subset U_{2\epsilon}(\mathbb{C} \setminus A)$, hence $d(\mathbb{C} \setminus A, \mathbb{C} \setminus B) < 2\epsilon$, as required. We can now proceed with the proof of Theorem \[mainthm3\]. Let $S$ be a nonempty proper subset of $\mathbb{C}$. As previously mentioned, if $S$ is bounded and $\mathbb{C} \setminus \operatorname{int}(S)$ is connected, then by Proposition \[TotApproxProp\] and Theorem \[mainthm1\], the set $S$ is totally approximable by polynomial filled Julia sets. Conversely, assume that $S$ is totally approximable by polynomial filled Julia sets. Then for any $\epsilon>0$, there exists a polynomial $P_\epsilon$ such that $d(S,\mathcal{K}(P_\epsilon))<\epsilon$ and $d(\partial S, \mathcal{J}(P_\epsilon)) < \epsilon$. Clearly, this implies that $S$ is bounded, since for each $\epsilon>0$, the set $\mathcal{K}(P_\epsilon)$ is compact. It remains to prove that $\mathbb{C} \setminus \operatorname{int}(S)$ is connected. Assume for a contradiction that $\mathbb{C} \setminus \operatorname{int}(S) = E \cup F$, where $E$ and $F$ are disjoint nonempty closed subsets of $\mathbb{C}$. Since $S$ is bounded, one of the sets $E,F$ must be unbounded, say $E$, while the other set, $F$, is bounded. Now, fix a point $z_0 \in F$, and let $\delta>0$ sufficiently small so that $\overline{\mathbb{D}}(z_0,2\delta) \subset \mathbb{C} \setminus E$. Since $F \cup \overline{\mathbb{D}}(z_0,2\delta)$ is a compact subset of the open set $\mathbb{C} \setminus E$, there is a bounded open set $V$ with $F \cup \overline{\mathbb{D}}(z_0,2\delta) \subset V \subset \overline{V} \subset \mathbb{C} \setminus E$. In particular, the boundary of $V$ is contained in $\mathbb{C} \setminus (E \cup F) = \operatorname{int}(S)$. Now, let $\epsilon>0$ be smaller than both $\delta$ and half the distance between $\partial V$ and $E \cup F$. Since $z_0 \in F \subset \mathbb{C} \setminus \operatorname{int}(S)$, we have that either $z_0 \in \partial S$ or $z_0 \in \mathbb{C} \setminus S$. In both cases, since $\partial S \subset U_\epsilon(\mathcal{J}(P_\epsilon))$ and $\mathbb{C} \setminus S \subset U_{2\epsilon}(\mathbb{C} \setminus \mathcal{K}(P_\epsilon))$ by Lemma \[lem2\], we get that the disk $\mathbb{D}(z_0,2\epsilon)$ contains a point $w_0 \in \mathbb{C} \setminus \mathcal{K}(P_\epsilon)$. This is enough to obtain a contradiction. Indeed, since $w_0 \in \mathbb{D}(z_0,2\delta) \subset V$ and $w_0 \in \mathbb{C} \setminus \mathcal{K}(P_\epsilon)$, the connectedness of $\mathbb{C} \setminus \mathcal{K}(P_\epsilon)$ (see [@MIL Lemma 9.4]) implies that $\partial V$ must have non-empty intersection with $\mathbb{C} \setminus \mathcal{K}(P_\epsilon)$. Since $\mathbb{C} \setminus \mathcal{K}(P_\epsilon) \subset U_{2\epsilon}(\mathbb{C} \setminus S)$, again by Lemma \[lem2\], it follows that the distance between $\partial V$ and $\mathbb{C} \setminus \operatorname{int}(S) = E \cup F$ is less than $2 \epsilon$, a contradiction. Therefore, $\mathbb{C} \setminus \operatorname{int}(S)$ is connected, as required. Approximation of connected sets {#sec7} =============================== In this section, we give a proof of Theorem \[t:connected\] by showing that our approximation method gives quasicircle Julia sets if the original set $E$ is connected. \[theoconnected\] Let $E$ be a connected compact set with connected complement, and assume as before that $0 \in \operatorname{int}(E)$. Then for any $s>0$ and $n \in \mathbb{N}$ such that $E \subset \operatorname{int}(\mathcal{K}(P_{n,s}))$, where $P_{n,s}$ is as defined in Theorem \[thmExplicit\], the Julia set $\mathcal{J}(P_{n,s})$ is a Jordan curve. Moreover, the polynomial $P_{n,s}$ is hyperbolic and $\mathcal{J}(P_{n,s})$ is a quasicircle. First, to prove that $\mathcal{J}(P_{n,s})$ is a Jordan curve, by [@CAR Theorem VI.5.3] it suffices to show that the Fatou set ${\widehat{\mathbb{C}}}\setminus \mathcal{J}(P_{n,s})$ consists of two completely invariant components. Since $E$ is connected and is contained in the Fatou set, it must be contained in a single Fatou component, say $A$. For polynomials, the basin of attraction of $\infty$ is completely invariant, so it suffices to show that $A$ is also completely invariant. Since $A$ contains the fixed point $0$, we have $P_{n,s}(A) \subset A$. If $P_{n,s}^{-1}(A)$ is not contained in $A$, then there exists a Fatou component distinct from $A$, say $A^{\prime}$, which is mapped onto $A$ by $P_{n,s}$. This is impossible, since all the zeros of $P_{n,s}$ belong to $E \subset A$. Therefore, $A$ is completely invariant and the Julia set is a Jordan curve. Now, since $A$ is completely invariant, all the critical points of $P_{n,s}$ belong to $A$, by [@CAR Theorem V.1.3]. It then follows from [@CAR Theorem V.2.2]) that $P_{n,s}$ is hyperbolic. Finally, by [@CAR Theorem VI.2.1], the Julia set $\mathcal{J}(P_{n,s})$ is not only a Jordan curve, but also a quasicircle. Numerical examples {#sec8} ================== In this section, we illustrate the method of Section \[sec3\] with some numerical examples. As shown in Theorem \[thmExplicit2\], the computation of Julia sets approximating a given compact set $E$ requires a numerical method for the computation of Fekete points for $E$. Unfortunately, the numerical computation of high-degree Fekete points is a hard large-scale optimization problem. Indeed, as far as we know, even for simple compact sets such as triangles, Fekete points have been computed only up to relatively small degrees, see e.g. [@TAY], [@PAS] and the references therein. In order to circumvent this issue, alternative numerical methods have been introduced for the computation of approximate Fekete points : see e.g. [@SOM] for a method based on QR factorization of Vandermonde matrices with column pivoting, and [@BOS2] for a discussion of the theoretical aspects of this method. However, for slightly complicated sets, we were not able to use the algorithm of [@SOM] to compute Fekete points of high degree, due to rank-deficiency and ill-conditioning issues. For this reason, we decided to replace Fekete points by other equidistributed points, Leja points, which are much more convenient from a computational point of view. This sequence of points, introduced by Leja [@LEJ], is defined inductively as follows. Choose an arbitrary point $z_1 \in \partial E$ and for $n \geq 2$, choose $z_n \in \partial E$ by maximizing the product $$\prod_{j=1}^{n-1}\|z-z_j\|$$ for $z \in E$, i.e. $$\prod_{j=1}^{n-1}\|z_n-z_j\| = \max_{z \in E} \prod_{j=1}^{n-1}\|z-z_j\|.$$ It is well-known that Leja points are easier to compute numerically than Fekete points, mainly because they are defined inductively by a univariate optimization problem. Furthermore, this optimization problem can be made discrete by replacing the compact set $E$ by a sufficiently refined discretization. We shall prove below that our approximation theorem \[thmExplicit2\] remains valid if Fekete points are replaced by Leja points, with small modifications. First, we need some lemmas. As previously mentioned, Leja points are also equidistributed with respect to harmonic measure on $\partial E$, in the sense of Proposition \[weakconvergence\]. In fact, Pritsker’s Theorem \[thmPRI\] also holds for Leja points, with the same constant. If $(z_n)$ is a sequence of Leja points for $E$, then $$\label{eqprit} \left\| \frac{1}{n} \sum_{j=1}^n \log{\|z-z_j\|} - p_\nu(z) \right\| \leq C(E) \frac{\log n}{\sqrt{n}} \qquad \left(z \in \Omega, \, g_\Omega(z,\infty)>\frac{1}{n}\right)$$ where $C(E)$ is Pristker’s constant as in Definition \[defPRI\]. See [@PRI Corollary 2.3]. We will also need an analogue of Theorem \[thmFekSze\] for Leja points. \[thmLeja\] For $n \in \mathbb{N}$, define $$a_n:= \prod_{j=1}^{n}\|z_{n+1}-z_j\|$$ Then $a_n^{1/n} \geq \operatorname{cap}(E)$ for each $n \in \mathbb{N}$, and $a_n^{1/n} \to \operatorname{cap}(E)$ as $n \to \infty$. See [@LEJ Lemme 1]. We can now prove the following variant of Theorem \[thmExplicit2\] for Leja points. As before, we assume that $E \subset \mathbb{C}$ is a uniformly perfect compact set with connected complement $\Omega$, and that $0 \in E$ is an interior point. Recall then that $r(E)$ and $R(E)$ are defined by $$r(E):= \operatorname{dist}(0,\partial E)$$ and $$R(E):= \sup_{w \in E} \|w\|.$$ \[thmExplicitLeja\] Let $s>0$, and suppose that $n$ is sufficiently large so that the following conditions hold : 1. $\displaystyle \frac{1}{n} \leq s$ 2. $\displaystyle C(E) \frac{\log n}{\sqrt{n}} \leq \frac{s}{4}$ 3. $\displaystyle e^{ns/4} \geq \frac{R(E)e^s}{r(E)}$ 4. $\displaystyle \left(\frac{R(E)e^s}{r(E)}\right)^{1/n} \frac{a_n^{1/n}}{\operatorname{cap}(E)} \leq e^{s/2}$. Then the polynomial $${\widetilde}{P}_n(z)={\widetilde}{P}_{n,s}(z):=z\frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \prod_{j=1}^n (z-z_j)$$ satisfies $$E \subset \operatorname{int}(\mathcal{K}({\widetilde}{P}_n)) \subset E_s,$$ where $E_s:= E \cup \{z \in \Omega : g_\Omega(z,\infty) \leq s\}$. The only difference between the assumptions of Theorem \[thmExplicitLeja\] and those of Theorem \[thmExplicit2\] is condition (iv), where $\delta_n(E)$ is replaced by $a_n^{1/n}$. The proof is exactly the same as in Theorem \[thmExplicit2\] except that in inequality (\[eqq\]), the use of Theorem \[thmFek\] is replaced by the inequality $$\prod_{j=1}^n \|z-z_j\|\leq \prod_{j=1}^n \|z_{n+1}-z_j\|=a_n \qquad (z \in E),$$ which follows directly from the definition of Leja points. We mention that using Leja points instead of Fekete points does not yield a better rate of approximation in Theorem \[mainthm2\]. The main improvement, however, is that the polynomials ${\widetilde}{P}_n$ can be computed numerically in a reasonable amount of time, even for high values of $n$, thereby giving an efficient algorithm to approximate a given shape by a polynomial Julia set. More precisely, let $E$ be a uniformly perfect compact set with connected complement, and assume that $0 \in \operatorname{int}(E)$. Take $s>0$ small and $n \in \mathbb{N}$ large. The algorithm consists of the following steps. - Compute $n$ Leja points for $E$ using a sufficiently refined discretization of the set. - Compute the polynomial $${\widetilde}{P}_n(z)=z\frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \prod_{j=1}^n (z-z_j),$$ where $\operatorname{cap}(E)$ is approximated by the quantity $a_n^{1/n}$, in view of Lemma \[thmLeja\]. - Plot the filled Julia set of the polynomial ${\widetilde}{P}_n$. The first two steps are easily carried out with <span style="font-variant:small-caps;">matlab</span>, for instance, even for large values of $n$. As for plotting the filled Julia set, it can be done in the obvious natural way, that is, dividing a square containing the set $E$ into a sufficiently large number of small squares (the pixels), iterating the center of each pixel a fixed number of times, and then coloring the pixel depending on whether the absolute value of the last iterate is big or not. This works relatively well in general, at least when $n$ is not too large, say $n \leq 2000$. We also tried a more precise method for plotting Julia sets based on distance estimation (see e.g. [@DRA]), but the difference in quality of the images obtained was negligible compared to the longer time required for the computations. We also mention that another difficulty in the implementation of the algorithm is to find how small $s$ can be taken, given $n$. By Theorem \[mainthm2\], we know that the ratio between the best possible $s$ and $\log{n}/\sqrt{n}$ is bounded. However, we observed that using $s=\log{n}/\sqrt{n}$ generally gives poor results compared to smaller values, say $s=1/n$. We now present several numerical examples to illustrate the method. We thank Siyuan Li, Xiao Li and Ryan Pachauri, three undergraduate students at the University of Washington who produced some of the examples as part of a research project supervised by the second author for the Washington Experimental Mathematics Lab (WXML). [The rabbit Julia set.]{} Our first example illustrates Theorem \[theoconnected\], which says that the Julia sets of the approximating polynomials are Jordan curves if the original set $E$ is connected. Figure 3 is a representation of the filled Julia set of the polynomial $${\widetilde}{P}_n(z)=z\frac{e^{-ns/2}}{\operatorname{cap}(E)^n} \prod_{j=1}^n (z-z_j),$$ with $n=700$ and $s=1/700$, as well as the original set $E$ (boundary in black). The Julia set was plotted using a resolution of $5000 \times 5000$, and the computations took approximately $120$ seconds. ![A filled Julia set in the shape of a rabbit.](rabbit2124_120s_1.pdf){width=".7\textwidth"} [The Batman Julia set.]{} Here is another example of a Jordan curve Julia set approximating a connected shape. Figure 4 is a representation of the filled Julia set of the polynomial ${\widetilde}{P}_n$ with $n=700$ and $s=1/700$, as well as the original set $E$ (boundary in black). The Julia set was plotted using a resolution of $6000 \times 6000$, and the computations took approximately $187$ seconds. ![A filled Julia set in the shape of Batman.](batman1_187s.pdf){width=".8\textwidth"} [The KLMY Julia set.]{} The following is an example of a disconnected Julia set representing the initials KLMY. Figure 5 is a representation of the filled Julia set of the polynomial ${\widetilde}{P}_n$ with $n=2000$ and $s=1/2000$. The Julia set was plotted using a resolution of $5000 \times 5000$, and the computations took approximately $340$ seconds. ![A filled Julia set in the shape of the initials KLMY.](klmy_16023_340s.pdf){width=".8\textwidth"} We note that the filled Julia set has infinitely many connected components, although only four of them are visible. This is most likely due to a lack of precision of the algorithm in <span style="font-variant:small-caps;">matlab</span>. In order to be able to see the smaller components, one can instead use the software <span style="font-variant:small-caps;">ultrafractal</span> to plot the Julia set. This is what we did in the following examples. [The fish-heart-diamond Julia set.]{} The following is an example of a disconnected Julia set representing the shapes of a fish, a heart and a diamond. Figure 6 is a representation of the filled Julia set of the polynomial ${\widetilde}{P}_n$ with $n=550$ and $s=1/550$, obtained with <span style="font-variant:small-caps;">ultrafractal</span>. ![A filled Julia set in the shape of a fish, a heart and a diamond.](Fractal2new.pdf){width=".7\textwidth"} Figure 7 is a zoomed portion of the boundary of the fish where one can see small distorted copies of the heart and the diamond. ![Zoomed portion of the boundary.](Fractal2newboundary.pdf){width=".7\textwidth"} [99]{} V. Andrievskii, On the Approximation of a Continuum by Lemniscates, J. Approx. Theory., 105 (2000), 292–304. I. Binder, N. Makarov and S. Smirnov, Harmonic measure and polynomial Julia sets, Duke Math. J., 117 (2003), 343–365. C.J. Bishop and K.M. Pilgrim, Dynamical dessins are dense, Rev. Mat. Iberoam., 31 (2015), 1033–1040. L. Bos and N. Levenberg, On the calculation of approximate Fekete points: the univariate case, Electron. Trans. Numer. Anal., 30 (2008), 377–397. L. Bos, N. Levenberg and S. Waldron, On the spacing of Fekete points for a sphere, ball or simplex, Indag. Math. (N.S.), 19(2008), 163–176. L. Carleson and T. W. Gamelin, Complex Dynamics, Springer-Verlag New York, New York, 1993. L. Carleson and P. W. Jones, On coefficient problems for univalent functions and conformal dimension, Duke Math. J., 66 (1992), 169–206. V. Drakopoulos, Comparing rendering methods for Julia sets, Journal of WSCG, 10 (2002), 155-–161. M. Fekete, Über den transfiniten Durchmesser ebener Punktrnengen, Math. Z., 32 (1930), 108–114. D. Hilbert, Über die Entwicklung einer beliebigen analytischen Funktion einer Variabeln in eine unendliche nach ganzen rationalen Funktionen fortschreitende Reihe, Göttinger Nachrichten, (1897), 63–70. Z. Laiyi, On the degree of convergence of lemniscates in finite connected domains, J. Approx. Theory, 131 (2004), 185–195. F. Leja, Sur certaines suites liées aux ensembles plans er leur application à la représentation conforme, Ann. Polon. Math. 4 (1957), 8–13. T. Kim, Quaternion julia set shape optimization, Computer Graphics Forum, 34 (2015), 157–176. K.A. Lindsey, Shapes of polynomial Julia sets, Ergodic Theory Dynam. Systems, 35 (2015), 1913–1924. J. Milnor, Dynamics in one complex variable, Princeton University Press, Princeton, 2006. R. Pasquetti and F. Rapetti, Spectral element methods on unstructured meshes: Comparisons and recent advances, J. Sci. Comput., 27 (2006), 377–387. C. Pommerenke, Univalent functions, Vandenhoeck & Ruprecht, Göttingen, 1975. I.E. Pritsker, Equidistribution of points via energy, Ark. Mat., 47 (2011), 149–173. T. Ransford, Potential theory in the complex plane, Cambridge University Press, Cambridge, 1995. E. Saff and V. Totik, Logarithmic potentials with external fields, Springer-Verlag, Berlin, 1997. A. Sommariva and M. Vianello, Computing approximate Fekete points by QR factorizations of Vandermonde matrices, Comput. Math. Appl., 57 (2009), 1324–1336. M.A. Taylor, B.A. Wingate, R.E. Vincent, An algorithm for computing Fekete points in the triangle, SIAM J. Numer. Anal., 38 (2000), 1707–1720. [^1]: First author supported by an NSF Mathematical Sciences Research Postdoctoral Fellowship [^2]: Second author supported by NSERC and NSF Grant DMS-1664807	{ "pile_set_name": "ArXiv" }
--- abstract: 'Transceivers operating in frequency division duplex experience a signal into the receiver due to the limited duplexer stop-band isolation. This signal in combination with the second-order nonlinearity of the receive mixer may lead to a with twice the transmit signal bandwidth. In direct conversion receivers, this nonlinear interference may cause a severe signal-to-interference-plus-noise ratio degradation of the wanted receive signal. This contribution presents a nonlinear Wiener model type adaptive filter for the cancellation of the interference in the digital . The included and DC-notch filter at the output of the proposed adaptive filter ensure that the provided replica includes the receiver front-end filtering. A second, robust version of the nonlinear algorithm is derived which provides numerical stability for highly correlated input signals which arise in e.g. LTE-A intra-band multi-cluster transmission scenarios. The performance of the proposed algorithms is evaluated by numerical simulations and by measurement data. [*Index terms —* second-order intermodulation, self-interference, adaptive filters, interference cancellation, LTE-A, 5G, RLS]{}' author: - bibliography: - 'references/global\_bibliography.bib' title: 'A Robust Nonlinear RLS Type Adaptive Filter for Second-Order-Intermodulation Distortion Cancellation in FDD LTE and 5G Direct Conversion Transceivers ' --- = \[draw, fill=blue!20, rectangle, minimum height=3em, minimum width=1em\] = \[draw, circle, node distance=1cm\] = \[coordinate\] = \[coordinate\] = \[pin edge=[to-,thin,black]{}\] Introduction ============ Modern transceivers are enhanced by digital signal processing to mitigate non-idealities in the analog front-end. One of the main reasons of receiver desensitization in transceivers is the limited duplexer isolation between the transmitter and the receiver which is around 50dB to 55dB [@Ericsson_1; @Vazny_1]. The resulting signal can be identified as the root cause of several receiver interferences. Especially in receivers multiple clock sources are needed to cover the different scenarios and band combinations. Due to cross-talk between the receivers on the chip and device nonlinearities, spurs appear in the receiver front-end. (img) at (0,0) ; If such a spur falls near the actual frequency, then the signal is down-converted into the where it causes a degradation of the wanted receive signal. The cancellation of this so called modulated spurs with adaptive filtering is demonstrated in [@Gebhard2016; @Kanumalli2016]. Another prominent interference caused by the signal and the second-order nonlinearity of the receiver is the . This second-order nonlinear distortion is caused by e.g. a coupling between the - and -ports in the I-, and Q-path of the IQ-mixer as indicated in Fig. \[fig:IMD2\_block\_diagram\] [@Razavi_1]. An interesting fact of this nonlinear interference is, that one part of the generated second-order intermodulation products always falls around zero-frequency independent of the -to- frequency offset (duplexing distance). In case of direct-conversion receiver architectures, this leads to a degradation of the wanted receive signal. The mathematical modeling in [@Gebhard2017; @Kiayani_1] shows that the interference contains the squared envelope of the equivalent signal. The resulting interference has twice the signal bandwidth and contains a DC due to the envelope-squaring. In the receiver front-end, the overall DC arising from a number of sources is canceled by a mixed-signal cancellation to prevent the from saturation. In the digital domain, the signal is filtered by a to reduce its bandwidth to the signal bandwidth. In the existing literature, the authors of [@Lederer_1; @Frotzscher_3; @Kahrizi_1] discussed adaptive type interference cancellation algorithms for frequency-flat duplexer stop-bands. In [@Frotzscher_2] a Volterra kernel based approach for frequency-selective Tx-Rx responses is proposed. The authors in [@Kiayani_1] presented a two-step approach for the cancellation and considered a static 3rd-order nonlinearity and IQ-imbalance in the transmit mixer. In [@Gheidi_1] a transceiver is considered where the transmit signal of both transmitters leaks through a diplexer into one unpaired receiver. The diplexer stop-band is modeled as a first-order system which states a nearly frequency-flat response. The authors incorporated a fourth-order nonlinearity without memory into the estimation process, which results in an problem with four unknown coefficients. This contribution presents a nonlinear Wiener model type adaptive filter (IM2RLS) with exponential forgetting factor which is suitable for highly frequency selective duplexer stop-band frequency responses like indicated in Fig. \[fig:spectrum\]. It targets the digital cancellation for high performance cellular base stations and mobile phones. The Wiener model uses a static nonlinearity at the output of the adaptive filter which has the advantage that less coefficients are needed in the estimation process compared to a Volterra kernel based adaptive filter [@mathews2000polynomial]. An additional version of the proposed algorithm is presented which enhances the algorithm by a DC-notch filter to cancel the DC in the interference replica. This is needed because direct-conversion receivers employ a DC cancellation to suppress the DC in order to prevent the from saturation. The DC in the received signal is time-variant and has many sources like e.g. LO-LO self mixing [@Razavi_1], and therefore must not be related explicitly to the DC which is generated by the interference. Consequently, the interference related DC is removed from the received signal which complicates the replica estimation. This DC removal is considered in [@Frotzscher_2; @Gebhard2017], and neglected in [@Lederer_1; @Lederer_3; @Frotzscher_3; @Kiayani_1]. The derived IM2RLS with DC-notch filter is extended by a regularization which makes the algorithm applicable for highly correlated transmit signals where the autocorrelation matrix can be close to singular. A high correlation in the transmit signal can be due to oversampling which happens e.g. in the case of multi-cluster transmissions (introduced in 3GPP LTE-A Release 11) where only a part of the available are allocated. The presented IM2RLS algorithm is an extension to the nonlinear type adaptive filter derived in [@Gebhard2017] with improved steady-state cancellation and convergence speed. The structure of the presented work is as follows: Section \[sec:problem\_statement\] explains the characterization and demonstrates the degradation of the performance due to the interference. Section \[sec:system\_model\] provides a detailed interference model which motivates the proposed structure of the nonlinear adaptive filter. In Section \[sec:IM2RLS\], the IM2RLS algorithm is derived and the impact of adding a DC-notch filter to the algorithm is evaluated. The R-IM2RLS alrorithm is derived in section \[sec:R-IM2RLS\] which is robust against highly correlated input signals as they occur in intra-band multi-cluster transmissions. Finally, in the sections \[sec:simulations\] and \[sec:measurements\], the performance of the R-IM2RLS algorithm is evaluated with simulations and measured data using RF components. Problem statement {#sec:problem_statement} ================= The receiver is characterized by using two cosine signals with the frequencies $f_1$ and $f_2$ of equal amplitude and the total power $P_{\text{in,2t}}$ at the input of the nonlinear mixer. The resulting total power generated at DC, $f_1+f_2$ and $f_2-f_1$ at the output of the mixer can be calculated by [@liu2009ip2], where $\text{IIP2}$ is the two-tone IIP2 value in dBm. Here, half of the total power falls to DC, and one quarter each to $f_1+f_2$ and $f_2-f_1$. To characterize the in a zero-IF receiver, the frequencies $f_1$ and $f_2$ are chosen such that $f_2-f_1$ falls within the bandwidth. Thereby the power at $f_2-f_1$ is measured and the is determined by . For modulated signals, the power is modulation dependent and further reduced by the . This is considered by a correction-factor which corrects the power calculated by the two-tone formula [@Walid_1; @Atalla_1]. Although the DC-, and channel-select filtering in the receiver reduces the interference power by 6dB in the two-tone signal case [@liu2009ip2], and by about 13.4dB [@Gebhard2017; @Walid_1; @Atalla_1] in the case of modulated signals, the left-over interference may lead to a severe degradation of the wanted signal in reference sensitivity cases [@3GPP_2]. Assuming a transmitter power of 23dBm at the antenna, and an average -to- duplexer isolation at the transmit frequency of 50dB, the signal power at the input of the receiver is . After amplification with the gain which is assumed as 20dB, the signal power increases to at the input of the nonlinear mixer. The two-tone value of typical mixers is between 50dBm and 70dBm [@Madadi_1; @Dufrene_1]. Assuming an of 60dBm, the resulting power with a full allocated LTE10 QPSK modulated transmission and the determined correction factor of is\ [@Gebhard2017]. In an LTE10 reference sensitivity case, the wanted signal power at the antenna can be as low as -97dBm [@3GPP_2]. The thermal noise power within 10MHz bandwidth is -104.5dBm and the assumed receiver is 4.5dB which results in a receiver noise floor at -100dBm. After amplification with 20dB gain, the wanted signal power is -77dBm and the noise floor at -80dBm corresponding to an of 3dB. The drops from 3dB to an of 2.27dB due to the interference assuming an of +60dBm. In case of an reduced IIP2 of 55dBm / 50dBm, the drops even further to 1dB / -1.4dB, respectively. Fig. \[fig:spectrum\] depicts the spectrum of the frequency selective equivalent signal $y_{\text{BB}}^{\text{TxL}}$ which generates the complex valued interference $y_{\text{BB}}^{\text{IMD2}}$ by a coupling between the RF-to-LO terminals of the I-, and Q-path mixer. The total received signal $y_{\text{BB}}^{\text{Tot}}$ contains the wanted signal $y_{\text{BB}}^{\text{Rx}}$ which is degraded by the interference and the noise. table\[x index =0, y index =3\] [plot\_data/spectrum\_full\_allocation\_IIP2\_50dBm.dat]{}; table\[x index =0, y index =1\] [plot\_data/spectrum\_full\_allocation\_IIP2\_50dBm.dat]{}; table\[x index =0, y index =2\] [plot\_data/spectrum\_full\_allocation.dat]{}; table\[x index =0, y index =4\] [plot\_data/spectrum\_full\_allocation\_IIP2\_50dBm.dat]{}; table\[x index =0, y index =6\] [plot\_data/spectrum\_full\_allocation\_IIP2\_50dBm.dat]{}; System model {#sec:system_model} ============ Interference Model ------------------- Based on the block diagram in Fig. \[fig:IMD2\_block\_diagram\] depicting an transceiver operating in mode, a detailed interference model is derived. The used mathematical operators $\left(.\right)^$, $\left(.\right)^T$, $\left(.\right)^H$, and $\ast$ denote the complex conjugate, transpose, Hermitian transpose, and convolution, respectively. The complex transmit signal is up-converted to the passband and amplified by the linearly assumed with gain $A_{\text{PA}}$ resulting in the transmit signal $$x_{\text{RF}}(t) = A_{\text{PA}}\Re\left\{x_{\text{BB}}(t)e^{j2\pi f_{\text{Tx}}t}\right\}.$$ This signal leaks through the duplexer stop-band impulse response $$h_{\text{RF}}^{\text{TxL}}(t)=2\Re\left\{h_{\text{BB}}^{\text{TxL}}(t) e^{j2\pi f_{\text{Tx}}t}\right\},$$ which is modeled by the equivalent duplexer impulse response $h_{\text{BB}}^{\text{TxL}}(t)$ into the receiver, thereby creating the signal $$\label{eq:y_RF_TxL} \begin{aligned} y_{\text{RF}}^{\text{TxL}}(t) &= x_{\text{RF}}(t) \ast h_{\text{RF}}^{\text{TxL}}(t)\\ &=A_{\text{PA}} \Re\left\{\left[x_{\text{BB}}(t) \ast h_{\text{BB}}^{\text{TxL}}(t)\right]e^{j2\pi f_{\text{Tx}}t}\right\}. \end{aligned}$$ The received signal at the output of the with gain $A_{\text{LNA}}$ $$\label{eq:y_Tot_after_LNA} y_{\text{RF,LNA}}^{\text{Tot}}(t)=A_{\text{LNA}}\left[y_{\text{RF}}^{\text{TxL}}(t)+y_{\text{RF}}^{\text{Rx}}(t)+v_{\text{RF}}(t)\right],$$ is composed by the amplified signal, the wanted Rx signal $y_{\text{RF}}^{\text{Rx}}(t)$ and the noise signal $v_{\text{RF}}(t)$. The output signal of the I-, and Q-path mixer is combined into the complex valued signal $y_{\text{RF,mixer}}^{\text{Tot}}(t)$ . It contains the wanted signal which is down-converted with the linear gain , and the second order interference with the mixer -to- terminal coupling coefficient $\alpha_2=\alpha_2^\text{I} + j \alpha_2^\text{Q}$. $$\label{eq:mix_out_sig} \begin{aligned} y_{\text{RF,mixer}}^{\text{Tot}}(t)&= y_{\text{RF,LNA}}^{\text{Tot}}(t)\alpha_1^{\text{I}} cos\left(2\pi f_{\text{Rx}}t\right) \\ & + y_{\text{RF,LNA}}^{\text{Tot}}(t) \left[\alpha_2^{\text{I}} y_{\text{RF,LNA}}^{\text{Tot}}(t)\right]\\ & - j y_{\text{RF,LNA}}^{\text{Tot}}(t)\alpha_1^{\text{Q}} sin\left(2\pi f_{\text{Rx}}t\right)\\ & + j y_{\text{RF,LNA}}^{\text{Tot}}(t) \left[\alpha_2^{\text{Q}} y_{\text{RF,LNA}}^{\text{Tot}}(t)\right] \\ & = y_{\text{RF,LNA}}^{\text{Tot}}(t)\alpha_1 e^{-j2\pi f_{\text{Rx}}t} + \alpha_2 \, y_{\text{RF,LNA}}^{\text{Tot}}(t)^2 \end{aligned}$$ Assuming a direct conversion receiver, and using the identity , the total mixer output signal by neglecting the signal content which falls outside the bandwidth becomes $$\label{eq:y_Tot_after_mixer} \begin{aligned} y_{\text{RF,mixer}}^{\text{Tot}}(t)&=\alpha_1\frac{A_{\text{LNA}}}{2} y_{\text{BB}}^{\text{Rx}}(t) +\alpha_1 \frac{A_{\text{LNA}}}{2} v_{\text{BB}}(t)\\ &+ \frac{\alpha_2}{2} \cdot \left(\left\|A_{\text{LNA}} A_{\text{PA}} \, x_{\text{BB}}(t)h_{\text{BB}}^{\text{TxL}}(t) \right\|^2 +\frac{1}{2}\left\|y_{\text{BB}}^{\text{Rx}}(t)\right\|^2 \right. \\ &\left. + \Re\left\{y_{\text{BB}}^{\text{Rx}}(t) v^_{\text{BB}}(t) \right\} +\frac{1}{2}\left\|v_{\text{BB}}(t)\right\|^2 \right). \end{aligned}$$ As $\left\|\alpha_2\right\|<<1$, the three last terms in may be neglected [@Kiayani_1; @Gebhard2017]. The total received discrete-time signal including the DC-cancellation and channel-select filtering becomes $$\label{eq:y_Tot_after_CSF} \begin{aligned} y_{\text{BB}}^{\text{Tot}}[n]&= \alpha_1 \frac{A_{\text{LNA}}}{2} y_{\text{BB}}^{\text{Rx}}[n] \ast \bar{h}_{\text{s}}[n] + \alpha_1 \frac{A_{\text{LNA}}}{2} v_{\text{BB}}[n] \ast \bar{h}_{\text{s}}[n]\\ &+ \underbrace{\frac{\alpha_2}{2}\left\| A_{\text{LNA}} A_{\text{PA}} x_{\text{BB}}[n]h_{\text{BB}}^{\text{TxL}}[n] \right\|^2 \ast \bar{h}_{\text{s}}[n]}_{y_{\text{BB}}^{\text{IMD2}}[n]}, \end{aligned}$$ where the DC-, and are combined in the impulse response . Here, is the impulse invariant , scaled and sampled version of the continuous-time duplexer impulse response $h_{\text{BB}}^{\text{TxL}}(t)$. Interference Replica Model -------------------------- For the adaptive filter development to cancel the interference in the digital , the interference model is rewritten to the form $$\label{eq:y_Tot_after_CSF_adaptive_filter} \begin{aligned} y_{\text{BB}}^{\text{Tot}}[n]&= \underbrace{\frac{\alpha_2^{\text{I}}}{2}\left\| A_{\text{LNA}} A_{\text{PA}} x_{\text{BB}}[n]h_{\text{BB}}^{\text{TxL}}[n] \right\|^2 \ast \bar{h}_{\text{s}}[n]}_{y_{\text{BB}}^{\text{IMD2,I}}[n]} \\ &+j \underbrace{\frac{\alpha_2^{\text{Q}}}{2}\left\| A_{\text{LNA}} A_{\text{PA}} x_{\text{BB}}[n]h_{\text{BB}}^{\text{TxL}}[n] \right\|^2 \ast \bar{h}_{\text{s}}[n]}_{y_{\text{BB}}^{\text{IMD2,Q}}[n]} + v_{\text{BB}}'[n] \end{aligned}$$ where the complex valued wanted signal and the noise signal are combined in $v_{\text{BB}}'[n]$. Assuming $\alpha_2^{\text{I}} > 0$, and approximating the duplexer impulse response $h_{\text{BB}}^{\text{TxL}}[n]$ by the impulse response vector ${\mathbf}{h}_{\text{BB}}^{\text{TxL}}$ of length $N_{{\mathbf}{w}}$, we can rewrite the model further to $$\label{eq:final_IMD2_model} \begin{aligned} y_{\text{BB}}^{\text{Tot}}[n]&= \left\|{\mathbf}{x}^T[n] {\mathbf}{h}_{\text{I}} \right\|^2 \ast \bar{h}_{\text{s}}[n] +j \left\|{\mathbf}{x}^T[n] {\mathbf}{h}_{\text{Q}} \right\|^2 \ast \bar{h}_{\text{s}}[n] + v_{\text{BB}}'[n] \\ &=y_{\text{BB}}^{\text{IMD2,I}}[n] + j \epsilon \, y_{\text{BB}}^{\text{IMD2,I}}[n] + v_{\text{BB}}'[n], \end{aligned}$$ where ${\mathbf}{h}_{\text{I}}$ and ${\mathbf}{h}_{\text{Q}}$ are incorporating ${\mathbf}{h}_{\text{BB}}^{\text{TxL}}$ and all scalar scaling factors in the I-, and Q-path respectively. The used vector ${\mathbf}{x}[n]$ is the complex valued tapped delay-line input signal vector , and the real valued scaling factor $\epsilon$ shows that the Q-path interference may be modeled as a scaled version of the interference. Motivated by the model we propose the I-path interference replica model $$\label{eq:IMD2_model} \begin{aligned} \hat{y}_{\text{AC,I}}[n]= \left\|{\mathbf}{x}^T[n] {\mathbf}{w}_{\text{I}}[n] \right\|^2 \ast \bar{h}_\text{s}[n], \end{aligned}$$ using the adaptive filter coefficient vector ${\mathbf}{w}_{\text{I}}[n]$. The index AC indicates the DC cancellation in the replica generation. The replica model comprises an adaptive Wiener model filter where the output signal is DC-, and channel-select filtered. The Q-path interference is generated by estimating the scaling parameter $\epsilon$ by a linear single-tap algorithm which uses the estimated I-path interference as reference input. This model is used to derive the adaptive filter structure shown in Fig. \[fig:IMD2\_block\_diagram\] to cancel the interference in the digital . For the case if $\alpha_2^{\text{I}} < 0$, the sign of the desired signal in the I-path $d_{\text{I}}$ and the replica signal of the adaptive filter need to be changed. Nonlinear Recursive Least-Squares algorithm {#sec:IM2RLS} =========================================== In this section, a nonlinear Wiener model type adaptive filter to estimate the channel-select filtered I-path interference is developed. In a first step the IM2RLS algorithm without DC-notch filter, which implies that the received signal contains the DC, is developed. Therefore, the replica model without DC cancellation $$\label{eq:I_patch_IMD2_replica} \begin{aligned} \hat{y}_{\text{I}}[n] &= \left\|{\mathbf}{x}^T[n] {\mathbf}{w}_{\text{I}}[n] \right\|^2 \ast h_\text{s}[n] \\ & = {\mathbf}{x}^T[n] {\mathbf}{w}_{\text{I}}[n] {\mathbf}{x}^H[n] {\mathbf}{w}_{\text{I}}^[n] \ast h_\text{s}[n] \end{aligned}$$ is used. The cost function up to the time index $n$ with the exponential forgetting factor is $$\label{eq:LS_cost_function_sum} \begin{aligned} J_{\text{LS}}[n] &= \sum^{n}_{i=0}\lambda^{n-i}\left\|d_{\text{I}}[i] - {\mathbf}{x}^T[i] {\mathbf}{w}_{\text{I}}[n] {\mathbf}{x}^H[i] {\mathbf}{w}_{\text{I}}^[n] \ast h_\text{s}[i]\right\|^2. \end{aligned}$$ This cost function is visualized in Fig. \[fig:LS\_cost\_function\_3d\_with\_DC\] for an example impulse response ${\mathbf}{h}_{\text{I}}=\left[1, 0.5\right]^T$ and $\lambda = 1$ where the estimated coefficients $w_{\text{I,0}}$ and $w_{\text{I,1}}$ are constrained to be real valued. Two equivalent global minimum points and a local maximum at the origin ${\mathbf}{w}_{\text{I}}={\mathbf}{0}$ can be observed. The two solutions , and minimize the cost function which can be explained with the absolute-squaring nature of the interference. Both solutions lead to the same replica signal. Assuming real valued impulse response coefficients $h_s[n]$, and observing that $d_{\text{I}}[i]$ is the desired signal in the I-path, and therefore real valued, the gradient of the cost function may be derived. The gradient of the cost function with respect to the conjugate coefficient vector ${\mathbf}{w}^_{\text{I}}$ using the Wirtinger calculus [@Brandwood_1; @Van_den_Bos_1; @mandic2009complex] becomes $$\label{eq:RLS1_gradient} \begin{aligned} \nabla_{{\mathbf}{w}^_{\text{I}}} J_{\text{LS}}&= \left[\frac{\partial J_{\text{LS}}[n]}{\partial {\mathbf}{w}^_{\text{I}}[n]}\right]^T\\ &=\sum^{n}_{i=0}\lambda^{n-i} \left[-2\, d_{\text{I}}[i]{{\mathbf}{x}}^T[i] {\mathbf}{w}_{\text{I}}[n] {{\mathbf}{x}}^[i] \ast h_s[i] \right.\\ &\left. \hspace{2cm} +2 \left({{\mathbf}{x}}^T[i] {\mathbf}{w}_{\text{I}}[n] {{\mathbf}{x}}^[i] \ast h_s[i]\right) \right.\\ & \left. \hspace{2cm} \cdot \left({{\mathbf}{x}}^H[i] {\mathbf}{w}_{\text{I}}^[n] {{\mathbf}{x}}^T[i] \ast h_s[i]\right) {\mathbf}{w}_{\text{I}}[n] \right]. \end{aligned}$$ By setting the gradient to zero, the Wiener Filter equation is obtained by $$\label{eq:RLS1_wiener_equation} \begin{aligned} \tilde{{\mathbf}{R}}\left({\mathbf}{w}_{\text{I}}[n]\right){\mathbf}{w}_{\text{I}}[n]&=\tilde{{\mathbf}{r}}\left({\mathbf}{w}_{\text{I}}[n]\right), \end{aligned}$$ where it can be observed that the autocorrelation matrix $\tilde{{\mathbf}{R}}$ and the cross-correlation vector $\tilde{{\mathbf}{r}}$ are functions of the unknown coefficient vector ${\mathbf}{w}_{\text{I}}[n]$. In a slowly varying or nearly stationary system environment it can be assumed that ${\mathbf}{x}^T[i] {\mathbf}{w}[n]\approx {\mathbf}{x}^T[i] {\mathbf}{w}[i-1]$ when the index $i$ is close to $n$ [@Chen_CMA_1; @Chen_CMA_2]. If the index $i << n$, the approximation introduces an error which is however attenuated by the forgetting factor. Defining the new cost function $$\label{eq:LS_cost_function_sum_new} \begin{aligned} J'_{\text{LS}}[n] &= \sum^{n}_{i=0}\lambda^{n-i}\left\|d_{\text{I}}[i] - {\mathbf}{x}^T[i] {\mathbf}{w}_{\text{I}}[i-1] {\mathbf}{x}^H[i] {\mathbf}{w}_{\text{I}}^[n] \ast h_\text{s}[i]\right\|^2 \\ &= \sum^{n}_{i=0}\lambda^{n-i}\left\|d_{\text{I}}[i] - {\mathbf}{z}^T[i] {\mathbf}{w}_{\text{I}}^[n] \ast h_\text{s}[i]\right\|^2 \\ &=\sum^{n}_{i=0}\lambda^{n-i}\left\|e_{\text{I}}[i] \right\|^2 \end{aligned}$$ and introducing the new input vector , we can overcome this limitation. Following the traditional derivation [@sayed2003fundamentals], the IM2RLS algorithm to estimate the I-path interference in the digital becomes -: $$\label{eq:I_patch_IMD2_replica_final_equation_2} \begin{aligned} \hat{y}_{\text{I}}[n] = {\mathbf}{z}^T[n] {\mathbf}{w}_{\text{I}}^[n-1] \ast h_\text{s}[n] \end{aligned}$$ $$\label{eq:error} \begin{aligned} e_{\text{I}}[n] = d_{\text{I}}[n] - \hat{y}_{\text{I}}[n] \end{aligned}$$ $$\label{eq:kalman_gain_final} \begin{aligned} {\mathbf}{k}[n]=\frac{{\mathbf}{P}[n-1] {\mathbf}{z}_{\text{f}}[n] }{\lambda + {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{P}[n-1] {\mathbf}{z}_{\text{f}}[n]} \end{aligned}$$ $$\label{eq:inv_autocorr_matrix_final} \begin{aligned} {\mathbf}{P}[n]=\frac{1}{\lambda}\left[{\mathbf}{P}[n-1] -{\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{P}[n-1]\right] \end{aligned}$$ $$\label{eq:coeff_update_2_final} \begin{aligned} {\mathbf}{w}_{\text{I}}[n]&={\mathbf}{w}_{\text{I}}[n-1]+e_{\text{I}}[n] {\mathbf}{k}[n] \end{aligned}$$ To avoid the channel-select filtering of each element in the vector which is mainly necessary to align the signals due to the CSF group delay, we introduce the signals and . Using the delay line vector\ , the vector\ ${\mathbf}{z}_{\text{f}}[n]$ may be approximated by . With this formulation, a fractional and non-constant group delay of the may be incorporated. In case if the group delay $\tau_{\text{g}}$ is constant, and an integer multiple of the sampling time (as e.g. in linear phase filters), the may be approximated by delaying the signal by . In both approximations, the band-limiting effect of the on ${\mathbf}{z}_{\text{f}}[n]$ is ignored. However, this may be tolerated because due to the envelope-squaring operation in which doubles the signal bandwidth, anyhow an of 2 is mandatory to avoid aliasing. Due to the fact, that the I-, and IMD2 interference differ only by a real valued scaling factor $\epsilon$ as derived in , the estimated I-path replica may be used as a reference to estimate the IMD2 replica. This may be done by a linear RLS algorithm which uses the estimated replica as reference input signal to estimate the IMD2 replica. In this case, the RLS estimates also a possible sign difference between the I-, and IMD2 interference. Consequently, only the sign of $\alpha_2^{\text{I}}$ has to be detected during calibration of the receiver which may be done by correlation. The replica signal generation is channel-select filtered which reduces the bandwidth of the replica signal to the bandwidth of the received signal. Second-Order Condition ---------------------- The complex Hessian [@Van_den_Bos_1; @Schreier_2010] of the cost function at the coefficient value becomes $$\label{eq:Hessian_matrix_with_DC} \begin{aligned} H_{\text{I}}&=\frac{\partial}{\partial {\mathbf}{w}_{\text{I}}}\left[\frac{\partial J_{\text{LS}}}{\partial {\mathbf}{w}_{\text{I}}^}\right]^T\|_{{\mathbf}{w}_{\text{I}}={\mathbf}{0}} \\ &= \sum^{n}_{i=0}\lambda^{n-i} \left[-2\, d_{\text{I}}[i] {{\mathbf}{x}}^[i] {{\mathbf}{x}}^T[i] \ast h_s[i]\right]. \end{aligned}$$ If the desired signal $d_{\text{I}}[n]$ contains the DC (when the receiver has no DC filtering), then and the Hessian matrix becomes negative semi-definite like depicted with the local maximum in Fig. \[fig:LS\_cost\_function\_3d\_with\_DC\]. The usual choice of the zero-vector as initialization of ${\mathbf}{w}_{\text{I}}[-1]$ results in a zero-gain vector ${\mathbf}{k}[n]$ for all $n$. This is reasoned in the cost function depicted in Fig. \[fig:LS\_cost\_function\_3d\_with\_DC\] which has a local maximum at ${\mathbf}{w}_{\text{I}}={\mathbf}{0}$ and therefore a vanishing gradient. Consequently, the algorithm is initialized with and the parameters , and with . DC Cancellation --------------- To employ an interference replica without DC, the replica signal is filtered by the DC-notch filter . The new error signal $e_{\text{AC,I}}[n] = d_{\text{AC,I}}[n] - \hat{y}_{\text{AC,I}}[n]$ with the DC-filtered signals is used in the update equation . Here, the introduced index AC indicates the DC filtered signals. The IM2RLS algorithm with DC-suppression can be summarized as -: $$\label{eq:I_patch_IMD2_replica_final_equation_3} \begin{aligned} \hat{y}_{\text{I}}[n] = {\mathbf}{z}^T[n] {\mathbf}{w}_{\text{I}}^[n-1] \ast h_\text{s}[n] \end{aligned}$$ $$\label{eq:DC_notch_filter_2} \begin{aligned} \hat{y}_{\text{AC,I}}[n] = a\, \hat{y}_{\text{AC,I}}[n-1] + \hat{y}_{\text{I}}[n] - \hat{y}_{\text{I}}[n-1]. \end{aligned}$$ $$\label{eq:error_1} \begin{aligned} e_{\text{AC,I}}[n] = d_{\text{AC,I}}[n] - \hat{y}_{\text{AC,I}}[n] \end{aligned}$$ $$\label{eq:kalman_gain_final_1} \begin{aligned} {\mathbf}{k}[n]=\frac{{\mathbf}{P}[n-1] {\mathbf}{z}_{\text{f}}[n] }{\lambda + {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{P}[n-1] {\mathbf}{z}_{\text{f}}[n]} \end{aligned}$$ $$\label{eq:inv_autocorr_matrix_final_1} \begin{aligned} {\mathbf}{P}[n]=\frac{1}{\lambda}\left[{\mathbf}{P}[n-1] -{\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{P}[n-1]\right] \end{aligned}$$ $$\label{eq:coeff_update_2_final_1} \begin{aligned} {\mathbf}{w}_{\text{I}}[n]&={\mathbf}{w}_{\text{I}}[n-1]+e_{\text{AC,I}}[n] {\mathbf}{k}[n] \end{aligned}$$ The parameter in determines the sharpness of the DC-notch filter and is chosen as $a=0.998$. In case of DC filtering in the main receiver , and the Hessian matrix at ${\mathbf}{w}_{\text{I}} = {\mathbf}{0}$ is not positive semi-definite anymore. In this case, the local maximum becomes a saddle-point like depicted in Fig. \[fig:LS\_cost\_function\_3d\_without\_DC\]. Using $N_{\text{CSF}}$ as the number of coefficients of the impulse response, the computational complexity of the IM2RLS with DC-notch filter is real multiplications and $2 N_{{\mathbf}{w}}$ real divisions per iteration. Multiple Solutions of the IM2RLS Algorithm ------------------------------------------ In the cost function shapes depicted in Fig. \[fig:LS\_cost\_function\_3d\_with\_DC\] and Fig. \[fig:LS\_cost\_function\_3d\_without\_DC\], the estimated impulse response coefficients $w_{0}$ and $w_{1}$ (omitting the index I for the I-path) are constrained to be real valued. It can be observed that the two solutions , and minimize the cost function. The existence of multiple solutions can be explained by the absolute-squaring nature of the interference. If the coefficients are allowed to be complex valued, all coefficient pairs $\left\{w_{0}, w_{1}\right\}$ converge to $\left\|w_{0}^{\text{end}}\right\|=\left\|h_{0}\right\|$ and $\left\|w_{1}^{\text{end}}\right\|=\left\|h_{1}\right\|$. This scenario is visualized in Fig. \[fig:IM2RLS\_multiple\_solutions\] where the convergence of the coefficients with the ten different initializations for is depicted. Furthermore, each of the estimated coefficient vectors ${\mathbf}{w}_i^{\text{end}}=\left[w_{0,i}^{\text{end}},w_{1,i}^{\text{end}}\right]^T$ after convergence reach the group delay of the real system impulse response ${\mathbf}{h}$. Performance of the IM2RLS with DC Suppression --------------------------------------------- In this section, the performance of the IM2RLS w/o and w/ DC cancellation is compared. In the first case, the receiver and the replica generation of the IM2RLS do not use a DC cancellation. In this hypothetical example it is assumed that the interference is the only DC source. In the second case, the receiver uses a DC suppression, and the IM2RLS the DC-notch filter. Both cases are compared within an scenario with full allocated signals using 10MHz bandwidth, QPSK modulation, short cyclic prefix, and an of 2. The frequency-selective duplexer stop-band impulse response shown in Fig. \[fig:dpx\_h\_coeff\] is used in for the interference generation. It is modeled with an system which has 15 complex valued coefficients (on the native LTE10 sampling rate of 15.36MHz) and a mean -to- isolation of 50dB [@Ericsson_1]. The resulting signal has a strong frequency-selectivity like indicated in Fig. \[fig:spectrum\]. The wanted signal power is at reference sensitivity level $P_{\text{Rx}}=-97\,\text{dBm}$ and the thermal noise floor is -104.5dBm within 10MHz bandwidth. The receiver is 4.5dB which results in an receiver noise floor of -100dBm. The LNA gain is 20dB, and the two-tone mixer is 50dBm. This results in an desensitization of the wanted signal from an to an of -1.4dB at $P_{\text{Tx}}\,=\,23\,\text{dBm}$. The I-path IMD2 interference is estimated by the IM2RLS using 15 taps, running at the sampling frequency of 30.72MHz . This means, the adaptive filter has less taps than the duplexer stop-band impulse response which has 30 complex valued coefficients at OSF = 2. The Q-path IMD2 replica is estimated by a linear 1-tap RLS (running at 30.72MHz sampling rate) which uses the I-path IMD2 replica as reference input. The IM2RLS algorithm uses the forgetting-factor and as suggested in [@Isermann_1]. The RLS in the uses the same forgetting factor and the initial coefficient $p[-1] = 1e7$. The coefficient vector of the IM2RLS algorithm is initialized with , and the RLS with zero. Fig. \[fig:SINR\_algo\_comparison\_DC\]. shows the steady state improvement at different transmit power levels for an of +50dBm. It can be observed, that in both cases (w/o and w/ DC cancellation) the is improved nearly up to the of 3dB. table\[x index =0, y index =2\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_50dBm\_no\_DC\_no\_reg.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_50dBm\_with\_DC\_no\_reg.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_50dBm\_no\_DC\_no\_reg.dat]{}; coordinates[(0,3) (23,3)]{}; (21,0.684) – (21,2.954); at (19,0.7) [SINR]{}; at (19,0.3) [improvement]{}; at (16,3.2) [Rx SNR=3dB]{}; The convergence behavior at the transmit power of 23dBm is depicted in Fig. \[fig:convergence\_algo\_comparison\_DC\]. For the hypothetical case that the receiver and the IM2RLS are using no DC suppression, the IM2RLS converges faster than with DC suppression. This is reasoned in the additional DC- power which supports the algorithm to converge faster. IM2RLS Algorithm $P_{\text{IMD2}}^{\text{CSF}}$ before $P_{\text{IMD2}}^{\text{CSF}}$ after IIP2 after canc. --------------------- --------------------------------------- -------------------------------------- ------------------ w/o DC cancellation -77.5dBm -95.8dBm 68.4dBm w/ DC cancellation -77.5dBm -94.5dBm 67dBm : IIP2 improvement by digital cancellation \[table:IIP2\_improvement\_with\_and\_without\_DC\_no\_reg\] The $\text{IIP2}$ improvement by the digital cancellation is summarized in Table \[table:IIP2\_improvement\_with\_and\_without\_DC\_no\_reg\] and may be calculated for the IM2RLS with DC-notch filter via $$\label{eq:IIP2_improvment_with_and_without_DC} \begin{aligned} \text{IIP2}_{\text{after canc.}}&= 2 P_{\text{RF}}^{\text{TxL}} - P_{\text{IM2, after canc.}}^{\text{CSF,LTE}} - 13.4\,\text{dB} \\ &= 2\cdot \left(23\,\text{dBm} - 50\,\text{dB} + 20\,\text{dB}\right)\\ & + 94.5\,\text{dBm} - 13.4\,\text{dB} = 67\,\text{dBm}. \end{aligned}$$ The $\text{IIP2}$ is improved from +50dBm to 68.4dBm and 67dBm by the digital cancellation with the IM2RLS w/o and w/ DC suppression, respectively. The correction factor of 13.4dB corrects the IMD2 power calculated with the formula, to the channel-select, and DC-filtered in-band power for the LTE10 full allocation case [@Gebhard2017]. For the calculation of the improvement, the power without DC is used in both cases. The derived IM2RLS algorithm with included DC-notch filter shows an excellent cancellation performance for a full allocated LTE10 transmit signal. However, for small bandwidth allocations like e.g. used in multi-cluster transmissions, the -type algorithm suffers from numerical instability due to the badly-conditioned autocorrelation matrix $\tilde{{\mathbf}{R}}$. To overcome this limitation, the regularized IM2RLS () is derived in the next section. table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_50dBm\_with\_DC.dat]{}; table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_50dBm\_without\_DC.dat]{}; Tikhonov Regularization of the nonlinear {#sec:R-IM2RLS} ========================================= To reduce the spectral emission of the signals, not all available subcarriers are allocated. A portion of the subcarriers at the band-edges (guard-band) are forced to zero which introduces correlation in the transmit samples. E.g. in a 10MHz LTE signal a maximum of 600 out of 1024 subcarriers may be occupied by data [@LTEPublicSafety]. This correlation in the signal $x_{\text{BB}}[n]$ leads to an badly-conditioned autocorrelation matrix ${\mathbf}{R}=E\left\{{{\mathbf}{x}}_{\text{BB}}[n] {{\mathbf}{x}}^H_{\text{BB}}[n]\right\}$ and respectively $\tilde{{\mathbf}{R}}=E\left\{{\mathbf}{z}_{\text{f}}[n] {\mathbf}{z}_{\text{f}}^H[n]\right\}$. Algorithms which need the estimation of the autocorrelation matrix or its inverse ${\mathbf}{P} = {\mathbf}{R}^{-1}$ to estimate the system coefficients either iteratively or in batch-mode, are sensitive to the condition number of ${\mathbf}{R}$ and may suffer from numerical instability if ${\mathbf}{R}$ is badly-conditioned. Because of this reason, a regularized version of the IM2RLS algorithm (R-IM2RLS) is derived in this section. A common method to overcome the problem of badly-conditioned autocorrelation matrices is regularization [@sayed2003fundamentals]. Adding a positive definite matrix to the estimated auto-correlation matrix in each iteration of the algorithm guarantees that the regularized autocorrelation matrix $\tilde{{\mathbf}{R}}'$ stays positive definite and maintains therefore the necessary condition for convergence and existence of [@Gunnarsson_1]. This method is commonly known as Tikhonov-regularization where a matrix ${\mathbf}{L}$ is used for the regularization [@huckle2012data]. By including a regularization term in the cost function , the new cost function $$\label{eq:regularized_cost} \begin{aligned} J'_{\text{R}}[n] &= \sum^{n}_{i=0}\lambda^{n-i}\left[\left\|e_{\text{I}}[i] \right\|^2 + \sigma \left\\|{\mathbf}{L} {\mathbf}{w}_{\text{I}}[n]\right\\|_2^2\right] \\ &=\sum^{n}_{i=0}\lambda^{n-i}\left[\left\|e_{\text{I}}[i] \right\|^2 + \sigma \, {\mathbf}{w}_{\text{I}}^T[n] {\mathbf}{L}^T {\mathbf}{L} {\mathbf}{w}_{\text{I}}^[n]\right] \end{aligned}$$ is defined where $e_{\text{I}}[i]=d_{\text{I}}[i] - {\mathbf}{z}^T[i] {\mathbf}{w}_{\text{I}}^[n] \ast h_\text{s}[i]$. The regularization parameter $\sigma \geq 0$ is used to adjust the regularization amount and the real valued matrix ${\mathbf}{L}$ is typically chosen as (standard Tikhonov regularization), (first order derivative), or $$\label{eq:smoothing_matrix_L} {\mathbf}{L}= \begin{bmatrix} -2 & 1 \\ 1 &-2 & 1 \\ & 1 &-2 & 1 & \\ & & \ddots & \ddots & \ddots \\ & & & & \\ & & & 1 & -2 \end{bmatrix}$$ (second order derivative) [@huckle2012data]. Using the Wirtinger calculus [@Brandwood_1] to obtain the gradient of the cost function , and setting the gradient to zero results in $$\label{eq:regularized_wiener_equation} \begin{aligned} \underbrace{\left[\sum^{n}_{i=0}\lambda^{n-i} \left({\mathbf}{z}_{\text{f}}[i] {\mathbf}{z}_{\text{f}}^H[i] + \sigma {\mathbf}{L}^T {\mathbf}{L} \right) \right]}_{\tilde{{\mathbf}{R}}'[n]} {\mathbf}{w}_{\text{I}}[n] = \underbrace{\sum^{n}_{i=0}\lambda^{n-i} d_{\text{I}}[i] {\mathbf}{z}_{\text{f}}[i]}_{\tilde{{\mathbf}{r}}[n]}. \end{aligned}$$ Reformulating the above equation leads to which is solved recursively using the algorithm. By expressing the cross-correlation vector $\tilde{{\mathbf}{r}}[n]$ by its previous estimate $\tilde{{\mathbf}{r}}[n-1]$, a recursive estimation of the form $$\label{eq:cross_corr_vec_regularized} \begin{aligned} \tilde{{\mathbf}{r}}[n] = \lambda \tilde{{\mathbf}{r}}[n-1] + d_{\text{I}}[n] {\mathbf}{z}_{\text{f}}[n] \end{aligned}$$ may be formulated. Similarly, a recursive estimation of the regularized autocorrelation matrix is obtained by $$\label{eq:iterative_update_autocorr_matrix_reg} \begin{aligned} \tilde{{\mathbf}{R}}'[n]&=\lambda \sum^{n-1}_{i=0}\lambda^{n-i-1} \left({\mathbf}{z}_{\text{f}}[i] {\mathbf}{z}_{\text{f}}^H[i] + \sigma {\mathbf}{L}^T {\mathbf}{L} \right) \\ &+ {\mathbf}{z}_{\text{f}}[n] {\mathbf}{z}_{\text{f}}^H[n] + \sigma {\mathbf}{L}^T {\mathbf}{L} \\ &= \lambda \tilde{{\mathbf}{R}}'[n-1] + \sigma {\mathbf}{L}^T {\mathbf}{L} + {\mathbf}{z}_{\text{f}}[n] {\mathbf}{z}_{\text{f}}^H[n]. \end{aligned}$$ Substituting ${\mathbf}{\Omega}[n]^{-1}=\lambda \tilde{{\mathbf}{R}}'[n-1] + \sigma {\mathbf}{L}^T {\mathbf}{L}$ into , the matrix ${\mathbf}{P}[n]=\tilde{{\mathbf}{R}}'^{-1}[n]$ becomes $$\label{eq:inv_autocorr_matrix_reg} \begin{aligned} {\mathbf}{P}[n]= \left[{\mathbf}{\Omega}[n]^{-1} + {\mathbf}{z}_{\text{f}}[n] {\mathbf}{z}_{\text{f}}^H[n]\right]^{-1}. \end{aligned}$$ After applying the matrix inversion lemma $$\label{eq:matrix_inversion_lemma} \begin{aligned} \left({\mathbf}{A}+{\mathbf}{B}{\mathbf}{C}{\mathbf}{D}\right)^{-1}={\mathbf}{A}^{-1}-{\mathbf}{A}^{-1}{\mathbf}{B}\left({\mathbf}{C}^{-1}+{\mathbf}{D} {\mathbf}{A}^{-1} {\mathbf}{B}\right)^{-1} {\mathbf}{D} {\mathbf}{A}^{-1} \end{aligned}$$ to avoid the matrix inversion, may be formulated as $$\label{eq:inv_autocorr_matrix_reg2} \begin{aligned} {\mathbf}{P}[n]= {\mathbf}{\Omega}[n] - {\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}[n] \end{aligned}$$ using the gain vector $$\label{eq:kalman_gain_reg} \begin{aligned} {\mathbf}{k}[n]=\frac{ {\mathbf}{\Omega}[n] {\mathbf}{z}_{\text{f}}[n]}{ 1+ {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}[n] {\mathbf}{z}_{\text{f}}[n] }. \end{aligned}$$ For the inversion $$\label{eq:omega} \begin{aligned} {\mathbf}{\Omega}[n]=\left[\lambda {\mathbf}{P}^{-1}[n-1] + \sigma {\mathbf}{L}^T {\mathbf}{L}\right]^{-1}, \end{aligned}$$ again the matrix inversion lemma is applyied which yields $$\label{eq:omega2} \begin{aligned} {\mathbf}{\Omega}[n]=\frac{1}{\lambda}\left({\mathbf}{P}[n-1] - {\mathbf}{\Sigma}[n] {\mathbf}{L} {\mathbf}{P}[n-1] \right) \end{aligned}$$ where the substitution $$\label{eq:S_matrix} \begin{aligned} {\mathbf}{\Sigma}[n]= \sigma {\mathbf}{P}[n-1] {\mathbf}{L}^T \left[ \lambda {\mathbf}{I} + \sigma {\mathbf}{L} {\mathbf}{P}[n-1] {\mathbf}{L}^T \right]^{-1} \end{aligned}$$ is used. After rearranging , the expression $$\label{eq:S_matrix_2} \begin{aligned} {\mathbf}{\Sigma}[n]&= \frac{\sigma}{\lambda} \left({\mathbf}{P}[n-1] - {\mathbf}{\Sigma}[n] {\mathbf}{L} {\mathbf}{P}[n-1] \right) {\mathbf}{L}^T \\ &= \sigma {\mathbf}{\Omega}[n] {\mathbf}{L}^T \end{aligned}$$ is obtained. Unfortunately, the calculation of ${\mathbf}{\Sigma}[n]$ in and therefore ${\mathbf}{\Omega}[n]$ still includes a matrix inversion after applying the matrix inversion lemma. However, by decomposing the matrix ${\mathbf}{L}^T {\mathbf}{L}$ in into a sum of $V$ dyads [@Dokoupil_1] $$\label{eq:omega_dyads} \begin{aligned} {\mathbf}{\Omega}[n]=\left[\lambda {\mathbf}{P}^{-1}[n-1] + \sigma \sum^{V}_{k=1} {\mathbf}{p}_{k,1} {\mathbf}{p}_{k,2}^T \right]^{-1}, \end{aligned}$$ applying the matrix inversion lemma results in the recursive calculation of via $$\label{eq:omega_dyads_recursive} \begin{aligned} {\mathbf}{\Omega}_{k}[n]={\mathbf}{\Omega}_{k-1}[n] - \frac{{\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}}{\frac{1}{\sigma}+{\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}} {\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] \end{aligned}$$ for $k=1 \ldots V$ in each iteration $n$ and ${\mathbf}{\Omega}_{0}[n]=\frac{1}{\lambda} {\mathbf}{P}[n-1]$. Reformulating yields $$\label{eq:kalman_gain_reformulated} \begin{aligned} {\mathbf}{k}[n] = {\mathbf}{P}[n] {\mathbf}{z}_{\text{f}}[n]. \end{aligned}$$ The recursive update of the coefficient vector ${\mathbf}{w}_{\text{I}}[n]$ is obtained by inserting , , , and into . The final nonlinear R-IM2RLS algorithm to estimate the I-path interference is summarized by -: $$\label{eq:I_patch_IMD2_replica_final_equation_4} \begin{aligned} \hat{y}_{\text{I}}[n] = {\mathbf}{z}^T[n] {\mathbf}{w}_{\text{I}}^[n-1] \ast h_\text{s}[n] \end{aligned}$$ $$\label{eq:error_2} \begin{aligned} e_{\text{I}}[n] = d_{\text{I}}[n] - \hat{y}_{\text{I}}[n] \end{aligned}$$ $$\label{eq:omega_dyads_recursive_1} \begin{aligned} {\mathbf}{\Omega}_{k}[n]={\mathbf}{\Omega}_{k-1}[n] - \frac{{\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}}{\frac{1}{\sigma}+{\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}} {\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] \end{aligned}$$ $$\label{eq:kalman_gain_reg_1} \begin{aligned} {\mathbf}{k}[n]=\frac{ {\mathbf}{\Omega}_{V}[n] {\mathbf}{z}_{\text{f}}[n]}{ 1+ {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}_{V}[n] {\mathbf}{z}_{\text{f}}[n] }. \end{aligned}$$ $$\label{eq:inv_autocorr_matrix_reg2_1} \begin{aligned} {\mathbf}{P}[n]= {\mathbf}{\Omega}_{V}[n] - {\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}_{V}[n] \end{aligned}$$ $$\label{eq:S_matrix_2_1} \begin{aligned} {\mathbf}{\Sigma}[n]= \sigma {\mathbf}{\Omega}_{V}[n] {\mathbf}{L}^T \end{aligned}$$ $$\label{eq:coeff_update_2_final_2} \begin{aligned} {\mathbf}{w}_{\text{I}}[n]&= \left[{\mathbf}{I} - \left({\mathbf}{I}-{\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n]\right){\mathbf}{\Sigma}[n] {\mathbf}{L}\right] {\mathbf}{w}_{\text{I}}[n-1] + {\mathbf}{k}[n] e_{\text{I}}[n] \end{aligned}$$ The proposed algorithm is initialized with ${\mathbf}{w}_{\text{I}}[-1] \neq {\mathbf}{0}$, and ${\mathbf}{P}[-1] = \nu \, {\mathbf}{I}$ with $\nu > 0$. When the DC suppression is used, then the R-IM2RLS update equations become -: $$\label{eq:I_patch_IMD2_replica_final_equation_5} \begin{aligned} \hat{y}_{\text{I}}[n] = {\mathbf}{z}^T[n] {\mathbf}{w}_{\text{I}}^[n-1] \ast h_\text{s}[n] \end{aligned}$$ $$\label{eq:DC_notch_filter_3} \begin{aligned} \hat{y}_{\text{AC,I}}[n] = 0.998\, \hat{y}_{\text{AC,I}}[n-1] + \hat{y}_{\text{I}}[n] - \hat{y}_{\text{I}}[n-1] \end{aligned}$$ $$\label{eq:error_regularization_with_DC_no_DC} \begin{aligned} e_{\text{AC,I}}[n] = d_{\text{AC,I}}[n] - \hat{y}_{\text{AC,I}}[n] \end{aligned}$$ $$\label{eq:omega_dyads_recursive_2} \begin{aligned} {\mathbf}{\Omega}_{k}[n]={\mathbf}{\Omega}_{k-1}[n] - \frac{{\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}}{\frac{1}{\sigma}+{\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] {\mathbf}{p}_{k,1}} {\mathbf}{p}_{k,2}^T {\mathbf}{\Omega}_{k-1}[n] \end{aligned}$$ $$\label{eq:kalman_gain_reg_2} \begin{aligned} {\mathbf}{k}[n]=\frac{ {\mathbf}{\Omega}_{V}[n] {\mathbf}{z}_{\text{f}}[n]}{ 1+ {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}_{V}[n] {\mathbf}{z}_{\text{f}}[n] }. \end{aligned}$$ $$\label{eq:inv_autocorr_matrix_reg2_2} \begin{aligned} {\mathbf}{P}[n]= {\mathbf}{\Omega}_{V}[n] - {\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n] {\mathbf}{\Omega}_{V}[n] \end{aligned}$$ $$\label{eq:S_matrix_2_2} \begin{aligned} {\mathbf}{\Sigma}[n]= \sigma {\mathbf}{\Omega}_{V}[n] {\mathbf}{L}^T \end{aligned}$$ $$\label{eq:coeff_update_2_final_3} \begin{aligned} {\mathbf}{w}_{\text{I}}[n]&= \left[{\mathbf}{I} - \left({\mathbf}{I}-{\mathbf}{k}[n] {\mathbf}{z}_{\text{f}}^H[n]\right){\mathbf}{\Sigma}[n] {\mathbf}{L}\right] {\mathbf}{w}_{\text{I}}[n-1] \\ &+ {\mathbf}{k}[n] e_{\text{AC,I}}[n] \end{aligned}$$ The DC-notch filter is used to remove the DC from the replica . The complexity of the R-IM2RLS with DC-notch filter and $L=\sigma {\mathbf}{I}$ is real multiplications and $2 N_{{\mathbf}{w}}^2+2N_{{\mathbf}{w}}$ real divisions per iteration. Simulation environment {#sec:simulations} ====================== The performance of the R-IM2RLS algorithm with the three above mentioned regularization matrices ${\mathbf}{L}$ is evaluated with an scenario using an LTE10 multi-cluster intra-band signal which has a native sampling frequency of $f_s=15.36\,\text{MHz}$, QPSK modulation and short cyclic prefix. The interference in the is estimated by the , while the is estimated by a linear RLS which uses the IMD2 replica as reference input. The resulting multi-cluster signal has a strong frequency-selectivity like indicated in Fig. \[fig:spectrum\_clustered\_Tx\]. The R-IM2RLS in the has 15 taps and runs on the higher sampling rate of 30.72MHz due to the of 2. This means, the adaptive filter has less taps than the impulse response which is estimated. The linear 1-tap RLS runs also on the sampling rate of 30.72MHz. The received signal $d[n]$ is DC filtered and the proposed algorithm is using the DC-notch filter to suppress the DC of the replica signal. The wanted signal has a power of at the antenna with an of 3dB. The assumed mixer is +60dBm which corresponds to an SNR desense of 1dB for the specific intra-band multi-cluster transmit signal at 23dBm power level. The thermal noise floor of the receiver is assumed at -104.5dBm per 10MHz and the receiver is 4.5dB. The resulting receiver noise floor and power with 20dB gain is at $\widehat{=}$ and -77dBm respectively. The spectrum of the signals at $P_{\text{Tx}}=23\,\text{dBm}$ is depicted in . It can be observed, that the resulting interference $y_{\text{BB}}^{\text{IMD2}}$ is mostly below the receiver noise floor but still leads to an degradation of 1dB. The depicted interference replica is estimated by the R-IM2RLS with the regularization ${\mathbf}{L}=3e-7\,{\mathbf}{I}$. table\[x index =0, y index =3\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; table\[x index =0, y index =1\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; table\[x index =0, y index =2\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; table\[x index =0, y index =4\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; table\[x index =0, y index =7\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; table\[x index =0, y index =6\] [plot\_data/spectrum\_clustered\_Tx\_IIP\_60dBm.dat]{}; The multi-cluster LTE10 signal uses 21/50 RBs (252 subcarriers from 1024), which means hat 3.78MHz of the available 9.015MHz are allocated. With an of 2 this corresponds to an allocated bandwidth-to-sampling-rate ratio of 3.78/30.72 = 0.12 which introduces a high correlation in the transmit samples. The resulting condition number of the dimensional autocorrelation matrix is in the order of which results in a bad conditioned estimation, and may lead to numerical problems. The regularization of the R-IM2RLS improves numerical estimation of the matrix ${\mathbf}{P}[n]$ by lowering the condition number of the regularized matrix $\tilde{{\mathbf}{R}}'$. IMD2 Self-Interference of a Multi-Cluster Tx Signal --------------------------------------------------- For the estimation of the resulting interference bandwidth, the bandwidth between the minimum and maximum allocated subcarrier in the multi-cluster signal is of interest. In the used clustered LTE10 transmit signal the allocated are $\left\{9-11,29-46\right\}$ with a numbering from left to right and the total number of 50 . For the bandwidth estimation the resulting bandwidth between the lowest allocated subcarrier ( 9) and the upper edge ( 46) of the allocated is . Each has 12 subcarriers and 15kHz subcarrier spacing. The resulting interference bandwidth is $2\times 6.84\,\text{MHz}=13.68\, \text{MHz}$ which means that a small portion of the interference is suppressed by the . The full interference including the DC, the interference after the and DC-removal, and the estimated replica are visualized in Fig. \[fig:spectrum\_clustered\_Tx\_IMD2\]. It can be observed, that the is able to estimate the interference down to 20dB below the receiver noise floor. table\[x index =0, y index =10\] [plot\_data/spectrum\_IMD2\_full\_BW\_IIP2\_60dBm\_clustered\_Tx\_reg\_with\_sigma\_I.dat]{}; table\[x index =0, y index =4\] [plot\_data/spectrum\_IMD2\_full\_BW\_IIP2\_60dBm\_clustered\_Tx\_reg\_with\_sigma\_I.dat]{}; table\[x index =0, y index =7\] [plot\_data/spectrum\_IMD2\_full\_BW\_IIP2\_60dBm\_clustered\_Tx\_reg\_with\_sigma\_I.dat]{}; table\[x index =0, y index =6\] [plot\_data/spectrum\_IMD2\_full\_BW\_IIP2\_60dBm\_clustered\_Tx\_reg\_with\_sigma\_I.dat]{}; Numerical Simulation Results ---------------------------- In the following simulation results, the self-interference cancellation performance in case of an intra-band multi-cluster Tx signal, using the R-IM2RLS algorithm using the DC-notch filter with different regularization matrices is evaluated. The forgetting factor of the is chosen as , , and the regularization constant . The RLS in the uses the same forgetting factor but the initial coefficient $p[-1]=1e7$. The coefficient vector of the is initialized with for the , and the RLS is initialized with zero. The performance is evaluated for the different regularization matrices (Tikhonov regularization), (first order derivative smoothing matrix), and (second order derivative smoothing matrix). The IM2RLS without regularization is not included in the comparison due to numerical instability reasoned by the extremely high condition number of $\tilde{{\mathbf}{R}}$ which is in the order of $10^7$. The performance of the is compared with the recently published LMS-type algorithm (IM2LMS) [@Gebhard2017]. table\[x index =0, y index =2\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_60dBm\_no\_DC\_cluster\_Tx\_smooth\_matrix\_L.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_60dBm\_no\_DC\_cluster\_Tx\_reg\_with\_sigma\_I\_1\_tap\_for\_Q\_IMD2.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_60dBm\_no\_DC\_cluster\_Tx\_reg\_with\_sigma\_1\_-1\_1\_tap\_for\_Q\_IMD2.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_60dBm\_no\_DC\_cluster\_Tx\_smooth\_matrix\_L\_1\_tap\_for\_Q\_IMD2.dat]{}; table\[x index =0, y index =3\] [plot\_data/Tx\_dBm\_vs\_Rx\_SNIR\_dB\_Rx\_-97dBm\_IIP2\_60dBm\_no\_DC\_cluster\_Tx\_IM2LMS\_mu\_1\_200\_gamma\_0\_001.dat]{}; coordinates[(0,3) (23,3)]{}; (21,2.581) – (21,2.983); at (18,2.675) [SINR]{}; at (18,2.6) [improvement]{}; at (16,3.1) [Rx SNR=3dB]{}; The IM2LMS uses the step-size $\mu=0.005$, the regularization parameter $\gamma = 0.001$, and the initial coefficient vector . The IMD2 replica is estimated by a linear normalized LMS which uses the IMD2 replica estimated by the IM2LMS as reference input. The normalized LMS uses a step-size of 1, the regularization parameter is set to 1e-7 and the initial coefficient is set to zero. The value of the step-size is set to the best compromise between steady-state cancellation and convergence time. The convergence of the algorithms is compared using the ensemble , and the steady-state cancellation by the . The improvement of the signal for the different algorithms and regularizations is depicted in Fig. \[fig:SINR\_algo\_clustered\_Tx\]. table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_60dBm\_no\_DC\_clustered\_Tx\_reg\_sigma\_I\_1\_tap\_for\_Q\_IM2.dat]{}; table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_60dBm\_no\_DC\_clustered\_Tx\_reg\_sigma\_1\_-1\_1\_tap\_for\_Q\_IM2.dat]{}; table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_60dBm\_no\_DC\_clustered\_Tx\_smooth\_matrix\_L\_1\_tap\_for\_Q\_IM2.dat]{}; table\[x index =0, y index =1\] [plot\_data/cancellation\_performance\_smoothed\_MA500\_IIP2\_60dBm\_no\_DC\_clustered\_Tx\_IM2LMS\_mu\_1\_200\_gamma\_0\_001\_w0\_1e-4.dat]{}; The convergence behaviour of the algorithms is depicted in Fig. \[fig:convergence\_clustered\_Tx\]. The shows a faster initial convergence than the IM2LMS algorithm which takes about twice as long to reach an of -10dB. The evolution of the condition number of is illustrated in Fig. \[fig:Rxx\_cond\_number\]. table\[x index =0, y index =2\] [plot\_data/Rxx\_condition\_number\_over\_time\_IIP2\_60dBm\_no\_DC\_clustered\_Tx\_no\_reg.dat]{}; table\[x index =0, y index =2\] [plot\_data/Rxx\_condition\_number\_over\_time\_IIP2\_60\_no\_DC\_clustered\_Tx\_reg\_with\_sigma\_I.dat]{}; table\[x index =0, y index =2\] [plot\_data/Rxx\_condition\_number\_over\_time\_IIP2\_60\_no\_DC\_clustered\_Tx\_reg\_with\_sigma\_1\_-1.dat]{}; table\[x index =0, y index =2\] [plot\_data/Rxx\_condition\_number\_over\_time\_IIP2\_60\_no\_DC\_clustered\_Tx\_smooth\_matrix\_L.dat]{}; The condition number of $\tilde{{\mathbf}{R}}$ estimated by the without regularization drastically increases up to values between $10^7$ and $10^8$. In contrast to that, the condition number of $\tilde{{\mathbf}{R}}'$ estimated by the with different regularization matrices ${\mathbf}{L}$ stays below 400 for the specific clustered Tx example. The achieved after the digital cancellation is summarized in Table \[table:IIP2\_improvement\_clustered\_Tx\]. The and IM2LMS algorithms are improving the from 60dBm to about 77dBm and 73dBm, respectively. Algorithm IIP2 after canc. ----------------------------------------------------------------------- ------------------ R-IM2RLS, ${\mathbf}{L}=3e-7 \, {\mathbf}{I}$ 77.2dBm R-IM2RLS, ${\mathbf}{L}=3e-7 \,\text{upperbidiag}\left(1,-1\right) $ 76.5dBm R-IM2RLS using and $\sigma = 3e-7$ 76.4dBm IM2LMS 73dBm : IIP2 improvement by digital cancellation for the clustered Tx signal \[table:IIP2\_improvement\_clustered\_Tx\] Verification of the derived algorithm with measurement data {#sec:measurements} =========================================================== The proposed R-IM2RLS algorithm is evaluated with measurement data and Matlab post-processing. The measurement setup (A) depicted in Fig. \[fig:IM2RLS\_measurement\_setup\] includes the band 2 duplexer model B8663 from TDK, the ZX60-2534MA+ with 41.3dB gain and 2.6dB and the ZAM-42 Level 7 mixer which has 25dB terminal isolation. The measurement is carried out for the I-path mixer and a full allocated LTE-A transmit signal with 10MHz bandwidth, QPSK modulation and short cyclic prefix. The transmit frequency is set to $f_{\text{Tx}}=1.855\,\text{GHz}$ and the mixer frequency is $f_{\text{Rx}}=1.935\,\text{GHz}$ (80MHz duplexing distance). The LTE transmit signal is generated with the R&S SMW 200A signal generator (B), and the signal which leaks into the receiver with 80MHz frequency offset to the signal is amplified by the gain. This amplified signal generates the interference at the output of the mixer which is measured with the real-time oscilloscope . The signal after the is measured by the R&S FSW26 spectrum analyzer (D), and the signal with 7dBm for the ZAM-42 mixer is generated by the R&S SMB 100A signal generator (E). (img) at (0,0) ; at (-0.9cm,1.9cm) [(E)]{}; at (0.9cm,1cm) [(B)]{}; at (2cm,0cm) [(C)]{}; at (-2.4cm,-1.8cm) [(A)]{}; at (0.9cm,-0.2cm) [(D)]{}; table\[x index =0, y index =1\] [plot\_data/spectrum\_measured\_ISP\_Lab\_IMD2\_data.dat]{}; table\[x index =0, y index =2\] [plot\_data/spectrum\_measured\_ISP\_Lab\_IMD2\_data.dat]{}; table\[x index =0, y index =3\] [plot\_data/spectrum\_measured\_ISP\_Lab\_IMD2\_data.dat]{}; table\[x index =0, y index =4\] [plot\_data/spectrum\_measured\_ISP\_Lab\_IMD2\_data.dat]{}; The transmit power is set to $P_\text{RF}^{\text{Tx}} =19.3\,\text{dBm}$, which leads in combination with the duplexer attenuation of 67.6dB (at $f_{\text{Tx}}=1.855\,\text{GHz}$) and the gain of 41.3dB to the typical signal power of . The measured mixer output data stream and the complex valued transmit samples are used for the Matlab post-processing. The spectrum of the signals before and after digital cancellation with the using a Tikhonov regularization and the parameters , and are depicted in Fig. \[fig:measured\_ISP\_IMD2\_data\]. The Matlab post-cancellation showed that 10 taps were sufficient to cancel the interference by 2.2dB down to the noise floor. The coefficient vector was initialized with , and the convergence of the coefficients is shown in Fig. \[fig:coeff\_measured\_data\] which indicates that the coefficients converged after about 5 LTE symbols. table\[x index =0, y index =1\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =2\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =3\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =4\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =5\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =6\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =7\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =8\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =9\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; table\[x index =0, y index =10\] [plot\_data/coeff\_measured\_ISP\_IMD2\_data.dat]{}; Conclusion ========== This paper presented a novel nonlinear type adaptive filter (IM2RLS) and its robust version (R-IM2RLS) for the digital self-interference cancellation in transceivers. The R-IM2RLS provides stability and numerical tractability for highly correlated transmit signals which may result in an ill-conditioned autocorrelation matrix. The proposed is able to cancel the interference generated by a highly frequency-selective leakage signal, and its performance is evaluated with different regularization matrices. Typical receivers use a DC cancellation to prevent the form saturation and a to limit the signal bandwidth. Therefore the interference which is generated by the second-order nonlinearity in the mixer is DC filtered and its bandwidth is reduced to the signal bandwidth. Consequently, the adaptive filter needs to provide a DC-filtered in-band replica. This contribution shows that the adaptive filter is able to reproduce the in-band interference without DC by including the and a DC-notch filter within the algorithm. It is shown, that the proposed algorithm may have multiple solutions of the estimated coefficient vector because of the envelope-squaring nature of the interference. The algorithm converges within a view symbols and the steady-state degradation by the self-interference in case of an multi-cluster transmit signal is improved in simulation from 1dB to less than 0.05dB. The performance of the R-IM2RLS is proved in an measurement scenario with discrete RF components. The interference in the received signal is canceled to the noise floor and a convergence of the coefficients within 5 LTE symbols is achieved. Acknowledgment {#acknowledgment .unnumbered} ============== The authors wish to acknowledge DMCE GmbH & Co KG, an Intel subsidiary for supporting this work carried out at the Christian Doppler Laboratory for Digitally Assisted RF Transceivers for Future Mobile Communications. The financial support by the Austrian Federal Ministry of Science, Research and Economy and the National Foundation for Research, Technology and Development is gratefully acknowledged.	{ "pile_set_name": "ArXiv" }
\[section\] \[thm\][Theorem]{} \[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Notation]{} \[thmAbs\][Theorem]{} \[thmAbs\][Lemma]{} \[thmAbs\][Corollary]{} \[thmAbs\][Proposition]{} \[thmAbs\][Definition]{} \[thmAbs\][Notation]{} \[thmAbs\][Claim]{} \[section\] \[thm2\][Claim]{} \[thm\][Remark]{} \[thm2\][Example]{} \[thm\][Théorème]{} \[thm\][Lemme]{} \[thm\][Corollaire]{} \[thm\][Définition]{} \[thm2\][Affirmation]{} \[thm\][Remarque]{} \[thm2\][Exemple]{} seccntformat\#1[ ]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study $\{0,\alpha\}$-sets, which are sets of unit vectors of $\mathbb{C}^m$ in which any two distinct vectors have angle 0 or $\alpha$. We investigate some distance-regular graphs that provide new constructions of $\{0,\alpha\}$-sets using a method by Godsil and Roy. We prove bounds for the sizes of $\{0,\alpha\}$-sets of flat vectors, and characterize all the distance-regular graphs that yield $\{0,\alpha\}$-sets meeting the bounds at equality.' author: - \| Junbo Huang\ \ \ bibliography: - 'mybib.bib' date: 'June 10, 2013' title: 'Sets of Complex Unit Vectors with Two Angles and Distance-Regular Graphs' --- Introduction ============ The angle between two unit vectors $x$ and $y$ in $\mathbb{C}^m$ is defined to be the number $\|x^y\|^2$, where $x^$ is the conjugate transpose of $x$. For $\alpha \in \mathbb{R}$ with $0 < \alpha \leq 1$, a set of unit vectors in $\mathbb{C}^m$ is called a $\{0,\alpha\}$-set if the angle between any two distinct vectors in the set is 0 or $\alpha$. The concept of $\{0,\alpha\}$-set is motivated by equiangular sets and mutually unbiased bases. A set of unit vectors in $\mathbb{C}^m$ is called equiangular if the angle between any two distinct vectors in the set is the same. Two orthonormal bases of $\mathbb{C}^m$ or $\mathbb{R}^m$ are called unbiased if the angle between any two vectors from different bases is a constant. Thus equiangular sets and sets of vectors from a collection of mutually unbiased bases are examples of $\{0,\alpha\}$-sets. Equiangular sets and mutually unbiased bases have various applications in quantum theory. For example, in quantum physics, it is desirable to recover the state of a physical system by means of measurements, and equiangular sets and mutually unbiased bases provide “good” measurements for such purposes. Because of the applications, equiangular sets and mutually unbiased bases have received much attention. Many constructions for such sets were found; see [@CCKS; @GodRoy09; @Hoggar98; @KlapRott04; @Konig99], for examples. Constructions of general $\{0,\alpha\}$-sets, however, received little attention. In 2005, Godsil and Roy [@RoyThesis] presented a method for constructing sets of unit vectors using certain bipartite graphs. Their method can be used to construct $\{0,\alpha\}$-sets. A vector from a set constructed using Godsil and Roy’s method is flat, that is, all of its entries have the same absolute value. Flat vectors have natural connections to combinatorics. For example, the columns of a complex Hadamard matrix are flat. In this paper, we investigate some new constructions of flat $\{0,\alpha\}$-sets using Godsil and Roy’s method, and study bounds related to such sets. We first present examples of distance-regular graphs, including an infinite family, that provide new constructions of $\{0,\alpha\}$-sets using Godsil and Roy’s method. We then show that certain upper bounds for the sizes of $\{0,\alpha\}$-sets proven by Delsarte, Goethals and Seidel [@DGS] can be improved for sets of flat vectors. We will see that the 8-cycle, the 4-cube, the folded 8-cube, and the coset graph of the extended binary Golay code yield $\{0,\alpha\}$-sets that satisfy the improved bounds at equality. Finally, we show that these are the only distance-regular graphs that provide flat $\{0,\alpha\}$-sets of maximum size with respect to the improved bounds. Godsil and Roy’s Construction ============================= In this section, we describe a method of constructing sets of unit vectors using certain bipartite graphs. The construction is due to Godsil and Roy, and appeared in Roy’s PhD thesis [@RoyThesis]. Let $G$ be a group acting on a set $V$. We say that the group $G$ acts on $V$ regularly if for every $a, b \in V$, there exists a unique $g \in G$ such that $a^g=b$. A character of an abelian group $G$ is a group homomorphism from $G$ to the multiplicative group of complex numbers of norm $1$. It is well known that characters of a finite abelian group $G$ form a group of size $\|G\|$. For the outline of a proof, see [@AlgCombCG Sect. 12.8]. In order to apply Godsil and Roy’s construction, we need a connected bipartite graph ${\Gamma}$ with colour classes $Y$ and $Z$ and an abelian group $G$ of automorphisms of ${\Gamma}$ acting regularly on each of $Y$ and $Z$. Let $u$ and $v$ be vertices of ${\Gamma}$ from the same colour class, and let $h$ be the element in $G$ such that $u^h=v$. Since $h$ is an automorphism of ${\Gamma}$, $h$ defines an injection from the set of neighbours of $u$ to the set of neighbours of $v$. Since the choice of $u$ and $v$ are arbitrary within a colour class, it follows that ${\Gamma}$ must be regular. Let $k$ denote the valency of ${\Gamma}$. Since $G$ acts on $Y$ and $Z$ regularly, we also have $\|Y\|=\|G\|=\|Z\|$. Let us use $n$ to denote the size of $G$, so ${\Gamma}$ has $2n$ vertices. To start, we pick two vertices $y \in Y$ and $z \in Z$, let $$D := \big\{g \in G: z^g \text{ is adjacent to } y \text{ in } {\Gamma}\big\},$$ and let $S'$ be the set of characters of $G$ restricted to $D$. Note that $\|D\|=k$. Since there are $n$ characters of $G$, the set $S'$ is a subset of $\mathbb{C}^k$ and has size $n$. It is shown in [@RoyThesis Ch. 4] that $$\big\{\|x^y\|^2: x,y \in S', x \neq y \big\} = \big\{\lambda^2: \lambda \text{ is an eigenvalue of } {\Gamma}, \lambda \neq \pm k \big\}.$$ Therefore, by normalizing the vectors in $S'$, we get a set of unit vectors with angle set determined by the eigenvalues of ${\Gamma}$. The construction is summarized in the following lemma. \[constr\] Suppose that a connected bipartite graph ${\Gamma}$ with colour classes $Y$ and $Z$ has an abelian group $G$ of automorphisms acting regularly on each of $Y$ and $Z$. Let $y \in Y$ and $z \in Z$, and let $$D := \big\{g \in G: z^g \text{ is adjacent to } y \text{ in } {\Gamma}\big\}.$$ Let $S'$ be the set of characters of $G$ restricted to $D$, and let $$S := \Big\{\frac{1}{\sqrt{k}}\chi: \chi \in S'\Big\},$$ where $k$ is the valency of ${\Gamma}$. If $2n$ is the number of vertices in ${\Gamma}$, then $S$ is a set of $n$ unit vectors in $\mathbb{C}^k$, with set of angles $$\Big\{\frac{\lambda^2}{k^2}: \lambda \text{ is an eigenvalue of } \Gamma, \lambda \neq \pm k \Big\}.$$ In Lemma \[constr\], if all the non-identity elements of $G$ have order two, then the range of a character of $G$ is a subset of $\{1,-1\}$; in this case, the set $S$ constructed from $G$ is a subset of $\mathbb{R}^k$. When searching for graphs for which Godsil and Roy’s construction can be applied, distance-regular graphs are natural candidates. In the next section, we review some basic facts about distance-regular graphs, focusing on those that are bipartite and have diameter four. Distance-Regular Graphs ======================= Let ${\Gamma}$ be a graph of diameter $d$ and let $v$ be a vertex of ${\Gamma}$. For any $i \in \mathbb{Z}$, let ${\Gamma}_i(v)$ be the set of vertices of ${\Gamma}$ that are at distance $i$ from $v$. Clearly ${\Gamma}_i(v) = \emptyset$ if $i<0$ or $i>d$. The graph ${\Gamma}$ is called distance-regular* if it is connected and, for any $i \in \{0,1,\dots,d\}$, the size of ${\Gamma}_i(u) \cap {\Gamma}_1(v)$ depends only on the distance between $u$ and $v$. Let ${\partial}$ denote the distance function of ${\Gamma}$. If ${\partial}(u,v)=i$, we write $$a_i := {\big\|}{\Gamma}_i(u) \cap {\Gamma}_1(v){\big\|}, \quad b_i := {\big\|}{\Gamma}_{i+1}(u) \cap {\Gamma}_1(v){\big\|}, \quad c_i := {\big\|}{\Gamma}_{i-1}(u) \cap {\Gamma}_1(v){\big\|}.$$ The numbers $a_i$, $b_i$ and $c_i$ are called the intersection numbers of the distance-regular graph. Since $b_0$ is the valency of any vertex, a distance-regular graph must be regular; we use $k$ to denote the valency $b_0$ of the graph. Let $k_i(v)$ be the size of ${\Gamma}_i(v)$. In particular, $k_0(v)=1$ for each vertex $v$. In fact, by double-counting the number of edges with one end in ${\Gamma}_{i-1}(v)$ and the other in ${\Gamma}_i(v)$, it follows that $k_i(v) c_i = k_{i-1}(v) b_{i-1}$, so inductively we see that $k_i(v)$ is independent of $v$, and we write $k_i$ for $k_i(v)$. The following lemma states some conditions that the intersection numbers satisfy. \[feas\] If $b_0$, $b_1$, $\dots$, $b_{d-1}$, $c_1$, $c_2$, $\dots$, $c_d$ are intersection numbers of a distance-regular graph with valency $k$ and diameter $d$, then both of the following conditions hold: - $k_i= k_{i-1} b_{i-1}/c_i$ for all $i \in \{1,2,\dots,d\}$. - $1 = c_1 \leq c_2 \leq \dots \leq c_{d-1} \leq c_d \leq k$. Condition (i) in the lemma has been explained before the lemma. For a proof of condition (ii) , see [@BiggsAG2 Ch. 20]. The intersection numbers determine many numerical data for the distance-regular graph. For example, from Lemma \[feas\], we see that the intersection numbers determine each $k_i$ and hence the number of vertices in the graph. The intersection numbers also determine the eigenvalues of the graph. In fact, the eigenvalues of a distance-regular graph are precisely those of the tridiagonal matrix $$B = \begin{bmatrix} a_0 & c_1 & & & &\\ b_0 & a_1 & c_2 & & & \\ & b_1 & a_2 & \ddots & & \\ & & b_2 & \ddots & \ddots & \\ & & & \ddots & \ddots & c_d \\ & & & & b_{d-1} & a_d \end{bmatrix}.$$ For a proof of this fact, see [@BiggsAG2 Ch. 21]. It is well known that any connected graph of diameter $d$ has at least $d+1$ distinct eigenvalues (see [@BiggsAG2 Ch. 2]), and therefore, by the size of the matrix $B$, any distance-regular graph of diameter $d$ has exactly $d+1$ distinct eigenvalues. For the rest of this section, we focus on bipartite distance-regular graphs of diameter four. For more general theory of distance-regular graphs, we refer the readers to [@BCN]. Let ${\Gamma}$ be a distance-regular graph with valency $k$. If ${\partial}(u,v)=i$ then any neighbour of $v$ must be at distance $i-1$, $i$ or $i+1$ from $u$. Consequently, $${\big\|}{\Gamma}_j(u) \cap {\Gamma}_1(v) {\big\|}= 0,$$ for $j \notin \{i-1,i,i+1\}$, and so $$a_i + b_i + c_i = k.$$ If ${\Gamma}$ is bipartite then for any $i$, all vertices in ${\Gamma}_i(v)$ must be in the same colour class, so $a_i=0$ for all $i$, whence $$b_i = k - c_i.$$ Therefore, for a bipartite distance-regular graph of diameter four, the numbers $k$, $c_2$ and $c_3$ determine all the intersection numbers for the graph. In fact, a bipartite graph of diameter four with $k$, $c_2$ and $c_3$ well-defined is automatically distance-regular. From now on, we use $k$, $c_2$ and $c_3$ to represent all the intersection numbers for a bipartite distance-regular graph of diameter four. For such a graph, the matrix $B$ above becomes $$\begin{bmatrix} 0&1&0&0&0 \\ k&0&c_2&0&0 \\ 0&k-1&0&c_3&0 \\ 0&0&k-c_2&0&k \\ 0&0&0&k-c_3&0 \end{bmatrix}$$ and has eigenvalues $\pm k$, $\pm \theta_1$ and 0, where $\theta_1 = \sqrt{k + c_2 (k-c_3-1)}$. We call $\theta_1$ the nontrivial eigenvalue for a bipartite distance-regular graph of diameter four. For such a graph, Brouwer, Cohen and Neumaier [@BCN Thm 5.4.1, p.173] proved a restriction on $c_2$ and $c_3$. \[c2bound\] Suppose that $c_2$ and $c_3$ are intersection numbers of a distance-regular graph of diameter at least four. If $c_2>1$ then $c_3 \geq 3c_2/2$. We conclude this section by stating an expression of the number of vertices in terms of the intersection numbers. \[vertices\] If a bipartite distance-regular graph of diameter four with intersection numbers $k$, $c_2$ and $c_3$ has $2n$ vertices, then $$n = \frac{k(k^2-(c_2+1)k+c_2(c_3 +1))}{c_2 c_3}.$$ Since $k_1=k$, by using the formulae in (i) from Lemma \[feas\], we deduce that $$k_2 = \frac{k(k-1)}{c_2}, \quad k_3 = \frac{k(k-1)(k-c_2)}{c_2c_3}, \quad k_4 = \frac{(k-1)(k-c_2)(k-c_3)}{c_2c_3}.$$ Since $2n = k_0+k_1+k_2+k_3+k_4$ and $k_0=1$, we have $$n = \frac{1+k+k_2+k_3+k_4}{2} = \frac{k(k^2-(c_2+1)k+c_2(c_3 +1))}{c_2 c_3}.$$ Constructing $\{0,\alpha\}$-Sets ================================ In this section, we present some distance-regular graphs that can be used to construct $\{0,\alpha\}$-sets using the construction described in Lemma \[constr\]. By the symmetry of the eigenvalues of a bipartite graph, in order to construction $\{0,\alpha\}$-sets, we need graphs that have exactly five eigenvalues 0, $\pm \theta$ and $\pm k$; distance-regular graphs satisfying this condition are precisely those with diameter four. Bipartite distance-regular graphs of diameter four have been used to construct sets of mutually unbiased bases by Godsil and Roy [@GodRoy09]. We describe some examples of such graphs that apply to Lemma \[constr\]. All the graphs described in this section give new constructions of $\{0,\alpha\}$-sets, except for the 8-cycle, which is a special case of the graphs used in [@GodRoy09]. Let $X$ be a set with an element called “zero”. The weight of an element in $X^m$ is defined to be the number of its nonzero entries. The Hamming distance between two elements in $X^m$ is the number of coordinates in which they differ. Let $\mathbb{F}_q$ be the finite field with $q$ elements. Then $\mathbb{F}_q^m$ is a vector space over $\mathbb{F}_q$. Let $C$ be a subspace of $\mathbb{F}_q^m$ such that every nonzero element in $C$ has weight at least two. The coset graph of $C$ (with respect to $\mathbb{F}_q^m$) is the graph with vertex set being the set $\mathbb{F}_q^m/C$ of all cosets of $C$, such that two cosets are adjacent if and only if their difference can be represented by a weight-one element in $\mathbb{F}_q^m$. $8$-Cycle --------- The $8$-cycle is the graph with vertex set $\mathbb{Z}_2 \times \mathbb{Z}_4$, such that $(x_1,y_1)$ and $(x_2,y_2)$ are adjacent if and only if $x_1 \neq x_2$ and $y_1-y_2 \in \{1,3\}$. The 8-cycle is bipartite, with $\{0\} \times \mathbb{Z}_4$ and $\{1\} \times \mathbb{Z}_4$ being the colour classes, and is the unique bipartite distance-regular graph of diameter four with intersection numbers $k=2$, $c_2=1$ and $c_3=1$. Its nontrivial eigenvalue is $\theta_1=\sqrt{2}$. The group $G:=(\mathbb{Z}_4,+)$ acts on the vertices of the 8-cycle by addition to the second coordinate (without changing the first), and each colour class is an orbit induced by the actions. Since $G$ is clearly a group of automorphisms of the graph, by Lemma \[constr\], we can construct a $\{0,1/2\}$-set of size 4 in $\mathbb{C}^2$, represented by the columns of the matrix $$\frac{1}{\sqrt{2}} \begin{bmatrix} 1 & i & 1 & i \\ 1 & -i & i & 1 \end{bmatrix}.$$ In fact, these vectors can be partitioned into two unbiased bases of $\mathbb{C}^2$, with a basis containing the first two columns. The construction of vectors from the 8-cycle has essentially been described in [@GodRoy09], in which the 8-cycle is viewed as the incidence graph of the affine plane of order two having the lines with infinite slope removed. $4$-Cube -------- The $4$-cube is the graph with vertex set $\mathbb{Z}_2^4$, such that two vertices are adjacent if and only if they have Hamming distance 1. It is straightforward to check that the 4-cube is a bipartite distance-regular graph of diameter four having 16 vertices, with intersection numbers $k=4$, $c_2=2$, $c_3=3$ and nontrivial eigenvalue $\theta_1=2$. Moreover, it is the unique bipartite distance-regular graph of diameter four with $k=4$, $c_2=2$ and $c_3=3$; see [@BCN Sect. 6.1]. The odd-weight elements form a colour class and the even-weight ones form the other. The set of even-weight elements is an additive abelian group $G$ acting regularly on the colour classes by addition, and is clearly a group of automorphisms of the graph. Since $G$ is a subgroup of $\mathbb{Z}_2^4$, every non-identity element of it has order two. Therefore, using Lemma \[constr\], we can construct a $\{0,1/4\}$-set of size 8 in $\mathbb{R}^4$, represented by the columns of the matrix $$\frac{1}{2} \begin{bmatrix} 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & -1 & -1 & 1 & 1 & -1 & -1 \\ 1 & -1 & 1 & -1 & 1 & -1 & 1 & -1 \\ 1 & -1 & -1 & 1 & -1 & 1 & 1 & -1 \end{bmatrix}.$$ Again, these vectors can be partitioned into two unbiased bases of $\mathbb{R}^4$, with a basis containing the first four columns. Folded $8$-Cube --------------- The folded $8$-cube is the graph with vertex set $\mathbb{Z}_2^7$, such that two vertices are adjacent if any only if they have Hamming distance $1$ or $7$. It is straightforward to check that the folded 8-cube is a bipartite distance-regular graph of diameter four having 128 vertices, with intersection numbers $k=8$, $c_2=2$, $c_3=3$ and nontrivial eigenvalue $\theta_1=4$. Moreoever, it is the unique bipartite distance-regular graph of diameter four with $k=8$, $c_2=2$ and $c_3=3$; see [@BCN Sect. 9.2D]. The odd-weight elements form a colour class of the even-weight ones form the other. Similar to the 4-cube, the even-weight strings form an abelian group $G$ of graph automorphisms acting regularly on the colour classes by addition. Using Lemma \[constr\], we can construct a $\{0,1/4\}$-set of size 64 in $\mathbb{R}^8$ (since $G \leq \mathbb{Z}_2^7$). Van-Lint Schrijver Partial Geometry ----------------------------------- Let $C$ be the subspace of $\mathbb{F}_3^6$ spanned by the all-one vector. Then any coset of $C$ in $\mathbb{F}_3^6$ have its elements sharing the same coordinate sum (computed in $\mathbb{F}_3$). For $i \in \{0,1,2\}$, let $V_i$ be the set of cosets of $C$ whose elements have coordinate sum $i$. It is easy to check that $V_0$, $V_1$ and $V_2$ all have the same size, which is $3^4=81$. Let ${\Gamma}'$ be the coset graph of $C$. Then ${\Gamma}'$ is tripartite, with colour classes $V_0$, $V_1$ and $V_2$. Let ${\Gamma}$ be the subgraph ${\Gamma}$ of ${\Gamma}'$ induced by $V_0$ and $V_1$. The incidence structure with point set $V_0$, line set $V_1$ and incidence graph ${\Gamma}$ is called the van Lint-Schrijver partial geometry; it was first introduced in [@LintSchrPG], and is also discussed in [@LintCam82] and [@BCN Sect. 11.5]. The incidence graph ${\Gamma}$ is a bipartite distance-regular graph of diameter four having 162 vertices, with intersection numbers $k=6$, $c_2=1$, $c_3=2$ and nontrivial eigenvalue $\theta_1=3$. See [@BCN Sect. 11.5]. The set $V_0$ is a subgroup of the abelian group $\mathbb{F}_3^6$; its action on $\mathbb{F}_3^6$ by addition induces orbits $V_0$, $V_1$ and $V_2$, so $V_0$ acts on the colour classes of ${\Gamma}$ regularly. Moreover, addition on the vertices of ${\Gamma}$ by an element of $V_0$ is clearly an automorphism of ${\Gamma}$. Therefore, using Lemma \[constr\], we can construct a $\{0,1/4\}$-set of size 81 in $\mathbb{C}^6$. Extended Binary Golay Code -------------------------- Let $I$ be the $12 \times 12$ identity matrix, and let A be the matrix $$\begin{bmatrix} 0&1&1&1&1&1&1&1&1&1&1&1 \\ 1&1&1&0&1&1&1&0&0&0&1&0 \\ 1&1&0&1&1&1&0&0&0&1&0&1 \\ 1&0&1&1&1&0&0&0&1&0&1&1 \\ 1&1&1&1&0&0&0&1&0&1&1&0 \\ 1&1&1&0&0&0&1&0&1&1&0&1 \\ 1&1&0&0&0&1&0&1&1&0&1&1 \\ 1&0&0&0&1&0&1&1&0&1&1&1 \\ 1&0&0&1&0&1&1&0&1&1&1&0 \\ 1&0&1&0&1&1&0&1&1&1&0&0 \\ 1&1&0&1&1&0&1&1&1&0&0&0 \\ 1&0&1&1&0&1&1&1&0&0&0&1 \end{bmatrix}.$$ The extended binary Golay code is the subspace of $\mathbb{F}_2^{24}$ generated by the rows of the matrix $[I\|A]$. In the extended binary Golay code, any element has even weight (and in fact, weight that is divisible by 4), and any nonzero element has weight at least eight. For more details about Golay codes, see [@macsloane77] and [@vanLintGTM]. Let $C$ be the extended binary Golay code, and let ${\Gamma}$ be the coset graph of $C$ (with respect to $\mathbb{F}_2^{24}$). Since the elements of $C$ have even weights, the weights of the elements in a coset of $C$ have the same parity; we call a coset of $C$ even if its elements have even weights, and odd otherwise. It then follows that ${\Gamma}$ is bipartite, with the even cosets forming a colour class and the odd cosets forming the other. In fact, ${\Gamma}$ is a distance-regular graph of diameter four having 4096 vertices, with intersection numbers $k=24$, $c_2=2$, $c_3=3$ and nontrivial eigenvalue $\theta_1 = 8$. Moreoever, ${\Gamma}$ is the unique bipartite distance-regular graph of diameter four with $k=24$, $c_2=2$ and $c_3=3$; see [@BCN Sect. 11.3D]. The set $G$ of even cosets of $C$ is a subgroup of the quotient group $\mathbb{F}_2^{24}/C$, and it acts on each colour class of ${\Gamma}$ regularly by addition. Note that $G$ acts on ${\Gamma}$ as graph automorphisms. Since the non-identity elements of $G$ have order two, by Lemma \[constr\], we can construct a $\{0,1/9\}$-set of size 2048 in $\mathbb{R}^{24}$. Extended Kasami Codes --------------------- Let $s$ and $t$ be powers of 2, with $t \leq s$, and let $F := \mathbb{F}_s$. Let $K(s,t)$ be the set of elements $x$ in $\mathbb{F}_2^F$ with even-weight that satisfy $$\sum\limits_{\alpha \in F}{x_\alpha \alpha} = \sum\limits_{\alpha \in F}{x_\alpha \alpha^{t+1}} = 0.$$ The subspace $K(s,t)$ of $\mathbb{F}_2^F$ is called the extended Kasami code (with parameters $s$ and $t$) if one of the following two conditions is satisfied: 1. $s=q^{2j+1}$, $t=q^m$, with $q=2^i$, $m \leq j$, and $gcd(m,2j+1)=1$. 2. $s=q^2$, $t=q$, with $q=2^i$. Every nonzero element in an extended Kasami code has weight at least four. Similar to the extended Golay code, the elements in a coset of $K(s,t)$ have the same parity, so the coset graph ${\Gamma}(s,t)$ of $K(s,t)$ is bipartite, with the even cosets forming a colour class and the odd ones forming the other. In fact, ${\Gamma}(s,t)$ is a distance-regular graph of diameter four having $2n$ vertices and $k$, $c_2$, $c_3$ as parameters, where 1. $(n,k,c_2,c_3) = (q^{4j+2},q^{2j+1},q,q^{2j}-1)$, 2. $(n,k,c_2,c_3) = (q^3,q^2,q,q^2-1)$, corresponding to the order above. See [@BCN Sect. 11.2]. The nontrivial eigenvalue of ${\Gamma}(s,t)$ is 1. $\theta_1=q^{j+1}$. 2. $\theta_1=q$. As in the extended binary Golay code, the even cosets of $K(s,t)$ form an abelian group $G$ of automorphisms of ${\Gamma}(s,t)$ that acts on each colour class of ${\Gamma}(s,t)$ regularly by addition. Since every non-identity element of $G$ has order two, by Lemma \[constr\], we can construct a $\{0,\alpha\}$-set of size $n$ in $\mathbb{R}^{k}$, where, corresponding to the order above, 1. $(\alpha,n,k)=(q^{-2j},q^{4j+2},q^{2j+1})$. 2. $(\alpha,n,k)=(q^{-2},q^3,q^2)$. Flat Bounds =========== In 1975, Delsarte, Goethals and Seidel [@DGS] proved bounds for sizes of sets of unit vectors with prescribed sets of angles. One type of the bounds they proved states the following when applied to $\{0,\alpha\}$-sets. \[DGS bounds\] Let $S$ be a $\{0,\alpha\}$-set in $\mathbb{C}^m$ with $0<\alpha<1$. Then $$\|S\| \leq \frac{(m+1)m^2}{2}.$$ If $S \subseteq \mathbb{R}^m$ then $$\|S\| \leq \frac{(m+2)(m+1)m}{6}.$$ Another proof of this theorem using elementary tensor algebra was later given by Calderbank, Cameron, Kantor and Seidel [@CCKS] in 1997. Recall that a vector in $\mathbb{C}^m$ is called flat if all of its entries have the same absolute value. It turns out that by using arguments similar to those by Calderbank, Cameron, Kantor and Seidel, the bounds in Theorem \[DGS bounds\] can be improved if the $\{0,\alpha\}$-set contains only flat vectors. \[flatbds\] Let $S$ be a $\{0,\alpha\}$-set of flat vectors in $\mathbb{C}^m$ with $0<\alpha<1$. Then $$\|S\| \leq \frac{(m^2-m+2)m}{2}.$$ If $S \subseteq \mathbb{R}^m$ then $$\|S\| \leq \frac{(m^2-3m+8)m}{6}.$$ Let $\beta:=\sqrt{\alpha}$. Then for distinct $x$ and $y$ in $S$, we have $\|x^* y\| \in \{0,\beta\}$. Let $M$ be the matrix whose columns are the vectors in $S$. Then $$M^* M = I+\beta C,$$ where $C$ is a Hermitian matrix with zero diagonal that has absolute value 0 or 1 for all off-diagonal entries. For each $x \in S$, let $$v_x := x \otimes x \otimes \overline{x}$$ be a tensor product, and let $S':=\{v_x: x \in S\}$. For more details about tensor algebra, see [@RomanLA]. If we let $N$ be the matrix whose columns are the vectors in $S'$, then since $$v_x^* v_y = (x \otimes x \otimes \overline{x})^* (y \otimes y \otimes \overline{y}) = (x^* y)^2 (\overline{x^* y}) = (x^* y) \|x^* y\|^2,$$ we have $$N^* N = I + \beta^3 C = (1-\beta^2)I + \beta^2(I+\beta C).$$ Since $\|\beta\|<1$, the matrix $(1-\beta^2)I$ is positive definite. Since $I+\beta C = M^* M$ is positive semidefinite, so is $\beta^2(I+\beta C)$. Hence $N^* N$ is positive definite and has full rank $\|S\|$. On the other hand, the rank of $N^* N$ is equal to the rank of $N$, which is the dimension of span$(S')$. Let $x \in S$. Since $x$ is a flat unit vector in $\mathbb{C}^m$, each of its entries has absolute value equal $1/\sqrt{m}$. Consider indices of $v_x$ that have forms $(i,j,j)$ or $(j,i,j)$. The entries of $v_x$ corresponding to these indices are $$x_i x_j \overline{x_j} = \frac{x_i}{m},$$ which depends only on $x_i$ (since $m$ is a constant). If an index of $v_x$ does not have one of the forms above, then it either has form $(i,i,j)$ or is equal to $(i,j,k)$ for some distinct $i$, $j$ and $k$. There are $m(m-1)$ indices of the first type, and there are $m(m-1)(m-2)$ indices of the second type. Since there are $m$ ways to choose $i$ for the indices $(i,j,j)$ and $(j,i,j)$, and since the entry of $v_x$ indexed by $(i,j,k)$ is equal to that by $(j,i,k)$, it follows that the dimension of span$(S')$ is at most $$m + m(m-1) + \frac{m(m-1)(m-2)}{2} = \frac{m(m^2-m+2)}{2}.$$ This is an upper bound for the rank of $N^* N$, which is equal to $\|S\|$, so $$\|S\| \leq \frac{m(m^2-m+2)}{2},$$ proving the first bound. Now assume further that $S \subseteq \mathbb{R}^m$. Then any $x \in S$ has entries equal to $1/\sqrt{m}$ or $-1/\sqrt{m}$, and $$v_x = x \otimes x \otimes x.$$ In this case, the entries of $v_x$ corresponding to the indices of the forms $(i,j,j)$, $(j,i,j)$ or $(j,j,i)$ are all equal to $x_i/m$. Since the entry of $v_x$ indexed by $(i,j,k)$ is invariant under permutations of the components $i$, $j$ and $k$, the dimension of span$(S')$ is at most $$m + \binom{m}{3} = \frac{m(m^2-3m+8)}{6},$$ proving the second bound $$\|S\| \leq \frac{m(m^2-3m+8)}{6}.$$ Note that Godsil and Roy’s construction in Lemma \[constr\] yields sets of flat vectors, so the bounds proven above are applicable to the $\{0,\alpha\}$-sets constructed using Godsil and Roy’s method. In fact, there are distance-regular graphs that produce $\{0,\alpha\}$-sets of optimal sizes with respect to these bounds. In particular, it is easy to check that the 4-cube, the folded 8-cube and the coset graph of the extended binary Golay code produce sets that meet the flat real bound, while the 8-cycle produces a set that meets the flat complex bound, all at equality. The related parameters for these graphs are summarized in Table 1. ------------------------ ----- ------- ------- ---------- ------ ------------------- Graph $k$ $c_2$ $c_3$ $\alpha$ $n$ Space \[0.3ex\] 4-cube 4 2 3 1/4 8 $\mathbb{R}^4$ Folded 8-cube 8 2 3 1/4 64 $\mathbb{R}^8$ Ext. binary Golay code 24 2 3 1/9 2048 $\mathbb{R}^{24}$ 8-cycle 2 1 1 1/2 4 $\mathbb{C}^2$ \[1ex\] ------------------------ ----- ------- ------- ---------- ------ ------------------- : $\{0,\alpha\}$-sets of largest size $n$ from graphs. Optimal Graphs ============== At the end of the previous section, we saw some graphs that produce $\{0,\alpha\}$-sets with the largest sizes. In this section, we shall see that those are the only distance-regular graphs that produce largest $\{0,\alpha\}$-sets. Let $S$ be a $\{0,\alpha\}$-set in $\mathbb{R}^k$ constructed from a distance-regular graph ${\Gamma}$ using the construction in Lemma \[constr\]. Then $S$ has the largest size $(k^2-3k+8)k/6$ if and only if ${\Gamma}$ is one of the following graphs: - $4$-cube. - Folded $8$-cube. - Coset graph of the extended binary Golay code. The graph ${\Gamma}$ has to be bipartite and have diameter four. Let $k$, $c_2$ and $c_3$ be its intersection numbers, and let $2n$ be the number of its vertices. By Lemma \[constr\] and Lemma \[vertices\], $$n = \frac{k(k^2-(c_2+1)k+c_2(c_3 +1))}{c_2 c_3}.$$ Since $\|S\|=(k^2-3k+8)k/6$, we have $$\frac{k(k^2-(c_2+1)k+c_2(c_3 +1))}{c_2 c_3} = \frac{(k^2-3k+8)k}{6},$$ which, after rearrangement, yields $$\label{realtight2} (6-c_2c_3) k^2 - (6c_2+6-3c_2c_3)k + (6c_2-2c_2c_3) = 0.$$ Suppose $6-c_2c_3=0$. Then (\[realtight2\]) becomes $$(6c_2-12)k + (6c_2-12) = 0.$$ Since $c_2 \leq c_3$ by Lemma \[feas\], we have $(c_2,c_3)=(1,6)$ or $(c_2,c_3)=(2,3)$. The case where $c_2=1$ is impossible, since $k=-1$ has to be true for the above equation to hold. On the other hand, $(c_2,c_3)=(2,3)$ satisfies (\[realtight2\]) no matter what $k$ is equal to. Now suppose $6-c_2c_3 \neq 0$. Applying the quadratic formula to (\[realtight2\]) yields $$k = \frac{6c_2+6-3c_2c_3 \pm (6c_2-6-c_2c_3)}{12-2c_2c_3}.$$ The case corresponding to the minus sign is discarded, since in that case $k=1$, which is impossible for a distance-regular graph of diameter four. Hence $$\label{realtight3} k = \frac{6c_2+6-3c_2c_3+6c_2-6-c_2c_3}{12-2c_2c_3} = \frac{6c_2-2c_2c_3}{6-c_2c_3}.$$ If $c_2=1$ then rewriting (\[realtight3\]) yields $$c_3 = \frac{6(k-1)}{k-2}.$$ Since $k-1$ and $k-2$ are coprime, $k-2$ must divide 6, whence $k \in \{3,4,5,8\}$. Since $c_3<k$, it follows $k=8$, whence $c_3=7$. If $c_2 \geq 2$ then let $\rho:=k/c_2$. By (\[realtight3\]) we have $$\rho = \frac{6-2c_3}{6-c_2c_3},$$ which implies $$\frac{\rho c_2 - 2}{\rho - 1} = \frac{6}{c_3}.$$ Since $\rho-1>0$, $$\frac{6}{c_3} = \frac{\rho c_2 - 2}{\rho - 1} \geq \frac{2\rho - 2}{\rho - 1} = 2,$$ whence $c_3 \leq 3$. Since $c_2 \geq 2$, Lemma \[c2bound\] implies that $c_2=2$ and $c_3=3$. But this contradictions our assumption that $6-c_2c_3 \neq 0$. The arguments above showed that, if $\|S\|=(k^2-3k+8)k/6$ then one of the following two conditions holds: - $c_2=2$ and $c_3=3$. - $k=8$, $c_2=1$ and $c_3=7$. However, graphs with $(k,c_2,c_3)=(8,1,7)$ do not produce $\{0,\alpha\}$-sets in $\mathbb{R}^8$. Indeed, such a graph has nontrivial eigenvalue $\theta_1=\sqrt{8}$ and hence $\alpha=1/8$. On the other hand, any flat unit vectors in $\mathbb{R}^8$ must have entries $\pm 1/\sqrt{8}$, and the only possible angles between such vectors are 9/16, 1/4, 1/16 and 0. So the real flat bound is not applicable to the graphs with $(k,c_2,c_3)=(8,1,7)$. Meanwhile, Brouwer, Cohen and Neumaier [@BCN Sect. 4.3, p.153-154] showed that any bipartite distance-regular graph of diameter four with $c_2=2$ and $c_3=3$ satisfies $k \in \{4,8,24\}$. These graphs are precisely the 4-cube, the folded 8-cube and the coset graph of the extended binary Golay code. The case for $\mathbb{C}^k$ is proven similarly. Let $S$ be a $\{0,\alpha\}$-set in $\mathbb{C}^k$ constructed from a distance-regular graph ${\Gamma}$ using the construction in Lemma \[constr\]. Then $S$ has the largest size $(k^2-k+2)k/2$ if and only if ${\Gamma}$ is the 8-cycle. The graph ${\Gamma}$ has to be bipartite and have diameter four. Let $k$, $c_2$ and $c_3$ be its intersection numbers. If $\|S\|=(k^2-k+2)k/2$ then by Lemma \[constr\] and Lemma \[vertices\], $$\frac{k(k^2-(c_2+1)k+c_2(c_3 +1))}{c_2 c_3} = \frac{(k^2-k+2)k}{2},$$ which is equivalent to $$\label{comptight1} (2-c_2c_3) k^2 - (2c_2+2-c_2c_3)k + 2c_2 = 0.$$ If $2-c_2c_3=0$ then $c_2=1$, $c_3=2$ and $k=1$, which is impossible, so $2-c_2c_3 \neq 0$. Applying the quadratic formula to (\[comptight1\]) yields $$k = \frac{2c_2+2-c_2c_3 \pm (2c_2+c_2c_3-2)}{4-2c_2c_3}.$$ The case with the minus sign is disgarded, for otherwise $k=1$, which is impossible. Therefore $$k = \frac{2c_2+2-c_2c_3+2c_2+c_2c_3-2}{4-2c_2c_3} = \frac{2c_2}{2-c_2c_3}.$$ The fact that $k$, $c_2$ and $c_3$ are all positive integers then implies that $$c_2 = c_3 = 1 \quad \text{and} \quad k=2.$$ The unique bipartite distance-regular graph of diameter four with $k=2$, $c_2=1$ and $c_3=1$ is the 8-cycle. Acknowledgment\ I thank Chris Godsil and Aidan Roy for many helpful discussions on the subject of complex lines.	{ "pile_set_name": "ArXiv" }
--- author: - \| Bijan Bagchi$^\circ$ and Andreas Fring$^\bullet$\ $\circ$ Department of Physics, Shiv Nadar University, Dadri, UP 201314, India\ $\bullet$ Department of Mathematics, City, University of London,\ $\,\,$ Northampton Square, London EC1V 0HB, UK\ E-mail: [email protected], [email protected] title: 'Quantum, noncommutative and MOND corrections to the entropic law of gravitation' --- Introduction ============ The weak equivalence principle is a well known concept, see e.g. [berry1989]{}, that identifies the inertial mass $m_{I}$ occurring in Newton’s second law of motion $\vec{F}=m_{I}\vec{a}$, with the gravitational mass $m_{G}$ in Newton’s inverse square law of gravitation. The latter accounts for the attractive force between a body of mass $m_{G}$ at the position $\vec{r}$ and $n$ different others specified by their masses $m_{i}$ occupying positions ${\vec{r}_{i}}$, $i=1,2,...,n$, as $$\vec{F}=-\sum_{i}\frac{Gm_{G}m_{i}(\vec{r}-\vec{r_{i}})}{\|\vec{r}-\vec{r_{i}}\|^{3}},\quad i=1,2,...,n, \label{1}$$ where $G$ is the gravitational constant. Equating these two expressions for the force when using $m_{I}=m_{G}$ readily yields an expression for the acceleration of a particle in a gravitational field $$\vec{a}=-G\sum_{i}\frac{m_{i}(\vec{r}-\vec{r_{i}})}{\|\vec{r}-\vec{r_{i}}\|^{3}},\quad i=1,2,...,n. \label{2}$$ A consequence of this is the curious feature, known since the time of Galileo, that objects that are dropped from some height, say from the top of a building, will arrive at the same time on the ground as long as their motion is not affected by air resistance or other disturbances, i.e. they fall at the same rate with equal accelerations. By invoking the holographic principle in the vicinity of a black hole E. Verlinde [@verlinde2011origin; @verlinde2017emergent] demonstrated recently that Newton’s second law of motion for a particle when confronted with the law of gravitation for celestial bodies, can be viewed as entropic in character. Employing the holographic argument that any information of the black hole, which is imagined as a sphere of Schwarzschild radius $R$, can emerge only from its surface (because what is inside the black hole remains totally intractable to an outside observer), and using the equipartition rule gave strong thermodynamical evidence to justify such a claim. Note that the boundary of the black hole sphere which is basically an equipotential surface acts as a holographic screen also popularly referred to as the event horizon. Assuming the change of entropy $\Delta S$ near the holographic screen to be linear in the displacement $\Delta x$ of a test particle, Verlinde suggested for $\Delta S$ the relation $$\Delta S=2\pi k_{B}\frac{mc}{\hbar }{\Delta x}=2\pi k_{B}\frac{\Delta x}{\lambdabar}, \label{V1}$$ with $\lambdabar:=\hbar /mc$ denoting the reduced Compton wavelength, $k_{B}$ the Boltzmann constant, $\hbar $ the reduced Planck constant and $c$ the speed of light. The change in entropy was also assumed to generate an entropic force $F$ acting on the particle to be of the form $$F\Delta x=T\Delta S, \label{V2}$$ where $T$ is the temperature. Taking $T$ to be given by Unruh’s temperature [@unruh1976notes] for an accelerated observer, namely $$k_{B}T=\frac{\hbar }{2\pi }\frac{a}{c}, \label{V3}$$ where $a$ stands for the acceleration of the particle, consistency gave Newton’s formula for the second law of motion $F=ma$ when combining (\[V1\])-(\[V3\]). Next, to arrive at the law of gravitation by restricting to the spherical boundary having an area $A=4\pi R^{2}$, $R$ being the radius of the sphere, he made use of the holographic principle that the total number of bits making up the maximally storage space is proportional to $A$. This gave the number $N$ of used bits as $N=Ac^{3}/G\hbar $, $G$ being identified as Newton’s gravitational constant as in (\[1\]). Using the equipartition rule for the average energy for every bit, $E=\frac{1}{2}Nk_{B}T=mc^{2}$, $M$ denoting the mass in the part of the space enclosed by the holographic screen, yielded the well known Newton’s law of gravitation: $F=-GMm/R^{2}$. It is worthwhile to recall some history behind Verlinde’s formulation. First, an early work by Jacobson [@jacobson1995] attempted to derive Einstein’s relativistic equations from pure thermodynamical considerations by making the constant of proportionality between the area and entropy universal following a preceding work of Bekenstein [@bekenstein1973] who in turn looked at the entropy of any isolated system to be bounded by its area. Second, Padmanabhan [@padmanabhan2004; @padmanabhan2010] arrived at a result of gravitational acceleration by reversing Unruh’s temperature-acceleration relation. Both Bekenstein as well as Padmanabhan’s discussions were carried out in a fully relativistic frameworks which apparently have no analogue in Verlinde’s non-relativistic formulation. An important observation made by Verlinde was that since the maximally allowed information stored in any continuum volume of space can only be finite, it is not sensible to talk of localizing a particle with an infinite degree of accuracy. Even though in the end he obtained the classical results which were devoid from the appearance of $\hbar $, it is pertinent to bear in mind that while the individual expressions of the change in entropy as well as Unruh’s temperature contain an explicit presence of ${\hbar }$, the latter fortunately cancels out when we look for a force-acceleration relationship. A question then naturally arises as to what happens if we seek higher corrections to the uncertainty principle as is needed to accommodate various modifications of the short distance structure in quantum theories that attempt to incorporate gravity [@Kempf2]. This in turn calls for an introduction of a so-called minimal length beyond which a localization of space-time events is no longer possible. With a minimal observable length $\Delta x=0$ that is characteristic of a physical quantum state, it is evident that an eigenstate with a zero-uncertainty in position can no longer depict a physical state. Models in string theory [@String1; @String2] as well as in quantum gravity [@Rovelli] do indeed support the existence of such a minimal length [@Hossenfelder]. Alternatively one may also view $\Delta x$ as a change in the black hole radius [@xiao2010]. Our manuscript is organized as follows: In section 2 we discuss how Verlinde’s argument can be modified by including quantum or noncommutative corrections in the energy and/or the entropy. We compute the correction terms to the gravitational force for various choices of the free parameters in the standard approach and generalized Heisenberg’s uncertainty relations. We use concrete expressions for the uncertainties obtained from different types of coherent states whose wave functions we derive explicitly in section 3 using Jackiw’s direct and analytic method. In section 4 we derive some Lagrangians and Hamiltonians for the corrected entropic force, which turn out to be explicitly time-dependent. Demanding that the damping to be small provides a further criterion that allows to exclude certain choices of the free parameters. We state our conclusions in section 5. Quantum corrections to Newton’s second law ========================================== Let us now see how the above effects might be incorporated into the above reasoning by modifying the equations (\[V1\])-(\[V3\]) and exploiting Verlinde’s observation that a strict localization of the test particle is not possible. As argued by Santos and Vancea [@santos2012] the total energy also depends on the momentum $p$ in form of the kinetic energy or possibly in a more general way. This means that the uncertainty in the total energy $\delta E$ could also acquire a term that depends on the uncertainty in the momentum $\delta p$. We assume here the form$$\delta E=F\delta x+\alpha \frac{p}{m}\delta p=T\delta S, \label{E1}$$to be valid at thermal equilibrium where $\alpha $ is dimensionless, possibly a constant. Note that $T\delta S$ is not a perfect differential. To counterbalance the additional term one also needs to modify the expression for $\delta S$. Here we take $$\delta S=2\pi k_{B}\left( \frac{1}{\lambdabar}\delta x+\frac{\beta }{mc}\delta p\right) , \label{E2}$$with a dimensionless parameter $\beta $ introduced to the equation. We keep the equation for the Unruh temperature (\[V3\]) unchanged. In the limit $\alpha ,\beta \rightarrow 0$ we recover the equations (\[V1\])-(\[V3\]). In [@santos2012] the options $\alpha =1$, $\beta =1$ and $\alpha =1$, $\beta =$ $p/mc$ were explored. Combining the equations (\[V3\]), (\[E1\]) and (\[E2\]) leads easily to a corrected expression for the force $$F=ma+\left( \beta \frac{\lambdabar}{c}a-\alpha \frac{p}{m}\right) \frac{\delta p}{\delta x}=ma+F^{\text{cor}}. \label{QF}$$Our task is now to interpret the additional term $F^{\text{cor}}$ and test which choices of $\alpha $ and $\beta $ are permissible. Keeping in mind that the variations $\delta p$ and $\delta x$ are interpreted as the uncertainties in a simultaneous measurement of $x$ and $p$ one can employ the standard Robertson version of Heisenberg’s uncertainty relation for a simultaneous measurement of two noncommuting operators $A$ and $B$ $$\left( \Delta A\right) ^{2}\left( \Delta B\right) ^{2}\geq \frac{1}{4}\left\vert \left\langle \left[ A,B\right] \right\rangle \right\vert ^{2}, \label{HU}$$ for $A=x$, $B=p$ with $[x,p]=i\hbar $ to make estimations about the ratio $\delta p/\delta x$. In [@santos2012; @ghosh2010planck] $\delta x\delta p\geq \hbar /2$ was used at saturation point of the lower bound, i.e. it was assumed that the test particle is in a coherent, possibly squeezed, state and $\delta p$ was traded for $\hbar /(2\delta x)$. This leaves the resulting expression with an unknown factor $\delta x^{-2}$. Furthermore it was suggested in [@santos2012] that the classical limit is obtained by the simultaneous limit $\hbar \rightarrow 0$, $\delta p\rightarrow 0$. This gives indeed the classical expression, but the proposed prescription lacks further justification. It addition, demanding $\delta p\rightarrow 0$ is ambiguous as one might as well require the simultaneous limit $\hbar \rightarrow 0$, $\delta x\rightarrow 0$. The question is why is the classical limit not obtainable simply from $\hbar \rightarrow 0$? Here we go a step further trying to achieve just that. Our main assumption is that we take the test particle to be in a specific state so that $\delta p/\delta x$ acquires a concrete value. Corrections from canonical coherent states and MOND dynamics ------------------------------------------------------------ Let us now assume at first the test particle to be in a standard canonical coherent state for which we have the well known expressions, see e.g. [gazeau2009coherent]{}, $$\delta x\delta p=\hbar /2,\qquad \ \ \ \ \ \ \text{with \ }\delta x=\sqrt{\frac{\hbar }{2m\omega }}\text{, ~\ }\delta p=\sqrt{\frac{m\omega \hbar }{2}.}$$Using these equalities we can evaluate the ratio $\delta p/\delta x$ for these states, such that the quantum corrected force (\[QF\]) becomes$$F_{\text{coherent}}=ma+\left( \beta \frac{\lambdabar}{c}a-\alpha \frac{p}{m}\right) m\omega =ma+F_{\text{coherent}}^{\text{cor}}. \label{Fco}$$We have now various options for the choice of $\alpha $ and $\beta $. We may take $\alpha \neq 0$, which suggests that the second term in (\[Fco\]) becomes a pure quantum correction with $\alpha \sim \hbar $. For instance, $\alpha =\omega \hbar /mc^{2}$ is an admissible choice corresponding to a quantum correction in the energy (\[E1\]). Having introduced an additional quantum correction it is not a surprise that we obtain also a quantum correction in $F$. Taking $\alpha =0$ the correction term $F_{\text{coherent}}^{\text{cor}}$ becomes a genuine quantum correction and the classical limit is simply reached by taking $\hbar \rightarrow 0$. One might take $\beta =1$ in this case, so that a classical correction in $\delta S$ has led us to a quantum correction in $F$. Finally one may wonder if one can reverse the setting of the previous example and obtain a classical correction to $F$ from an additional quantum term in $\delta S$. This is similar to Verlinde’s original argument in which also the $\hbar $ from the expression for $\delta S$ has cancelled the $\hbar $ appearing in the Unruh temperature. An example for such a classical theory of modified Newtonian dynamics (MOND) was proposed in $1983$ by Milgrom [@milgrom1983; @milgrom2015road] for situations when the gravitational force shows a marked departure from the conventional Newtonian expectation at low acceleration. The MOND theory, or so it is called, is typically applicable to scales of acceleration (less than the threshold value of $a_{0}\thickapprox 1.2\times 10^{-10}ms^{-2}$) which are rather small compared to what is observed in the solar system and perhaps relevant towards explaining galactic scale phenomena [@smolin2017mond]. It has been noted [@tortora2017] that the Milgrom scheme might be justified as an alternative means to solve for the dark matter problem which is still to find any experimental support, the prime reason being its rather poor coupling with visible matter. A recent paper by Verlinde [verlinde2017emergent]{} has sought to explore this issue using the standard thermodynamical arguments as a basis. A MOND theory has the force form given by a deformed acceleration [gozzi2017newton]{}$$F_{\text{MOND}}=ma\mu (a_{0}/a),\qquad \text{with \ }\mu (a_{0}/a)=\frac{1}{1+a_{0}/a}, \label{MF}$$where $a_{0}$ is the aforementioned small acceleration. It is, however, only a phenomenological form but worthwhile to note that in place of the usual expression of the acceleration as is implied by (\[1\]) namely, $a=MGr^{-2} $, in MOND, a test particle which is at a distance $r$ from a large mass $M$ is subject to the acceleration a given by $a^{2}/a_{0}=MGr^{-2}$, where $a\ll a_{0}$. Other variants of a modified Newtonian equation have been proposed in the literature [@bhat2017m], but we do not discuss them here. Choosing now $\alpha =0$ and $\beta =-c(\omega \lambdabar)(1+a/a_{0})$ we obtain precisely the form of the MOND force (\[MF\]) with modified acceleration. Remarkably this means with a corrected entropy, just taking the uncertainty of a particles position into account, we may interpret the force in a MOND theory as a classically emerging entropic force. Corrections from minimal value and minimal product coherent states {#corrmvmp} ------------------------------------------------------------------ Let us now assume our system to be in a noncommutative space on which the canonical Heisenberg commutation relations are generalized to [Kempf2,AFBB]{}$$\left[ x,p\right] =i\hbar \left( 1+\tau p^{2}\right) , \label{defHeisenberg}$$with $\tau \in \mathbb{R}^{+}$ denoting the dimensionful, i.e. $\left[ \tau \right] =s^{2}/m^{2}$, noncommutative constant. Such a generalized commutation relation arises as a particular case of an extended q-deformation given by the commutator [@AFBB] $$\lbrack x,p]=i\hbar q^{g(N)}+\frac{i\hbar }{4}(q^{2}-1)\left( \frac{x^{2}}{\delta ^{2}}+\frac{p^{2}}{\gamma ^{2}}\right) ,\quad \delta ,\gamma \in \mathbb{R} \label{qdef}$$ where $g$ is some arbitrary function of the number operator $N$ defined as the product of the creation and annihilation operator for the harmonic oscillator. Taking $g(N)=0$ and parametrizing the deformation parameter $q$ in the form $q=e^{2\tau \gamma ^{2}}$, we found in the limit $\gamma \rightarrow 0$, the $\tau $-corrected form (\[defHeisenberg\]). On a noncommutative space one needs to make an important distinction between what we refer to as minimal factor coherent states (mfco) and minimal product coherent states (mpco). The former are the states for which the minimal value is reached for one of the factors in the uncertainty relation, e.g. $\delta x$ in which case it is referred to as minimal length coherent state (mlco). In contrast,the mpco-states is a state for which the entire product in the uncertainty relation, e.g. $\delta x\delta p$, is minimized. Assuming now the test particle to be in a mlco-state, we have $$\delta x_{\min }\delta p=\hbar ,\qquad \ \ \ \ \ \ \text{with \ }\delta x_{\min }=\hbar \sqrt{\tau }\text{, ~\ }\delta p=\frac{1}{\sqrt{\tau }}. \label{xpmin}$$These values are easily obtained for (\[defHeisenberg\]) with (\[HU\]), see the next section for the derivation. We will comment also in more detail on the construction of meaningful explicit states that produce these values. Using (\[xpmin\]) in (\[QF\]) the noncommutatively corrected force becomes$$F_{\text{mlco}}=ma+\left( \beta \frac{\lambdabar}{c}a-\alpha \frac{p}{m}\right) \frac{1}{\hbar \tau }=ma+F_{\text{mlco}}^{\text{cor}}.$$Since $\hbar $ as well as $\tau $ are very small, $\hbar \tau \ll 1$, the correction term becomes very large, which does not make sense as we expect only a small modification. However, by demanding that $\alpha \sim (\hbar \tau )^{2}$ and $\beta \sim \hbar \tau $ this can be achieved. For instance, $\alpha =(\hbar \tau m\omega )^{2}$ and $\beta =\hbar \tau m\omega $ is an admissible choice from a dimensional point of view. When using this option the modifying terms proportional to $\delta p$ in (\[E1\]) and (\[E2\]) acquire a new interpretation. They are now noncommutative deformations that give rise to an entropically emergent noncommutative space-time structure. As is clear from the argument above, the deformed equation ([defHeisenberg]{}) is only one of many possibilities obtainable from ([qdef]{}) or other approaches. In [@AFBB] also higher order $\tau $-deformations were explored as for instance$$\left[ x,p\right] =i\hbar \left( 1+\tau p^{2}+\frac{1}{2}\tau ^{2}p^{4}\right) , \label{t2}$$with $\delta x_{\min }=1.14698\hbar \sqrt{\tau }$, $\delta p=0.740664/\sqrt{\tau }$ going up to$$\left[ x,p\right] =i\hbar e^{\tau p^{2}}, \label{t3}$$with $\delta x_{\min }=1.16582\hbar \sqrt{\tau }$, $\delta p=1/\sqrt{2\tau }$. It was noted in [@dorsch2012] that $1.16582\ldots $ can be expressed as $\sqrt{e/2}$. This means for the higher deformations the corrected force becomes $$F_{\text{mlco}}(\tau )=ma+\kappa _{\tau }F_{\text{mlco}}^{\text{cor}},$$where $\tau $ in $\kappa _{\tau }$ indicates the order in the deformation. So we have $\kappa _{1}=1$, $\kappa _{2}=0.6458$, …, $\kappa _{\infty }=e^{-1/2}=0.6065$. One might suspect a different outcome in regard to the previous argument when using mpco-states. Assuming still the emerging space-time structure to be noncommutative with deformed uncertainty relation (\[defHeisenberg\]) we obtain in this case $$\left( \delta x\delta p\right) _{\min }=\frac{2}{3}\hbar ,\qquad \ \ \ \ \ \ \text{with \ }\delta x=\frac{2}{\sqrt{3}}\hbar \sqrt{\tau }\text{, ~\ }\delta p=\frac{1}{\sqrt{3}\sqrt{\tau }}. \label{xpminmin}$$We note that indeed for these values the product $\delta x\delta p$ is smaller and $\delta x$ is larger in (\[xpminmin\]) when compared to ([xpmin]{}). As it is less obvious how to derive these values, we will comment in detail on the derivation in the next section. The correction term to the force is in this case half the correction term $F_{\text{mpco}}=ma+F_{\text{mlco}}^{\text{cor}}/2$ which is a further reduction when compared to $\kappa _{\infty }$. As the overall dependence on $\hbar $ and $\tau $ is unchanged the general discussion and interpretation is the same as for the mlco-states resulting from (\[defHeisenberg\]). Minimal length and minimal value coherent states ================================================ We will now provide the details on how the values for (\[xpmin\]) and ([xpminmin]{}) are obtained including a derivation of the associated explicit wavefunctions. Since the discussion for the values in (\[xpminmin\]) has not been presented elsewhere before, we provide also a general discussion on the appropriate method to be used. We distinguish here two fundamental questions regarding the measurement of an observable $A$ in any quantum mechanical system: a) what minimal value can the variance $\left( \Delta A\right) ^{2}:=$ $\left\langle A^{2}\right\rangle -\left\langle A\right\rangle ^{2}=\left\langle \hat{A}^{2}\right\rangle $ with $\hat{A}=A-\left\langle A\right\rangle $ or the standard deviation $\Delta A$ take and b) what is the associated state $\left\vert \psi \right\rangle _{\min ,A}$ in $\left\langle A\right\rangle :=\left\langle \psi \right\vert A\left\vert \psi \right\rangle $ that realises that minimum? These questions are then naturally extended to simultaneous measurements related to two or more operators. For two operators $A$ and $B$ the questions a) and b) have now three variants, i.e. what are the minimal values and corresponding states for $\Delta A$, $\Delta B$ or the product $\Delta A\Delta B$ within the simultaneous measurement of $A$ and $B$? For three operators $A$, $B$ and $C$ this extends to seven variants, i.e. what are the minimal values and corresponding states for $\Delta A$, $\Delta B$, $\Delta C$, $\Delta A\Delta B$, $\Delta A\Delta C$, $\Delta B\Delta C$ or $\Delta A\Delta B\Delta C$ within the simultaneous measurement of the expectation values of all three operators $A$, $B$ and $C$? Naturally the states that realise these different possibilities are usually non-identical. A typical example for the measurement of two operators $A$ and $B$ are the aforementioned position and momentum operators $x$ and $p$, respectively. To measure the position of a particle in space is an example for three operators corresponding to the coordinate components $x$, $y$ and $z$. This measurement is of course trivial in conventional space, but becomes nontrivial and interesting when one considers a noncommutative space in which coordinate component operators do not commute, see e.g. [AFLGBB,DFG]{} for concrete examples and [@Laurerev] for a recent review. The question regarding the minimal values of $\Delta A$ and $\Delta B$ is not challenging when the commutator on the right hand side of (\[HU\]) is a constant, as one can always achieve $\Delta A\rightarrow 0$ or $\Delta B\rightarrow 0$ by taking $\Delta B\rightarrow \infty $ or $\Delta A\rightarrow \infty $, respectively, and still respect the lower bound. This is the standard scenario of Heisenberg’s uncertainty relation in which one must give up all the information about $A$ or $B$ to measure $B$ or $A$, respectively, with absolute precision. However, when the resulting commutator on the right hand side of the inequality involves operators, i.e. when the lower bound becomes a function of $A$ and/or $B$ it is no longer possible to carry out the limits in this trivial manner. Such a scenario arises for theories formulated on certain noncommutative spaces as discussed in section \[corrmvmp\]. In that case one may assume that the minima are reached for coherent states that is at equality in (\[HU\]). By defining the function $f(\Delta A,\Delta B):=\Delta A\Delta B-\frac{1}{2}\left\vert \left\langle \left[ A,B\right] \right\rangle \right\vert $, the critical values are simply obtained by simultaneously solving $f(\Delta A,\Delta B)=0$ and $\partial _{\Delta B}f(\Delta A,\Delta B)=0$ for $\Delta A $ or $\partial _{\Delta A}f(\Delta A,\Delta B)=0$ for $\Delta B$ from which one can identify the minimum $\Delta A_{\min }$ or $\Delta B_{\min }$, respectively, see e.g. [@AFBB]. We note that there is no reason to expect that the product of the individual minimal values $\Delta A_{\min }$ and $\Delta B_{\min }$ is equal to the minimal value of the product $\left( \Delta A\Delta B\right) _{\min }$. From this argument we do not obtain any information about the states involved. The direct versus the analytic method ------------------------------------- Let us now see how to derive the associated minimizing states $\left\vert \psi \right\rangle _{\min ,A}$, $\left\vert \psi \right\rangle _{\min ,B}$ and $\left\vert \psi \right\rangle _{\min ,AB}$ for which these minima are reached. Following Jackiw [@jackiwmin] we recall the difference between the direct and analytic method that determine the state $\left\vert \psi \right\rangle $ minimizing the uncertainty product. Making the same assumption as above, namely that the minimum is reached for equality, the direct method follows from a comparison of the Schwartz and triangle inequality for $\left\vert \left\langle \hat{A}\hat{B}\right\rangle \right\vert ^{2}$. It then consists of solving $$\left[ A-\alpha +\frac{\left\langle \left[ A,B\right] \right\rangle }{2b^{2}}(B-\beta )\right] \left\vert \psi \right\rangle =0 \label{DM}$$for $\left\vert \psi \right\rangle $ involving the three free parameters $\alpha :=\left\langle A\right\rangle $, $\beta =:\left\langle B\right\rangle $ and $b^{2}:=\left\langle B^{2}\right\rangle -\left\langle B\right\rangle ^{2}$. Once the eigenvalue problem in (\[DM\]) is solved, these parameters are just computed in a self-consistent manner via their defining relations. They may be used to convert the solutions into proper square integrable functions and to minimize the desired quantities, that are either the separate minimal values $\Delta A_{\min }$ , $\Delta B_{\min }$ or the minimal product $\left( \Delta A\Delta B\right) _{\min }$. In the derivation of (\[DM\]) one makes two assumptions: First that the minimal state is reached for the equality sign in (\[HU\]) and second that the commutator $\left[ A,B\right] $ is a c-number rather than a q-number, that is a constant and not an operator. The analytic method on the other hand makes no assumptions about the right hand side in the inequality (\[HU\]). In that scheme one treats the left hand side as a functional, minimizing it together with the supplementary assumption that $\left\vert \psi \right\rangle $ is normalizable, i.e. one solves $$\frac{\delta }{\delta \left\langle \psi \right\vert }\left[ \left( \left\langle \psi \right\vert A^{2}\left\vert \psi \right\rangle -\left\langle \psi \right\vert A\left\vert \psi \right\rangle ^{2}\right) \left( \left\langle \psi \right\vert B^{2}\left\vert \psi \right\rangle -\left\langle \psi \right\vert B\left\vert \psi \right\rangle ^{2}\right) -m\left( \left\langle \psi \!\right. \left\vert \psi \right\rangle -1\right) \right] =0, \label{func}$$with Lagrange multiplier $m$, for $\left\vert \psi \right\rangle $. In a straightforward manner this leads to the eigenvalue problem$$\left[ \frac{\left( A-\alpha \right) ^{2}}{a^{2}}+\frac{\left( B-\beta \right) ^{2}}{b^{2}}\right] \left\vert \psi \right\rangle =2\left\vert \psi \right\rangle . \label{AM}$$As no assumption is made about the right hand side in (\[HU\]) an additional parameter $a^{2}:=\left\langle A^{2}\right\rangle -\left\langle A\right\rangle ^{2}$ enters the scheme when compared to the direct method. By re-expressing the direct method as$$\left[ \frac{\left( A-\alpha \right) ^{2}}{a^{2}}+\frac{\left( B-\beta \right) ^{2}}{b^{2}}\right] \left\vert \psi \right\rangle =2\frac{\left[ A,B\right] }{\left\langle \left[ A,B\right] \right\rangle }\left\vert \psi \right\rangle ,$$Jackiw [@jackiwmin] demonstrated that the two schemes coincide if and only if $\left\vert \psi \right\rangle $ is an eigenstate of the commutator $\left[ A,B\right] $. Thus when this is not the case at least one of the assumptions on which the direct method is based ceases to be valid. As there are no obvious analogs to the Schwartz and triangle inequality for three operators, it is not evident how to formulate the direct method for that situation. However, it is straightforward to generalize the analytic method to three observables, that is to minimize triple products of variances $\left( \Delta A\right) ^{2}\left( \Delta B\right) ^{2}\left( \Delta C\right) ^{2}$. Using the analogue to (\[func\]), simply with an additional factor $\left( \left\langle \psi \right\vert C^{2}\left\vert \psi \right\rangle -\left\langle \psi \right\vert C\left\vert \psi \right\rangle ^{2}\right) $ on the first term, one easily derives $$\left[ \frac{\left( A-\alpha \right) ^{2}}{a^{2}}+\frac{\left( B-\beta \right) ^{2}}{b^{2}}+\frac{\left( C-\gamma \right) ^{2}}{c^{2}}\right] \left\vert \psi \right\rangle =3\left\vert \psi \right\rangle , \label{triple}$$now with two additional free parameters $\gamma =\left\langle C\right\rangle $ and $c^{2}=\left\langle C^{2}\right\rangle -\left\langle C\right\rangle ^{2}$, see [@tripleWeigert] for an example computation and [ma2017exp]{} for an experimental verification. The advantage of the analytic over the direct method is that it is applicable a) irrespective of the nature of $\left[ A,B\right] $, i.e. resulting to a number or an operator b) even when the minimum is not reached for equality in (\[AM\]) and c) to generalizations of uncertainty relations involving more than two observables. Minimal value coherent states from direct method ------------------------------------------------ When one wishes to obtain more information about the wavefunctions $\left\vert \psi \right\rangle _{\min ,A}$, $\left\vert \psi \right\rangle _{\min ,B}$ and $\left\vert \psi \right\rangle _{\min ,AB}$ via the direct or analytic method one needs to specify a concrete representation for the operators involved. For the algebra (\[defHeisenberg\]) there are various meaningful representations $\Pi _{(i)}$ that we label by $i$. For instance, with regard to the standard inner product a non-Hermitian and a Hermitian one are $$x_{(1)}=(1+\tau p_{s}^{2})x_{s},~p_{(1)}=p_{s},\quad \text{and\quad }x_{(2)}=x_{s},~p_{(2)}=\frac{1}{\tau }\tan \left( \sqrt{\tau }p_{s}\right) ,~~~ \label{22}$$respectively. Here $x_{s}$ and $p_{s}$ are standard canonical variables satisfying $[x_{s},p_{s}]=i\hbar $. Naturally for $\tau \rightarrow 0$ one recovers the standard Heisenberg commutation relations. Models in terms of these variables and further representations have been studied in more detail in [@AFBB; @HrepvsnHrep]. As demonstrated in [@HrepvsnHrep] the corresponding physical quantities, namely expectation values for adjoint operators, are representation independent and one may simply chose the most suitable one for one’s purpose. Using now the non-Hermitian representation $\Pi _{(1)}$ in the equation for the direct method (\[DM\]) for the observables $A=x$ and $B=p$, we obtain in momentum space simply a first order differential equation $$\left[ i\hbar \left( 1+\tau p_{s}^{2}\right) \partial _{p_{s}}+i\hbar \frac{1+\tau b^{2}+\tau \beta ^{2}}{2b^{2}}(p_{s}-\beta )-\alpha \right] \psi _{d}(p_{s})=0, \label{ddiif}$$for the minimal state $\psi _{d}(p_{s})$. Setting the constant $\beta =0$, equation (\[ddiif\]) is easily solved to$$\psi _{d}(p_{s})=\left[ \frac{\sqrt{\tau }\Gamma \left( \frac{3}{2}+\frac{1}{2\tau b^{2}}\right) }{\sqrt{\pi }\Gamma \left( \frac{1}{2}+\frac{1}{2\tau b^{2}}\right) }\right] ^{1/2}\left( 1+\tau p_{s}^{2}\right) ^{-\frac{1}{4\tau b^{2}}-\frac{1}{4}}\exp \left[ -\frac{i\alpha \arctan \left( p_{s}\sqrt{\tau }\right) }{\hbar \sqrt{\tau }}\right] .$$The constant pre-factor is chosen so that these states are normalized with regard to the quasi-Hermitian inner product$$\left\langle \psi \!\right. \left\vert \psi \right\rangle _{\rho }:=\int\nolimits_{-\infty }^{\infty }\rho (p_{s})\psi ^{\ast }(p_{s})\psi (p_{s})dp_{s}=1, \label{product}$$with metric operator $\rho =\left( 1+\tau p_{s}^{2}\right) ^{-1}$. As this is by now well established in the literature we do not justify the choice of $\rho $ here any further, but instead refer the reader to [Urubu,Benderrev,Alirev,siegl2012metric,bagarello2013self,FabFring1,krejvcivrik2015pseudospectra]{} for the general formalism on how to construct meaningful metric operators and how to define consistent inner products for non-Hermitian systems. A well-known argument in [@Urubu] states that the metric becomes unique when two observables with specific properties are specified, see also [MGH]{}[^1]. Using the states $\psi _{d}$ in the expression $\left\langle .\right\rangle _{\rho }=\left\langle \psi _{d}\right\vert .\left\vert \psi _{d}\right\rangle _{\rho }$, we then easily compute the relevant expectation values$$\left\langle x\right\rangle _{\rho }=\alpha ,\quad \quad \left\langle x^{2}\right\rangle _{\rho }=\alpha ^{2}+\frac{\hbar ^{2}(1+\tau b^{2})^{2}}{4b^{2}},\quad \quad \left\langle p\right\rangle _{\rho }=0,\quad \quad \left\langle p^{2}\right\rangle _{\rho }=b^{2}. \label{expect}$$The values for $\left\langle x\right\rangle _{\rho }$, $\left\langle p\right\rangle _{\rho }~$and $\left\langle p^{2}\right\rangle _{\rho }$ are to be expected by definition from the formalism and the explicit computations are just consistency checks. Minimizing $\left( \Delta x\right) ^{2}$ as a function of $b$ we find $b=1/\sqrt{\tau }$, such that the minimal length becomes $\Delta x_{\min }=\hbar \sqrt{\tau }$, which coincides with the findings reported in (\[xpmin\]) and those in [@Kempf2]. For this value of $b$ we have $\Delta p=1/\sqrt{\tau }$ so that the product $\Delta x_{\min }\Delta p=\hbar $ does not saturate the lower bound. In the light of the discussion in the previous section this was to be expected and in fact the minimal value for $\Delta x\Delta p=\hbar /2(1+\tau b^{2})$, as well for $\Delta p=b$, would be obtained for $b=0$. However, while the corresponding minimal length state $\left\vert \psi \right\rangle _{\min ,p}=$ $\psi _{d}(b=1/\sqrt{\tau })$ is well-defined, $\psi _{d}(p_{s})$ is ill-defined for $b=0$. So the direct method does not allow us to compute the product coherent states $\left\vert \psi \right\rangle _{\min ,xp}$. Let us therefore employ the analytical method to determine them. Minimal product coherent states from the analytical method ---------------------------------------------------------- For the representation $\Pi _{(1)}$ the eigenvalue equation for the analytical method (\[AM\]) in momentum space becomes the second order differential equation $$\left[ -\frac{\hbar ^{2}\left( 1+\tau p_{s}^{2}\right) ^{2}}{a^{2}}\partial _{p_{s}}^{2}-\frac{2\hbar (i\alpha +\hbar p_{s}\tau )\left( 1+\tau p_{s}^{2}\right) }{a^{2}}\partial _{p_{s}}+\frac{\alpha ^{2}}{a^{2}}+\frac{(p_{s}-\beta )^{2}}{b^{2}}-2\right] \psi _{a}(p_{s})=0.$$One may verify that $\psi _{d}(p_{s})$ does not satisfy this equation, which is to be expected. Instead, setting $\beta =0$ this equation is solved in terms of associated Legendre polynomials of $P_{\ell }^{m}(x)$ and Legendre functions of the second kind $Q_{\ell }^{m}(x)$ $$\psi _{a}(p_{s})=\exp \left[ -\frac{i\alpha \arctan \left( p_{s}\sqrt{\tau }\right) }{\hbar \sqrt{\tau }}\right] \left[ c_{1}P_{\ell }^{m}(ip_{s}\sqrt{\tau })+c_{2}Q_{\ell }^{m}(ip_{s}\sqrt{\tau })\right] ,$$with $\ell =\sqrt{4a^{2}+\hbar ^{2}\tau ^{2}b^{2}}/(2b\tau \hbar )-1/2$ and $m=a\sqrt{1+2\tau b^{2}}/(b\tau \hbar )$. Setting $\ell $ and $m$ to small integer values we find the first meaningful solutions, in the sense of being nonvanishing real numbers, for the parameters $a$ and $b$ for $\ell =1$ and $m=2$. For those values we have $P_{1}^{2}(ip_{s}\sqrt{\tau })=0$ and $Q_{1}^{2}(ip_{s}\sqrt{\tau })=2/(1+\tau p_{s}^{2})$. Suitably normalized with regard to the inner product (\[product\]) the minimizing solution becomes$$\psi _{a}(p_{s})=\sqrt{\frac{8}{3\pi }}\frac{\tau ^{1/4}}{1+\tau p_{s}^{2}}\exp \left[ -\frac{i\alpha \arctan \left( p_{s}\sqrt{\tau }\right) }{\hbar \sqrt{\tau }}\right] .$$Using this solution we then easily compute the expectation values$$\left\langle x\right\rangle _{\rho }=\alpha ,\quad \quad \left\langle x^{2}\right\rangle _{\rho }=\alpha ^{2}+\frac{4\hbar ^{2}\tau }{3},\quad \quad \left\langle p\right\rangle _{\rho }=0,\quad \quad \left\langle p^{2}\right\rangle _{\rho }=\frac{1}{3\tau },$$for $\left\langle .\right\rangle _{\rho }=\left\langle \psi _{a}\right\vert .\left\vert \psi _{a}\right\rangle _{\rho }$. For these values the product of uncertainties $\left( \Delta x\Delta p\right) _{\min }=2\hbar /3$ is minimized by construction. Indeed this value is smaller than $\Delta x_{\min }\Delta p$ obtained from the direct method by just minimizing $\Delta x.$ On the other hand the uncertainty in $x$ computed from these states $\Delta x=2/\sqrt{3}\hbar \sqrt{\tau }$ is corrected by a factor $1.15$ and therefore slightly larger than the one computed from the direct method or the minimization of $f(\Delta x,\Delta p)$. Since the contributions from the two observables to the minimal product is not the same, i.e. $\Delta x=2\hbar \sqrt{\tau /3}$ and $\Delta p=1/\sqrt{3\tau }$, the states $\psi _{a}(p_{s})$ are squeezed coherent states for all values of $\alpha $. Lagrangian and Hamiltonian formulation of the entropic force law ================================================================ Let us now investigate the modified force equation (\[QF\]) further with a special focus on the question of which choices for $\alpha $ and $\beta $ are meaningful. We observed in the previous section that an important feature of the entropic force law is that it comes in two parts - one corresponding to the inertial term which is the so-called Newtonian or inertial mass $m_{I}$ times acceleration and the other due to quantum or classical correction $F^{\text{cor}}$ which influences it. The entire expression was derived purely on the basis of thermodynamical arguments and the use of the uncertainty principle for particular states. The latter sets a lower limit on the product of the uncertainties of position and canonical momentum but a canonical momentum (defined as the partial derivative of the Lagrangian with respect to the time-derivative of the generalized coordinate) has no place in Newtonian theory which is concerned only with the ordinary momentum given simply by mass times velocity. To keep things in order it would be reasonable to identify the $p$ appearing in (\[QF\]) with $\dot{x}$ whose rate of change is the acceleration $a=\ddot{x}$. With this understanding we re-express the modified force equation(\[QF\]) as $$F=P\ddot{x}-Q\dot{x}=-\frac{\partial V}{\partial x}, \label{FF}$$ where the quantities $P$ and $Q$ stand for $$P:=m+\beta \frac{\hbar }{mc^{2}}\frac{\delta p}{\delta x},\quad Q:=-\alpha \frac{\delta p}{\delta x}.$$ We also used in (\[FF\]) the definition $F=-\partial V/\partial x$, where $V$ is a potential. It is straightforward to see that when taking $$\mathcal{L}=\left( \frac{P}{2}\dot{x}^{2}-V\right) e^{-\frac{Q}{P}t} \label{LL}$$ to be a Lagrangian, the corresponding Euler-Lagrange equation of motion $$\frac{d}{dt}\left( \frac{\partial \mathcal{L}}{\partial \dot{x}}\right) =\frac{\partial \mathcal{L}}{\partial x}$$ is equivalent to (\[QF\]) or (\[FF\]) for time-independent $P$ and $Q$. We notice that $\mathcal{L}$ depends explicitly on the damping factor$$\frac{Q}{P}=-\frac{\alpha \omega }{m\omega \frac{\delta x}{\delta p}+\beta \frac{\hbar \omega }{mc^{2}}},$$which evidently needs to very small or vanishing altogether. We can use this fact to constrain possible choices for $\alpha $ and $\beta $ even further. For convenience we summarize previously discussed scenarios in table 1 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ trial states $\alpha $ $\beta $ $\delta x/\delta p$ $\left\vert Q/P\right\vert $ --------------------------- ------------------------------------------ ------------------------- ------------------------------- --------------------------------------------------------------- canonical coherent states $1$ $1$ $\frac{1}{m\omega }$ $\frac{\omega }{1+\hbar \omega /mc^{2}}\gg 1$ canonical coherent states $\frac{\hbar \omega }{mc^{2}}$ $1$ $\frac{1}{m\omega }$ $\frac{\omega }{1+mc^{2}/\hbar \omega }\gg 1$ canonical coherent states $0$ $1$ $\frac{1}{m\omega }$ $0$ MOND $0$ $-c(\omega $\frac{1}{m\omega }$ $0$ \lambdabar)(1+a/a_{0})$ mlco states $\left( \hbar \tau m\omega \right) ^{2}$ $\hbar \tau m\omega $\hbar \tau /\kappa _{\tau }$ $\frac{\hbar \tau m\omega ^{2}}{1/\kappa $ _{\tau }+\hbar \omega /mc^{2}}\ll 1$ mvco states $\left( \hbar \tau m\omega \right) ^{2}$ $\hbar \tau m\omega $2\hbar \tau $ $\frac{\hbar \tau m\omega ^{2}}{2+\hbar \omega /mc^{2}}\ll 1$ $ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ${\small {(}}$ The values in table 1 suggest that the framework discussed in [santos2012]{}, based on a quantum correction for the energy is inconsistent. However, two choices survive this simple test. First of all any choice with $\alpha =0$ and arbitrary nonvanishing but dimensionally acceptable $\beta $ is consistent. In particular that includes the values leading to a classical MOND theory. Furthermore, the deformations of the energy and entropy expressions corresponding to noncommutative deformations also yield consistent damping factors. Finally let us also comment on a Hamiltonian $\mathcal{H}$ that is readily derived from the Langrangian $\mathcal{L}$ when taking the associated canonical momentum $p$ (not to be confused with $\dot{x}$) to be $$p=\frac{\partial \mathcal{L}}{\partial \dot{x}}=\dot{x}Pe^{-\frac{Q}{P}t}.~~$$ This leads us to $$\mathcal{H}=\dot{x}\frac{\partial \mathcal{L}}{\partial \dot{x}}-L=\frac{p^{2}}{2P}e^{\frac{Q}{P}t}+Ve^{-\frac{Q}{P}t},$$ so that depending on the sign of $Q/P$ we obtain a damped kinetic term and amplified potential term, or vice versa, unless $Q=0$. Evidently the explicit time dependence in the coefficients provides a major hindrance for the quantization of this system. However, one might follow for instance recent work, using a modified Prelle-Singer approach, in which explicit time-independent first integrals have been identified in different parameter regimes for the damped linear harmonic oscillator problem in order to facilitate the quantization procedure [@chandrasekar2007]. Conclusions =========== We have re-examined Verlinde’s argument on the derivation of the entropic law of gravitation by taking into account a possible modification in the form of adding a term linearly dependent on the momentum uncertainty $\delta p$ in the energy (\[E1\]) and/or the entropy (\[E2\]) uncertainty. These terms involve two free parameters $\alpha $ and $\beta $ that are only constraint by dimensional arguments. We then computed the general correction term for the entropic force resulting as a consequences of this modification for various choices for the parameters. The uncertainty ratio $\delta p/\delta x$, on which these terms depend, was computed exactly for some concrete coherent states with underlying standard and generalized Heisenberg uncertainty relation. In addition, we derive explicit expressions for the normalizable wavefunctions associated to minimal length and minimal value coherent states. Furthermore, we found that the Lagrangian and Hamiltonian associated to the entropic force are explicitly time-dependent in an exponential form. Demanding for obvious physical reasons that this exponential is damped provided us with a further criterion to select possible values for the parameters $\alpha $ and $\beta $. On a standard space with conventional commutation relations for $x$ and $p$ we found consistent correction terms for $\alpha =0$ accompanied by any choice for $\beta $. The latter can be chosen to implement classical or quantum corrections in the entropy. Remarkably a specific version of the latter option allows for the emergence of a MOND theory (\[MF\]). Keeping $\alpha =0$ other choices for $\beta $ can lead to classical as well as quantum corrections. Any scenario with $\alpha \neq 0 $ and $\beta \neq 0$ leads to exponentially growing Langrangians and Hamiltonians and therefore appear to be inconsistent from a physical perspective. This includes the scenario advocated in [@santos2012]. However, when considering the equations in a noncommutative setting with deformed canonical commutation relations (\[defHeisenberg\]), (\[t2\]), (\[t3\]) dimensional consistency together with the requirement that the classical and noncommutative limit can be carried out by $\hbar \rightarrow 0 $ and $\tau \rightarrow 0$, respectively, leads to consist solutions with a very mild damping or amplifying factor in the Langrangian. One may say that this setting leads to an emergent theory of noncommutative space-time. There are many interesting issues left to investigate, as for instance to clarify the link to a relativistic theory and to extend the analysis to higher dimensions, using (\[triple\]), to name only two. Acknowledgments: One of us (BB) thanks Siddharth Seetharaman for useful comments. [99]{} M. V. Berry, , CRC Press, 1989. E. Verlinde, On the Origin of Gravity and the Laws of Newton, J. High Energy Phys. 2011(4), 29 (2011). E. Verlinde, Emergent gravity and the dark universe, SciPost Physics 2(3), 016 (2017). W. G. Unruh, Notes on black-hole evaporation, Phys. Rev. D 14(4), 870 (1976). T. Jacobson, Thermodynamics of spacetime: the Einstein equation of state, Phys. Rev. Lett. 75(7), 1260 (1995). J. D. Bekenstein, Black holes and entropy, Phys. Rev. D 7(8), 2333 (1973). T. Padmanabhan, Entropy of static spacetimes and microscopic density of states, Class. Quant. Gravity 21(18), 4485 (2004). T. Padmanabhan, Equipartition of energy in the horizon degrees of freedom and the emergence of gravity, Mod. Phys. Lett. A 25(14), 1129–1136 (2010). A. Kempf, G. Mangano, and R. B. Mann, Hilbert space representation of the minimal length uncertainty relation, Phys. Rev. D52, 1108–1118 (1995). D. Gross and P. Mende, String Theory Beyond the Planck Scale, Nucl. Phys. B303, 407 (1988). D. Amati, M. Ciafaloni, and G. Veneziano, Can Space-Time Be Probed Below the String Size?, Phys. Lett. B216, 41 (1989). C. Rovelli, Loop Quantum Gravity, Living Rev. Relativity 11, 5 (2008). S. Hossenfelder, Self-consistency in theories with a minimal length, Class. Quant. Grav. 23, 1815–1821 (2006). X.-G. He and B.-Q. Ma, Black holes and photons with entropic force, Chin. Phys. Lett. 27(7), 070402 (2010). M. A. Santos and I. V. Vancea, Entropic law of force, emergent gravity and the uncertainty principle, Mod. Phys. Lett. A 27(02), 1250012 (2012). S. Ghosh, Planck scale effect in the entropic force law, arXiv preprint arXiv:1003.0285 (2010). J.-P. Gazeau, , Wiley, 2009. M. Milgrom, A modification of the Newtonian dynamics as a possible alternative to the hidden mass hypothesis, The Astroph. Jour. 270, 365–370;371–383; (1983). M. Milgrom, Road to MOND: A novel perspective, Physical Review D 92(4), 044014 (2015). L. Smolin, MOND as a regime of quantum gravity, arXiv preprint arXiv:1704.00780 (2017). C. Tortora, L. Koopmans, and N. Napolitano, Testing Verlinde’s emergent gravity in early-type galaxies, arXiv preprint arXiv:1702.08865 (2017). E. Gozzi, Newton’s trajectories versus MOND’s trajectories, Phys. Lett. B 766, 112–116 (2017). A. Bhat, S. Dey, M. Faizal, C. Hou, and Q. Zhao, Modification of Schr[ö]{}dinger–Newton equation due to braneworld models with minimal length, Phys. Lett. B (2017). B. Bagchi and A. Fring, Minimal length in Quantum Mechanics and non-Hermitian Hamiltonian systems, Phys. Lett. A 373, 4307–4310 (2009). G. C. Dorsch and J. A. Nogueira, Maximally localized states in quantum mechanics with a modified commutation relation to all orders, Int. J. Mod. Phys. A 27(21), 1250113 (2012). A. Fring, L. Gouba, and B. Bagchi, Minimal areas from q-deformed oscillator algebras, J. Phys. A 43, 425202 (2010). S. Dey, A. Fring, and L. Gouba, PT-symmetric noncommutative spaces with minimal volume uncertainty relations, J. Phys. A45, 385302 (2012). L. Gouba, A comparative review of four formulations of noncommutative quantum mechanics, Int. J. of Mod. Phys. A 31(19), 1630025 (2016). R. Jackiw, Minimum Uncertainty Product, Number-Phase Uncertainty Product, and Coherent States, J. of Math. Phys. 9(3), 339–346 (1968). S. Kechrimparis and S. Weigert, Heisenberg uncertainty relation for three canonical observables, Phys. Rev. A 90(6), 062118 (2014). W. Ma, B. Chen, Y. Liu, M. Wang, X. Ye, F. Kong, F. Shi, S.-M. Fei, and J. Du, Experimental Demonstration of Uncertainty Relations for the Triple Components of Angular Momentum, Phys. Rev. Lett. 118(18), 180402 (2017). S. Dey, A. Fring, and B. Khantoul, Hermitian versus non-Hermitian representations for minimal length uncertainty relations, J. Phys. A: Math. and Theor. 46(33), 335304 (2013). F. G. Scholtz, H. B. Geyer, and F. Hahne, Quasi-Hermitian Operators in Quantum Mechanics and the Variational Principle, Ann. Phys. 213, 74–101 (1992). C. M. Bender, Making sense of non-Hermitian Hamiltonians, Rept. Prog. Phys. 70, 947–1018 (2007). A. Mostafazadeh, Pseudo-Hermitian Representation of Quantum Mechanics, Int. J. Geom. Meth. Mod. Phys. 7, 1191–1306 (2010). P. Siegl and D. Krej[č]{}i[ř]{}[í]{}k, On the metric operator for the imaginary cubic oscillator, Phys. Rev. D 86(12), 121702 (2012). F. Bagarello, From self-adjoint to non-self-adjoint harmonic oscillators: Physical consequences and mathematical pitfalls, Phys. Rev. A 88(3), 032120 (2013). F. Bagarello and A. Fring, Non-self-adjoint model of a two-dimensional noncommutative space with an unbound metric, Phys. Rev. A 88(4), 042119 (2013). D. Krej[č]{}i[ř]{}[í]{}k, P. Siegl, M. Tater, and J. Viola, Pseudospectra in non-Hermitian quantum mechanics, J. of Math. Phys. 56(10), 103513 (2015). D. P. Musumbu, H. B. Geyer, and W. D. Heiss, Choice of a metric for the non-Hermitian oscillator, J. Phys. A40, F75–F80 (2007). M. Znojil, I. Semor[á]{}dov[á]{}, F. R[ž]{}i[č]{}ka, H. Moulla, and I. Leghrib, Problem of the coexistence of several non-Hermitian observables in PT-symmetric quantum mechanics, Phys. Rev. A 95(4), 042122 (2017). V. Chandrasekar, M. Senthilvelan, and M. Lakshmanan, On the Lagrangian and Hamiltonian description of the damped linear harmonic oscillator, J. of Math. Phys. 48(3), 032701 (2007). [^1]: In [@znojil2017] this argument was incorrectly attributed our previous work [@AFBB].	{ "pile_set_name": "ArXiv" }
16.5cm 22.5cm 0.cm 0.cm 0.cm PRA-HEP-93/05\ May 1993 \ The importance of properly taking into account the factorization scheme dependence of parton distribution functions is emphasized. A serious error in the usual handling of this topic is pointed out and the correct procedure for transforming parton distribution functions from one factorisation scheme to another recalled. It is shown that the conventional $\overline{\rm {MS}}$ and DIS definitions thereof are ill-defined due to the lack of distinction between the factorisation scheme dependence of parton distribution functions and renormalisation scheme dependence of the strong coupling constant $\alpha_s$. A novel definition of parton distribution functions is suggested and its role in the construction of consistent next-to-leading order event generators briefly outlined. Introduction ============ During recent years significant progress in the determination of parton distribution functions (p.d.f.) in the nucleon has been achieved, basically as a result of new data [@CCFR; @NMC], combined with more sophisticated and reliable theoretical analyses [@TO; @CTEQ; @MRS]. In the CTEQ Collaboration [@CTEQ] a number of theorists, phenomenologists and experimentalists have combined their efforts in order to deal properly with all experimental and theoretical subtleties of quantitative QCD analysis of vast amount of data from various experiments and processes. In [@CTEQ; @MRS] p.d.f. are determined with high accuracy, unheard of just a few years ago. In such circumstances a careful reanalysis of various theoretical uncertainties is clearly needed. Although most of such uncertainties are discussed in sufficient detail in review papers like [@TO; @Levy], there is one which has not been so far satisfactorily covered in either these or in any other paper I am aware of. It concerns the factorization scheme (FS) dependence of finite order QCD predictions in processes involving hadrons in the initial state. The treatment of this ambiguity presented in [@TO; @Levy; @Tung] is incomplete and moreover contains an error in the very central point of the factorization mechanism. As the FS dependence of p.d.f. has so far obtained much less attention than it probably deserves, I discuss in this note several of its aspects, drawing on analogy with the much more publicised case of the renormalization scheme (RS) ambiguity of the running coupling constant (couplant) $\alpha_s$. I think that much of the confusion and misunderstanding that surrrounds this topic stems from the failure to distinguish these related but in principle separate uncertainties. The organization of this paper is as follows. In Section 2 the notation is introduced and some basic facts about the RS dependence of the couplant are recalled. The crucial point, i.e. the dependence of p.d.f. on the choice of the FS is discussed in Section 3, followed in Section 4 by a few critical remarks on the currently used p.d.f.. In Section 5 the merits of the so-called “zero” FS are discussed, and in particular it is shown how it can be used for the construction of consistent next-to-leading order (NLO) event generators. The results are summarized and conclusions drawn in Section 6. Remarks on renormalization scheme dependence ============================================ Before coming to the ambiguity in the definition of p.d.f. let me recall a few basic facts about RS dependence of the renormalized couplant $a(\mu)\equiv g^2(\mu)/4\pi^2$. In massles QCD (to which I restrict my attention) it obeys the equation $$\frac{{\mbox{d}}a(\mu)}{{\mbox{d}}\ln\mu}= -ba^2\left(1+c_1a(\mu)+ c_2a^2(\mu)+\cdots\right), \label {da}$$ where the coefficients $b,c_1$, are fixed by the number of quark flavours, while all the higher ones are essentially free, defining the so called [renormalization convention]{} (RC), RC=$\{c_i,i\geq2\}$. In the simplest case this couplant enters the perturbation expansion of a physical quantity $R$, depending on a single external momemtum $Q$, in the form $$R(Q)=a(\mu)\left(1+r_{1}(\mu/Q,c_1)a(\mu)+ r_{2}(\mu/Q,c_1,c_2)a^{2}(\mu)+\cdots\right). \label {r}$$ Although not written out explicitly, also the couplant $a(\mu)$ (when (\[da\]) is considered to the k-th order) depends on all $c_i,i\leq k-2$. Moreover, both the couplant and the coefficients $r_k$ depend also on the specification which of the infinite number of solutions to (\[da\]) we have in mind. Each of these solutions can be labelled, for instance, by the familiar $\Lambda$ parameter. Combining this last information with that on $c_i$ defines what is usually called the renormalization scheme. Only if this RS is fixed does the specification of the scale $\mu$ (together with $c_i,i\geq 2$) uniquely determine both the couplant and the coefficients $r_k$. Although I prefer the terminology advocated in [@PMS], where the the term “RS” is reserved for a unique specification of both $a$ and $r_k$ (for detailed discussion of this point, see [@my]), I adopt in the following the more conventional notation in order to stay in close contact with [@TO; @CTEQ; @MRS]. In this notation $\mu$ is set equal to some “natural” scale in the problem and the variation of the RS is parametrized by means of the corresponding $\Lambda_{\rm RS}$ and the coefficients $c_i,i\geq 2$. Considering now (\[da\],\[r\]) to the NLO, we are left with only one degree of freedom, corresponding to the variation of $\Lambda_{\rm RS}$. To make the following considerations as transparent as possible, let me set $c_1=0$ in the rest of this note. [^1] The internal consistency of perturbation theory implies, using (\[da\]), the following relation between $\mu, a$ and $r_1$: $$r_1(\mu/Q,{\rm RS})=b\ln\left(\frac{\mu} {\Lambda_{\rm RS}}\right)-\rho= \frac{1}{a}-\rho\;\;\Rightarrow a=\frac{1}{r_1+\rho} \label{a(r)}$$ where $\rho$ $$\rho=b\ln\left(\frac{Q} {\Lambda_{\rm RS}}\right)-r_1(\mu=Q,{\mbox{RS}}) \label{rho}$$ is the renormalization group (RG) invariant [@PMS], which contains the $Q$-dependence of $R(Q)$. As the variations of $\mu$ and the RS are actually two sides of the same coin, it is redundant to vary both of them. Without the los of generality we can fix $\mu$ (for instance by setting it equal to some external momentum, like $\mu=Q$) in (\[a(r)\]) and elsewhere and vary the RS only. On the other hand, as the “natural” scale in the problem is usually not so unambiguously defined after all, many authors vary both the scale $\mu$ and the RS. Although unnecessary complication, this procedure is certainly legal. One has, however, to keep in mind that without the specification of the RS the choice of $\mu$ doesn’t fix either the couplant $a(\mu)$ or the coefficients $r_k$. In different RS’s the same $\mu$ implies different $a(\mu/\Lambda_{\rm RS})$ and $r_k$ and thus the choice of the RS is as important as that of $\mu$. With this in mind let me continue to label the RS by means of $\Lambda_{\rm RS}$, but keep $\mu$ still as a free parameter. Eq. (\[a(r)\]) furthermore suggests that instead of $\Lambda_{\rm RS}$, the value of $r_{\rm 1}$ can equally well serve the purpose of labelling the various RS. Substituting (\[a(r)\]) into (\[r\]) and truncating it to the NLO we get $$R^{\rm NLO}(\rho,r_1)=\frac{2r_1+\rho}{(r_1+\rho)^2} =\frac{1}{\rho} \left(\frac{1+2r_1/\rho}{(1+r_1/\rho)^2}\right) \label{rnlo}$$ as an explicit function of $r_1$ and $\rho$. The obvious consequence of the nontrivial dependence of $R^{\rm NLO}$ on the RS is that it would be a profound mistake to “transform” the couplant from one RS into another by means of equating the NLO (in fact any finite order) approximation to (\[r\]) in two different RS, say RS$^{(1)}$, RS$^{(2)}$, i.e. by solving the equation $$a({\rm RS}^{(1)})(1+r_1({\rm RS}^{(1)})a({\rm RS}^{(1)}))= a({\rm RS}^{(2)})(1+r_1({\rm RS}^{(2)})a({\rm RS}^{(2)})) \label{error}$$ Assuming (\[a(r)\]) in RS$^{(1)}$ and expressing $r_1({\rm RS}^{(2)})$ in terms of $a^{(1)}=a({\rm RS}^{(1)}), a^{(2)}=a({\rm RS}^{(2)})$ we get from (\[error\]) $$r_1({\rm RS}^{(2)})=\frac{2a^{(1)}/a^{(2)}-1}{a^{(2)}}- \rho\frac{a^{(1)}}{a^{(2)}} \label{trans}$$ which yields the correct relation (\[a(r)\]) between $r_1({\rm RS}^{(2)})$ and $a^{(2)}$ only for the trivial case $a^{(1)}=a^{(2)}$. In other words imposing the relation (\[error\]) leads to inconsistency for any nontrivial RG transformation! In the particular case when RS$^{(2)}$ is defined by means of the effective charges approach of [@Grunberg], corresponding to $r_1^{\rm ECH}=0$, (\[error\]) implies $$a^{\rm ECH}=\frac{1}{\rho}\left(\frac{1+2r_1^{(1)}/\rho} {(1+r_1^{(1)}/\rho)^2} \right) \label{aech}$$ while the correct relation reads $a^{\rm ECH}=1/\rho$. Parton distributions in general factorization scheme ==================================================== In processes involving hadrons in the initial state, there is, beside the mentioned RS dependence of the couplant, another kind of ambiguity, concerning the definition of p.d.f. beyond the leading order [@Politzer]. Again, to simplify the discussion as much as possible, I restrict the discussion to the nonsinglet (NS) quark distribution functions as revealed in the lepton-nucleon deep inelastic scattering (DIS). Contrary to the RS dependence of the couplant [@my; @Patrick; @Pich], this latter ambiguity has so far received much less attention [@ja]. This is somewhat surprising, taking into account that its consequences may actually be even more important than those discussed in Section 2. In DIS the factorization theorem [@Politzer] implies that the generic NS structure function $F^{\rm NS}(x,Q^2)$ can be written as a convolution (I drop the label “NS” in the following) $$F(x,Q^2)=\int_x^1\frac{dy}{y}q(y,M) C\left(\frac{Q}{M},\frac{x}{y},a(\mu)\right) \label{conv}$$ of the perturbatively uncalculable quark distribution function $q(x,M)$, defined at the factorization scale $M$, and obeying the evolution equation $$\frac{{\rm d}q(x,M)}{{\rm d}\ln M}=\int_x^1\frac{dy}{y} \left[a(M)P^{(0)}\left(\frac{x}{y}\right)+a^2(M)P^{(1)} \left(\frac{x}{y}\right)+\cdots\right] \label{AP}$$ and the hard scattering cross-section $C(Q/M,z,a(\mu))$ admitting perturbation expansion in powers of the couplant at the hard scattering scale $\mu$, generally different from $M$ [^2] $$C(Q/M,z,a(\mu))=\delta(1-z)+a(\mu)C^{(1)}(Q/M,z)+\cdots \label{C}$$ While $P^{(0)}(z)$ is unique, both $P^{(1)}$ in (\[AP\]) and $C^{(1)}$ in (\[C\]) are ambiguous, but internal consistency of the factorization procedure links the variation of the NLO hard scattering cross-section $C^{(1)}$ with that of the NLO branching function $P^{(1)}$ [@Politzer]: $$C^{(1)}(Q/M,z,\mbox{RS})=P^{(0)}(z)\ln (Q/M)+P^{(1)}(z)/b+ \kappa (z,{\rm RS}) \label{C1}$$ where the FS-invariant function $\kappa(z,{\rm RS})$ still, however, depends on the RS of the couplant $a(M)$. Similar consistency conditions do exist at each order of perturbation theory. In the rest of this note I stay within the NLO approximation. The dependence of $C^{(1)}$ on the RS of the couplant appears as a consequence of the fact, that the r.h.s. of (\[AP\]) is given as an expansion in powers of this couplant. I shall return to this point in the next section. For fixed RS of the couplant $a(M)$, $P^{(1)}$, or via (\[C1\]), $C^{(1)}$, defines at the NLO the FS of p.d.f. : FS:={$P^{(1}(z)$}. In the above relations the factorization scale M is again, as in the case of $\mu$, kept as a free parameter. The first point I want to emphasize is that the renormalization of the couplant, to the NLO order fully described by variations of $\Lambda_{\rm RS}$ in both the couplant $a$ and the coefficients $r_k$, is [independent]{} of the FS of p.d.f. as specified by the NLO branching function $P^{(1)}$ in the evolution equation (\[AP\])! [Any combination]{} of the RS={$\Lambda_{\rm RS}$} and FS={$P^{(1)}$} represents in principle equally legal definition of p.d.f.. Moreover, as $P^{(1)}$ is a function of $z$, it represents in fact an infinite number of of degrees of freedom, and thus its variations can be expected to be at least as important as that of the factorization scale $M$. Also this point will be elucidated in the next section. Secondly, let me recall the obvious fact that although the physical observable $F(x,Q^2)$ is, when (\[AP\],\[C1\]) are taken to all orders, independent of the parameters describing the renormalization and factorization schemes, any finite order approximation to these expansions inevitably leads to nontrivial dependence of $F(x,Q^2)$ on $M$, RS and FS. I mention it here as it is related to the basic question I want to address in this section and which concerns the way the p.d.f. transform when the FS={$P^{(1)}$} is varied. As all the considerations are much more transparent in terms of conventional moments, defined, for a generic function $f(x)$, as $$f(N)=\int_0^1x^{N-1}f(x)dx \label{moments}$$ let me rewrite (\[conv\]-\[C1\]) in terms of them, explicitly writing out the dependence on both the RS of the couplant $a(M)$ and the FS of the p.d.f. ($d_N, d_N^{(1)}, \kappa(N)$ are the moments of $P^{(0)}(z), P^{(1)}(z), \kappa(z)$ respectively): $$\frac{dq(N,M,{\mbox {RS,FS}})}{d\ln M}= q(N,M,{\mbox {RS,FS}})\gamma_N; \;\;\; \gamma_N\equiv d_Na(M,{\mbox {RS}})+ d_N^{(1)}(\mbox{FS})a^2(M,{\mbox {RS}}). \label{APN}$$ This is easily solved: $$q(N,M,{\mbox {RS,FS}})= A_N(a(M,{\mbox{RS}}))^{-d_N/b}\exp \left[-a(M,{\mbox{RS}}) d_N^{(1)}({\mbox {FS}})/b\right], \label{q(N)}$$ where the constants $A_N$, introduced in [@Politzer], are independent of $M$ as well as $d^{(1)}_N$. For moments of the structure function (\[conv\]) we get, excplicitly writing out the dependence of the NLO approximation to $F(N,Q^2)$ on M and FS={$d_N^{(1)}$},: $$F(N,Q^2,M,{\rm RS},d_N^{(1)})=q(N,M,{\mbox{RS}},d^{(1)}_N) \left(1+a(\mu,{\mbox{RS}})C^{(1)}(Q/M,N,{\mbox{RS}},d^{(1)}_N)\right), \label{convmom}$$ with the following consistency condition, implied by (\[C1\]), $$C^{(1)}(Q/M,N,\mbox{RS},d^{(1)}_N)=d_N\ln \frac{Q}{M}+\frac{d_N^{(1)}}{b}+\kappa(N,{\mbox{RS}}). \label{C1N}$$ For each moment $N$, the expression (\[convmom\]) is a function of $M$ which, however, still depends on two futher parameters, one specifying the RS of the couplant and the other ($d^{(1)}_N$) the FS of p.d.f.. If we now want to transform $q(N,M,{\mbox{RS}},{\mbox {FS}})$ from one FS={$d^{(1)}_N$} into another, specified by FS={$\overline{d^{(1)}_N}$}, we again cannot do so by imposing the relation (the RS-dependence is suppressed in the rest of this section) $$q(N,M,d^{(1)}_N)\left(1+a(M)C^{(1)}(Q/M,N,d^{(1)}_N)\right)= q(N,M,\overline{d^{(1)}_N})\left(1+a(M)C^{(1)}(Q/M,N, \overline{d^{(1)}_N}\right) \label{errorn}$$ As emphasized above, $F(N,Q^2,M,d_N^{(1)})$ does nontrivially depend on $M, d_N^{(1)}$, as well as the RS and thus postulating equation like (\[errorn\]) would violate this basic feature of finite order approximations. Formally this is clear from inserting (\[q(N)\]) into (\[errorn\]) and solving the resulting equation. Similarly to (\[trans\]), we get the following relation between $\overline{C^{(1)}}\equiv C^{(1)}(Q/M,N,\overline{d^{(1)}_N})$ and $\overline{d^{(1)}_N}$: $$\overline{d^{(1)}_N}=d^{(1)}_N-\frac{b}{a(M)}\ln \left( \frac{1+a(M)C^{(1)}} {1+a(M)\overline{C^{(1)}}}\right) \label{transd}$$ which reduces to the correct one, as given in (\[C1\]), only for the trivial case $\overline{d^{(1)}_N}=d^{(1)}_N$! For any other case the equation (\[transd\]) is incompatible with the the consistency condition (\[C1\]) and thereby wrong. I discuss this point in detail as in many papers, including [@TO; @Tung; @Levy], the equation (converted into moments and restricted to the NS channel) used to transform the quark distribution function between the so-called DIS and $\overline{\rm MS}$ “schemes” (more on them in the next section) $$q_{\rm DIS}(N,M)=q_{\overline{\rm MS}}(N,M) \left(1+a(M)C^{(1)}_{\overline{\rm MS}}(N,M)\right) \label{chyba}$$ is ($C^{(1)}_{\rm DIS}=0$ by definition) precisely of the incorrect form (\[errorn\])! The only theoretically consistent way of transforming $q(N,M,{\mbox {FS}})$ from one FS={$d^{(1)}_N$} into another is given explicitly in (\[q(N)\]) with, as emphasized, the constants $A_N$ held fixed. In [@ja] I have discussed the whole procedure, based on the use of Jacobi polynomials [@Shaw; @jarames], in $x-$space. Although currently other, superior, methods of solving the evolution equations are available [@TO], the fact that Jacobi polynomials are constructed from conventional moments (\[q(N)\]) for which we know how the FS transformations operate, makes them invaluable in this kind of considerations. Finally a remark. The constants $A_N$ represent the most natural way of parametrizing the uncalculable nonperturbative properties of the nucleon. They are not related to any particular “initial” $M_0$, nor to any FS={$d^{(1)}_N$}, but determine the asymptotic behaviour of $q(N,M,{\mbox{RS}},d^{(1)}_N)$ as $M\rightarrow\infty$, which is unique. Remarks on current phenomenology ================================ In the preceding Section I have discussed the central question of the FS dependence of the p.d.f.. Let me now turn to the current phenomenology of DIS, related to this subject. The first remark concerns the meaning of the words “$\overline{\mbox{MS}}$” and “DIS”, when used in the connection with the p.d.f. at the NLO. The $\overline{\mbox {MS}}$ p.d.f. are defined in [@TO] as those “which appear in the equation such as (\[conv\]) with hard scattering part $C^{(1)}$ calculated with the $\overline{\mbox{MS}}$ subtraction prescription.” Although correct, this definition is obviously incomplete, as it specifies merely the RS of the couplant but tell us nothing about the FS={$P^{(1)}(z)\}$ to be used in (\[AP\])! Recall that the term $\ln 4\pi-\gamma_{\rm E}$ defining the $\overline{\mbox{MS}}$ “subtraction scheme”, is an artifact of extending the definition of the couplant into 4-$\epsilon$ dimensions. As any RS of the couplant may be combined with any FS of p.d.f., there is an infinite set of $\overline{\mbox{MS}}$-like p.d.f., sharing the same definition of the couplant, and therefore the right to be called “$\overline{\mbox{MS}}$”, but arbitrarily differing in $P^{(1)}$. The one usually understood in the literature under the label “$\overline{\mbox{MS}}$” corresponds to the FS used in [@Flor; @Curci], which should be called “MS” as it sets (within the dimensional regularization), the finite parts of the counterterms, renormalizing the appropriate composite operators, to zero. Altough the name for a FS is basically a matter of semantics, the current use of the term “$\overline{\mbox{MS}}$” is misleading as it fails to specify the FS of the p.d.f. used. Ambiguity of a different kind is associated with the use of the label “DIS”. This scheme is supposed to be defined by the condition: $$C^{(1)}_{\rm DIS}(Q=M,z,{\mbox{RS,FS}})=0 \;\;\;\Rightarrow P^{(1)}_{\rm DIS}(z,{\rm RS})=-b\kappa(z,{\mbox{RS}}) \label{DIS}$$ which, however, is again not unique, due to the fact that the FS invariant $\kappa(z,\mbox{RS})$ still depends on the RS of the couplant. In fact it is the the combination $$\varepsilon(z)\equiv\kappa (z,{\mbox{RS}})+ P^{(0)}(z)\ln (Q/\Lambda_{\rm RS}) \label{invariant}$$ which is independent of $M$, FS={$P^{(1)}(z)\}$ as well as the RS of the couplant [@ja]. Consequently in a given RS of the couplant $a(M,{\rm RS})$ we find $$P^{(1)}_{\rm DIS}(z,{\rm RS})=-b\varepsilon(z)+bP^{(0)}(z)\ln \frac{Q}{\Lambda_{\rm RS}}. \label{PDIS}$$ For example we find $$P^{(1)}_{\rm DIS}(z,\overline{\mbox{MS}})= P^{(1)}_{\rm DIS}(z,\mbox{MS})+ bP^{(0)}(z)\ln \frac{\Lambda_{\rm MS}}{\Lambda_{\overline{\rm MS}}} \label{MSBMS}$$ In general, there is again an infinite set of “DIS”-like p.d.f., which share the property $C^{(1)}=0$, but differ in the NLO branching function $P^{(1)}(z)$ and the couplant, thereby leading to different numerical predictions when inserted into (\[conv\]). P.d.f. bearing the name “DIS”, like those of [@DFLM], or some of [@CTEQ], tacitly assume $\overline{\mbox{MS}}$ as the RS of the couplant. This, however, is not a must and thus for an unambiguous specification of the “DIS”-like FS the RS of the couplant should always be specified. The second remark concerns the practical aspect of exploiting the vast freedom in the definition of p.d.f. at the NLO. As already mentioned, little phenomenological attention has so far been payed to the FS dependence of p.d.f.. This may be due in part to the failure to appreciate the independence of these two renormalization procedures. It is also true that to apply, for instance, the idea of “optimization” [@PMS] to the FC dependence of p.d.f. in $x$-space is technically much more involved. In [@ja] I have, however, argued that at least the FS defined by setting $P^{(1)}=0$ (at the NLO; at higher orders it would generalize by setting all higher order AP branching functions to zero) should seriously be considered. In this “zero” FS the full NLO correction is put into the hard scattering cross-section $C^{(1)}$, thereby representing in some sense the opposite of the DIS FS, which sets $C^{(1)}(z)=0$. Moreover, it turns out [@ja] that when the moments of p.d.f. are considered, this FS is very close to that obtained via the Principle of Minimum Sensitivity of [@PMS] (for $c_1=0$ they even coincide). Although this results doesn’t automatically imply the same close relation for the p.d.f. themselves, it seems reasonable to add this FS to the list of those used in phenomenological applications. “Zero” FS and NLO event generators ================================== There is, in fact, another reason why this FS could be of considerable interest. Recall, that all currently used event generators, like HERWIG, PYTHIA, JETSET, LEPTO etc., are based on essentially leading-log parton showers. Although they are sometimes combined with NLO hard scattering cross-sections, the overall description remains only LO. The simple picture of LO parton showers becomes much more complicated when one attempts to generalize them to the NLO. To get a consistent NLO description of any hard scattering process in the “zero” FS we, however, need merely the LO parton showers as $P^{(1)}(z)=0$ in this FS! As the example of moments of structure functions shows, this choice may be quite reasonable and should definitely be tried. Although simple at first glance, one has to be careful in taking for the NLO cross-section that corresponding to this “zero” FS. From (\[C1\]) we easily find its form: $$C^{(1)}_{zero}(Q/M,z)\equiv P^{(0)}(z)\ln (Q/M)+\kappa(z,\mbox{RS}) \label{zero}$$ Using the results of [@Curci] on $P^{(1)}$ and $C^{(1)}$, $\kappa(z,\mbox{RS})$ can straightforwardly be evaluated in the $\overline{\mbox{MS}}$ RS. Transformation to any other RS is then trivial. Let me stress that this procedure is not equivalent to the so-called matching of parton showers to fixed order matrix elements, as recently implemented in LEPTO event generator [@LEPTO]. There, the exact $O(\alpha_s)$ matrix element is matched to the parton shower at some particular value of incoming parton virtuality $t_m$ (see Fig.1) in the sense that below $t_m$ only parton showers are used while above $t_m$ the matrix element takes fully over. In the case of “zero” FS, the situation is different and the NLO cross-section ${\rm d}\sigma^{NLO}(Q/M,z,t)/dt$ which, when integrated over $t$, yields (\[zero\]) contributes at any virtuality, even below that given by the factorization scale $t=M^2$, as only the pole term $1/t$ plus some finite part is factorized into the parton p.d.f.. As a result, the NLO hard scattering cross-section ${\rm d}\sigma^{NLO}(Q/M,z,t)/dt$ becomes a discontinuous function of $t$ at $t=M^2$, this discontinuity being cancelled by a similar discontinuity of the parton shower contribution, which is restricted by definition to the domain $t\leq M^2$! Work on practical implementation of this idea is in progress. The situation is sketched in Fig.2a, where the full ${\rm d}\sigma^{NLO}(Q/M,z,t)/dt$, corresponding to the sum of the diagram in Fig.1 and the one with the gluon radiated off the outgoing quark, is plotted as a function of $\tau=-t$ for fixed $Q^2$ and $z$. The full result is a sum of three terms: the pole term of the form $A(z)/\tau$, the residue of which is proportional to $P^{(0)}(z)$, the $\tau$ independent constant $B(z,Q)$ and $C(z,Q)\tau$, linearly rising with $\tau$. Notice that for $\tau\rightarrow0$, the finite part of ${\rm d}\sigma^{NLO}(z,Q,t)$, i.e. the sum of the last two terms, becomes negative. The result of factorization, i.e. the separation of the full NLO cross-section, containing all infrared and parallel singularities, into a part included in the quark distribution function and the remaining, finite, hard scattering cross-section, is represented in Fig.2b by the dashed and dotted curves, discontinuous at the factorization scale $\tau=M^2$. For $\tau>M^2$ the hard scattering cross-section ${\rm d}\sigma^{NLO}(z,Q,t)$ coincides with the full result, but for $\tau<M^2$ its definition is ambiguous as it depends on how much of the finite part will accompany the sigular pole pole term into the definition of the quark distribution function. Fig.2b corresponds to the case that only the pole term $A/\tau$ is subtracted. Recall that, for instance, the term $P^{(0)}(z)\ln(Q/M)$ appearing in (\[C1\]) is essentially the integral of the pole term $A/\tau$ from $M^2$ to the upper kinematically allowed value of $\tau$, proportional to $Q^2$. Summary and conclusions ======================= In this note I have discussed several aspects of factorization scheme dependence of parton distribution functions. I have emphasized potential importance of proper treatment of this ambiguity for the reliablity of theoretical analyses of ever better data. Special attention has been payed to the correct transformation of p.d.f. between different factorization schemes and the ambiguities in the meaning of some of the currently most popular definition of p.d.f. have been brought to light. Finally the so-called “zero” FS has been proposed and shown to be potentialy useful in the construction of NLO event generators. 0.01cm I am grateful to P. Kolář for careful reading of the manuscript amd many stimulating discussions. [99]{} S.A. Rabinowitz et al., Phys. Rev. Lett. [70]{}, 134 (1993)\ M. Mishra et al., Nevis Laboratory preprints 1459, 1460, 1461 A. Amaudruz et al., Phys. Lett. [B294]{}, 120 (1992)\ A. Amaudruz et. al, Phys. Lett. [B295]{}, 159 (1992)\ A. Amaudruz et. al, Nucl. Phys. [B371]{}, 3 (1992) J.F. Owens and W.K. Tung, Fermilab-PUB-92/59-T J. Botts et al, Fermilab-PUB-92/371 A. Martin, R.G. Roberts and W.J. Stirling, Phys. Rev. [D47]{}, 867 (1993)\ A. Martin, R.G. Roberts and W.J. Stirling, RAL-92-076\ A. Martin, R.G. Roberts and W.J. Stirling, RAL-93-014 W. K. Tung, Nucl. Phys. [B315]{}, 378 (1989) K. Charchula, M. Krawczyk, H. Abramowicz and A. Levy, DESY 90-019 P.M. Stevenson, Phys. Rev. [D23]{}, 2916 (1981) J. Chýla and A. Kataev, Phys. Lett. [B297]{}, 385 (1992) G. Grunberg, Phys. Rev. [D29]{}, 2315 (1984) H.D. Politzer, Nucl. Phys. [B194]{}, 493 (1982) P. Aurenche, P. Chiapetta, M. Fontannaz and D. Schiff, Nucl. Phys. [B286]{}, 509 (1987)\ P. Aurenche, P. Chiapetta, A. Douiri, M. Fontannaz and D. Schiff, Nucl. Phys. [B286]{}, 553 (1987)\ P. Aurenche and P. Chiapetta, Z.Phys.C-Particles and Fields [34]{}, 201 (1987) F. Le Diberder and A. Pich, Phys. Lett. [B289]{}, 165 (1992) J. Chýla, Z.Phys.C-Particles and Fields [43]{}, 431 (1989) I. Barker, C. Langensiepen and G. Shaw, Nucl. Phys. [B186]{}, 61 (1981) J. Chýla and J. Rameš, Z.Phys.C-Particles and Fields [31]{}, 151 (1986) E.G. Floratos,D.A. Ross and C.T. Sachrajda, Nucl. Phys. [B129]{}, 66 (1977) G. Curci, W. Furmanski and R. Petronzio, Nucl. Phys. [B175]{}, 27 (1980)\ W. Furmanski and R. Petronzio, Z.Phys.C-Particles and fields [11]{}, 293 (1982) M. Diemoz, F. Ferroni, E. Longo and G. Martinelli, Z.Phys.C-Particles and Fields [39]{}, 21 (1988) G. Ingelman, in [Physics at HERA]{}, Proceedings of the Workshop, Hamburg 1991, ed. G. Ingelman, W. Buchmüller 0.01cm [Figure captions]{}\ Fig.1: Feynman diagram describing the process e$^-$+q$\rightarrow$e$^-$+q+g with parallel singularity in the $t$-channel. Fig.2a: A typical shape of ${\rm d}\sigma(Q,z,t)/dt$ as a function of $\tau=-t$ for fixed $x,Q^2$. In this example the pole term (dashed curve) is added to a linearly rising finite part (dotted line) to give the full NLO contribution (solid curve). Fig.2b: Separation of the full NLO contribution (solid curve) into the part absorbed in the quark distribution function (dashed curve) and the finite NLO hard scattering cross-section (dotted curve). The discontinuities of the last two curves at $\tau=M^2$ cancel in the sum. [^1]: No essential conclusion obtained in the following does, however, depend on this purely technical simplification. [^2]: If considered to all orders of $a(\mu)$, $C(Q/M,z,a(\mu))$ doesn’t actually depend on $\mu$ [@Politzer]. Contrary to the $M$-dependence of $q(N,M)$ and $C^{(1)}(N,M)$, which is the basic feature of the factorization theorem and holds to all orders, the dependence of $C(Q/M,z)$ on $\mu$ is merely a consequence of truncating expansion (\[C\]) to a finite order.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In the previous papers in this series, we have measured the stellar and [ <span style="font-variant:small-caps;">HI</span> ]{} content in a sample of edge-on galaxies. In the present paper, we perform a simultaneous rotation curve and vertical force field gradient decomposition for five of these edge-on galaxies. The rotation curve decomposition provides a measure of the radial dark matter potential, while the vertical force field gradient provide a measure of the vertical dark matter potential. We fit dark matter halo models to these potentials. Using our [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption results, we find that a typical dark matter halo has a less dense core ($0.094\pm0.230$M$_\odot$/pc$^3$) compared to an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model ($0.150\pm0.124$M$_\odot$/pc$^3$). The [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption dark matter halo has a longer scale length $R_c$ of $1.42\pm 3.48$kpc, versus $1.10\pm 1.81$kpc for the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model. The median halo shape is spherical, at $q=1.0\pm0.6$ (self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}), while it is prolate at $q=1.5\pm0.6$ for the optically thin. Our best results were obtained for ESO274-G001 and UGC7321, for which we were able to measure the velocity dispersion in Paper III. These two galaxies have drastically different halo shapes, with one oblate and one strongly prolate. Overall, we find that the many assumptions required make this type of analysis susceptible to errors.' author: - \| S. P. C. Peters$^{1}$, P. C. van der Kruit$^{1}$[^1], R. J. Allen$^{2}$ and K. C. Freeman$^{3}$\ $^{1}$Kapteyn Astronomical Institute, University of Groningen, P.O.Box 800, 9700AV Groningen, the Netherlands\ $^{2}$Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA\ $^{3}$Research School of Astronomy and Astrophysics The Australian National University, Cotter Road Weston Creek, ACT 2611,\ Australia bibliography: - 'refsV.bib' date: 'Accepted 2015 month xx. Received 2015 Month xx; in original form 2015 Month xx' title: \| The Shape of Dark Matter Haloes\ V. Analysis of observations of edge-on galaxies --- \[firstpage\] galaxies: haloes, galaxies: kinematics and dynamics, galaxies: photometry, galaxies: spiral, galaxies: structure Introduction {#sec:dmintro} ============ Modern day cosmological dark matter simulations predict that dark matter has a significant impact on the formation and evolution of galaxies. The dark matter clumps into haloes, which serve as gravitational sinks for baryonic matter to fall in to. Once inside, these baryons form the galaxies and other visible structures in the universe. The size and shape of these haloes is influenced by the type of dark matter particle and its merger history. As such, getting a grip on the shape of the halo offers a potential constraint on the dark matter model [@Davis1985A]. The shape of dark matter haloes can be classified by the shape parameter $q$, using the ratios between the vertical axis $c$ and radial axis $a$, such that $q = c/a$. This divides the potential halo shapes up into three classes: prolate ($q > 1$), oblate ($q < 1$) and spherical ($q \sim 1$). This is, of course, only a simplified version of reality, where we can also expect triaxial shapes and changes of shape with radius and history [@VeraCiro2014]. However, for haloes with masses ${\lower.5ex\hbox{$\; \buildrel > \over \sim \;$}}10^{12.3} h^{-1} M_\odot$ the haloes can be adequately described by a single vertical-to-radial axis ratio [@Schneider2012A]. Direct observations of the shape of haloes is tricky, as there are only few tracers that offer a clear view on the vertical gravitational potential of the halo. The flat rotation curves of galaxies, while an excellent tracer of the radial potential of the halo, provides no information in the vertical direction. Luckily, some tracers do exist. The stellar streams of stripped, in-falling galaxies can be used to measure the potential the stream is traversing. @Helmi2004A analysed the stream of Sagittarius, as found in the 2MASS survey, around the Galaxy and found that the data best fits a model with a prolate shape, with an axis ratio of 5/3. In a similar fashion, only further out, do the satellites of the Galaxy offer such a tracer. Globular cluster NGC5466 was modelled by @Lux2012A and found to favor an oblate or triaxial halo, while excluding spherical and prolate haloes with a high confidence. The stellar stream has been reanalysed by @VeraCiro2013, who report an oblate halo with $q=0.9$ for $r\leq10$kpc. Gravitational lensing offers another measure of the vertical gravitational potential. Strong lensing uses the Einstein lens of background sources around a single galaxy or cluster. By modelling the lens, it is possible to create a detailed mass map of the system. By combining this with gas and stellar kinematics, it is possible to calculate the dark matter mass distribution [@Treu2010A]. For example, @Barnabe2012A applied this method to lensed galaxy SDSS J2141, to find a slightly oblate halo ($q = 0.91^{+0.15}_{-0.13}$). Weak lensing lacks the clear gravitation lens seen in strong lensing. As such, it is unable to measure the halo of a single galaxy. Instead, it provides an average halo shape from a statistically large sample of sources, by modelling the alignment of background galaxies to a large series of foreground galaxies. This allows the sources to probe the outer edges of haloes. A recent analysis by @vanUitert2012 finds, on a sample of $2.2\times10^7$ galaxies, that the halo ellipticity distribution favors oblate, with $q=0.62^{+0.25}_{-0.26}$. Polar ring galaxies are also of interest in the study of the dark matter halo shape, as the orbits of the stars in the polar rings are very sensitive to the gravitational potential. The method was pioneered by @Schweizer1983A who noted that the polar ring in galaxy A0136-0801 indicated a massive halo that was more spherical than flat. @Whitmore1987A studied this galaxy in more detail and found $q=0.98\pm0.20$ for this galaxy. These authors also studied galaxies NGC4650A and ESO415-G26, for which they report $q=0.86\pm0.21$ and $q=1.05\pm0.17$. Galaxy NGC4650A has also been studied by @Sackett1994A, who reported a flattened halo with $q$ between 0.3 and 0.4. Another way to place constraints on the halo shape is by carefully modelling local edge-on galaxies. The thickness of the [ <span style="font-variant:small-caps;">HI</span> ]{}layer in spiral galaxies corresponds directly on the local hydrostatic equilibrium. The [ <span style="font-variant:small-caps;">HI</span> ]{}layer flares out at large radii, as there is less matter gravitationally binding it to the central plane. Because of this, flaring provides a sensitive tracer of the vertical potential as a function of radius in the disc. By combining this with the horizontal potential as gained from the rotation curve, and estimates for the gas and stellar mass distributions, one can fit the potential well created by the dark matter halo. This method was first applied to the Galaxy by @Celnik1979A and by @vdk81c on edge-on galaxy NGC891. In the latter it was found that the halo was spherical rather than flattened as the stellar disc. NGC4244 was analysed by @Olling1995A, who found a highly flattened halo of $q=0.2^{+0.3}_{-0.1}$. @OBrien2010D also set out to measure the velocity dispersion as function of radius. Using that approach to measure the halo shape of [UGC7321]{}, they found a spherical halo ($q = 1.0 \pm 0.1$). This paper is the fifth in a series. We will refer to earlier papers of the series simply as Papers I through IV. In this paper V, we will perform a similar analysis, using the measured parameters for the [ <span style="font-variant:small-caps;">HI</span> ]{}disc (Paper III) and stellar disc (Paper IV). The series is based on a large part of the PhD thesis of the first author [@Peters2014]. The paper has the following outline. In Section \[sec:hydrostaticmodels\], we will provide a detailed description of hydrostatic models we use and detail the fitting strategy. Section \[sec:hydrostaticresults\] presents and discusses the results. We summarise the results in Section \[sec:hydrostaticssummary\] Modelling Strategy {#sec:hydrostaticmodels} ================== Overall Strategy & Sample ------------------------- @Celnik1979A proposed a new strategy in which the flaring of the [ <span style="font-variant:small-caps;">HI</span> ]{} layers is used as a tracer of the vertical gravitational potential in the Galaxy. This method was extended to edge-on galaxy NGC891 by @vdk81c. In combination with a traditional rotation curve decomposition, this offers a view on both the radial and vertical directions of the gravitational potential. The dark matter halo model can be fit to this. Further previous work on this subject, using the flaring of gas layers, includes that by @Olling1995A [@Olling1996A], @Olling2000, @bc97, @nsj05, @k03, @kdkh07 and @bj08, and others. We will use this same strategy to model the dark matter halo shape for the sample of eight galaxies defined in Paper I and extensively studied in further papers in this series. In Paper III, we have measured the structure and kinematics of the [ <span style="font-variant:small-caps;">HI</span> ]{}in these galaxies, using modelling procedures that allow for the correction of self-absorption in the [ <span style="font-variant:small-caps;">HI</span> ]{}for assumed spin temperatures. The stellar discs were modelled in Paper IV. Based on the quality of the results from Papers III and IV, we have decided to model five out of the original eight galaxies in our sample. These are IC5249, ESO115-G021, ESO138-G014, ESO274-G001 and UGC7321. All five galaxies are late-type Sd, with self-absorption corrected [ <span style="font-variant:small-caps;">HI</span> ]{}masses between $4.1\!\pm\!0.1\!\times\!10^8$ and $7.8\!\pm\!0.8\!\times10^9$M$_\odot$ (Table 1 in Paper III). The distances vary greatly between these galaxies. ESO274-G001 is only 3.0Mpc away, while the distance to IC5249 is estimated at 32.1Mpc. In Paper III, we successfully measured the velocity dispersion of the [ <span style="font-variant:small-caps;">HI</span> ]{}as function of radius in ESO274-G001 and UGC7321. In ESO115-G021, the velocity dispersion appeared to increase with radius from about 7 [[km s$^{-1}$]{}]{}in the inner parts to 12-14 [[km s$^{-1}$]{}]{}at the outer measured point. As we are skeptical of this result, we will adopt a constant velocity dispersion of 10km/s for this galaxy. We also adopt this constant velocity dispersion for IC5249 and ESO138-G014. For UGC7321 the velocity dispersion drops from 10 [[km s$^{-1}$]{}]{}in the central parts to 8 [[km s$^{-1}$]{}]{}at 8 kpc, then remains constant at this level out to 10 kpc, after wich it increases to 12 [[km s$^{-1}$]{}]{}at 14 kpc. We have sufficient confidence in it to adopt it in our analysis. Galaxy UGC7321 has a total B-band magnitude of 13.75 and V-band magnitude of 13.19 [@Taylor2005], giving a B-V band difference of 0.56.\[sec:estimateML\] To estimate the stellar mass to light ($M_/L_R$) ratio, we use the model interpolation engine[^2], which is based on the @Worthey1994A and @Bertelli1994A stellar population models. Using a single burst, Salpeter (slope -2.35) IMF we find a good fit using a single population of 950Myr and \[Fe/H\]=0.0, with B-V=0.554. This would place $M_/L_R$ in the Rc-band at 0.55. We have repeated this exercise for our other galaxies, each time finding $M_/L_R\approx0.5$. We therefore adopt a mass to light $M_/L_{R}=0.5$ as a lower boundary to the stellar disc mass, and use the reported R-band total luminosities reported in Table 3 of Paper IV. As an upper boundary, we adopt $M_/L_{R}<3$. At solar metallicity, this would imply a single stellar population of approximately 8Myr. At multiplicities lower than solar, this would increase to an even large age. \[sec:MLdiscussion\] Decomposition Strategy ---------------------- We start with the assumption that there are three components, stars, gas+dust and dark matter. Each of these components add to the gravitational potential, so ideally one would need to write down and solve the combined Poisson-Boltzmann equation. This is internally consistent, but requires simplifying assumptions such as on the properties of the stellar velocity tensor (e.g. that it is Gaussian and isothermal or a superposition of such components) and its variation in space. This approach, which indeed is self-consistent, has been used by the earlier workers referred to above, starting with @Olling1996A. Our stategy is the following. We infer the radial force near the plane of the galaxy from the [ <span style="font-variant:small-caps;">HI</span> ]{}rotation curve, assuming that due to the relative low velocity dispersion of order 10 [[km s$^{-1}$]{}]{}the asymetric drift term in the Jeans equation can be ignored. Next we model the light distribution and assuming a constant mass-to-light ratio estimate the contribution of the disk up the a free parameter $M/L$. We do the same from the inferred mass distribution of gas, adding the usual contribution (25% of the total) for helium. We assume also that the molecular gas content in these late-type dwarfs is small. This gives us (except for the parameter $M/L$) the gradient in the radial force due to the halo. We next estimate the vertical force from the thickness and velocity dispersion of the [ <span style="font-variant:small-caps;">HI</span> ]{}, using the vertical Jeans equation. Since we evaluate the forces at $z = 100$ pc, this should be an excellent approximation, even though the equations used are not internally self-consistent as in the approach in the first paragraph. This strategy has previously been used by @OBrien2010D. ### Radial Tracer The radial force gradient is calculated using a classic rotation curve decomposition performed at the mid-plane of the galaxy [@vanAlbada1985A]: $$v_\textrm{total}^2(R) = v_\textrm{gas}^2(R) + v_\textrm{stellar}^2(R) + v_\textrm{halo}^2(R)\,\,.\label{eqn:vresidual}$$ The total rotation $v_\textrm{total}(R)$ is the observed rotation curve of the [ <span style="font-variant:small-caps;">HI</span> ]{}gas, as we measured previously in Paper III. The theoretical rotation curves of the gas $v_\textrm{gas}(R)$ and stars $v_\textrm{stellar}(R)$ represent the contributions due to the stellar and gaseous mass components. We calculate the theoretical rotation curve of the stellar and gaseous components, using equation (A.17) of @Casertano83 $$v^2(R) = -8GR \int_0^\infty r \int_0^\infty \frac{\partial \rho(r,z)}{\partial r} \frac{K(p)-E(p)}{(Rrp)^{1/2}} \,dz\,dr \,\,,$$ with $$\begin{aligned} p &=& x - (x^2 -1)^{1/2}\,\,,\\ x &=& \frac{R^2 + r^2 + z^2}{2Rr}\,\,.\end{aligned}$$ Here $K(p)$ and $E(p)$ are the complete elliptical integrals of the first and second kind. The equations are evaluated numerically. The equation of the density distribution $\rho(r,z)$ of the [ <span style="font-variant:small-caps;">HI</span> ]{}disc has been previously presented in Equation 15 in Paper II. We have presented the measured densities in Paper III. The equations for the stellar disc and bulge luminosity distributions have been previously presented in Equations 1 and 3 in Paper IV. We convert to a combined stellar mass distribution from these equations using $$\rho(r,z) = (M_/L_R)\left(j_\textrm{disc}(R,z) + j_\textrm{bulge}(R,z)\right)\,\,.$$ The measurements for the stellar disc have been presented in Paper IV. Note that we adopt a single mass-to-light ratio $M_/L_R$. While this may be a bit unrealistic, we noted in Paper IV that the fits for some of the bulges appear to supplement the stellar disc, rather than model a central component. Treating the two as distinct would therefore be invalid. A fixed $M_/L_R$ ratio also has the advantage of limiting the complexity of the parameter space we will be fitting in. ### Vertical Tracer We follow the method of @OBrien2010D to calculate the gradient of the vertical force. The gradients due to each mass component add up as $$\begin{aligned} \frac{dF_{z,\textrm{total}}(R,z)}{dz} &=& \nonumber \\ \frac{dF_{z,\textrm{gas}}(R,z)}{dz}&+&\frac{dF_{z,\textrm{stellar}}(R,z)}{dz}+\frac{dF_{z,\textrm{halo}}(R,z)}{dz} \label{eqn:zresidual}\end{aligned}$$ Following @OBrien2010D, the disc is assumed to be in vertical hydrostatic equilibrium, such that the vertical gas pressure gradient and total vertical gravitational force of the galaxy potential $\Phi_\textrm{total}$ due to all mass components balance perfectly, $$\nabla\left(\sigma_\textrm{gas}^2 \,\rho_\textrm{gas}\right) = \rho_\textrm{gas} \nabla \Phi_\textrm{total}\label{eqn:hydro}\,\,.$$ We next assume that the gas velocity dispersion is isothermal[^3] in $z$. In that case, Equation \[eqn:hydro\] reduces to $$\frac{\sigma_\textrm{gas}^2 \,\partial\left[ \log \rho_\textrm{gas}(R,z)\right]}{\partial z^2} = - \frac{\partial F_{z,\textrm{total}} (R,z)}{\partial z}\,\,.\label{assumption:isothermal}$$ Since our [ <span style="font-variant:small-caps;">HI</span> ]{}disc is modelled as a Gaussian distribution (see Equation 15 of Paper II), this becomes $$\frac{\partial F_{z,\textrm{total}} (R,z)}{\partial z} = -\frac{\sigma^2_\textrm{gas}(R)}{{z_0}^2(R)}\label{eqn:Ftotal}\,\, ,$$ such that the vertical gradient of $F_\textrm{z,total}$ is constant with height $z$. The gradients of the stellar and gas force components was calculated using the Poisson equation, where we assume the disc is axi-symmetric and the circular rotation is constant with height [@OBrien2010D], $$\frac{\partial F_{z} (R,z)}{\partial z} = -4\pi G \rho(R,z) + \frac{1}{R} \frac{\partial \, v^2(R)}{\partial R},\label{eqn:poissonz}$$ where we use the density $\rho$ and rotation $v$ of each of the two components. The squared velocity gradient is calculated numerically. Our modelling of the vertical tracer will use a plane at a height $z$ of 100pc. Halo Potential -------------- We make use of the flattened, but axi-symmetric, pseudo-isothermal halo model proposed by @Sackett1994A, where we assume that the equatorial plane of the halo matches that of the galaxy. In this model the density is stratified in concentric ellipsoids, as given by $$\rho_\textrm{halo}(R,z) = \frac{\rho_{0,\textrm{halo}}\,R_c^2}{R_c^2 + R^2 + z^2/q^2}.$$ The ellipsoids formed by this density distribution have axis ratio $q\equiv c/a$, with core radius $R_c$. The potential due to this density distribution is given by @Sackett1990A as $$\begin{aligned} \Phi_\textrm{halo}(R,z) &=\ \ 2 \pi G q \,\rho_{0,\textrm{halo}} \,R_c^2 \int_0^{1/q} \left[\frac{1}{x^2(1-q^2) + 1}\right]&\nonumber \\ &\times \log\left\{1+\frac{x^2}{R_c^2} \left[\frac{R^2}{x^2(1-q^2) +1}+z^2\right]\right\}\,dx.& \end{aligned}$$ @Sackett1994A provides the solution to the forces associated with this halo in spherical coordinates: $$\begin{aligned} F_R(R,z) &=& \frac{-v_H^2 R \gamma}{h \arctan \gamma} \left[ \frac{\mu^2}{\mu^2-1} \left(\frac{\arctan \gamma\mu}{\gamma\mu} - \frac{\arctan\gamma}{\gamma}\right)\right.\label{eqn:halor} \nonumber \\ &&\left.- \frac{\nu^2}{\nu^2-1} \left(\frac{\arctan \gamma\nu}{\gamma \nu} - \frac{\arctan \gamma}{\gamma} \right) \right] \,,\\ F_z(R,z) &=& \frac{-v_H^2 z \gamma}{h \arctan \gamma}\left(\frac{\arctan \gamma\mu}{\gamma \mu} - \frac{\arctan \gamma \nu}{\gamma \nu}\right)\,,\label{eqn:haloz}\end{aligned}$$ where $$\gamma \equiv \frac{\sqrt{1-q^2}}{q}\,\, ,\,\, \nu\equiv \sqrt{\frac{2c}{b-h}}\,\,,\,\, \mu\equiv\sqrt{\frac{2c}{b+h}},$$ and $$\begin{aligned} h&\equiv&\sqrt{b^2-4ac} \, ,\\ a &\equiv& (1-q^2) R_c^2 \, ,\\ b &\equiv& z^2 + R^2 + (1-q^2) R_c^2 \, ,\\ c &\equiv& z^2.\end{aligned}$$ Note that to calculate $q>1$ one requires the use of complex numbers, although these return to real numbers at the end of the calculation. Only using real numbers will force a fit to be constrained to $q<1$. The calculation has a singularity at $q=1$, which we automatically substitute with $q=0.999$ where required. The asymptotic halo velocity $v_H$ is defined as $$v_H^2 = \frac{4\pi G\,\rho_{0,\textrm{halo}} \,R_c^2\, q \arccos q}{\sqrt{1-q^2}}\label{eqn:dmhalovelocity}$$ Fitting Strategy ---------------- Using Equation \[eqn:vresidual\], we calculate the observed circular rotation curve due to the dark matter halo $v_\textrm{halo}^2$ at the mid-plane of the galaxy over the full range of $R$. In a similar way we calculate the observed vertical force gradient $dF_{z,\textrm{halo}}/dz(R,z)$ due to the dark matter halo at a height $z$ of 100pc, using Equation \[eqn:zresidual\]. We inspect the results to determine in which range of radii $R$ they are sufficiently reliable. In contrast to @OBrien2010D we fit both the vertical and radial tracers simultaneously. Since the rotation curve decomposition is less sensitive to noise than the vertical force gradient decomposition, we can often fit a larger radial range for the rotation curve decomposition than the vertical force gradient decomposition. We will fit the dark matter halo using Equations \[eqn:halor\] and \[eqn:haloz\], where we calculate the gradients numerically, and use the mid-plane approximation $v^2=-RF$ [@Kuijken1989A]. Both tracers operate in drastically different numerical regimes. The total observed rotation curve is often near 100km/s, while the observed vertical force gradient has a value of -0.004 km$^2$ s$^{-2}$ pc$^{-2}$. As these numbers are so drastically different, the combined $\chi^2$ would be dominated by the rotation curve decomposition. We therefore normalise the data of each force by its maximum value in that range, such that the total $\chi^2$ error is calculated as $$\begin{aligned} \chi^2 &=& \chi^2_R + \chi^2_z\,\label{eqn:halochi}\, ,\\ \chi^2_R &=& \sum_{R_i} \frac{v_\textrm{total}^2 - v_{ \relax \ifmmode \textrm{\textsc{HI}} \else \textsc{H{\smaller}I} \fi }^2 - v_\textrm{stellar}^2 - v_\textrm{halo}^2}{\max({v_\textrm{total})^2}}\,\,,\label{eqn:halochi1}\\ \chi^2_z &=& \sum_{R_i} \frac{\left( dF_{z,\textrm{diff}}/dz \right)^2}{\min(dF_{z,\textrm{total}}/dz)^2}\,\,,\label{eqn:halochi2}\\ \frac{dF_{z,\textrm{diff}}}{dz} &=& \frac{dF_{z,\textrm{total}}}{dz} \nonumber \\ &-& \frac{dF_{z,\textrm{gas}}}{dz}+\frac{dF_{z,\textrm{stellar}}}{dz}+\frac{dF_{z,\textrm{halo}}}{dz}\,\,,\end{aligned}$$ where the $R_i$ values are all data points within the respective fitting ranges used for the vertical and radial directions. We have tested many variations of Equations \[eqn:halochi1\] and \[eqn:halochi2\], including converting and integrating both the tracers back into forces so that they could compared more directly. The overall problem however remained, as the two forces were too different in strength and the errors in the radial component would dominate the fit. We have decided to stick to the units adopted by @OBrien2010D, as this offered us the best way to compare results. In Paper III, we performed a Monte-Carlo Markov-Chain (MCMC) fit to the neutral hydrogen, such that a set of samples from the so-called chain together cover the likelihood distribution of the parameters. Our fitting strategy here makes use of this likelihood distribution. We take the last 1000 samples from the chain and perform a fit of the dark matter halo to each individual sample. In total, we thus get 1000 solutions for the halo. We make use of the <span style="font-variant:small-caps;">PSwarm</span> particle swarm optimization algorithm [@Vaz09], as implemented through the <span style="font-variant:small-caps;">OpenOpt</span> library [@Kroshko]. The <span style="font-variant:small-caps;">PSwarm</span> is an example of a global optimization routine, which can avoid being stuck in local optima in the solution space. The fit is performed directly to $\chi^2$ as defined in Equation \[eqn:halochi\]. In some cases, the models converge to unrealistic solutions. We therefore base our results on the 25% of the samples with the lowest $\chi^2$ errors. The halo model uses three free parameters: the halo central density $\rho_{0,\textrm{halo}}$, the scale length $R_c$ and the halo shape $q$. Together with the mass-to-light conversion $M_/L_{R}$ for the stellar disc, we thus have four free parameters. We have considered using an additional mass-to-light conversion for the bulge, but found that with these four parameters the solutions were already becoming degenerate. Adding an additional free parameter would have worsened this problem. We constrain $\rho_{0,\textrm{halo}}$ between 0 and 3 M$_\odot$ pc$^{-3}$, $R_c$ between 100pc and 10kpc, $q$ between 0.1 and 2.0, and $M_/L_{R}$ between 0.5 and 3.0. ![image](Petersetal-Vfig1a.jpg){width="40.00000%"} ![image](Petersetal-Vfig1b.jpg){width="40.00000%"} ![image](Petersetal-Vfig1c.jpg){width="40.00000%"} ![image](Petersetal-Vfig1d.jpg){width="40.00000%"} ![image](Petersetal-Vfig2a.jpg){width="49.00000%"} ![image](Petersetal-Vfig2b.jpg){width="49.00000%"} ![image](Petersetal-Vfig3a.jpg){width="40.00000%"} ![image](Petersetal-Vfig3b.jpg){width="40.00000%"} ![image](Petersetal-Vfig3c.jpg){width="40.00000%"} ![image](Petersetal-Vfig3d.jpg){width="40.00000%"} ![image](Petersetal-Vfig4a.jpg){width="49.00000%"} ![image](Petersetal-Vfig4b.jpg){width="49.00000%"} ![image](Petersetal-Vfig5a.jpg){width="40.00000%"} ![image](Petersetal-Vfig5b.jpg){width="40.00000%"} ![image](Petersetal-Vfig5c.jpg){width="40.00000%"} ![image](Petersetal-Vfig5d.jpg){width="40.00000%"} ![image](Petersetal-Vfig6a.jpg){width="49.00000%"} ![image](Petersetal-Vfig6b.jpg){width="49.00000%"} ![image](Petersetal-Vfig7a.jpg){width="40.00000%"} ![image](Petersetal-Vfig7b.jpg){width="40.00000%"} ![image](Petersetal-Vfig7c.jpg){width="40.00000%"} ![image](Petersetal-Vfig7d.jpg){width="40.00000%"} ![image](Petersetal-Vfig8a.jpg){width="49.00000%"} ![image](Petersetal-Vfig8b.jpg){width="49.00000%"} Results {#sec:hydrostaticresults} ======= IC5249 ------ We have been moderately successful in modelling the dark matter halo of galaxy IC5249. The decomposition of both the optically thin and self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{} results is shown in Figure \[fig:IC5249decomposition\]. As is clear from the figure, the uncertainties in the stellar halo contribution, and subsequently the dark matter halo contribution are quite severe. This is mostly due to the fact that the measurements of both the vertical force gradient, as well as the rotation curve, start relatively far out (near 5.0-5.5kpc). The data of the inward parts of the galaxy were too uncertain for a reliable measurement of the tracers. As the dark matter halo shape $q$ can most accurately be constrained from the vertical force gradient in the inner parts of the galaxy [see @OBrien2010D], this lack of data does not allow us to constrain $q$ significantly. The optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model yields $q=1.5^{+0.5}_{-0.5}$, while the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}yields $q=1.0^{+0.4}_{-0.3}$. The lack of a significant constraint on $q$ leads to strong correlations between the other parameters. This is reflected in the cross-correlation diagrams for the parameters, seen in Figure \[fig:IC5249-crosscorr\] (left) for the optically thin case, and Figure \[fig:IC5249-crosscorr\] (right) for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}case. An oblate dark matter halo shape ($q<1$) produces a less massive stellar disc, with a shorter dark matter halo scale length $R_c$ and higher dark matter halo core density $\rho_0$. This behavior holds in both models. Over the whole dataset, we find that for an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model the halo is found at $\rho_0 =0.007^{+0.005}_{-0.001}$M$_\odot$/pc$^3$, $R_c=5.45^{+0.21}_{-0.52}$kpc. The stellar disc is found with $M_/L_R=2.62^{+0.38}_{-1.19}$. The self-absorption [ <span style="font-variant:small-caps;">HI</span> ]{}models return $\rho_0 = 0.004^{+0.003}_{-0.001}$M$_\odot$/pc$^3$, a scale length $R_c$ of $9.79^{+0.21}_{-2.06}$, and a stellar disc with $M_/L_R=2.98^{+0.02}_{-1.17}$. As we discussed in Section \[sec:MLdiscussion\], the most likely $M_/L_R$ values lie close to 0.5. If we thus limit ourselves to the data points at $M_/L_R<0.55$, we find for the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}an oblate halo with $q =0.76_{-0.03}^{+0.04}$. The core density of the dark matter halo is $\rho_0 = 0.017_{-0.001}^{+0.001}$M$_\odot$/pc$^3$ and its scale length is $R_c = 4.49_{-0.05}^{+0.02}$kpc. The self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}model returns an even more oblate halo, with a shape of $q =0.55_{-0.03}^{+0.14}$, a core density of $\rho_0 = 0.014_{-0.003}^{+0.003}$M$_\odot$/pc$^3$ and a scale length of $h_0 = 5.19_{-1.01}^{+1.30}$kpc. Comparing the two models, we see that the dark matter halo of the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}requires a more oblate halo, with longer scale length $R_c$ and less massive central density $\rho_0$. ESO115-G021 ----------- We have not been very successful in modelling ESO115-G021. As can be seen from the results in Figure \[fig:ESO115decomposition\], the observed vertical force gradient has a nearly flat slope near the inner parts of the galaxy ($R<4$kpc). This is problematic to fit to, as we always expect the vertical force gradient to get increasingly strong near the inner parts. We have attempted fitting only beyond $R=3.5$kpc, but this left only 2kpc in which we could fit the data, which did not result in a stable fit. Smoothing was applied on the input parameters, but this did not improve the quality of the observed vertical force gradient. As such, we hereby present our best fit, but encourage the reader to have a skeptical treatment of them. The self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}model of the galaxy results in a dark matter core density $\rho_0$ of $0.015^{+0.001}_{-0.001}$M$_\odot$/pc$^3$, a scale length $R_c$ of $3.02^{+0.03}_{-0.08}$kpc and an oblate shape of $q=0.5^{+0.1}_{-0.1}$. The stellar disk is found to have a high $M_/L_R=2.89^{+0.11}_{-0.44}$. The optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model produces a more massive central core density $\rho_0=0.022^{+0.004}_{-0.001}$M$_\odot$/pc$^3$, a shorter scale length of $R_c=2.30^{+0.10}_{-0.15}$ and a less oblate halo shape $q=0.7^{+0.1}_{-0.1}$. We again find a high $M_/L_R=2.84^{+0.16}_{-1.15}$\ The cross-correlation diagrams of the models are shown in Figure \[fig:ESO115-crosscorr\]. ESO138-G014 {#sec:DM138} ----------- Galaxy ESO138-G014 was initially hard to model, as the total observed vertical force gradient was already weaker than the contribution from the neutral hydrogen alone. As we noted before in Paper III, the galaxy seems to have quite a thick [ <span style="font-variant:small-caps;">HI</span> ]{}layer (see Figure 13 of Paper III). The most likely explanation for this is that the galaxy is not seen completely edge-on. This is consistent with the observed stellar disc from Paper IV, where we measured $i=86.8^\circ$ (Table 2 of Paper IV). We have attempted to correct for this by lowering the observed thickness by 30%, but will treat the results are uncertain in view of this inclination issue. The noise in this galaxy was too high for the velocity dispersion to be measure, so we keep this fixed at $\sigma=10$km/s. The results for the optically thin and the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}models are shown in Figure \[fig:ESO138decomposition\]. The cross-correlation diagram of the optically thin model is shown in Figure \[fig:ESO138-crosscorr\] (left), while the one for the self-absorption model is shown in Figure \[fig:ESO138-crosscorr\] (right). In both cases, the rotation curve and the vertical force gradient have been successfully fitted. Only at the larger radii do observed and theoretical curves of the vertical force gradient start to deviate. The optically thin model produces a halo core density $\rho_{0,\textrm{halo}}$ of $0.261^{+0.002}_{-0.003}$ M$_\odot$/pc$^3$ and a scale-length of $R_c =0.70^{+0.02}_{-0.01}$kpc. The halo is distinctly prolate, with the optimal solution located at the boundary condition of $q =2.0^{+0.1}_{-0.2}$. The stellar disc is very bright, with $M_/L_R=0.50^{+0.02}_{-0.01}$. Compared to this, the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}model finds a halo with a higher core density of $\rho_{0,\textrm{halo}} = 0.620^{+0.009}_{-0.007}$M$_\odot$/pc$^3$ and a scale-length of $R_c = 0.41_{-0.01}^{+0.02}$kpc. Again, the optimal solution favors a prolate halo at $q =1.9^{+0.1}_{-0.3}$. In this case the mass-to-light is $M_/L_{R}= 0.69^{+0.04}_{-0.06}$. ESO274-G001 ----------- Galaxy ESO274-G001 is one of the two galaxies from Paper III for which we could accurately measure the velocity dispersion. We model the galaxy in the standard way, setting the $M_/L_{R}$ lower boundary at 0.5. Both the rotation curve and vertical force decomposition are shown in Figure \[fig:ESO274decomposition\], where we show the results for both the optically thin and self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}models. Both models have reproduced the rotation curve and the vertical force gradient reasonably well, although the self-absorbing model was more successful at the vertical force gradient. We show the cross-correlation diagram for the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model in Figure \[fig:ESO274-crosscorr\] (left). There is a clear correlation between the various parameters, which is mostly due to the uncertainty in $M_/L_{R}= 0.96^{+0.46}_{-0.42}$. The halo is oblate with $q =0.7^{+0.1}_{-0.1}$, $\rho_{0,\textrm{halo}} = 0.150^{+0.020}_{-0.022}$M$_\odot$/pc$^3$ and $R_c = 1.10^{+0.04}_{-0.03}$kpc. The mass to light ratio $M_/L_{R}$ was $0.96^{+0.46}_{-0.42}$. The cross-correlation diagram of the [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption model is shown in Figure \[fig:ESO274-crosscorr\] (right). There is again scatter in $M_/L_{R}$, although the value has dropped compared to the optically thin model. It is now at $0.76^{+0.57}_{-0.26}$. The shape of the halo is identical to the optically thin model, with an oblate shape of $q=0.7^{+0.1}_{-0.1}$. The other parameters are $\rho_{0,\textrm{halo}} = 0.094^{+0.009}_{-0.019}$M$_\odot$/pc$^3$ and $R_c = 1.42^{+0.09}_{-0.03}$kpc. Compared to the optically thin model, the self-absorption [ <span style="font-variant:small-caps;">HI</span> ]{}model produces a dark matter halo with longer scale length $R_c$ and lower central density $\rho_0$. If we limit the analysis to $M_/L_R<0.55$, we find that the halo becomes even more oblate, $q = 0.64_{-0.01}^{+0.01}$ for the optically thin and $q =0.67_{-0.01}^{+0.01}$ for the self-absorption model. The central density of the halo also goes up to $\rho_0 = 0.171_{-0.001}^{+0.002}$M$_\odot$/pc$^3$ for the optically thin model, and $\rho_0 = 0.103_{-0.001}^{+0.001}$M$_\odot$/pc$^3$ for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}. The scale lengths goes down to $h_0 = 1.07_{-0.01}^{+0.01}$kpc and $h_0 = 1.39_{-0.01}^{+0.01}$kpc respectively. A lower mass in the stellar disc thus results in haloes that are more oblate, have higher central core densities and slightly shorter scale lengths. ![image](Petersetal-Vfig9a.jpg){width="40.00000%"} ![image](Petersetal-Vfig9b.jpg){width="40.00000%"} ![image](Petersetal-Vfig9c.jpg){width="40.00000%"} ![image](Petersetal-Vfig9d.jpg){width="40.00000%"} ![image](Petersetal-Vfig9e.jpg){width="40.00000%"} ![image](Petersetal-Vfig9f.jpg){width="40.00000%"} ![image](Petersetal-Vfig10a.jpg){width="40.00000%"} ![image](Petersetal-Vfig10b.jpg){width="40.00000%"} ![image](Petersetal-Vfig10c.jpg){width="40.00000%"} ![image](Petersetal-Vfig10d.jpg){width="40.00000%"} ![image](Petersetal-Vfig10e.jpg){width="40.00000%"} ![image](Petersetal-Vfig10f.jpg){width="40.00000%"} ![image](Petersetal-Vfig11a.jpg){width="49.00000%"} ![image](Petersetal-Vfig11b.jpg){width="49.00000%"} UGC7321 ------- Galaxy UGC7321 has the highest signal to noise ratio for the [ <span style="font-variant:small-caps;">HI</span> ]{} data from our sample. The galaxy was previously modelled by @OBrien2010D, who found that the halo flattening $q$ was round ($q=1.0\pm0.1$). Their modelling strategy consisted of a two-pass scheme, in which they first performed a rotation curve decomposition, and only then performed a separate fit to the vertical force gradient. This second fit however failed, and the authors had to drastically deviate from the results from the rotation curve decomposition, and use a very low mass stellar disc, in order to reproduce the observed vertical force gradient. We have performed an inspection of the codes used by in the analysis of @OBrien2010D. It appears there was a restriction which allowed only models with $q\leq1$ and it would have been impossible for them to fit a prolate halo. @Banerjee2010A also analysed [UGC7321]{} and found a spherical halo. These authors assumed a constant velocity dispersion, or at most a decreasing gradient, in their work, and use a different potential than us. In Figure \[fig:UGC7321rotation\], we demonstrate our own rotation curve decomposition of this galaxy, and in Figure \[fig:UGC7321zforce\] we show the vertical force gradient decomposition. Rather than present decompositions for only the optically thin and self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}models that we measured, as for the previous galaxies, we show the results for six fits. Since @OBrien2010D found a good fit at a negligible stellar mass, the first panel in both figures demonstrates a fit in which the $M_/L_{R}$ can range between zero and three, for an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}disc. This fit should therefore be the closest to the results obtained by @OBrien2010D. The 1000 samples produce a range of solutions. Both the rotation curve and the vertical force gradient are well reproduced. The mass-to-light ratio has a median of $M_/L_{R}= 1.58_{-0.68}^{+0.45}$, which is significantly higher than measured by @OBrien2010D. The halo is very prolate, $q =1.90_{-0.32}^{+0.10}$ and has a high central density of $\rho_{0,\textrm{halo}} = 0.324_{-0.043}^{+0.067}$M$_\odot$/pc$^3$ and short scale-length of $R_c = 0.64_{-0.02}^{+0.05}$kpc. Previously, @OBrien2010D have reported $R_c = 0.52\pm0.02$kpc and $ \rho_{0,\textrm{halo}} =0.73\pm0.05$M$_\odot$/pc$^3$. As we estimate a minimum of $M_/L_{R}=0.5$, the second panel raises the boundary condition for the minimal stellar mass-to-light ratio to $0.5$. The mass-to-light is found to be $M_/L_{R}= 1.63_{-0.66}^{+0.48}$, still very similar to the previous model. The observed rotation curve and vertical velocity gradients are well reproduced, as shown in Figures \[fig:UGC7321rotation\] and \[fig:UGC7321zforce\]. We find $\rho_{0,\textrm{halo}} = 0.318_{-0.042}^{+0.064}$M$_\odot$/pc$^3$ and $R_c = 0.64_{-0.02}^{+0.03}$kpc, roughly similar values as the previous fit. The halo shape runs firmly towards $q=2$, which is also the boundary condition. We have tested the effect of lifting this boundary condition. When we do this, the model tends to run towards even greater values of $q$. However, since the current research question focuses primarily on prolate versus oblate, we have decided to stick to an upper boundary of $q=2$. We present a cross correlation diagram of this fit in Figure \[fig:UGC7321crosscorrelation\] (left). Our next fit uses the self-absorption [ <span style="font-variant:small-caps;">HI</span> ]{}model rather than the optically thin model. We again let $M_/L_{R}$ run from zero to three. As shown in Figures \[fig:UGC7321rotation\] and \[fig:UGC7321zforce\], the stellar disc in this fit gets a negligible mass assigned ($M_/L_{R}= 0.00_{-0.00}^{+0.04}$). The other parameters are $\rho_{0,\textrm{halo}} =0.307_{-0.003}^{+0.002}$M$_\odot$/pc$^3$, $R_c = 0.71_{-0.01}^{+0.01}$kpc and $q=1.84_{-0.09}^{+0.12}$. Compared to the optically thin model, the dark matter halo is again strongly prolate, but has a longer scale-length and lower central density. Similar to the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}case, we again increase the lower $M_/L_{R}$ boundary to $0.5$. The results are shown in Figures \[fig:UGC7321rotation\] and \[fig:UGC7321zforce\]. The observed rotation curve has been modelled well, but the model fails to account for the vertical force gradient and produces too strong a vertical force gradient. This directly illustrates why $M_/L_{R}$ was zero in the previous fit, as this was the only way for the vertical force gradient to be fit. The parameters found are $\rho_{0,\textrm{halo}} = 0.28_{-0.01}^{+0.01}$M$_\odot$/pc$^3$, $R_c =0.72_{-0.01}^{+0.01}$kpc, $q=2.00_{-0.03}^{+0.01}$ and $M_/L_{R}=0.5$. The cross-correlation diagram for this fit is shown in Figure \[fig:UGC7321crosscorrelation\] (right).\[sec:UGC7321-decomposition\] In their revious study, @OBrien2010D were unable to fit the rotation curve and the vertical force gradient simultaneously. They successfully started with a rotation curve decomposition, in which the stellar disc mass was a free parameter. However, in order to subsequently perform their vertical force gradient decomposition, they were forced to drastically lower the stellar mass. They eventually found a spherical halo, but only when they allowed very small $M_/L_{R}$ (the best fit was actually for $M_/L_{R}=0$), smaller than we allowed here. As two final tests, we ran fits to the optically thin and self-absorption [ <span style="font-variant:small-caps;">HI</span> ]{}results, in which we constrained $q=1$. The results are shown in Figures \[fig:UGC7321rotation\] and \[fig:UGC7321zforce\]. The rotation curve decomposition does not depend strongly on $q$ [@OBrien2010D]. As such, it is again reproduced well. Clearly, however, the vertical force gradient is fit poorly. A spherical halo simply does not work for this galaxy. Name [ <span style="font-variant:small-caps;">HI</span> ]{}model $\rho_0$ \[M$_\odot$/pc$^3$\] $R_c$ \[kpc\] $q$ $M_* / L_R$ ------------- ------------------------------------------------------------- ------------------------------- ------------------------ --------------------- ------------------------ IC5249 SA $0.004^{+0.003}_{-0.001}$ $9.79^{+0.21}_{-2.06}$ $1.0^{+0.4}_{-0.3}$ $2.98^{+0.02}_{-1.17}$ IC5249 OT $0.007^{+0.005}_{-0.001}$ $5.45^{+0.21}_{-0.52}$ $1.5^{+0.5}_{-0.5}$ $2.62^{+0.38}_{-1.19}$ ESO115-G021 SA $0.015^{+0.001}_{-0.001}$ $3.02^{+0.03}_{-0.08}$ $0.5^{+0.1}_{-0.1}$ $2.89^{+0.11}_{-0.44}$ ESO115-G021 OT $0.022^{+0.004}_{-0.001}$ $2.30^{+0.10}_{-0.15}$ $0.7^{+0.1}_{-0.1}$ $2.84^{+0.16}_{-1.15}$ ESO138-G014 SA $0.620^{+0.009}_{-0.007}$ $0.41^{+0.02}_{-0.01}$ $1.9^{+0.1}_{-0.3}$ $0.69^{+0.04}_{-0.06}$ ESO138-G014 OT $0.261^{+0.002}_{-0.003}$ $0.70^{+0.02}_{-0.01}$ $2.0^{+0.1}_{-0.2}$ $0.50^{+0.02}_{-0.01}$ ESO274-G001 SA $0.094^{+0.009}_{-0.019}$ $1.42^{+0.09}_{-0.03}$ $0.7^{+0.1}_{-0.1}$ $0.76^{+0.57}_{-0.26}$ ESO274-G001 OT $0.150^{+0.020}_{-0.022}$ $1.10^{+0.04}_{-0.03}$ $0.7^{+0.1}_{-0.1}$ $0.96^{+0.46}_{-0.42}$ UGC7321 SA $0.270^{+0.001}_{-0.001}$ $0.72^{+0.01}_{-0.01}$ $2.0^{+0.1}_{-0.1}$ $0.50^{+0.02}_{-0.01}$ UGC7321 OT $0.318^{+0.064}_{-0.042}$ $0.64^{+0.03}_{-0.02}$ $1.9^{+0.1}_{-0.3}$ $1.63^{+0.48}_{-0.66}$ In the previous section, we have presented the results for the individual galaxies. So how do the results compare to each other? In Table \[tbl:DMhaloes\], we present an overview of all the derived parameters. For ESO138-G014, we only present the results where the thickness of the [ <span style="font-variant:small-caps;">HI</span> ]{}layer has been reduced by 30%. For galaxy UGC7321, we present the results for the default model, in which the mass-to-light ratio $M_/L_R$ was allowed to vary between 0.5 and 3.0, and the halo shape $q$ between 0.1 and 2.0. We present an overview of the average of these parameters in Table \[tbl:globalhalo\]. There is an interesting difference between optically thin and self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}models. Overall, we see that the halo of an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model has a core density that is overestimated by $150\%$. The scale length of the dark matter halo is $28\%$ longer in the self-absorption model, compared to the optically thin model. In addition, where the optically thin models have a median shape that is prolate, the median shape is spherical in self-absorption [ <span style="font-variant:small-caps;">HI</span> ]{}models. The mass-to-light ratio of the stellar disc drops by more than half when self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}is accounted for. All of our five discs are sub-maximal. This was already demonstrated by @OBrien2010D and @Banerjee2010A for UGC7321, who reported a stellar disc M/L$_R$ at a maximum of 2.5, although their final decomposition found a maximum of M/L$_R$ = 0.2. The result is consistent with the work by @Bershady2011A, who argued that all galaxies are sub-maximal based on an analysis of the central vertical velocity dispersion of the discs stars and the maximum rotation of 30 face-on galaxies. Similar conclusions have previously been reached by @bot97 and @Kregel2005. @Martinsson2013A confirmed these results after performing dynamically determined rotation curve mass decompositions for all these 30 galaxies. Name $\rho_0$ \[M$_\odot$/pc$^3$\] $R_c$ \[kpc\] $q$ $M_ / L_R$ Citation ----------------------- ------------------------------- ---------------------- ------------------------ ------------- ---------------- UGC7321 $0.73\pm0.05$ $0.52\pm0.02$ $1.0\pm0.1$ $0.2$ @OBrien2010D UGC7321 $0.048\pm0.009$ $2.7\pm0.2$ 1 @Banerjee2010A NGC4244 $0.6595\pm0.6495$ $6.15\pm5.85$ 0.2 @Olling1996A M31 0.011 21 0.4 @Banerjee2008A Galaxy (NFW) $0.0413_{-0.016}^{+0.013}$ $9.26_{-4.2}^{+5.6}$ @Nesti2013 Galaxy $0.0147_{-0.0093}^{+0.029}$ $16.1_{-7.8}^{+17}$ @Nesti2013 Galaxy 12 5/3 @Helmi2004A Galaxy 7.1 0.7 @Olling2000 Lensing $23.9_{-0.5}^{+0.2}$ $0.62_{-0.25}^{+0.25}$ @vanUitert2012 ESO138-G014 (NFW) $0.013\pm0.002$ $7.5\pm0.5$ @Hashim2014 ESO138-G014 (Burkert) $0.077\pm0.042$ $10.7\pm2.7$ @Hashim2014 In Figure \[fig:DMdiscussion\], we demonstrate the correlation between the four free parameters from our fit. We have also included a range of points from other authors in this view (see Table \[tbl:otherDM\] for an overview, note that multiple halo models are used an as such the core radius can be expected to vary). Inspecting the figure, the most notable correlations is the one between core radius $R_c$ and halo core density $\rho_0$. With the exception of one point, ESO138-G14 by @Hashim2014 using a NFW halo, all of the points seem to follow a relation of $R_c\approx1/\rho_0$. Our result for ESO138-G14 are uncertain due to the possiblity of a r4esisual inclination compared to edge-on and any results on this galaxy, including ours at $q\sim 2$ should be treated with caution. This relation is similar to the degeneracy between the two parameters in an individual galaxy, as for example in Figure 11, and it interesting to observe a similar trend visible across multiple galaxies and halo models. If this is a true relation, then it implies that there are two families of haloes: one compact halo family with high core density $\rho_0$ and scale length $R_c$, and a second non-compact halo family with low core density $\rho_0$ and scale length $R_c$. name optically thin self absorbing ----------- ------------------ ------------------ $\rho_0$ $0.150\pm 0.124$ $0.094 \pm0.230$ $R_c$ $1.10\pm 1.81$ $1.42\pm 3.48$ $q$ $1.5\pm 0.6$ $1.0\pm 0.6$ $M_/L_R$ $1.63\pm 0.91$ $0.76\pm 1.12$ : Overview of the global parameters of our halo sample. Shown are the median and standard deviations of the parameters. The units of $\rho_0$ are in M$_\odot$/pc$^3$ and those of the radius $R_c$ in kpc.[]{data-label="tbl:globalhalo"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12a.jpg "fig:"){width="23.00000%"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12b.jpg "fig:"){width="23.00000%"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12c.jpg "fig:"){width="23.00000%"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12d.jpg "fig:"){width="23.00000%"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12e.jpg "fig:"){width="23.00000%"} ![Overview of the correlations between the various halo parameters. Blue crosses represent the halo parameters for the self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}, while green circles represent the optically thin models. Black markers are the data points from Table \[tbl:otherDM\][]{data-label="fig:DMdiscussion"}](Petersetal-Vfig12f.jpg "fig:"){width="23.00000%"} Our best results are for ESO274-G001 and UGC7321, where we have been able to include a measurement of the velocity dispersion of the [ <span style="font-variant:small-caps;">HI</span> ]{}as function radius in the analysis. Focusing on these two, we find two very different haloes. In ESO274-G001, the halo is oblate with a shape $q=0.7\pm0.1$ (regardless of the [ <span style="font-variant:small-caps;">HI</span> ]{}model); while in UGC7321, the halo is distinctly prolate with a shape of $q=1.9^{+0.1}_{-0.3}$ (optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}) and $q=2.0\pm0.1$ (self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}). So how do these shapes compare to other galaxies? Looking at our own Galaxy, @Law2010A propose a triaxial dark matter halo for the Milky Way, in which $(c/a)_\phi=0.72$ and $(b/a)_\phi=0.99$. @Banerjee2011A proposed a dark matter halo shape for the Galaxy which becomes progressively more prolate with radius. @VeraCiro2013 report an oblate halo with $q=0.9$ for the inner 10kpc, based on stellar streams. Using lensing, @Barnabe2012A also found a slightly oblate halo at $q=0.91^{+0.15}_{-0.13}$ for galaxy SDSSJ2141. The large weak lensing galaxies sample of @vanUitert2012, in which $2.2\times10^7$ galaxies were studied, produced a halo ellipticity distribution that also favors oblate haloes. The distributions of the halo shape $q$ was $0.62^{+0.25}_{-0.25}$. The three polar ring galaxies studied by @Whitmore1987A, A0136-0801, NGC4650A and ESO415-G026, had slightly oblate to spherical halo shapes $q$: $0.98\pm0.20$, $0.86\pm0.21$ and $1.05\pm0.17$. From this selected sample of papers, it becomes apparent that our result for ESO274-G001, with $q=0.7\pm0.1$, is consistent with other papers. While ESO274-G001 clearly matches up with measurements in other galaxies, the other case of our two best fits, UGC7321, is a more ‘problematic’ one (Figures 13, 14, 15 and 16). With a halo shape of $q=1.9_{-0.3}^{+0.1}$ for the self-absorption model, the dark matter halo shape is very strongly prolate. As we commented before in Section \[sec:UGC7321-decomposition\], our upper boundary condition for the halo shape is $q\leq2$. If we had removed this boundary, some of the fits returned results as high as $q\sim5000$, which are clearly not physical. The galaxy has been previously analysed by @Banerjee2010A, whom successfully modelled the dark matter halo shape for a spherical halo. @OBrien2010D had problems fitting the dark matter halo shape. They had to lower their initially measured asymptotic halo rotation (see Equation \[eqn:dmhalovelocity\]) in order to get a successful fit to their data at $q=1.0\pm0.1$, although they were limited to $q\leq1$ in their analysis. Had their boundary condition been higher, it would have been likely that they too would have found higher values for $q$. Concerns regarding reliability and degeneracy --------------------------------------------- Given that our two best galaxies produce such drastically different results, how reliable is our methodology? To answer this question, let us recap the underlying assumptions from this paper and Paper IV. ### Concerning the neutral hydrogen We start with the neutral hydrogen. In Paper I, we argued that the [ <span style="font-variant:small-caps;">HI</span> ]{}in edge-on galaxies could suffer from significant self-absorption. To model the [ <span style="font-variant:small-caps;">HI</span> ]{}more accurately, we developed a new tool that allowed the neutral hydrogen in galaxies to be fit automatically, while incorporating a treatment for the self-absorption of the gas. Indeed, we saw in Section 7 of Paper II that the visible mass of a galaxy drops as one rotates it from face-on to fully edge-on. In Paper II, we developed a method to model the [ <span style="font-variant:small-caps;">HI</span> ]{}content of a galaxy that was edge-on. In Section 3 of Paper III, we tested this method on a series of simulated galaxies, showing that we could reproduce the input parameters reasonably well using our method. We also demonstrated that assuming an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}disc, which in reality was self-absorbing, could lead to a wrong measure of face-on surface density, thickness of the [ <span style="font-variant:small-caps;">HI</span> ]{}layer and the velocity dispersion. We have continued the use of the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}results into this paper to demonstrate how the dark matter halo measurement is affected by this. As discussed in the previous section, the results are drastic. How accurate are our results now? One of the key assumptions made in Paper II was the effective spin temperature of the neutral hydrogen of $100$K. While this was has proven a very successful value on which to base our results, it is an assumption based purely on what seemed to work best. In reality the neutral hydrogen most likely consists of multiple phases, such as the cold neutral medium (CNM) and warm neutral medium (WNM). The effective spin temperature is a result of the mix of the phases of the CNM, which has a median spin temperature of 80K and the WNM, with temperatures between 6000 and 10000K. In Section 4 of Paper II, we demonstrated how the interplay of [ <span style="font-variant:small-caps;">HI</span> ]{}gas phases could lead to an effective median spin temperature. So what would be the consequence of a wrong estimate of the spin temperature? Suppose that the spin temperature would have been $T_\textrm{spin}=90$K rather than 100K. In that case the face-on surface density of the neutral hydrogen will be higher, which in this paper would lead to higher theoretical rotation from the gas components in the galaxy and stronger vertical force gradients then are currently found. Simultaneously, the thickness of the disc would be smaller, and thus the total observed vertical force gradient would be larger (Equation \[eqn:Ftotal\]). Although hard to estimate the exact effect, the phases of the [ <span style="font-variant:small-caps;">HI</span> ]{}all have different distributions, together producing the observed thickness of the disc [@Lockman1991]. The effective spin temperature could thus be height dependent as well. Another assumption is the uniform density of the [ <span style="font-variant:small-caps;">HI</span> ]{}as a function of radius. In reality of course, galaxies have spiral arms, supernova, shocks, gravitational collapse, and other features, all of which create a drastically non-uniform [ <span style="font-variant:small-caps;">HI</span> ]{}disc. The question thus remains how strongly the parameters are affected by this. @Kamphuis2013A made a valiant attempt to model the density waves in galaxies NGC5023 and UGC2082, demonstrating that these could be detected in edge-on galaxies[^4]. Indeed, as we discussed in Section 5 of Paper I, the position-velocity (XV-)diagrams are not symmetric on both side of the galaxies. This problem would most strongly affect the velocity dispersion, which is dependent upon small-scale features. In most cases, this leads to an overestimation of the [ <span style="font-variant:small-caps;">HI</span> ]{}velocity dispersion, as the fitting algorithm tries to ‘smooth over’ the small-scale fluctuations. The result would be an overestimation of the observed vertical force gradient through Equation \[eqn:Ftotal\]. This could be a likely reason why UGC7321 has such a distinctive halo shape. We have also assumed that the [ <span style="font-variant:small-caps;">HI</span> ]{}has an isotropic velocity dispersion tensor, i.e has the same value in the $R$, $z$ and $\theta$ directions (see Section 2.2 of Paper II). While there currently is no observational proof that this is an invalid assumption, it remains untested. If the velocity dispersion tensor were in reality anisotropic, it would imply that our observed vertical force gradient is wrong (Equation \[eqn:Ftotal\]). Simultaneously, it would affect the amount of observed [ <span style="font-variant:small-caps;">HI</span> ]{}in the self-absorption mode, as gas with low velocity dispersion would give a alrger effect than gas with high velocity dispersion (Equation 19 of Paper II). It would affect the rotation curve measurements to a minor degree. We also have assumed that the velocity dispersion is isothermal in $z$, i.e. does not vary with height. In Section 4.7 of Paper III, we have attempted to measure this in ESO274-G001. We found a very small increase of 1km/s in the slice above the central 290pc of the disc. If this were confirmed in other galaxies, it would mean that Equation \[assumption:isothermal\] is false, and thus our observed total force gradient would have been wrong. Previously, @Lockman1991 has attempted to model the vertical structure of the Galaxy using multiple [ <span style="font-variant:small-caps;">HI</span> ]{}phases, which had different scale heights and velocity dispersions. The phases with the highest scale heights also have the highest velocity dispersions. It is quite possible that the effective spin temperature of the different phases would increase for those with larger scale heights. Assuming that our currently observed velocity dispersion is due to the combination of high and low $z$ gas, the total vertical force gradient than we are currently reporting would be weaker in the mid-plane, and stronger at high values of $z$. The thickness of the [ <span style="font-variant:small-caps;">HI</span> ]{}disc has been assumed to follow a Gaussian form (see Equation 15 of Paper II) for mathematical convenience in Equation \[eqn:Ftotal\]. Other possible candidates for the [ <span style="font-variant:small-caps;">HI</span> ]{}model would have been the sech and sech$^2$ functions. @Olling1995A previously discussed the various types of discs and concluded that the changes due to this would be minor. A sech function has more extended wings and steeper inner slopes, compared the Gaussian function. If our galaxies have high-projected latitude gas, such as due to warps or [ <span style="font-variant:small-caps;">HI</span> ]{}haloes, then it is possible that a fit with a Gaussian function would find the FWHM of the disc to be unrealistically large. A fit with a sech or even sech$^2$ function could then be a better approximation to the shape of the [ <span style="font-variant:small-caps;">HI</span> ]{}disc. Another assumption in our model is the perfect edge-on nature of the [ <span style="font-variant:small-caps;">HI</span> ]{}disc. In Section 3.5 of Paper III, we tested how our [ <span style="font-variant:small-caps;">HI</span> ]{}fitting strategy worked on a galaxy at $i=88.8^\circ$. We found that the parameters were well recovered. However, suppose that some galaies are even further from edge-on than that, as formally indicated by our stellar decompositions (Table 4 of Paper IV). In that case, the thickness of the [ <span style="font-variant:small-caps;">HI</span> ]{}is probably overestimated and the circular rotation underestimated. This has a profound effect on the rotation curve decomposition, which would require a larger rotational contribution due to the dark matter. In a similar vein, the total observed vertical force gradient would be underestimated, and would thus require a more massive dark matter halo in the decomposition. On the other hand, the density of the [ <span style="font-variant:small-caps;">HI</span> ]{}would be lower at a height of 100pc, such that in Equation \[eqn:poissonz\] the vertical force gradient due to the [ <span style="font-variant:small-caps;">HI</span> ]{}would be lower. ### Concerning the stellar disc In Paper IV, we set out to use the <span style="font-variant:small-caps;">FitSKIRT</span> tool to model the stellar disc of our sample of galaxies. We modelled the galaxies using a stellar disc and a bulge component. While the results were acceptable in most cases, these galaxies were selected to be relatively bulge-less (Section 3 of Paper I). Because of this, the fitting routine was found to ‘misuse’ the bulge component as a tool to better model the stellar disc. Thus, the bulges in most of the galaxies serve more like an extension of the stellar disc, rather than like a separate central component (see Table 4 of Paper IV for the parameters). In some cases, the amount of light from the bulge component is similar to that of the stellar disc itself (Table 3 of Paper IV). Due to the massive amount of processing power required to perform the stellar decompositions, we have been unable to test how well the fitting of just an exponential disc to the data would have worked. Most likely, the results are similar to the combined parameters adopted here. The stellar models in Table 4 of Paper IV were often found to deviate from complete edge-on. In the absence of dust lanes, which could prove an independent check, it remains unclear how accurate this result is. If the galaxies were in reality more edge-on than measured, the stellar discs would have shorter scale heights. In Equation \[eqn:poissonz\], this would imply a lower stellar density at a height of $z=100$pc, and thus would require a stronger dark matter halo vertical force gradient. ### Concerning the cross-correlation between parameters An advantage of our MCMC method is that we have not one, but a whole range of parameter sets for each galaxy. This allows us to explore the interplay between the cross-correlations, such as demonstrate Figure 12 for the self-absorption model of ESO274-G001. As is clear from this figure, a whole range of solutions can be valid. For example, one parameter set returned a core density of 0.1M$_\odot$/pc$^3$, with scale length of 1.4kpc, halo shape $q$ of $0.67$ and $M_/L_R$ of 0.4. A different but equal parameter set returned a core density of $0.06$,M$_\odot$/pc$^3$, with scale length of nearly 1.6kpc, shape $q=0.84$ and $M_/L_R$ of 1.7. These results are drastically different, yet both are accepted parameter sets. The largest source of uncertainty is the stellar disc $M_/L_R$. It was beyond the scope of this project to measure this parameter in each of our galaxies, which is why we have adopted it as a free parameter. A different solution would have been to adopt the maximum value of $M_/L_R$ permitted in the rotation curve decomposition, the so-called maximum disc approach (e.g. @Carignan1985 [@vanAlbada1985A]). However, these galaxies were selected to be dark matter dominated at all radii and as such this approach would have been invalid (Section 3 of Paper I). In addition, the applicability of the maximum disc criteria has already been questioned by @Kregel2005 and @Martinsson2013, whom both report sub-maximal stellar discs. Although beyond the scope of this project, the best approach would be to perform a full stellar population synthesis analysis of each galaxy. For examples see @Bruzual2003 and @Maraston2005. By measuring M$_\odot$/pc$^3$ rather than using it as a free parameter, the solution space becomes far less degenerate and the parameters can thus be fixed far more accurately. Another cause of concern is the boundary conditions imposed upon our data. We have done our best to impose realistic boundary conditions. For the dark matter halo shape, we adopted $0.1<q<2.0$, as we believe that even more oblate or prolate haloes would be unrealistic. In Section \[sec:MLdiscussion\] we adopted $0.5<M_/L_R<3.0$ as a likely boundary, based on the stellar population models by @Worthey1994A and @Bertelli1994A. As can clearly be seen from the various cross-correlation diagrams, the models often still run into the boundary conditions. While it is possible to raise or remove the boundary conditions, we do not believe that this would lead to realistic results and we have therefore refrained from doing so. ### Halo model In this work, we have adopted the dark matter halo model by @Sackett1994A. With this model, we can create flattened, axi-symmetric, pseudo-isothermal haloes. In this model, the density is stratified in concentric ellipsoids. We have chosen this model to be able to compare directly to @OBrien2010D, who also use this model. There are many other halo model. @Carignan1985 [@Carignan1988] used isothermal, rather than pseudo-isothermal haloes to model their galaxies. @Kormendy2004IAU compared the merits between isothermal and pseudo-isothermal haloes. As is shown in that paper [and reproduced in @OBrien2010D], the rotation curve of an isothermal halo initially rises above the asymptotic velocity $v_H$, before dropping towards it again. In contrast, the pseudo-isothermal rotation curve approaches the asymptotic velocity gradually from lower values. This behavior would affect the results for the rotation curve decomposition. There are more models, such as the NFW and Burkert halo model, each of which has some mathematical or theoretical advantage [@Burkert1995; @Navarro1996]. Even more exotic models exist in which the dark matter halo shape can vary with radius [@VeraCiro2013]. This can for example lead to haloes that get progressively more prolate with radius [@Banerjee2011A]. While all these haloes are very interesting, we believe that the quality of the data, as discussed in this section, does not warrant such a detailed exploration of the properties of the various halo models. An altogether different solution would have been the use of MOND, which would have removed the need for a dark matter halo altogether [@Milgrom1983A]. We find that in many of our vertical force decompositions, a slight increase in the mass of the [ <span style="font-variant:small-caps;">HI</span> ]{}and stellar disc would be sufficient to account for the total observed vertical force gradient. As we argued before, additional mass in both the stellar and [ <span style="font-variant:small-caps;">HI</span> ]{}discs is allowed for by the data. While it is beyond the scope of this project to test MOND on our data, it is an interesting avenue for further research. Conclusions {#sec:hydrostaticssummary} =========== We have attempted to measure the shape of the dark matter halo in five galaxies, using a simultaneous decomposition of the rotation curve and of the vertical force gradient at the mid-plane. For a dark matter halo model, we have adopted the @Sackett1994A dark matter halo. Both optically thin and self-absorbing [ <span style="font-variant:small-caps;">HI</span> ]{}models were used. We find that this leads to drastically different results. As we have argued in Papers I, II and III, the [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption models are the more accurate representation of galaxies. Using [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption, we found that a typical dark matter halo has a less dense core ($0.094\pm0.230$M$_\odot$/pc$^3$)[^5] compared to an optically thin [ <span style="font-variant:small-caps;">HI</span> ]{}model ($0.150\pm0.124$M$_\odot$/pc$^3$). The [ <span style="font-variant:small-caps;">HI</span> ]{}self-absorption dark matter halo had a longer scale length $R_c$ of $1.42\pm 3.48$kpc, versus $1.10\pm 1.81$kpc for the optically thin [ <span style="font-variant:small-caps;">HI</span> ]{} model. The median halo shape was spherical, at $q=1.0\pm0.6$ (self-absorbing), while it was prolate at $q=1.5\pm0.6$ for the optically thin. Our best results were obtained for ESO274-G001 and UGC7321, for which we were able to measure the velocity dispersion in Paper III. These two galaxies have drastically different halo shapes. ESO274-G001 was found to be oblate at $q=0.7\pm0.1$ (both [ <span style="font-variant:small-caps;">HI</span> ]{}models), while UGC7321 returns a distinctly prolate halo at $q=1.9_{-0.3}^{+0.1}$ (optically thin) and $q=2.0\pm0.1$ (self-absorbing). The halo of ESO274-G001 iss similar to those found in other studies, but UGC7321 is more problematic. In UGC7321, the most likely cause of concern is the presence of spiral arms and an [ <span style="font-variant:small-caps;">HI</span> ]{}halo. With these drastically different results, we concluded that the question whether haloes are oblate or prolate is not settled. The results for both of our best galaxies appear to be fine. A larger set of galaxies needs to be analysed, before it can become clear if one of these galaxies is an outlier, or if prolate and oblate haloes are equally likely in nature. We extensively discussed the various assumptions and sources of uncertainty in our models, of which there are many. While we have done our best to treat for these assumptions, for example using MCMC fits to the [ <span style="font-variant:small-caps;">HI</span> ]{}cube, we found that fitting the hydrostatics of the dark matter halo using the vertical force gradient near the mid-plane of the galaxy will always be tricky. Acknowledgments {#acknowledgments .unnumbered} =============== SPCP is grateful to the Space Telescope Science Institute, Baltimore, USA, the Research School for Astronomy and Astrophysics, Australian National University, Canberra, Australia, and the Instituto de Astrofisica de Canarias, La Laguna, Tenerife, Spain, for hospitality and support during short and extended working visits in the course of his PhD thesis research. He thanks Roelof de Jong and Ron Allen for help and support during an earlier period as visiting student at Johns Hopkins University and the Physics and Astronomy Department, Krieger School of Arts and Sciences for this appointment. PCK thanks the directors of these same institutions and his local hosts Ron Allen, Ken Freeman and Johan Knapen for hospitality and support during many work visits over the years, of which most were directly or indirectly related to the research presented in this series op papers. Work visits by SPCP and PCK have been supported by an annual grant from the Faculty of Mathematics and Natural Sciences of the University of Groningen to PCK accompanying of his distinguished Jacobus C. Kapteyn professorhip and by the Leids Kerkhoven-Bosscha Fonds. PCK’s work visits were also supported by an annual grant from the Area of Exact Sciences of the Netherlands Organisation for Scientific Research (NWO) in compensation for his membership of its Board. \[lastpage\] [^1]: For more information, please contact P.C. van der Kruit by email at [email protected]. [^2]: Available at astro.wsu.edu/worthey/dial/dial\_a\_pad.html. [^3]: Note that we have also discussed this assumption in Section 7 of Paper IV. [^4]: These authors did not model the [ <span style="font-variant:small-caps;">HI</span> ]{}as self-absorbing, which most likely is hampering their results. [^5]: Central value is the median, error is the standard deviation.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We explore the interspecies interaction quench dynamics of ultracold spin-polarized few-body mass balanced Fermi-Fermi mixtures confined in a double-well with an emphasis on the beyond Hartree-Fock correlation effects. It is shown that the ground state of particle imbalanced mixtures exhibits a symmetry breaking of the single-particle density for strong interactions in the Hartree-Fock limit, which is altered within the many-body approach. Quenching the interspecies repulsion towards the strongly interacting regime the two species phase separate within the Hartree-Fock approximation while remaining miscible in the many-body treatment. Despite their miscible character on the one-body level the two species are found to be strongly correlated and exhibit a phase separation on the two-body level that suggests the anti-ferromagnetic like behavior of the few-body mixture. For particle balanced mixtures we show that an intrawell fragmentation (filamentation) of the density occurs both for the ground state as well as upon quenching from weak to strong interactions, a result that is exclusively caused by the presence of strong correlations. Inspecting the two-body correlations a phase separation of the two species is unveiled being a precursor towards an anti-ferromagnetic state. Finally, we simulate in-situ single-shot measurements and showcase how our findings can be retrieved by averaging over a sample of single-shot images.' author: - 'J. Erdmann' - 'S. I. Mistakidis' - 'P. Schmelcher' bibliography: - 'bibliography.bib' title: \| Phase Separation Dynamics Induced by an Interaction Quench\ of a Correlated Fermi-Fermi Mixture in a Double Well --- Introduction ============ Ultracold Fermi gases offer an excellent testbed for simulating and exploring exotic quantum phases of matter [@giorgini2008theory; @chevy2010ultra; @bloch2008many]. Recent experimental advances constitute a valuable resource for disclosing the intricate complexity of condensed matter systems. Indeed several key quantities can be adjusted in the laboratory including the interparticle interaction strength via Feshbach resonances [@inouye1998observation; @chin2010feshbach], the particle number [@serwane2011deterministic; @zurn2012fermionization; @wenz2013few] and the external potential landscape [@bloch2008many; @greiner2002quantum]. Besides single species also mixtures of fermions can nowadays be experimentally prepared with neutral fermionic atoms e.g. in different hyperfine states such as $\prescript{40}{}{K}$ [@wu2012ultracold; @wille2008exploring], $\prescript{6}{}{Li}$ [@wille2008exploring; @moerdijk1995resonances] and $\prescript{87}{}{Sr}$ [@takamoto2006improved]. In this context, impressive features have been revealed evincing for instance superfluidity [@chen2005bcs; @chin2006evidence], quantum magnetism [@sowinski2018ground; @hung2011exotic; @yannouleas2016ultracold; @koutentakis2018probing], insulating phases [@lisandrini2017topological; @nataf2016chiral; @zhou2016mott], phase separation [@shin2006observation; @partridge2006pairing; @shin2008phase], fermi polarons [@massignan2014polarons; @scazza2017repulsive; @schmidt2018universal; @mistakidis2018repulsive] and Josephson junctions [@macri2013tunneling; @valtolina2015josephson; @spuntarelli2007josephson; @erdmann2018correlated]. A major focus has been the phase diagram of Fermi-Fermi (FF) mixtures ranging from the strongly attractive to the strongly repulsive regime of interactions [@partridge2006pairing; @iskin2007mixtures; @cazalilla2009ultracold; @jo2009itinerant; @cazalilla2009ultracold; @greif2013short; @cherng2007superfluidity; @sanner2012correlations]. For instance, referring to attractive particle imbalanced FF mixtures it has been shown that beyond a critical polarization the mixture forms a superfluid paired core being surrounded by a shell of unpaired fermions [@partridge2006pairing; @iskin2007mixtures]. Turning to the repulsive regime of interactions magnetization effects emerge. For increasing repulsion, a first order phase transition [@cazalilla2009ultracold] between paramagnetism and itinerant ferromagnetism [@jo2009itinerant; @cazalilla2009ultracold; @greif2013short; @cherng2007superfluidity; @sanner2012correlations] has been revealed. It has been argued that this transition can be described by the mean-field model of Stoner [@stoner1933lxxx; @snoke2009solid] for strongly short-range repulsively interacting fermions. The majority of the above-mentioned studies has been focussed on the static properties of FF mixtures within a Hartree-Fock (HF) i.e. mean-field description in higher dimensions. Most importantly, the dynamical properties of FF mixtures are largely unexplored and especially the role of many-body (MB) effects is much less understood. An intriguing prospect here is whether magnetization or phase separation effects emerge during the nonequilibrium dynamics of FF mixtures. A widely used technique to induce the nonequilibrium dynamics is the so-called quantum quench [@polkovnikov2011colloquium; @langen2015ultracold], where the quantum evolution is generated following a sudden change of an intrinsic system’s parameter such as the interaction strength [@kollath2007c; @mistakidis2014interaction; @mistakidis2015negative; @mistakidis2017correlation; @mistakidis2017mode]. For instance, it has been recently shown that the interaction quench dynamics of a Bose-Bose mixture crossing the miscibility-immiscibility threshold leads to the dynamical phase separation of the two clouds which exhibit domain-wall structures [@mistakidis2017correlation]. Turning to FF mixtures a natural question that arises is whether such a phase separation can be observed and what is its dependence on the particle number of each species [@ozawa2010population; @combescot2001bcs]. Another interesting aspect here is whether any instabilities occuring in the HF approximation [@stoner1933lxxx] are altered due to the presence of correlations as well as the crucial role of the latter [@mazurenko2017cold; @partridge2006pairing; @sanner2012correlations; @pekker2011competition] in the course of the evolution. Motivated also by the experimental capability to prepare few-fermion mixtures in one-dimension [@serwane2011deterministic; @zurn2012fermionization; @wenz2013few; @murmann2015antiferromagnetic], we study here the interaction quench dynamics of a spin-polarized FF mixture confined in a double-well. To simulate the correlated quantum dynamics of the FF mixture we employ the Multi-Layer Multi-Configurational Time-Dependent Hartree Method for Atomic Mixtures (ML-MCTDHX) [@ML-MCTDHX], which is a variational method capturing all the important particle correlations. We find that the ground state of particle imbalanced species exhibits a symmetry breaking, for strong interactions, on the single-particle density level within the HF approximation. This behavior is a manifestation of the Stoner instability [@stoner1933lxxx; @snoke2009solid] and renders the mixture immiscible. The presence of higher-order quantum correlations alters this instability and an intrawell fragmentation of the one-body density arises, i.e. the density profile breaks into several density branches (filaments) [^1], while the two species remain miscible. Performing an interspecies interaction quench from weak-to-strong coupling we find that within the HF approximation the $\sigma$-species (with $\sigma=A,B$ denoting each species) single-particle density filamentizes and subsequently the two species phase separate. In sharp contrast, in the presence of quantum correlations the filamentation of the one-body density becomes suppressed and the fermionic components show a miscible behavior on the one-body level. Remarkably enough, Mott-like one-body correlations [@sherson2010single; @larson2008mott; @katsimiga2017dark; @mistakidis2017correlation] between the filaments formed are revealed, indicating their tendency for localization. Most importantly, both the intra- and interspecies two-body correlation functions show that two fermions of the same or different species cannot populate the same filament but only distinct ones. The latter, which is arguably one of our main results, unveils that a phase separation process occurs only on the two-body level suggesting the formation of few-body anti-ferromagnetic like order [@murmann2015antiferromagnetic; @koutentakis2018probing]. Turning to particle balanced FF mixtures we find that the single-particle density of the ground state exhibits a miscible behavior at weak and strong interactions in both the HF and MB approaches. Moreover, an intrawell fragmentation occurs only within the MB approach. Quenching the interspecies interaction from weak-to-strong coupling we observe that in the HF approximation the FF mixture remains miscible throughout the evolution, while performing an overall breathing motion. Within the MB approach the two species besides undergoing a breathing mode while remaining miscible, further exhibit an intrawell fragmentation (filamentation) of their single-particle density. Also in this case Mott-like one-body correlations appear between the distinct filaments formed. Moreover, two fermions of the same or different species exhibit an anti-correlated behavior in a single filament, whilst they are strongly correlated when residing in distinct filaments indicating the tendency towards an anti-ferromagnetic state. Finally, we simulate single-shot absorption measurements and showcase that by averaging a sample of in-situ images we can adequately reproduce the MB fermionic quench dynamics on the single-particle density level. This work is structured as follows. Section \[sec:theory\] presents our setup and the basic observables of interest. The nonequilibrium dynamics induced by an interspecies interaction quench for particle imbalanced and balanced species within a double-well is analyzed in Secs. \[sec:particle\_imbalanced\] and \[sec:particle\_balanced\] respectively. We summarize our findings and provide an outlook in Section \[sec:conclusion\]. In Appendix \[single\_shots\_details\] we provide a brief discussion regarding our numerical implementation of the single-shot procedure. Finally, in Appendix \[sec:numerics\] we present further details of our numerical simulations and demonstrate the convergence of the results discussed in the main text. Theoretical Framework {#sec:theory} ===================== Setup {#sec:setup} ----- We consider a FF mixture consisting of $N_A$ and $N_B$ spin polarized fermions with equal masses $M_A=M_B\equiv M$ for the $A$ and $B$ species respectively. Such a mass balanced fermionic mixture can be experimentally realized by two different hyperfine states e.g. of $\prescript{40}{}{K}$ or $\prescript{6}{}{Li}$ [@wang2000ground; @dieckmann2002decay]. These internal states could refer, for instance, to the $\ket{F=9/2,m_F-9/2}$ and $\ket{F=9/2,m_F=-7/2}$ of $\prescript{40}{}{K}$ [@zwierlein2006fermionic]. The mixture is confined in an one-dimensional double-well external potential [@bloch2005ultracold] which is composed by a harmonic oscillator with frequency $\omega$ and a centered Gaussian with height $V_0$ and width $w$. The resulting MB Hamiltonian reads $$\begin{aligned} \mathcal{H}&=\sum\limits_{\sigma=A,B} \sum\limits_{i=1}^{N_\sigma}\left[ -\frac{\hbar^2}{2M}\left( \frac{d}{d x_i^\sigma}\right)^2 +\frac{1}{2}M\omega_\sigma^2(x_i^\sigma)^2 \right. \nonumber\\ &+ \left.\frac{V_0}{w \sqrt{2\pi}} e^{-\frac{(x_i^\sigma)^2 }{2w^2}}\right]\nonumber \\ &+ \sum\limits_{i=1}^{N_A} \sum \limits_{j=1}^{N_B}g_{AB}\delta(x_i^{A}-x_j^{B}).\label{Hamilt} \end{aligned}$$ We operate in the ultracold regime, hence $s$-wave scattering is the dominant interaction process. Consequently the interspecies interactions can be adequately modeled by contact interactions, which scale with the effective one-dimensional coupling strength $g_{AB}$ for the different fermionic species. Since $s$-wave scattering is forbidden for spin-polarized fermions [@pethick2002bose; @lewenstein2012ultracold], due to the antisymmetry of the fermionic wavefunction, fermions of the same species are considered to be non-interacting. Therefore, only interspecies interactions are relevant in the MB Hamiltonian. The effective interspecies one-dimensional coupling strength [@olshanii1998atomic] is given by ${g_{AB}} =\frac{{2{\hbar ^2}{a^s_{AB}}}}{{\mu a_ \bot ^2}}{\left( {1 - {\left\|{\zeta (1/2)} \right\|{a^s_{AB}}}/{{\sqrt 2 {a_ \bot }}}} \right)^{ - 1}}$, where $\zeta$ refers to the Riemann zeta function and $\mu=\frac{M}{2}$ is the corresponding reduced mass. ${a_\bot } = \sqrt{\hbar /{\mu{\omega _ \bot }}}$ is the transversal length scale with transversal confinement frequency ${{\omega _ \bot }}$ and ${a^s_{AB}}$ is the three-dimensional $s$-wave scattering length between the two distinct species. We note that $g_{AB}$ can be experimentally adjusted either by means of ${a^s_{AB}}$ with the aid of Feshbach resonances [@kohler2006production; @chin2010feshbach] or by manipulating ${{\omega _ \bot }}$ via confinement-induced resonances [@olshanii1998atomic; @kim2006suppression]. In the following our Hamiltonian is rescaled in units of $\hbar \omega_{\perp}$. Thus, the corresponding length, time, and interaction strength scales are expressed in terms of $\sqrt{\frac{\hbar}{M \omega_{\perp}}}$, $\omega_{\perp}^{-1}$ and $\sqrt{\frac{\hbar^3 \omega_{\perp}}{M}}$ respectively. Moreover, the amplitude of the Gaussian barrier $V_0$, its width $w$ and the frequency of the harmonic oscillator $\omega$ are given in units of $\sqrt{ \frac{ \hbar^3 \omega_{\perp} }{ M }}$, $\sqrt{\frac{\hbar}{M \omega_{\perp}}}$, and $\omega_{\perp}$. To limit the spatial extension of our system we impose hard-wall boundary conditions at $x_\pm=\pm40$. Throughout this work, our system is initially prepared in the MB ground state of the Hamiltonian (\[Hamilt\]) within the weak interspecies interaction regime, namely $g_{AB}=0.1$. The corresponding double-well potential is characterized by the harmonic oscillator frequency $\omega=0.1$, barrier height $V_0=2$ and width $w=1$. Thus, in a non-interacting single-particle picture four doublets are included below the maximum of the barrier. To induce the nonequilibrium dynamics of the FF mixture in the double-well we quench at $t=0$ the interspecies interaction strength towards the strongly correlated regime, e.g. $g_{AB}=4.0$, and let the system evolve in time. Quenching the interspecies repulsion towards the strongly interacting regime favors the occurrence of a breathing mode [@mistakidis2017correlation] and the appearance of strong intra- and interspecies correlations \[see Secs. \[sec:particle\_imbalanced\] and \[sec:particle\_balanced\]\] due to the quench imported interaction energy into the system [@fang2014quench]. Below, we first analyze the dynamics of a particle imbalanced mixture with $N_A=3$ ($N_A=5$) and $N_B=1$ fermions respectively, and subsequently examine the corresponding particle balanced case with $N_A=N_B=2$ and $N_A=N_B=5$. Many-Body Approach {#sec:wfn} ------------------ To solve the underlying MB Schr[ö]{}dinger equation that governs the quech-induced dynamics of the FF mixture we utilize ML-MCTDHX [@ML-MCTDHX]. It is based on an expansion of the MB wavefunction with respect to a time-dependent and variationally optimized MB basis. Such a treatment enables us to take into account both the inter- and intraspecies correlations inherent in the system. In order to include the inter- and intraspecies correlations, we first introduce $M$ distinct species functions, $\Psi^{\sigma}_k (\vec x^{\sigma};t)$. Here, $\vec x^{\sigma}=\left( x^{\sigma}_1, \dots, x^{\sigma}_{N_{\sigma}} \right)$ refer to the spatial $\sigma=A,B$ species coordinates of each component consisting of $N_{\sigma}$ fermions. Then the MB wavefunction, $\Psi_{MB}$, is expressed as a truncated Schmidt decomposition [@horodecki2009quantum] of rank $D$ $$\Psi_{MB}(\vec x^A,\vec x^B;t) = \sum_{k=1}^D \sqrt{ \lambda_k(t) }~ \Psi^A_k (\vec x^A;t) \Psi^B_k (\vec x^B;t). \label{Eq:WF}$$ In this expression $D\le \min(\dim(\mathcal{H}^A),\dim(\mathcal{H}^B))$ and $\mathcal{H}^{\sigma}$ is the Hilbert space of the $\sigma$-species (see also the discussion below). The Schmidt coefficients $\lambda_k(t)$ in decreasing order are denoted as the natural species populations of the $k$-th species function $\Psi^{\sigma}_k$ of the $\sigma$-species. They serve as a measure of the system’s entanglement or interspecies correlations. Specifically, the system is called entangled or interspecies correlated [@roncaglia2014bipartite] when at least two distinct $\lambda_k(t)$ are nonzero, since in this latter case the total MB state \[Eq. (\[Eq:WF\])\] cannot be expressed as a direct product of two states. To explicitly incorporate the interparticle correlations each of the species functions $\Psi^{\sigma}_k (\vec x^{\sigma};t)$ is expanded using the determinants of $m^{\sigma}$ distinct time-dependent fermionic single-particle functions (SPFs), $\varphi_1,\dots,\varphi_{m_{\sigma}}$. In particular $$\begin{split} &\Psi_k^{\sigma}(\vec x^{\sigma};t) = \sum_{\substack{l_1,\dots,l_{m_{\sigma}} \\ \sum l_i=N}} C_{k,(l_1, \dots,l_{m_{\sigma}})}(t)\\& \sum_{i=1}^{N_{\sigma}!} {\rm sign}(\mathcal{P}_i) \mathcal{P}_i \bigg[ \prod_{\substack{j\in\{1,\dots,m^{\sigma}\} \\ {\rm with}~ l_j=1}} \varphi_j(x_{K(j)};t)\bigg]. \label{Eq:SPFs} \end{split}$$ Here, $\mathcal{P}$ refers to the permutation operator which exchanges the particle positions $x_{\mu}$, $\mu=1,\dots,N_{\sigma}$ within the SPFs. Also $K(j)\equiv \sum_{\nu=1}^{j}l_{\nu}$ with $l_{\nu}$ denoting the occupation of the $\nu$th SPF and $j\in\{1,2,\dots,m^{\sigma}\}$. The symbol $\rm{sign}(\mathcal{P}_i)$ denotes the sign of the corresponding permutation and $C_{k,(l_1,\dots,l_{m_{\sigma}})}(t)$ are the time-dependent expansion coefficients of a certain determinant. The eigenfunctions of the one-body reduced density matrix of the $\sigma$-species $\rho_\sigma^{(1)}(x,x^\prime;t)=\langle\Psi_{MB}(t)\|\hat{\Psi}^{\sigma,\dagger}(x)\hat{\Psi}^\sigma(x^\prime)\|\Psi_{MB}(t)\rangle$ are termed natural orbitals $\phi^{\sigma}_i(x;t)$, where $\hat{\Psi}^{\sigma}(x)$ refers to the fermionic field operator of the $\sigma$-species. The eigenvalues of $\phi^{\sigma}_i(x;t)$ are the so-called natural populations $n^{\sigma}_i(t)$. If more than $N_\sigma$ natural populations, $n_i(t)$, possess a non-negligible occupation ($0<n_i(t)<1$ with $N_{\sigma}<i<m^{\sigma}$), the fermionic $\sigma$-species is termed intraspecies correlated, otherwise the MB state reduces to the HF ansatz [@pethick2002bose; @giorgini2008theory; @pitaevskii2016bose]. Indeed ML-MCTDHX enables us to operate within different approximation orders [@ML-MCTDHX], and we e.g. retrieve the HF ansatz [@pethick2002bose] in the limit of $D=1$ and $m^{\sigma}=N_{\sigma}$ $$\begin{split} &\Psi_{HF}(\vec x^{A},\vec x^{B};t) =\\ &\prod_{\sigma=A,B}\sum_{i=1}^{N_{\sigma}!} {\rm sign}(\mathcal{P}_i) \mathcal{P}_i \left[\varphi_1(x_1^{\sigma};t) \cdots \varphi_{N_{\sigma}}(x_{N_{\sigma}}^{\sigma};t) \right]. \label{Eq:HF} \end{split}$$ Furthermore, employing the Dirac-Frenkel variational principle [@frenkel1932wave; @dirac1930note] for the MB ansatz \[see Eqs. (\[Eq:WF\]), (\[Eq:SPFs\])\] we obtain the ML-MCTDHX equations of motion [@ML-MCTDHX] for the fermionic mixture. These equations correspond to $D^2$ linear differential equations of motion for the coefficients $\lambda_i(t)$ coupled to a set of $D$\[${m_A}\choose{N_A}$+$ {m_B}\choose{N_B}$\] non-linear integro-differential equations for the species functions and $m^A+m^B$ integro-differential equations for the SPFs. Correlation Functions {#def_cor_functions} --------------------- To unveil the degree of intraspecies correlations at the one-body level during the quench dynamics we employ the normalized spatial first order correlation function [@naraschewski1999spatial; @sakmann2008reduced; @mistakidis2017correlation] $$\begin{aligned} g^{(1)}_\sigma(x,x^\prime;t)=\frac{\rho_\sigma^{(1)}(x,x^\prime;t)}{\sqrt{\rho_\sigma^{(1)}(x;t)\rho_\sigma^{(1)}(x^\prime;t)}}.\label{one_body_cor}\end{aligned}$$ Here, $\rho_\sigma^{(1)}(x,x^\prime;t)=\langle\Psi(t)\|\hat{\Psi}^{\sigma,\dagger}(x)\hat{\Psi}^\sigma(x^\prime)\|\Psi(t)\rangle$ refers to the one-body reduced density matrix of the $\sigma$ species and $\rho_\sigma^{(1)}(x;t)\equiv\rho_\sigma^{(1)}(x,x^\prime=x;t)$ is the one-body density. $\hat{\Psi}^{\sigma,\dagger}(x)$ \[$\hat{\Psi}^{\sigma}(x)$\] is the fermionic field operator that creates \[annihilates\] a $\sigma$ species fermion at position $x$. $\|g^{(1)}_{\sigma}(x,x';t)\|$ is bounded within the interval $[0,1]$ and measures the proximity of the MB state to a product state for a fixed set of coordinates $x$, $x'$. Two different spatial regions $R$, $R'$, with $R \cap R' = \varnothing$, possessing $\|g^{(1)}_{\sigma}(x,x';t)\|= 0$ with $x\in R$ and $x'\in R'$ are referred to as perfectly incoherent, whilst for $\|g^{(1)}_{\sigma}(x,x';t)\|= 1$, $x\in R$, $x'\in R'$ the regions are said to be fully coherent. When at least two distinct spatial regions are partially incoherent, i.e. $\|g^{(1)}_{\sigma}(x,x';t)\|<1$ this signifies the emergence of one-body intraspecies correlations, while their absence is designated by $\|g^{(1)}_{\sigma}(x,x';t)\|=1$ for every $x$, $x'$. Most importantly, the situation where a certain spatial region $R$ is fully coherent, i.e. $\|g^{(1),\sigma}(x,x';t)\|^2 \approx 1$ $x,x'\in R$, and perfect incoherence occurs between different spatial regions $R$, $R'$, i.e. $\|g^{(1),\sigma}(x,x';t)\|^2\approx 0$, $x\in R$, $x'\in R'$ with $R \cap R' = \varnothing$, indicates the appearance of Mott-like correlations [@sherson2010single; @larson2008mott; @katsimiga2017dark; @mistakidis2017correlation]. To estimate the degree of second order intra- and interspecies correlations in the course of the dynamics, we inspect the normalized two-body correlation function [@sakmann2008reduced; @mistakidis2017correlation] $$\begin{aligned} g^{(2)}_{\sigma\sigma\prime}(x,x^\prime;t)=\frac{\rho_{\sigma\sigma\prime}^{(2)}(x,x^\prime;t)}{\rho_\sigma^{(1)}(x;t)\rho_{\sigma\prime}^{(1)}(x^\prime;t)}. \label{two_body_cor}\end{aligned}$$ In Eq. (\[two\_body\_cor\]), $\rho^{(2)}(x,x';t)=\langle\Psi(t)\|\hat{\Psi}^{\sigma,\dagger}(x')\hat{\Psi}^{\dagger,\sigma^\prime}(x)\hat{\Psi}^{\sigma^\prime} (x)\\\hat{\Psi}^\sigma(x')\|\Psi(t)\rangle$ denotes the diagonal two-body reduced density matrix which provides the probability of measuring two particles of species $\sigma$ and $\sigma^\prime$ located at $x$ and $x^\prime$ respectively at time $t$. Referring to the same (different) species, i.e. $\sigma=\sigma'$ ($\sigma \not=\sigma'$), $\|g^{(2)}_{\sigma \sigma'}(x,x^\prime;t)\|$ accounts for the intraspecies (interspecies) two-body correlations. We remark that if $g^{(2)}_{\sigma\sigma\prime}(x,x^\prime;t)=1$ holds, the state is termed fully second order coherent, while in case that $g^{(2)}_{\sigma\sigma\prime}(x,x^\prime;t)>1$ \[$g^{(2)}_{\sigma\sigma\prime}(x,x^\prime;t)<1$\] it is termed correlated \[anti-correlated\]. $g^{(2)}_{\sigma\sigma\prime}(x,x^\prime;t)$ is experimentally accessible by in-situ density-density fluctuation measurements [@tavares2016chaotic; @tsatsos2017granulation; @endres2013single]. Interaction Quench Dynamics of a Particle Imbalanced Mixture {#sec:particle_imbalanced} ============================================================ ![One-body density $\rho^{(1)}_\sigma(x)$ of the ground state of the $\sigma$-species of a FF mixture for different interspecies repulsions $g_{AB}$ (see legend) within (a) the MB approach and (b) the HF approximation. The mixture consists of $N_A=3$ and $N_B=1$ fermions and it is trapped in a double-well potential. The rectangles in (a) indicate the intrawell fragmentation (filamentation) of $\rho^{(1)}_\sigma(x)$ occuring for strong interspecies interactions. The inset in (b) shows the corresponding energetically degenerate configuration of $\rho^{(1)}_\sigma(x)$ with the $\rho^{(1)}_\sigma(x)$ of the main panel for $g_{AB}=4.0$ within the HF approximation.[]{data-label="abb:gs3-1"}](Fig1.eps){width="48.00000%"} Initial State {#in_par_imb} ------------- We consider an interspecies repulsively interacting ($g_{AB}$) mass balanced ($M_A=M_B$) FF mixture with spin-polarized and particle imbalanced components consisting of $N_A=3$ and $N_B=1$ fermions. The mixture is confined within a double-well and it is initialized in its corresponding interspecies interacting ground state as described by the Hamiltonian of Eq. (\[Hamilt\]), using either imaginary time propagation or improved relaxation [@ML-MCTDHX] within ML-MCTDHX. The double-well is characterized by frequency $\omega=0.1$, barrier height $V_0=2$ and width $w=1$. To inspect the ground state of the FF mixture we invoke the $\sigma$-species single-particle density $\rho_\sigma^{(1)}(x;t)$ \[see also Eq. (\[one\_body\_cor\])\]. Within the weakly interacting regime, $g_{AB}=0.1$, we observe that each $\rho_\sigma^{(1)}(x)$ shows an equal population in the two wells of the double-well and it is distributed in a symmetric manner both in the HF approximation as well as on the MB level, see Figs. \[abb:gs3-1\] (a) and (b). Also, $\rho_A^{(1)}(x;t)$ and $\rho_B^{(1)}(x;t)$ feature a miscible behavior in both approaches. Note that due to the particle imbalance the $A$-species which contains the higher particle number exhibits a broader single-particle density distribution within each well when compared to the $B$-species. Turning to the strong interaction regime, $g_{AB}=4.0$, an intrawell fragmentation of $\rho_\sigma^{(1)}(x)$ occurs on the MB level, see Fig. \[abb:gs3-1\] (a), while $\rho_A^{(1)}(x;t)$ and $\rho_B^{(1)}(x;t)$ show again a miscible behavior. Intrawell fragmentation refers to the filamentation tendency of the one-body density, i.e. to the appearance of several local maxima occurring in $\rho_\sigma^{(1)}(x)$ within each well, see the dashed rectangle in Fig. \[abb:gs3-1\] (a) where two sub maxima are present. Interestingly enough, within the HF approximation the mixture becomes immiscible with $\rho_A^{(1)}(x)$ and $\rho_B^{(1)}(x)$ being phase separated as it can be observed by their asymetric distribution with respect to $x=0$ illustrated in Fig. \[abb:gs3-1\] (b). This latter behavior can be thought of as the few-body analog of the Stoner’s instability [@jacquod2000supression; @jacquod2001ground] which is a well-known phenomenon in solid state physics being responsible for magnetization effects emerging in itinerant systems. Indeed, within the HF approximation in the strongly interacting regime the energy of a miscible state is larger when compared to the energy of a phase separated (immiscible) one due to the strong impact of the interaction energy [@stoner1933lxxx; @snoke2009solid]. Thus, the particle number assymetry favors a phase separated state with $\rho_B^{(1)}(x)$ being localized in one of the wells and $\rho_A^{(1)}(x)$ distributing around it, see Fig. \[abb:gs3-1\] (b). Since the double-well is symmetric, the same occupation structure of $\rho_A^{(1)}(x)$ and $\rho_B^{(1)}(x)$ with interchanged wexlls possesses an equal energy, i.e. the two configurations are energetically degenerate, see the inset of Fig. \[abb:gs3-1\] (b). Recall that this phenomenon occurs, in the one-dimensional spin-independent case considered here, only within the HF approximation and not at the MB level in accordance to the Lieb-Mattis theorem [@lieb1962lieb]. In the latter approach the Stoner instability ceases to exist due to the involvement of higher-order quantum superpositions [@koutentakis2018probing; @volosniev2014strongly]. Next, we examine the quantum dynamics of the above-mentioned weakly interacting, $g_{AB}=0.1$, FF mixture by quenching the interspecies repulsion at $t=0$ towards the strongly correlated regime of interactions, $g_{AB}=4.0$. ![Evolution of the $\sigma$-species one-body density $\rho^{(1)}_\sigma(x;t)$ of a FF mixture within the (a), (b) HF approximation and (c)-(f) MB approach following an interaction quench from $g_{AB}=0.1$ to $g_{AB}=4.0$. The FF mixture consists of (a)-(d) $N_A=3$ and $N_B=1$ particles and (e), (f) $N_A=5$ and $N_B=1$ fermions. The left and right columns correspond to the densities of the $A$ and the $B$ species respectively. The rectangles indicate the number of filaments formed of the corresponding $\rho^{(1)}_\sigma(x;t)$ within the left well. (g), (h) Overlap integral $\Lambda(t)$ between the species of the FF mixture during the evolution within different approximations and varying postquench interactions (see legend) for (g) $N_A=3$, $N_B=1$ and (h) $N_A=5$, $N_B=1$. []{data-label="abb:den3-1"}](Fig2.eps){width="49.00000%"} Single-Particle Density Evolution {#one_body_par_imb} --------------------------------- To visualize the nonequilibrium dynamics of the particle imbalanced FF mixture on the one-body level we employ the single-particle density evolution $\rho_\sigma^{(1)}(x;t)$ for each of the species after the quench. Focusing on the HF approximation, see Figs. \[abb:den3-1\] (a) and (b), we observe that an overall breathing mode [@abraham2012quantum; @abraham2014quantum; @abraham2014quantum] of both fermionic clouds takes place manifested as a contraction and expansion dynamics of $\rho_\sigma^{(1)}(x;t)$. The frequency of this breathing mode is $\omega_{br}=0.2=2\omega$ which is in accordance with the corresponding theoretical prediction [@abraham2014quantum]. Most importantly, a phase separation process between the two species occurs and each $\rho_\sigma^{(1)}(x;t)$ exhibits an intrawell fragmentation. This phase separation is a consequence of the Stoner instability that exists in this strongly interacting regime even in the ground state of the system \[see also our discussion in Sec. \[in\_par\_imb\].\] Regarding the intrawell fragmentation we observe that $\rho_A^{(1)}(x;t)$ forms two filaments in each well, while $\rho_B^{(1)}(x;t)$ exhibits one filament in each well and one (of lower amplitude) located at the position of the barrier of the double-well \[see also the dashed rectangles in Figs. \[abb:den3-1\] (a) and (b)\]. In sharp contrast to the above, utilizing the correlated approach the single-particle density evolution shows a completely different behavior, see Figs. \[abb:den3-1\] (c) and (d). The two components remain miscible throughout the evolution in accordance to the ground state properties discussed in Sec. \[in\_par\_imb\]. Moreover, an intrawell fragmentation emerges with the two filaments formed in $\rho_\sigma^{(1)}(x;t)$ within each well being more pronounced for the $B$-species, while the filamentary structure of the $A$-species is suppressed and hardly discernible \[see also the rectangles in Figs. \[abb:den3-1\] (c) and (d)\]. Finally, both clouds undergo a breathing motion with approximately the same frequency as the one observed in the HF approximation. To infer about the effect of the majority species particle number on the nonequilibrium dynamics, we next consider a mass balanced FF mixture with $N_A=5$ and $N_B=1$. The corresponding single-particle density evolution following an interspecies interaction quench from $g_{AB}=0.1$ to $g_{AB}=4.0$ is shown in Figs. \[abb:den3-1\] (e) and (f) within the MB approach. As it can be deduced, a larger particle number of the majority component leads to an increased number of filaments within each well for each of the species as compared to the case of a smaller particle number \[compare Figs. \[abb:den3-1\] (c), (d) and (e), (f)\]. It is also worth mentioning that the filament formation of both species is washed out for higher particle numbers \[see also the rectangles in Figs. \[abb:den3-1\] (e) and (f)\]. However, the particle number does not significantly alter the breathing frequency of each species and their miscible character. To expose the degree of spatial phase separation, namely the degree of miscibility or immiscibility of the mixture, occuring on the one-body level during the quench dynamics, we employ the overlap integral function $\Lambda(t)$ [@mistakidis2017correlation; @bandyopadhyay2017dynamics; @jain2011quantum] between the two species $$\begin{aligned} \Lambda (t)=\frac{\left[\int d x \, \rho_A^{(1)}(x;t)\rho_B^{(1)}(x;t)\right]^2}{\left[\int d x\, (\rho_A^{(1)}(x;t))^2\right]\left[\int d x\, (\rho_B^{(1)}(x;t))^2\right]}.\end{aligned}$$ This quantity being normalized to unity takes values between $\Lambda=0$ and $\Lambda=1$ corresponding to zero and complete spatial overlap of the two species on the single-particle level. Figures \[abb:den3-1\] (g) and (h) present $\Lambda(t)$ for the setups $N_A=3$, $N_B=1$ and $N_A=5$, $N_B=1$ respectively for different interaction quench amplitudes. Regarding the evolution in the HF approximation, $\Lambda(t)$ drops close to zero at short time scales ($t>6$) for both systems. After this initial drop the overlap remains almost constant exhibiting small amplitude oscillations which reflect the breathing motion of each cloud. Notice that the maxima of these small amplitude oscillations appear at time intervals of the contraction of the cloud, see e.g. Figs. \[abb:den3-1\] (a) and (g) at $t\approx40$. In sharp contrast to the above behavior, $\Lambda(t)$ shows small fluctuations around 0.9 within the MB approach during the entire evolution. The aforementioned evolution of $\Lambda(t)$ reflects the miscible character of the dynamics on the single-particle level. The same overall phenomenology in terms of $\Lambda(t)$ holds equally, in both approaches, for other postquench interaction strengths, see e.g. Figs. \[abb:den3-1\] (g) and (h) for $g_{AB}=4.0$. We further remark that for postquench interaction strengths $g_{AB}>2.0$ the overlap function features a similar dynamics, while for quenches to $g_{AB}<2.0$ the mixture remains miscible in both the HF and the MB approach (results not shown here for brevity). ![One-body correlation function $g_\sigma^{(1)}(x,x^\prime;t)$ shown for different time instants (see legends) during the interaction quench dynamics of a FF mixture for ($a_1$)-($a_4$) the $A$-species and ($b_1$)-($b_4$) the $B$-species. ($c_1$)-($c_4$) Snapshots of the two-body intraspecies correlation function $g_{AA}^{(2)}(x,x^\prime;t)$ and ($d_1$)-($d_4$) the interspecies two-body correlation function $g_{AB}^{(2)}(x,x^\prime;t)$. In all cases, the FF mixture consists of $N_A=3$, $N_B=1$ particles and it is initialized in the weakly interacting ground state, $g_{AB}=0.1$, of the double-well. To induce the dynamics we perform an interaction quench from $g_{AB}=0.1$ to $g_{AB}=4.0$. []{data-label="abb:coh3-1"}](Fig3.eps){width="49.00000%"} Correlation Dynamics {#cor_par_imb} -------------------- To unveil the underlying correlation mechanisms [@sanner2012correlations; @pekker2011competition] that lead to the intrawell fragmentation during the MB quench dynamics, we investigate the one-body $g_{\sigma}^{(1)}(x,x^\prime,t)$ \[Eq. (\[one\_body\_cor\])\] and the two-body $g_{\sigma\sigma'}^{(2)}(x,x',t)$ \[Eq. (\[two\_body\_cor\])\] intra- and interspecies correlation functions during evolution, see Fig. \[abb:coh3-1\]. As it is expected the intraspecies two-body correlation function for the $B$-species is zero, since this species contains only a single particle. Below we examine $g_{\sigma}^{(1)}(x,x^\prime,t)$ and $g_{\sigma\sigma'}^{(2)}(x,x',t)$ following an interaction quench of the FF mixture from $g_{AB}=0.1$ to $g_{AB}=4.0$. Figures \[abb:coh3-1\] (a~1~)-(a~4~) and (b~1~)-(b~4~) show $g_{A}^{(1)}(x,x^\prime,t)$ and $g_{B}^{(1)}(x,x^\prime;t)$ respectively for selected time instants of the MB evolution. Overall, we observe that throughout the evolution the off-diagonal elements of $g_{\sigma}^{(1)}(x,x^\prime,t)$ are supressed. Indeed on the one-body level, each filament of both species is fully coherent with itself \[see e.g. $g_A^{(1)}(x=-2.5,x^\prime=-2.5;t=24)\approx1$ in Fig. \[abb:coh3-1\] (a~1~)\] and mainly incoherent with any of the other filaments \[see e.g. $g_A^{(1)}(x=-2.5,x^\prime=2.5;t=24)\approx0$ in Fig. \[abb:coh3-1\] (a~1~)\]. We note that this behavior of $g_{\sigma}^{(1)}(x,x^\prime,t)$ is more pronounced for the $A$-species, while in the $B$-species two distinct filaments appear to be partially incoherent \[e.g. $g_B^{(1)}(x=-2.5,x^\prime=2.5;t=76)\approx0.3$ in Fig. \[abb:coh3-1\] (b~3~)\], as shown in Figs. \[abb:coh3-1\] (b~1~)-(b~4~). This structure of $g_{\sigma}^{(1)}(x,x^\prime,t)$ indicates the occurrence of Mott-like correlations [@sherson2010single; @larson2008mott; @katsimiga2017dark] in the system \[see also the corresponding discussion in Sec. \[def\_cor\_functions\]\] and suggests the tendency of the observed filaments to be localized structures. Moreover, we can infer that each $\sigma$-species fermion, and especially the $A$-species ones, is more likely to be localized in one filament and do not reside in two or more filaments. We next study the two-body intraspecies correlation function $g_{AA}^{(2)}(x,x';t)$, see Figs. \[abb:coh3-1\] (c~1~)-(c~4~). A strongly anti-correlated behavior within each filament, see the depleted diagonal behavior, occurs for every time instant \[e.g. $g_{AA}^{(2)}(x=-2.5,x^\prime=-2.5;t=24)\approx0$ in Fig. \[abb:coh3-1\] (c~1~)\], while two different filaments appear to be correlated \[e.g. $g_{AA}^{(2)}(x=-2.5,x^\prime=2.5;t=24)\approx1.3$ in Fig. \[abb:coh3-1\] (c~1~)\]. As a consequence, two particles of the $A$-species cannot reside in the same filament but they are more likely to be found in any pair of distinct filaments. The corresponding interspecies correlation function $g_{AB}^{(2)}(x,x')$, shown in Figs. \[abb:coh3-1\] (d~1~)-(d~4~), displays similar characteristics to $g_{AA}^{(2)}(x,x';t)$. Namely, a correlation hole exists \[see e.g. $g_{AB}^{(2)}(x=-2.5,x^\prime=-2.5;t=24)\approx0$ in Fig. \[abb:coh3-1\] (d~1~)\] which excludes the possibility of an $A$ and a $B$ particle to be in the same filament. However, the off-diagonal elements of $g_{AB}^{(2)}(x,x';t)$ exhibit a correlated behavior \[see e.g. $g_{AB}^{(2)}(x=-2.5,x^\prime=2.5;t=24)\approx2.2$ in Fig. \[abb:coh3-1\] (d~1~)\] providing the possibility for an $A$ and a $B$-species fermion to be located at different filaments. We remark, that similar correlation structures have also been observed for the $N_A=5$, $N_B=1$ case (not shown here for brevity). An important conclusion that can be extracted from the above analysis is that on the MB level phase separation between the species can be inferred only on the two-body level and not by simply observing the corresponding single-particle densities. Recall that the single-particle density evolution does not exhibit any phase separation within the MB approach, see Figs. \[abb:den3-1\] (c), (d), which is in sharp contrast to the HF approximation where the fermionic components are evidently immiscible, see Figs. \[abb:den3-1\] (a), (b). Combining also the results of $g_{\sigma}^{(1)}(x,x^\prime,t)$ and $g_{\sigma\sigma'}^{(1)}(x,x^\prime,t)$, it becomes apparent that all $N_A+N_B$ fermions reside in distinct filaments. Therefore regarding the spatially resolved distribution of the system, a superposition state consisting of all permutations of possible fermionic configurations concerning the four filaments formed, i.e $(B-A-A-A)$, $(A-B-A-A)$, $(A-A-B-A)$ and $(A-A-A-B)$, is permitted. This latter behavior suggests the tendency towards an anti-ferromagnetic like state of the few-body system [@koutentakis2018probing; @murmann2015antiferromagnetic]. ![Single-shot images of each species, at distinct time instants of the interaction quench dynamics (see legends), obtained by averaging over (a), (e) $N_{shots}=1$, (b), (f) $N_{shots}=50$, and (c), (g) $N_{shots}=500$. (d), (h) The corresponding $\rho^{(1)}_{\sigma}(t)$ is evaluated within the MB approach. In all cases the system consists of $N_A=3$ and $N_B=1$ fermions confined in a double-well potential and the dynamics is induced by quenching the interspecies interaction from $g_{AB}=0.1$ to $g_{AB}=4.0$.[]{data-label="abb:shots3-1"}](Fig4.eps){width="49.00000%"} Single-Shot Images {#shots_par_imb} ------------------ In order to offer further possible experimental evidences of the correlated quantum quench dynamics of the FF mixture we simulate in-situ single-shot absorption measurements [@sakmann2016single; @mistakidis2017correlation; @lode2017fragmented]. These measurements essentially probe the spatial configuration of the atoms and they are based on the MB probability distribution which is available within ML-MCTDHX [@ML-MCTDHX]. To simulate the corresponding experimental procedure we perform a convolution of the spatial particle configuration with a point spread function being determined by the corresponding experimental resolution. For more details regarding the numerical implementation of this procedure in binary systems we refer the interested reader to Appendix \[single\_shots\_details\], while more elaborated discussions are provided in Refs. [@mistakidis2017correlation; @erdmann2018correlated]. The point spread function used here possesses a Gaussian shape with width $w_{PSF}=1 \ll l\approx 3.2$, where $l=\sqrt{1/\omega}$ denotes the corresponding harmonic oscillator length. We note that in few-body experiments [@zurn2012fermionization; @wenz2013few] fluorescence imaging is another promising technique to probe the state of the system since it eliminates unavoidable noise sources that might destroy the experimental signal [@serwane2011deterministic]. However, the simulation of this experimental technique lies beyond our current scope. Here, by simulating single-shot measurements we aim to show how in-situ imaging can be used to adequately monitor the nonequilibrium quantum dynamics of the aforementioned few-body particle imbalanced FF mixture. Utilizing the MB wavefunction of the system, obtained within ML-MCTDHX, we simulate in-situ single-shot images at each time instant $t$ of the MB evolution. Consecutively imaging first the $A$ and then the $B$ species at time $t\equiv t_{im}$, these images are designated by $\mathcal{A}^A(\tilde{x};t_{im})$ and $\mathcal{A}^B(\tilde{x'}\|\mathcal{A}^A(\tilde{x});t_{im})$ for the $A$ and $B$ species respectively. In the following, we focus on the dynamics of a FF mixture with $N_A=3$ and $N_B=1$ within the double-well upon quenching the interspecies interaction strength from $g_{AB}=0.1$ to $g_{AB}=4.0$. Figures \[abb:shots3-1\] (a), (e) show the first simulated in-situ single-shot images for each species, $\mathcal{A}^A(\tilde{x};t)$ and $\mathcal{A}^B(\tilde{x'}\|\mathcal{A}^A(\tilde{x});t)$, at two distinct time instants during evolution, namely at $t_1=8$ and $t_2=25$. As it can be seen the images for both species, and especially the $\mathcal{A}^A(\tilde{x};t)$, exhibit a filamentized structure resembling this way the overall tendency observed in the one-body density evolution, see also Figs. \[abb:den3-1\] (c), (d). Moreover, let us comment that the spatial position of these images is in accordance with our previous discussion, regarding the spatial distribution of the particles of each species, based on the correlation functions \[Fig. \[abb:coh3-1\]\]. For instance, $\mathcal{A}^B(\tilde{x'}\|\mathcal{A}^A(\tilde{x});t_2)$ shows a population of a right well filament \[Fig. \[abb:shots3-1\] (e)\] which does not contradict the analysis obtained from the correlation function that a possible particle configuration is the $(A-A-A-B)$. Furthermore, we should emphasize that a direct correspondence between the one-body density and one single-shot image is not possible due to the small particle number of the considered FF mixture, $N_A=3$ and $N_B=1$, and the presence of multiple orbitals in the system. In particular, the MB state is a superposition of multiple orbitals \[see Eqs. (\[Eq:WF\]) and (\[Eq:SPFs\])\] and thus imaging an atom alters the MB state of the remaining atoms and consequently the relevant one-body density. For a more elaborated discussion on this topic see [@mistakidis2017correlation; @katsimiga2017many; @katsimiga2018many]. To obtain the one-body density of the system we average over several single-shot images for each of the species, namely $\bar{\mathcal{A}}^A(\tilde{x};t)=1/N_{shots}\sum_{k=1}^{N_{shots}} \mathcal{A}_k^A(\tilde{x};t)$ and $\bar{\mathcal{A}}^B(\tilde{x}^{'}\|\mathcal{A}^A(\tilde{x});t)=1/N_{shots}\sum_{k=1}^{N_{shots}} \mathcal{A}_k^B(\tilde{x}^{'}\|\mathcal{A}^A(\tilde{x});t)$, see also Eq. (\[averaging\]) in Appendix \[single\_shots\_details\]. In particular, Figs. \[abb:shots3-1\] (b)-(c) and (f)-(g) present $\bar{\mathcal{A}}^A(\tilde{x};t)$ and $\bar{\mathcal{A}}^B(\tilde{x}^{'}\|\mathcal{A}^A(\tilde{x});t)$ at time instants $t=t_1$ and $t=t_2$ for an increasing number of single-shots $N_{shots}$. It becomes evident that upon increasing $N_{shots}$ the averaged images, $\bar{\mathcal{A}}^A(\tilde{x};t)$ and $\bar{\mathcal{A}}^B(\tilde{x}^{'}\|\mathcal{A}^A(\tilde{x});t)$, display progressively the actual profile of the one-body density $\rho^{(1)}_{A}(x)$ and $\rho^{(1)}_{B}(x)$ obtained within ML-MCTDHX \[Figs. \[abb:shots3-1\] (d) and (h)\]. Interaction Quench Dynamics of a Particle Balanced Mixture {#sec:particle_balanced} ========================================================== ![One-body densities $\rho^{(1)}_\sigma(x)$ of the $\sigma$-species ground state of a FF mixture for different interspecies interaction strengths $g_{AB}$ (see legend) within (a) the MB approach and (b) the HF approximation. The mixture consists of $N_A=N_B=2$ fermions and is trapped in a double-well potential. The rectangle in (a) indicates the intrawell fragmentation (filamentation) of $\rho^{(1)}_\sigma(x)$ for strong interactions.[]{data-label="abb:gs2-2"}](Fig5.eps){width="48.00000%"} Ground state {#in_par_bal} ------------ To further elaborate on the interaction quench dynamics of FF mixtures trapped in a double-well potential we next examine particle balanced systems. In particular, we study a FF mixture with $N_A=N_B=2$ fermions and follow the same quench scenario as in the above Section \[sec:particle\_imbalanced\]. To this end, we first obtain the ground state of the system described by the Hamiltonian of Eq. (\[Hamilt\]) with interspecies interaction $g_{AB}=0.1$. The dynamics is subsequently induced by performing an interaction quench to the strongly interacting regime $g_{AB}=4.0$. The double-well possesses a frequency $\omega=0.1$, barrier height $V_0=2$ and width $w=1$. The corresponding single-particle density $\rho_\sigma^{(1)}(x;t)$ of the $\sigma$-species ground state \[see also Eq. (\[one\_body\_cor\])\] of the FF mixture is shown in Figs. \[abb:gs2-2\] (a) and (b) for different interspecies repulsions, $g_{AB}$, for both the MB and the HF approach. Note that since the FF mixture is particle balanced both equal mass species exhibit exactly the same one-body density, i.e. $\rho^{(1)}_A(x;t)=\rho^{(1)}_B(x;t)$, for both approaches. For weak interactions, $\rho_\sigma^{(1)}(x;t)$ populate the two wells in a symmetric manner (with respect to reflections at $x=0$), while $\rho_A^{(1)}(x;t)$ and $\rho_B^{(1)}(x;t)$ are miscible in both approaches, see Figs. \[abb:gs2-2\] (a) and (b). Turning to the strong interspecies interaction regime we observe that a broadening of $\rho_\sigma^{(1)}(x;t)$ occurs in the HF approximation as a result of the enhanced repulsion. In contrast within the MB approach $\rho_\sigma^{(1)}(x;t)$ besides being broadened shows an intrawell fragmentation, indicated by the dashed rectangle in Fig. \[abb:gs2-2\] (a). Let us also mention at this point that a symmetry breaking of $\rho_\sigma^{(1)}(x;t)$ by means of the Stoner instability does not take place in the current setup, since both species, besides being mass balanced, contain the same number of particles $N_A=N_B$ and therefore exhibit exactly the same behavior. As a consequence phase separation between the species is not favored on the one-body level. To induce the nonequilibrium dynamics, in the following, we quench the interspecies repulsion from $g_{AB}=0.1$ to $g_{AB}=4.0$. ![Time-evolution of the one-body density $\rho^{(1)}_\sigma(x;t)$ for the $\sigma$-species for the $A$- (left column) and the $B$-species (right column) of the FF mixture within the (a), (b) HF approximation and (c)-(f) MB approach following an interaction quench from $g_{AB}=0.1$ to $g_{AB}=4.0$. The FF mixture consists of (a)-(d) $N_A=N_B=2$ atoms and (e), (f) $N_A=N_B=5$ fermions. At $t=0$ it is prepared in the ground state of the double-well for $g_{AB}=0.1$. The rectangles in (a), (b) indicate a contraction event of the fermionic cloud and the resulting interference patterns, while in (c)-(f) they mark the number of filaments formed in the left well during the evolution. []{data-label="abb:den2-2"}](Fig6.eps){width="49.00000%"} Evolution on the Single-Particle level {#den_par_bal} -------------------------------------- The spatially resolved quench dynamics of the FF mixture can be investigated via the $\sigma$-species one-body density $\rho^{(1)}_\sigma(x;t)$, see Figs. \[abb:den2-2\] (a)-(d). Within the HF approximation the quenched one-body density evolution \[Figs. \[abb:den2-2\] (a) and (b)\] exhibits an overall breathing motion comprising both wells. This breathing motion of the fermionic cloud is, of course, characterized by an expansion and contraction of the symmetric density branches located in each well. Notice that during the contraction process these density branches collide on top of the barrier, i.e. at $x=0$, giving rise to several interference patterns \[see the dashed rectangles in Figs. \[abb:den2-2\] (a) and (b)\]. These interference patterns become even more pronounced for stronger interactions (not shown here). Inspecting $\rho^{(1)}_\sigma(x;t)$ within the MB approach, Figs. \[abb:den2-2\] (c) and (d), we observe that both species undergo a breathing mode comprising the double-well but most importantly an intrawell fragmentation of the fermionic cloud takes place within each well. In particular, for the $N_A=N_B=2$ case two filaments appear in each well, see here the dashed rectangles in Figs. \[abb:den2-2\] (c), (d). It is worth mentioning at this point that the existence of these filaments is a consequence of beyond HF correlations that built in the system [^2] To conclude upon the dependence of the above-described MB dynamics on the number of fermions in particle balanced FF mixtures, Figs. \[abb:den2-2\] (e) and (f) present $\rho^{(1)}_\sigma(x;t)$ for the same quench amplitude as before (i.e. from $g_{AB}=0.1$ to $g_{AB}=4.0$) but for a system containing $N_A=N_B=5$ fermions. $\rho^{(1)}_\sigma(x;t)$ possesses a broader distribution when compared to the $N_A=N_B=2$ case and performs an overall breathing motion with the same frequency as in the case $N_A=N_B=2$. Strikingly enough, the emergent intrawell fragmentation results in five distinct filaments \[see the rectangles in Figs. \[abb:den2-2\] (e), (f)\] of the $\sigma$-species fermionic cloud within each well. Therefore, we can infer that the number of filaments formed $N_f$ is proportional to the particle number $N_f=N_\sigma$. We note that we have checked this conclusion also for other particle numbers, e.g. $N_{\sigma}=3,6$ (results not shown here). Correlation Properties {#cor_par_bal} ---------------------- In order to expose the role of correlations in the above-discussed interaction quench dynamics of particle balanced FF mixtures we resort to the corresponding $g_{\sigma}^{(1)}(x,x^\prime;t)$ \[Eq. (\[one\_body\_cor\])\] and the $g_{\sigma\sigma'}^{(2)}(x,x';t)$ \[Eq. (\[two\_body\_cor\])\] correlation functions, see Fig. \[abb:coh2-2\]. Since the considered FF mixture is particle balanced it holds that $g_A^{(1)}(x,x^\prime;t)= g_B^{(1)}(x,x^\prime;t)$ and $g_{AA}^{(2)}(x,x^\prime;t)=g_{BB}^{(2)}(x,x^\prime;t)$. To this end, below we discuss only $g_A^{(1)}(x,x^\prime;t)$, $g_{AA}^{(2)}(x,x^\prime);t$ and $g_{AB}^{(2)}(x,x^\prime;t)$ following an interaction quench of the $N_A=N_B=2$ FF mixture from $g_{AB}=0.1$ to $g_{AB}=4.0$. Inspecting $g_A^{(1)}(x,x^\prime;t)$ we deduce that each filament is fully coherent with itself throughout the evolution, since $g_A^{(1)}(x,x^\prime\approx x;t)\approx1$ (with $x$ varying on the spatial scale of each filament) as shown in Figs. \[abb:coh2-2\] (a~1~-(a~4~). Regarding the coherence of two distinct filaments we discern between the cases of expansion \[e.g. at $t_1=34$, $t_2=98$ in Figs. \[abb:den2-2\] (c), (d)\] and contraction \[e.g. at $t_3=56$, $t_4=73$ in Figs. \[abb:den2-2\] (c), (d)\] of the fermionic cloud. Referring to contraction events, see Figs. \[abb:coh2-2\] (a~1~) and (a~2~), we observe that two filaments residing in distinct wells are fully incoherent between each other \[see e.g. $g_A^{(1)}(x=-5.5,x^\prime=2.5;t=34)\approx0$ in Fig. \[abb:coh2-2\] (a~1~)\]. However, two filaments located within the same well are partially coherent, see for instance $g_A^{(1)}(x=-5.5,x^\prime=-2.5;t=34)\approx0.5$ in Fig. \[abb:coh2-2\] (a~1~). In contrast, during the expansion of the cloud, see Figs. \[abb:coh2-2\] (a~3~) - (a~4~), every two filaments independently of their location appear to be fully incoherent among each other \[e.g. $g_A^{(1)}(x=-7.0,x^\prime=-2.0;t=56)\approx0.1$ in Fig. \[abb:coh2-2\] (a~3~))\]. Concluding we can infer that during the particle balanced FF quench dynamics Mott-like one-body correlations [@sherson2010single; @larson2008mott; @katsimiga2017dark] emerge either for filaments located at distinct wells (contraction events of the femionic cloud) or for all filaments formed (expansion events of the fermionic cloud). ![($a_1$)-($a_4$) One-body correlation function $g_A^{(1)}(x,x^\prime;t)$ of the $A$-species for different time instants (see legends) during the interaction quench dynamics from $g_{AB}=0.1$ to $g_{AB}=4.0$ of the particle balanced FF mixture. ($b_1$)-($b_4$) The corresponding intraspecies two-body correlation function $g_{AA}^{(2)}(x,x^\prime;t)$ and ($c_1$)-($c_4$) the interspecies two-body correlation function $g_{AB}^{(2)}(x,x^\prime;t)$. The mixture consists of $N_A=N_B=2$ fermions and is initially prepared in the weakly interacting, $g_{AB}=0.1$, ground state of the double-well.[]{data-label="abb:coh2-2"}](Fig7.eps){width="50.00000%"} To gain insight into the two-body character of the dynamics we first study the second order intraspecies correlation function $g_{AA}^{(2)}(x,x^\prime);t$ depicted in Figs. \[abb:coh2-2\] (b~1~)-(b~4~). Overall, strong anti-correlations occur within each well \[e.g. $g_{AA}^{(2)}(x=-5.5,x^\prime=-5.5;t=34)\approx0$ in Fig. \[abb:coh2-2\] (b~1~)\], whilst between the different wells a correlated behavior takes place \[e.g. $g_{AA}^{(2)}(x=-5.5,x^\prime=2.5;t=34)\approx1.3$ in Fig. \[abb:coh2-2\] (b~1~)\] in the course of the evolution. Moreover, the probability of two fermions of the same species to populate non-symmetric (with respect to $x=0$) filaments is favored when compared to the probability of occupying symmetric ones \[compare $g_{AA}^{(2)}(x=-7.5,x^\prime=2.5;t=56)\approx2.4$ and $g_{AA}^{(2)}(x=-2.5,x^\prime=2.5;t=56)\approx1.8$ respectively in Fig. \[abb:coh2-2\] (b~3~)\]. To obtain a further understanding of the FF mixture dynamics, we finally inspect the interspecies correlation function $g_{AB}^{(2)}(x,x^\prime;t)$, see Figs. \[abb:coh2-2\] (c~1~)-(c~4~). A correlation hole emerges on the diagonal elements of $g_{AB}^{(2)}(x,x^\prime;t)$ \[see e.g. $g_{AB}^{(2)}(x=-2.5,x^\prime=-2.5;t=56)\approx0.1$ in Fig. \[abb:coh2-2\] (c~3~)\] indicating that fermions of different species can not populate the same filament. Furthermore, two filaments within the same well are found to be strongly correlated \[e.g. $g_{AB}^{(2)}(x=-7.5,x^\prime=-2.5;t=56)\approx2.0$ in Fig. \[abb:coh2-2\](c~3~)\] which means that it is likely to be occupied by fermions of $A$ and $B$-species. On the other hand, regarding filaments located at different wells it is more preferable for two fermions of different species to reside in symmetric (with respect to $x=0$) filaments \[$g_{AB}^{(2)}(x=-2.5,x^\prime=2.5;t=56)\approx1.2$ in Fig. \[abb:coh2-2\] (c~1~)\] rather than non-symmetric ones \[$g_{AB}^{(2)}(x=-7.5,x^\prime=2.5;t=56)\approx0.8$ in Fig. \[abb:coh2-2\] (c~3~)\]. Combining the knowledge gained from the one- and two-body correlations we can conclude that the spatial distribution of the system in terms of the $A$- and $B$-species fermions in the four emerging filaments is either $A-B-A-B$ or $B-A-B-A$. The above is a few-body precursor of anti-ferromagnetic order. ![Single-shot images of each species, at distinct time instants (see legends) of the interaction quench dynamics from $g_{AB}=0.1$ to $g_{AB}=4.0$, obtained by averaging over (a), (e) $N_{shots}=1$, (b), (f) $N_{shots}=50$ and (c), (g) $N_{shots}=500$. (d), (h) The corresponding $\rho^{(1)}_{\sigma}(t)$ calculated within the MB approach. In all cases the system consists of $N_A=N_B=5$ fermions trapped in a double-well potential.[]{data-label="abb:shots2-2"}](Fig8.eps){width="49.00000%"} Single-Shot Simulations {#shots_par_bal} ----------------------- To showcase further experimental links of the above-discussed MB quench dynamics of particle balanced FF mixtures we next briefly discuss the outcome of the corresponding single-shot simulations. Notice that more details of this procedure can be found in Sec. \[shots\_par\_imb\], in Appendix \[single\_shots\_details\] and in [@mistakidis2017correlation; @erdmann2018correlated; @koutentakis2018probing]. In particular, we consider the setup containing $N_A=N_B=5$ fermions in each component and perform single-shot simulations during the dynamics induced by a quench from $g_{AB}=0.1$ to $g_{AB}=4.0$, see also Figs. \[abb:den2-2\] (e), (f). The first single-shot images for both species at $t_1=8$ and $t_2=32$ \[see Figs. \[abb:shots2-2\] (a) and (e)\] resemble the filamentized structure of $\rho_\sigma^{(1)}(x;t)$ \[see Figs. \[abb:shots2-2\] (a) and (e)\] within the double-well at both time instants. However, as also discussed in Sec. \[shots\_par\_imb\], an adequate correspondence between a single-shot image and the corresponding one-body density is not possible due to the small particle number. To capture the structures building upon $\rho_\sigma^{(1)}(x)$ we average over several single-shot realizations depicted in Figs. \[abb:shots2-2\] (b)-(c) and (f)-(g) for each time instant. As it can be seen, the averaged images $\bar{\mathcal{A}}^A(\tilde{x};t)$ and $\bar{\mathcal{A}}^B(\tilde{x}^{'}\|\mathcal{A}^A(\tilde{x});t)$ gradually approach $\rho_\sigma^{(1)}(x;t)$ \[Figs. \[abb:shots2-2\] (d) and (h)\] as $N_{shots}$ is increased. Conclusions {#sec:conclusion} =========== We have investigated the nonequilibrium quantum dynamics of a spin-polarized FF mixture confined in a double-well potential upon quenching the interspecies repulsion from the weak to the strong interaction regime and for both particle imbalanced and balanced mixtures. Comparing the dynamics within the HF approximation and the MB level enables us to infer about the crucial role of correlations on both the one- and two-body level in the course of the dynamics. In particular, we reveal a variety of interesting phenomena with MB origin such as phase separation processes, alteration of the Stoner instability and filamentation of the single-particle density. Regarding the ground state of particle imbalanced species a symmetry breaking of the single-particle density occurs for strong interspecies interactions within the HF approximation being related to the Stoner’s instability that renders the two fermionic clouds immiscible. Alteration of this instability is observed at the MB level due to the existence of higher-order correlations rendering the two components miscible and leading to a prominent intrawell fragmentation of the one-density. To induce the dynamics we suddenly change the interspecies interaction from weak-to-strong values. It is found that within the HF approximation the $\sigma$-species single-particle density filamentizes, i.e. the initial Gaussian-like density profile breaks into several localized density branches called filaments while the two species exhibit a dynamical phase separation. In sharp contrast, when correlations are included the filamentation of the one-body density becomes more faint and the two species show a miscible behavior on the one-body level. To provide further insights into the MB character of the dynamics we utilize the one- and two-body correlation functions. On the one-body level Mott-like correlations between the filaments are revealed, indicating their tendency for localization. Most importantly, both the intra- and interspecies correlation functions show a correlation hole in their diagonal elements suggesting that two fermions of the same or different species can not populate the same filament. However, the occurence of strong correlations between two distinct filaments indicates that two fermions of the same or different species can reside in distinct filaments. It is these observations that unveil the phase separated character of the MB dynamics on the two-body level while consisting a precursor of anti-ferromagnetic order. Turning our attention to particle balanced FF mixtures and their relevant ground state properties, we are able to showcase that while intrawell fragmentation occurs at the MB level within the HF approach only a broadening is present. In this case the species remain miscible both for weak and strong interspecies interactions independently of the considered approach. Performing an interspecies interaction quench from weak-to-strong coupling we observe that in the HF approximation the two fermionic clouds remain miscible throughout the evolution. Furthermore, they undergo an overall breathing motion over the double-well while in the course of the contraction events of this motion prominent interference patterns appear. Within the MB approach the two species are miscible and perform an overall breathing mode. Most importantly and in sharp contrast to the HF approximation, the clouds exhibit an intrawell fragmentation (filamentation) visible in their single-particle density with the number of filaments formed being proportional to the number of fermions of each species. Inspecting the one-body correlation function in the course of the evolution we deduce that Mott-like one-body correlations appear either for filaments located at distinct wells (contraction events) or for all filaments (expansion events of the fermionic cloud). Referring to the two-body correlations we find that two fermions of the same or different species exhibit an anti-correlated behavior in a single filament while they are strongly correlated when residing in distinct filaments with the non-symmetric ones (with respect to the center) being more favorable. The above indicate that the two species phase separate suggesting the formation of anti-ferromagnetic like order in the few-body system. Finally, we provide possible experimental realizations for both the particle imbalanced as well as the particle balanced cases by simulating single-shot measurements. In particular, we show how an averaging process of the obtained in-situ images can be used to adequately retrieve the MB fermionic quench dynamics. There are several promising research directions that are of interest for future investigations along the lines of the current effort. An imperative prospect is to simulate the corresponding radiofrequency spectrum [@mistakidis2018repulsive] in the case of particle imbalanced FF mixtures in order to reveal possibly emerging polaronic states and subsequently examine their properties. Another straightforward direction in particle imbalanced setups would be to consider a larger particle number for the minority species, e.g. $N_A=5$, $N_B=3$, and unveil whether phase separation processes and magnetization effects occur in such systems. Certainly the study of the interspecies interaction quench dynamics of mass imbalanced FF mixtures in order to induce a dynamical phase separation of the two species and showcase the role of correlations is an intriguing perspective. The Single-Shot Algorithm {#single_shots_details} ========================= The numerical simulation of the single-shot procedure relies on a sampling of the MB probability distribution [@sakmann2016single; @katsimiga2017many; @katsimiga2018many; @mistakidis2017correlation]. We remark that the implementation of this experimental measurement process has already been reported for single-component bosons and fermions [@sakmann2016single; @katsimiga2017many; @katsimiga2018many; @koutentakis2018probing; @erdmann2018correlated] as well as for binary bosonic and fermionic mixtures [@mistakidis2017correlation]. Below we provide a brief sketch of the corresponding numerical procedure but for more details we refer the interested reader to [@sakmann2016single; @katsimiga2017many; @katsimiga2018many; @koutentakis2018probing; @erdmann2018correlated]. The single-shots for binary mixtures depend strongly on the system specific inter- and intraspecies correlations [@mistakidis2017correlation]. For a MB system the presence of entanglement \[see Eq. (\[Eq:WF\])\] between the species is very important for the image ordering. In the following, we analyze the corresponding numerical process when the imaging is performed first on the $A$ and then to the $B$ species providing this way the absorption images $\mathcal{A}^A(\tilde{x})$ and $\mathcal{A}^B(\tilde{x}'\|\mathcal{A}^A(\tilde{x}))$. An important remark here is that in order to image first the $B$ and then the $A$ species we need to follow the same procedure, obtaining the corresponding images $\mathcal{A}^B(\tilde{x})$ and $\mathcal{A}^A(\tilde{x}'\|\mathcal{A}^B(\tilde{x}))$. To achieve the imaging of the $A$ and subsequently of the $B$ species we sequentially annihilate all $A$-species fermions. In particular at a certain time instant of the imaging, for instance $t_{im}$, a random position is drawn satisfying $\rho_{N_A}^{(1)}(x_1')>q_1$ with $q_1$ being a random number belonging to the interval \[$0$, $ \max\lbrace{\rho^{(1)}_{N_A}(x;t_{im})\rbrace}$\]. To proceed, the ($N_A+N_B$)-body wavefunction is projected onto the ($N_A-1+N_B$)-body one by the virtue of the projection operator $\frac{1}{\mathcal{N}}(\hat{\Psi}_A(x_1')\otimes \hat{\mathbb{I}}_B)$. Here $\hat{\Psi}_A(x_1')$ denotes the fermionic field operator annihilating an $A$ species fermion at position $x_1'$, while $\mathcal{N}$ is the normalization constant. As it can be easily deduced, this process directly affects the Schmidt weights, $\lambda_k$. In this way both $\rho^{(1)}_{N_A-1}(t_{im})$ and $\rho^{(1)}_{N_B}(t_{im})$ are changed. Recall that the $B$ species have not been imaged yet. Indeed, the Schmidt decomposition of the MB wavefunction after this first measurement reads $$\begin{split} &\ket{\tilde{\Psi}_{MB}^{N_A-1,N_B}(t_{im})}=\\ &\sum_i \sqrt{\tilde{\lambda}_{i,N_A-1}(t_{im})}\ket{\tilde{\Psi}_{i,N_A-1}^A(t_{im})}\ket{\Psi_i^B(t_{im})}. \label{Eq:A1} \end{split}$$ The $N_A-1$ species wavefunction is $\ket{\tilde{\Psi}_{i,N_A-1}^A}=\frac{1}{N_i}\hat{\Psi}_A(x_1')\ket{\Psi_i^A}$, and the normalization factor $N_i=\sqrt{\bra{\Psi_i^A}\hat{\Psi}_A^{\dagger}(x_1')\hat{\Psi}_A(x_1')\ket{\Psi_i^A}}$. Also the Schmidt coefficients of the ($N_A-1+N_B$)-body wavefunction read $\tilde{\lambda}_{i,N_A-1}=\lambda_i N_i/\sum_i \lambda_i N_i^2$. To complete the imaging process we repeat the above steps $N_A-1$ times and then obtain the distribution of positions ($x'_1$, $x'_2$,...,$x'_{N_A-1}$). The latter is subsequently convoluted with a point spread function resulting in the single-shot image of the $A$-species $\mathcal{A}^A(\tilde{x})=\frac{1}{\sqrt{2\pi}w_{PSF}}\sum_{i=1}^{N_A}e^{-\frac{(\tilde{x}-x'_i)^2}{2w_{PSF}^2}}$. In this expression $\tilde{x}$ denote the spatial coordinates within the image and $w_{PSF}$ is the width of the employed point spread function. The MB wavefunction after annihilating all $N_A$ fermions reads $$\begin{split} &\ket{\tilde{\Psi}_{MB}^{0,N_B}(t_{im})}=\\ &\ket{0^A} \otimes\sum_i \frac{\sqrt{\tilde{\lambda}_{i,1}(t_{im})} \braket{x'_{N_A}\|\Phi_{i,1}^A}}{\sum_j{\sqrt{\tilde{\lambda}_{j,1}(t_{im})\|\braket{x'_{N_A}\|\Phi_{j,1}^A}\|^2}}}\ket{\Psi_i^B(t_{im})}. \label{Eq:A3} \end{split}$$ In this expression $\braket{x'_{N_A}\|\Phi_{j,1}^A}\equiv\braket{0^A\|\hat{\Psi}_A(x'_{N_A})\|\Phi_{j,1}^A}$ is the single-particle orbital of the $j$-th mode, while the $B$-species wavefunction, i.e. $\ket{\Psi_{MB}^{N_B}(t_{im})}$, is the second term in the cross product of the right-hand side. As it can be seen, $\ket{\Psi_{MB}^{N_B}(t_{im})}$ refers to a non-entangled $N_B$-particle wavefunction. Therefore the subsequent single-shot procedure of the $B$ species is the same as for a single-species ensemble [@sakmann2016single; @katsimiga2017many; @katsimiga2018many]. This proccedure has been extensively tested for different single-component setups, see for more details [@sakmann2016single; @katsimiga2017many; @katsimiga2018many] and references therein. Therefore we only brief discuss it below. Referring to $t=t_{im}$ i.e. the imaging time, we compute $\rho^{(1)}_{N_B}(x;t_{im})$ from $\ket{\Psi^{N_B}_{MB}}\equiv \ket{\Psi(t_{im})}$ and a random position $x''_1$ is drawn obeying $\rho^{(1)}_{N_B}(x''_1;t_{im})>q_2$, where $q_2$ is a random number in the interval \[$0$, $\rho^{(1)}_{N_B}(x;t_{im})$\]. Consequently, one particle is annihilated at $x''_1$ and we calculate $\rho^{(1)}_{N_B-1}(x;t_{im})$ from $\ket{\Psi^{N_B-1}_{MB}}$. Then, a new random position $x''_2$ is drawn from $\rho^{(1)}_{N_B-1}(x;t_{im})$. Repeating the above procedure $N_B-1$ times we obtain the distribution of positions ($x''_1$, $x''_2$,...,$x''_{N_B-1}$). This distribution is finally convoluted with a point spread function providing a single-shot image $\mathcal{A}^B(\tilde{x'}\|\mathcal{A}^A(\tilde{x}))$. Finally, it can be shown that the average image of the $\sigma$ species, i.e. $\mathcal{\bar{A}}^{\sigma}(\tilde{x})$, over several ($N_{shots}$) single-shot images \[$\mathcal{A}^{\sigma}(\tilde{x})$\] is directly related to the $\sigma$ species one-body density, $\rho_{\sigma}^{(1)} (x_{\sigma}')$, since $$\mathcal{\bar{A}}^{\sigma}(\tilde{x})=\frac{N_{\sigma}}{\sqrt{2\pi}w_{PSF}}\int dx_{\sigma}' e^{-\frac{(\tilde{x}-x_{\sigma}')^2}{2w^2_{PSF}}} \rho_{\sigma}^{(1)} (x_{\sigma}'). \label{averaging}$$ Here, $\tilde{x}$ denote the spatial coordinates within the image and $x_{\sigma}'$ is the spatial coordinate of the $\sigma$ species. Also $w_{PSF}$ is the width of the employed point spread function and $N_{\sigma}$ refers to the particle number of the $\sigma$ species. Convergence and Further Details of the Many-Body Simulations {#sec:numerics} ============================================================ Let us briefly discuss the ingerdients of our MB simulations and showcase their numerical convergence. As it has been already argued in Sec. \[sec:wfn\], ML-MCTDHX [@ML-MCTDHX] is a variational method for solving the time-dependent MB Schr[ö]{}dinger equation for atomic mixtures consisting either of bosonic [@mistakidis2017correlation; @katsimiga2017dark; @mistakidis2018effective] or fermionic [@cao2017collective; @koutentakis2018probing; @erdmann2018correlated; @mistakidis2018repulsive] species. Within this approach, the MB wavefunction is expanded in terms of a time-dependent variationally optimized MB basis. Such a treatment, allows us to take into account the relevant intra- and interspecies correlation effects utilizing a computationally feasible basis size. In this way, the number of basis states can be significantly reduced as compared to methods which rely on a time-independent basis. The latter is achieved by choosing the relevant subspace of the Hilbert space at each time instant of the evolution in a more efficient manner. ![Evolution of the one-body density deviation $\Delta\rho^{(1)}_{A;C,C^\prime}(t)$ between the $C^\prime=10-(8,8)$ and other orbital configurations $C=M-(m^A,m^B)$ (see legend). The FF mixture consists of $N_A=N_B=2$ fermions and to induce the dynamics we perform a quench from $g_{AB}=0.1$ to $g_{AB}=4.0$.[]{data-label="abb:convergence"}](Fig9.eps){width="40.00000%"} The Hilbert space truncation refers to the employed numerical configuration space designated by $C=D-(m^A,m^B)$. In this notation $D=D^A=D^B$ and $m^A$, $m^B$ correspond to the number of species and single-particle functions respectively for each of the species \[see also Eqs. (\[Eq:WF\]) and (\[Eq:SPFs\])\]. For our simulations we invoke a primitive basis based on a sine discrete variable representation including 400 grid points. To conclude upon the convergence of our MB simulations we assure that variations of the numerical configuration space $C=D-(m^A,m^B)$ do not essentially affect the observables of interest. Note that all MB calculations presented in the main text are based on the numerical configuration space $C=6-(6,6)$ for $N_A=3$, $N_B=1$, on the $C=10-(10;10)$ in the case of $N_A=5$, $N_B=1$ and $N_A=N_B=5$ and on the $C=10-(8;8)$ when $N_A=N_B=2$. Therefore, the available Hilbert space for the corresponding simulation includes 4992 (10720) and 13140 (7060) coefficients for the $N_A=3$, $N_B=1$ ($N_A=5$, $N_B=1$) and the $N_A=N_B=5$ ($N_A=N_B=2$) cases respectively. This is in sharp contrast to an exact diagonalization procedure which should take into account 4.2 $10^9$ (3.3 $10^{13}$) and 6.9 $10^{21}$ (6.3 $10^{9}$) coefficients for the $N_A=3$, $N_B=1$ ($N_A=5$, $N_B=1$) and the $N_A=N_B=5$ ($N_A=N_B=2$) cases, rendering these simulations infeasible. Finally, let us briefly showcase the convergence of our results for a varying number of species and single-particle functions. For this investigation we resort to the $\sigma$-species one-body density, $\rho^{(1)}_{\sigma;C}(x;t)$, during the nonequilibrium dynamics and calculate its spatially integrated absolute deviation for each of the species between the $C'=10-(8,8)$ and other numerical configurations $C=D-(m^A,m^B)$. Namely $$\Delta\rho^{(1)}_{\sigma;C,C^\prime}(t) =\frac{1}{2N_\sigma}\int dx\|\rho_C^{(1),\sigma}(x;t) -\rho_{C^\prime}^{(1),\sigma}(x;t)\| \label{convergence}$$ Figure \[abb:convergence\] presents $\Delta\rho^{(1)}_{A;C,C^\prime}(t)$ for a FF mixture with $N_A=N_B=2$ fermions following an interspecies interaction quench from $g_{AB}=0.1$ to $g_{AB}=4.0$. We remark, that $\Delta\rho^{(1)}_{B;C,C^\prime}(t)=\Delta\rho^{(1)}_{A;C,C^\prime}(t)$ at all times and for all configurations due to the particle balanced mixture. Therefore the results of the $A$-species are representative for both species. Inspecting Fig. \[abb:convergence\], it becomes evident that a systematic convergence of $\Delta\rho^{(1)}_{\sigma;C,C^\prime}(t)$ can be achieved. More specifically, comparing $\Delta\rho^{(1)}_{A;C,C^\prime}(t)$ between the $C=8-(8,8)$ and $C'=10-(8,8)$ approximations we observe that the corresponding relative difference is below $0.15\%$ throughout the evolution. Finally, we remark that a similar analysis has been performed for all other particle configurations, i.e. particle number imbalanced systems as well as higher particle numbers, discussed within the main text and found to be adequately converged (not shown here for brevity). Acknowledgements {#acknowledgements .unnumbered} ================ S.I.M. and P.S. gratefully acknowledge financial support by the Deutsche Forschungsgemeinschaft (DFG) in the framework of the SFB 925 “Light induced dynamics and control of correlated quantum systems”. [^1]: Due to the presence of strong interspecies interactions the Gaussian-like density profile within each well deforms and either exhibits several local maxima or breaks into distinct density branches [@mistakidis2017correlation]. Throughout this work we refer to both these local maxima or density branches as filaments. [^2]: For instance, inspecting the corresponding orbital densities we observe that the filaments building upon higher than the second populated orbitals used (not shown here) are much more pronounced than those developed in the first two orbitals.	{ "pile_set_name": "ArXiv" }
--- author: - \| [^1]\ Bethe Center for Theoretical Physics and Physikalisches Institut, Universität Bonn\ Nussallee 12, D-53115 Bonn, Germany\ Centro de Física Teórica de Partículas, Instituto Superior Técnico\ Avenida Rovisco Pais, 1049-001 Lisboa, Portugal\ E-mail: title: 'Reconstructing Dark Matter Properties via Gamma-Rays with [Fermi]{}-LAT' --- Introduction ============ If Dark Matter (DM) is detected and identified, the measurement of its properties like mass, annihilation cross-section and annihilation channels plays a central role in the determination of the particle nature of the DM. It will allow us to constrain models of particle physics beyond the Standard Model, for instance supersymmetry and universal extra dimensions. Furthermore a convincing DM discovery may require consistent signals in multiple experiments in multiple channels (direct, indirect, collider). We discuss the capabilities of the [Fermi]{}–LAT instrument for identifying particle DM properties with gamma-ray observations from the Galactic Center (GC). The differential intensity of the photon signal from a given observational region in the galactic halo from the annihilation of DM particles has different possible origins: internal bremsstrahlung and secondary photons (prompt) as well as Inverse Compton Scattering (ICS). External bremsstrahlung and synchrotron emission also contribute to the photon flux; however, for the energies of interest here and for typical DM masses, both bremsstrahlung and synchrotron emission are expected to be subdominant with respect to ICS. For the sake of simplicity we will neglect these sources in what follows. The differential flux of prompt gamma-rays from DM annihilations and coming from a direction within a solid angle $\Delta\Omega$ is given by $$\left(\frac{d\Phi_{\gamma}}{dE_\gamma}\right)_{{\rm prompt}} (E_{\gamma},\, \Delta\Omega) = \frac{\langle\sigma v\rangle}{2\,m_\chi^2}\sum_i\frac{dN_{\gamma}^i}{dE_{\gamma}}\, \textrm{BR}_i \, \frac{1}{4\,\pi} \, \int_{\Delta\Omega}d\Omega \, \int_\textrm{los}\rho\big(r(s,\,\Omega)\big)^2 \, ds\,, \label{Eq:promptflux}$$ where $\langle\sigma v\rangle$ is the total thermally averaged annihilation cross section, $m_\chi$ the mass of the DM particle, $\textrm{BR}_i$ the annihilation fraction into channel $i$, $dN_\gamma^i/dE_\gamma$ the differential gamma-ray yield of standard model particles into photons of energy $E_\gamma$, $\rho(r)$ the DM density profile and $r$ the distance from the GC. Here we will focus on the NFW halo profile [@Navarro:1995iw]; the dependence on the DM halo profile has been studied in reference [@Bernal:2010ip]. An abundant population of energetic electrons and positrons produced in DM annihilations either directly or indirectly from the hadronization, fragmentation, and subsequent decay of the SM particles in the final states, gives rise to secondary photons at various wavelengths via ICS off the diffuse radiation fields in the galaxy. We approximate this photon background as a superposition of three black-body spectra consisting of the CMB, the optical starlight and the infrared radiation due to rescattering of starlight by dust [@ics]. The differential flux of high energy photons produced by the ICS processes is given by [@Blumenthal:1970gc] $$\left(\frac{d\Phi_{\gamma}}{dE_\gamma}\right)_{{\rm ICS}} (E_{\gamma},\, \Delta\Omega) = \frac{1}{E_{\gamma}} \, \frac{1}{4\pi} \, \int_{\Delta\Omega}d\Omega \, \int_\textrm{los}ds \, \int_{m_e}^{m_\chi}dE\,\mathcal{P}(E_{\gamma},\,E) \, \frac{dn_e}{dE}\big(E,\,r,\,z\big) ~, \label{Eq:ICSflux}$$ where $\mathcal{P}(E_{\gamma},\,E)$ is the differential power emitted into scattered photons of energy $E_{\gamma}$ by an electron with energy $E$. The minimal and maximal energies of the electrons are determined by the electron mass $m_e$ and the DM particle mass. The quantity $dn_e/dE$ is the electron plus positron spectrum after propagation in the Galaxy, which will differ from the energy spectrum produced at the source. We determine the propagated spectrum by solving the diffusion-loss equation that describes the evolution of the energy distribution for electrons and positrons assuming steady state [@propaga]. Regarding the propagation parameters (like diffusion coefficient, energy losses and thickness of the diffusion zone), we take their values from the commonly used MED model [@propaga]. Again, the dependence on the propagation model has been studied in reference [@Bernal:2010ip]. There are three main components contributing to the high-energy gamma-ray background: the diffuse galactic emission has been estimated by taking the conventional model of the GALPROP code [@galprop]. On the other hand, another source of background particularly important when looking at the GC is that of resolved point sources. We consider all the point sources detected by the first 11 months of [Fermi]{}–LAT [@pointsources] lying in the region of interest. Finally, for the isotropic extragalactic gamma-ray background we used the recent measurements by the [Fermi]{}–LAT collaboration [@Abdo:2010nz]. For a $10^\circ \times 10^\circ$ region around the GC, the diffuse galactic emission dominates below $\sim 20$ GeV. Above that value, the emission coming from point sources is the most important. The isotropic extragalactic gamma-ray background is at the percent level. The Large Area Telescope ([Fermi]{}–LAT) is the primary instrument on board of the [Fermi Gamma-ray Space Telescope]{}. It performs an all-sky survey, covering a large energy range for gamma-rays, with an effective area $\simeq 8000$ cm$^2$ and a field of view of $2.4$ sr. In the following analysis, we consider a 5-year mission run, and an energy range from 1 GeV extending up to 300 GeV. We divide this energy interval into 20 evenly spaced logarithmic bins. In order to maximize the signal-to-noise ratio, it has been pointed out that for a NFW profile the best strategy is to focus on a region around the GC of $\sim10^\circ\times 10^\circ$ [@maxsb]. Hence, this is our choice. Reconstructing Dark Matter properties ===================================== Once gamma-rays are identified as having been produced in DM annihilations, the next step concerns the possibilities of constraining DM properties [@papas]. ![[Fermi]{}–LAT abilities to constrain DM properties. We consider DM annihilation into a pure $b\bar b$ final state and two DM masses: $m_\chi=80$ GeV (left panels) and $m_\chi=270$ GeV (right panels). Dark blue (light orange) regions represent $68\%$ CL ($90\%$ CL) contours. We assume a $10^\circ \times 10^\circ$ observational region around the GC, a NFW DM halo profile, the MED propagation model and $\langle\sigma v\rangle=3\cdot 10^{-26}$ cm$^3$ s$^{-1}$. The black crosses indicate the values of the parameters for the simulated observed “data”.[]{data-label="Fig:b10deg"}](b_79_3-26.eps "fig:"){width="7.5cm"} ![[Fermi]{}–LAT abilities to constrain DM properties. We consider DM annihilation into a pure $b\bar b$ final state and two DM masses: $m_\chi=80$ GeV (left panels) and $m_\chi=270$ GeV (right panels). Dark blue (light orange) regions represent $68\%$ CL ($90\%$ CL) contours. We assume a $10^\circ \times 10^\circ$ observational region around the GC, a NFW DM halo profile, the MED propagation model and $\langle\sigma v\rangle=3\cdot 10^{-26}$ cm$^3$ s$^{-1}$. The black crosses indicate the values of the parameters for the simulated observed “data”.[]{data-label="Fig:b10deg"}](b_268_3-26.eps "fig:"){width="7.5cm"} In figure \[Fig:b10deg\] we depict the [Fermi]{}–LAT reconstruction prospects after 5 years for DM annihilation into a pure $b\bar b$ final state reconstructed as either $\tau^+\tau^-$ or $b\bar b$ and two possible DM masses: $m_\chi=80$ GeV (left panels) and $m_\chi=270$ GeV (right panels). We also assume DM particle with a typical thermal annihilation cross section $\langle\sigma v\rangle=3\cdot 10^{-26}$ cm$^3$ s$^{-1}$, the MED propagation model, a NFW DM halo profile and a $10^\circ \times 10^\circ$ observational region around the GC. These benchmark points are represented in the figure by black crosses. The dark blue regions and the light orange regions correspond to the $68\%$ CL and $90\%$ CL contours respectively. The different panels show the results for the planes $(m_\chi,\,\langle\sigma v\rangle)$, $($BR$_{\tau(b)},\,\langle\sigma v\rangle)$ and $(m_\chi,\,$BR$_{\tau(b)})$, marginalizing with respect to the other parameter in each case. BR$_{\tau(b)}=100\%$ ($0\%$) corresponds to an annihilation into a pure $\tau^+\tau^-$ $\left(b\bar b\right)$ final state. For the first model chosen in figure \[Fig:b10deg\] (left panels), $m_\chi=80$ GeV and in general for light DM masses, the reconstruction prospects seem to be promising, allowing the determination of the mass, the annihilation cross section and the annihilation channel at the level of $\sim$20% or better. On the other hand, for heavier DM particles, the regions allowed by data grow considerably worsening the abilities of the experiment to reconstruct DM properties. This is shown for the second model in figure \[Fig:b10deg\] (right panels), $m_\chi=270$ GeV. In this case, [Fermi]{}–LAT would only be able to constrain the DM mass to be in the range $\sim (30-500)$ GeV and determine the annihilation cross section within an order of magnitude. Let us note the appearance of a second spurious minima corresponding to a lighter mass $m_\chi\sim 60$ GeV, an annihilation cross section $\langle\sigma v\rangle\sim 10^{-26}$ cm$^3$ s$^{-1}$ and annihilating mainly ($\sim 75\%$) on $\tau^+\tau^-$ pairs. ![[Fermi]{}–LAT abilities to constrain DM properties. We assume the measured signal is due to DM annihilating into $\mu^+\mu^-$, but the fit is obtained assuming DM annihilates into either $\tau^+\tau^-$ or $b\bar b$. We assume two DM masses: $m_\chi=50$ GeV (left panels) and $m_\chi=105$ GeV (right panels). Dark blue (light orange) regions represent the $68\%$ CL ($90\%$ CL) contours. See the text for the rest of the parameters. The black cross in the left-top panel in each plot indicates the values of the parameters for the simulated observed “data”. Note that the other panels have no cross as they lie outside the parameter space of the simulated observed “data”. The squares indicate the best-fit point.[]{data-label="Fig:mubtau"}](mu_b-tau_49.eps "fig:"){width="7.5cm"} ![[Fermi]{}–LAT abilities to constrain DM properties. We assume the measured signal is due to DM annihilating into $\mu^+\mu^-$, but the fit is obtained assuming DM annihilates into either $\tau^+\tau^-$ or $b\bar b$. We assume two DM masses: $m_\chi=50$ GeV (left panels) and $m_\chi=105$ GeV (right panels). Dark blue (light orange) regions represent the $68\%$ CL ($90\%$ CL) contours. See the text for the rest of the parameters. The black cross in the left-top panel in each plot indicates the values of the parameters for the simulated observed “data”. Note that the other panels have no cross as they lie outside the parameter space of the simulated observed “data”. The squares indicate the best-fit point.[]{data-label="Fig:mubtau"}](mu_b-tau_105.eps "fig:"){width="7.5cm"} In figure \[Fig:mubtau\] we assume that DM actually annihilates into $\mu^+\mu^-$ pairs, but we analyze the data assuming DM annihilations into either $\tau^+\tau^-$ or $b\bar b$, for $\langle\sigma v\rangle=3\cdot 10^{-26}$ cm$^3$ s$^{-1}$ and for two DM masses: $m_\chi = 50$ GeV (left panels) and $m_\chi = 105$ GeV (right panels). Again, the black crosses indicate the values of the parameters for the simulated observed “data”; the squares indicate the best-fit points. Naïvely, one would expect that the $\mu^+\mu^-$ (leptonic) channel is identified as being closer to the $\tau^+\tau^-$ (leptonic) channel than to the $b\bar b$ (hadronic) channel. Let us remember that DM annihilation channels are commonly classified into two broad classes: hadronic and leptonic channels. Leptonic channels typically give rise to a harder spectrum and, in particular for $e^+e^-$ and $\mu^+\mu^-$, the cutoff is be very sharp and around a maximum energy (i.e. the mass of the DM particle in the case on annihilating DM). Contrary to what was expected, the reconstructed composition of the annihilation channels tends to be dominated by $b\bar b$, instead of $\tau^+\tau^-$. Indeed, the contribution due to ICS in the case of the $\mu^+\mu^-$ (and also the $e^+e^-$) channel could substantially alter the different prompt spectra. Hence, when taking into account the contribution of ICS to the gamma-ray spectrum, the annihilation channels cannot be generically classified as hadronic or leptonic, as DM annihilations into $\mu^+\mu^-$ pairs are better reproduced with the $b\bar b$ channels than with the $\tau^+\tau^-$ channel. ![[Fermi]{}–LAT abilities to constrain DM properties. We assume the measured signal is due to DM annihilating into $\mu^+\mu^-$ and the fit is obtained assuming DM annihilates into $\mu^+\mu^-$ or $b\bar b$. We assume ICS+prompt photons (left panels) or only prompt photons (right panels) for the reconstructed signal, for $m_\chi=50$ GeV. Dark blue (light orange) regions represent the $68\%$ CL ($90\%$ CL) contours. See the text for the rest of the parameters. The black crosses indicate the values of the parameters for the simulated observed “data”. The squares in the right panels indicate the best-fit point.[]{data-label="Fig:mumub"}](mucon.eps "fig:"){width="7.5cm"} ![[Fermi]{}–LAT abilities to constrain DM properties. We assume the measured signal is due to DM annihilating into $\mu^+\mu^-$ and the fit is obtained assuming DM annihilates into $\mu^+\mu^-$ or $b\bar b$. We assume ICS+prompt photons (left panels) or only prompt photons (right panels) for the reconstructed signal, for $m_\chi=50$ GeV. Dark blue (light orange) regions represent the $68\%$ CL ($90\%$ CL) contours. See the text for the rest of the parameters. The black crosses indicate the values of the parameters for the simulated observed “data”. The squares in the right panels indicate the best-fit point.[]{data-label="Fig:mumub"}](musin.eps "fig:"){width="7.5cm"} The results just discussed can be illustrated in a different way by analyzing the simulated observed signal “data” from DM annihilation into $\mu^+\mu^-$ assuming DM annihilates into either $\mu^+\mu^-$ or $b\bar b$. This is depicted in figure \[Fig:mumub\] where we show the results for the case that we try to reconstruct the full signal (prompt and ICS) generated by a $50$ GeV DM particle adding the ICS contribution (left panels) or with only prompt photons (right panels). As can be seen in the left panels, if ICS is taken into account, DM properties can be reconstructed with good precision. However, if the ICS contribution is not added to the simulated signal events (the simulated observed “data” always has the ICS included), DM annihilation into a pure $\mu^+\mu^-$ channel would be excluded at more $90\%$ CL, providing thus a completely wrong result. Conclusions =========== In this work we have studied the abilities of the [Fermi]{}–LAT instrument to constrain Dark Matter properties by using the current and future observations of gamma-rays from the Galactic Center produced by DM annihilations. Unlike previous works, we also take into account the contribution to the gamma-ray spectrum from ICS of electrons and positrons produced in DM annihilations off the ambient photon background. We show that the inclusion of the ICS contribution for hadronic channels and for the $\tau^+\tau^-$ channel does not give rise to important differences in the reconstruction process. This is not the case if DM annihilates into the $\mu^+\mu^-$ and the $e^+e^-$ channel. In this latter case, adding the ICS contribution to the prompt gamma-ray spectrum turns out to be crucial in order not to obtain completely wrong results. On the other hand, we found that for Dark Matter masses below $\sim 200$ GeV and for typical thermal annihilation cross sections, it will be possible to obtain stringent bounds on the Dark Matter properties such as its mass, annihilation cross section and annihilation channels. [99]{} J. F. Navarro, C. S. Frenk and S. D. M. White, Astrophys. J. [462]{} (1996) 563 \[arXiv:astro-ph/9508025\]. N. Bernal and S. Palomares-Ruiz, arXiv:1006.0477 \[astro-ph.HE\]. T. A. Porter and A. W. Strong, arXiv:astro-ph/0507119; T. A. Porter, I. V. Moskalenko, A. W. Strong, E. Orlando and L. Bouchet, Astrophys. J. [682]{} (2008) 400 \[arXiv:0804.1774 \[astro-ph\]\]; M. Cirelli and P. Panci, Nucl. Phys. B [821]{} (2009) 399 \[arXiv:0904.3830 \[astro-ph.CO\]\]. G. R. Blumenthal and R. J. Gould, Rev. Mod. Phys. [42]{} (1970) 237. V. L. Ginzburg and S. L. Syrovatskii, New York, Gordon and Breach, 1969; E. A. Baltz and J. Edsjö, Phys. Rev. D [59]{} (1998) 023511 \[arXiv:astro-ph/9808243\]; T. Delahaye, R. Lineros, F. Donato, N. Fornengo and P. Salati, Phys. Rev. D [77]{} (2008) 063527 \[arXiv:0712.2312 \[astro-ph\]\]. A. W. Strong and I. V. Moskalenko, Astrophys. J. [509]{} (1998) 212 \[arXiv:astro-ph/9807150\]; A. E. Vladimirov [et al.]{}, arXiv:1008.3642 \[astro-ph.HE\]. [Fermi-LAT]{} Collaboration, A. A. Abdo [et al.]{}, Astrophys. J. Suppl. [188]{} (2010) 405 \[arXiv:1002.2280 \[astro-ph.HE\]\]. Collaboration, A. A. Abdo [et al.]{} Phys. Rev. Lett. [104]{} (2010) 101101 \[arXiv:1002.3603 \[astro-ph.HE\]\]. T. E. Jeltema and S. Profumo, JCAP [0811]{} (2008) 003 \[arXiv:0808.2641 \[astro-ph\]\]; P. D. Serpico and G. Zaharijas, Astropart. Phys. [29]{} (2008) 380 \[arXiv:0802.3245 \[astro-ph\]\]. S. Dodelson, D. Hooper and P. D. Serpico, Phys. Rev. D [77]{} (2008) 063512 \[arXiv:0711.4621 \[astro-ph\]\]; N. Bernal, A. Goudelis, Y. Mambrini and C. Muñoz, JCAP [0901]{} (2009) 046 \[arXiv:0804.1976 \[hep-ph\]\]; N. Bernal, arXiv:0805.2241 \[hep-ph\]; T. E. Jeltema and S. Profumo, JCAP [0811]{} (2008) 003 \[arXiv:0808.2641 \[astro-ph\]\]; S. Palomares-Ruiz and J. M. Siegal-Gaskins, JCAP [1007]{} (2010) 023 \[arXiv:1003.1142 \[astro-ph.CO\]\]. [^1]: The author was supported by the EU project MRTN-CT-2006-035505 HEPTools and the DFG TRR33 ‘The Dark Universe’.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Charmonia spectral functions at finite temperature are studied using QCD sum rules in combination with the maximum entropy method. This approach enables us to directly obtain the spectral function from the sum rules, without having to introduce any specific assumption about its functional form. As a result, it is found that while $J/\psi$ and $\eta_c$ manifest themselves as significant peaks in the spectral function below the deconfinement temperature $T_c$, they quickly dissolve into the continuum and almost completely disappear at temperatures between 1.0 $T_c$ and 1.1 $T_c$.' author: - Philipp Gubler - Kenji Morita - Makoto Oka title: Charmonium spectra at finite temperature from QCD sum rules with the maximum entropy method --- Since QCD was established to be the theory of strong interactions, charmonium has often been used as a suitable probe of its dynamics, owing to the fact that in this system both perturbative and nonperturbative aspects of QCD play equally important roles [@Novikov]. The behavior of charmonia in a hot or dense medium has also attracted much interest, as it was suggested some time ago, that in the color-deconfined medium with a temperature above $T_c$ charmonia will dissolve due to the color Debye screening, and thus serve as a signal for the formation of quark-gluon plasma [@Matsui]. Testing these early suggestions from first principles of QCD has become feasible only recently, as new developments in lattice QCD have made it possible to access the charmonium spectral functions with the help of the maximum entropy method (MEM) [@Asakawa; @Datta; @Umeda; @Jakovac]. These studies found that the lowest charmonium states ($J/\psi$ and $\eta_c$) survive up to temperatures as high as $\sim$ 1.5 $T_c$ or higher. Besides lattice QCD, the method of QCD sum rules [@Shifman] provides another tool for investigating the properties of hadrons at finite temperature [@Bochkarev; @Hatsuda]. Using this approach various charmonium channels were studied recently [@Morita1; @Morita3], and evidence for a considerable change of the spectral functions just above $T_c$ was found. To specify the nature of this change is the major goal of this study. For this task we employ MEM, which is applicable to QCD sum rules [@Gubler] and has the advantage that one does not have to introduce any strong assumption about the functional form of the spectral function, such as the “pole + continuum" ansatz, which is often used in QCD sum rule studies. Let us first recapitulate what sort of information QCD sum rules can provide on the charmonium spectral function at finite temperature [@Bochkarev; @Hatsuda]. One considers the time-ordered correlator at finite temperature $$\Pi^{\mathrm{J}}(q) = i \displaystyle \int d^4x e^{iqx} \langle T [j^{\mathrm{J}}(x) j^{\mathrm{J}}(0) ] \rangle_T, \label{eq:correlator}$$ where $j^{\mathrm{J}}(x)$ stands for $\bar{c} \gamma_{\mu} c(x)$ and $\bar{c} \gamma_{5} c(x)$ in the vector ($\mathrm{V}$) and pseudoscalar ($\mathrm{PS}$) channel, respectively. The expectation value $\langle \mathcal{O} \rangle_T$ is defined as $\langle \mathcal{O} \rangle_T \equiv \mathrm{Tr}( e^{-H/T} \mathcal{O} ) / \mathrm{Tr}( e^{-H/T} )$. Throughout this work, we will set the spatial momentum of the charmonium system relative to the thermal medium to be $\textbf{0}$; thus, $q^{\mu} = (\omega, \textbf{0})$. In this circumstance, there is only one independent component in the correlator of the vector channel. In what follows, we will use the dimensionless functions $\tilde{\Pi}^{\mathrm{V}}(q^2) \equiv \Pi^{\mu,\mathrm{V}}_{\mu}(q)/(-3q^2)$ and $\tilde{\Pi}^{\mathrm{PS}}(q^2) \equiv \Pi^{\mathrm{PS}}(q)/q^2$ for the analysis. Going to the deep Euclidean region $q^2 \equiv -Q^2 \ll 0$, one can calculate the correlation functions using the operator product expansion (OPE), giving an expansion in local operators $O_n$ with increasing mass dimension $n$: $\tilde{\Pi}^{\mathrm{J}}(q^2) = \sum_n C^{\mathrm{J}}_n(q^2) \langle O_n \rangle_T$. As was first discussed in [@Hatsuda], as long as the temperature $T$ lies below the separation scale of the OPE, which is of the order of $\sim 1$ GeV, all the temperature effects can be included into the expectation values of the local operators $\langle O_n \rangle_T$, while the Wilson coefficients $C^{\mathrm{J}}_n(q^2)$ are independent of $T$. Furthermore, to improve the convergence of the OPE and suppressing the influence of high energy states onto the sum rule, we apply the Borel transform to the correlator, leading to the final result of the OPE for $\nu \equiv 4m_c^2/M^2$, $M$ being the Borel mass: $$\begin{split} \mathcal{M}^{\mathrm{J}}(\nu) = & e^{-\nu}A^{\mathrm{J}}(\nu)[1 + \alpha_s(\nu) a^{\mathrm{J}}(\nu) + b^{\mathrm{J}}(\nu) \phi_b(T) \\ &+ c^{\mathrm{J}}(\nu) \phi_c(T) + d^{\mathrm{J}}(\nu) \phi_d(T)]. \end{split} \label{eq:OPE}$$ The first two terms in Eq.(\[eq:OPE\]) are the leading order perturbative term and its first order $\alpha_s$ correction. The third and fourth terms contain the scalar and twist-2 gluon condensates of mass dimension 4: $\phi_b(T) = \frac{4\pi^2}{9(4m_c^2)^2} G_0$ and $\phi_c(T) = \frac{4\pi^2}{3(4m_c^2)^2} G_2$, where $G_0 = \langle \frac{\alpha_s}{\pi} G^a_{\mu\nu} G^{a\mu\nu}\rangle_T$ and $G_2$ is defined as $\langle \frac{\alpha_s}{\pi} G^{a\mu\sigma} G^{a\nu}_{\sigma}\rangle_T = (u^{\mu}u^{\nu} - \frac{1}{4}g^{\mu\nu})G_2$, $u^{\mu}$ being the four velocity of the medium. For the detailed expressions of the Wilson coefficients of these terms, see [@Morita3]. To evaluate the possible influence of higher order contributions, we include one more term, which is proportional to the scalar gluon condensate of dimension 6, $\phi_d(T) = \frac{1}{(4m_c^2)^3}\langle g^3 f^{abc} G^{a\nu}_{\mu} G^{b\lambda}_{\nu} G^{c\mu}_{\lambda} \rangle_T$. The Wilson coefficient of this term can be found in [@Marrow]. The correlator can also be expressed by a dispersion relation, in terms of the spectral function $\rho^{\mathrm{J}}(\omega)$ of the channel specified by the operator $j^{\mathrm{J}}(x)$. After the Borel transform one obtains $$\mathcal{M}^{\mathrm{J}}(\nu) = \displaystyle \int_0^{\infty}dx^2 e^{-x^2 \nu} \rho^{\mathrm{J}}(2m_c x). \label{eq:dispersion}$$ Equating Eqs.(\[eq:OPE\]) and (\[eq:dispersion\]) then gives the final form of the sum rules. In the vector channel, an additional constant term contributes to Eq.(\[eq:OPE\]), which originates from a pole at $\omega = 0$ in $\rho^{\mathrm{V}}(\omega)$ [@Bochkarev]. As this so-called scattering term considerably complicates the analysis, we eliminate it by taking the derivative of Eqs.(\[eq:OPE\]) and (\[eq:dispersion\]) with respect to $\nu$ and analyze only the resulting derivative sum rule in this channel. For a discussion on the validity of this procedure in the heavy quark sum rules, see [@Morita4]. The usual strategy of analyzing QCD sum rules is to make some reasonable assumptions on the functional form of the spectral function, and then extract information on the lowest lying peak from Eqs.(\[eq:OPE\]) and (\[eq:dispersion\]). This method, however, has several shortcomings. First of all, the widely used “pole + continuum" ansatz, which certainly works well at $T=0$, may not be appropriate at temperatures above $T_c$, where the lowest lying state is expected to be modified and eventually melt into the $c$-$\bar{c}$ continuum, which could become the dominant contribution. Furthermore, it is not always possible to unambiguously fit a specific ansatz to the OPE results, because of the occurrence of equally valid solutions. Such a situation arose in [@Morita1; @Morita3], where it was not possible to determine a unique solution for the used parametrization of the spectral function. To handle these problems, we propose to use MEM, which allows us to extract the spectral function from Eqs.(\[eq:OPE\]) and (\[eq:dispersion\]) without prejudice on its functional form. Moreover, it can be proven that this method provides a unique solution for the spectral function [@Asakawa2]. Let us now briefly summarize the basic ideas of MEM, which helps us to carry out the task of inverting the integral of Eq.(\[eq:dispersion\]). This is, however, an ill-posed problem as we have only access to $\mathcal{M}^{\mathrm{J}}(\nu)$ in the region of $\nu$ where the OPE shows sufficient convergence and, furthermore, have only information on $\mathcal{M}^{\mathrm{J}}(\nu)$ with limited precision due to the uncertainties of the values of $m_c$, $\alpha_s$ and the various condensates. Nonetheless, using Bayesian probability theory, MEM makes it possible to at least obtain the most probable form of the spectral function $\rho(\omega)$. To do this, one defines a probability $P[\rho\|\mathcal{M}\mathcal{I}]$ for $\rho$ to have a specific form given $\mathcal{M}$ and additional information $\mathcal{I}$ on $\rho$ such as positivity and asymptotic values. Using Bayes’ theorem, this probability can be rewritten as $P[\rho\|\mathcal{M}\mathcal{I}] \propto P[\mathcal{M}\|\rho\mathcal{I}] P[\rho\|\mathcal{I}] = e^{- L + \alpha S}$, where $e^{-L}$ is the likelihood function, used in standard $\chi^2$ methods. $e^{\alpha S}$ stands for the prior probability, given by the parameter $\alpha$ and the Shannon-Jaynes entropy, $$S = \displaystyle \int_0^{\infty} d\omega \Bigr[ \rho(\omega) - m(\omega) - \rho(\omega) \log \Bigl( \frac{\rho(\omega)}{m(\omega)} \Bigr) \Bigl]. \label{eq:SJe}$$ $m(\omega)$ is called the default model and can be used to incorporate prior knowledge on the asymptotic values of $\rho(\omega)$ into the analysis. Determining now the most probable $\rho(\omega)$ corresponds to the numerical problem of finding the maximum of the functional $- L + \alpha S$, for which we use the Bryan algorithm [@Bryan]. As a last step, the spectral function $\rho_{\alpha}(\omega)$ maximizing $- L + \alpha S$ for a fixed value of $\alpha$ is averaged using $\displaystyle \int d \alpha \rho_{\alpha}(\omega) P[\alpha \|\mathcal{M}\mathcal{I}]$, which gives the final output of the MEM analysis. For more details see [@Asakawa2], while specific issues concerning the application of this method to QCD sum rules are discussed in [@Gubler]. Next, we describe how the temperature dependencies of the gluonic condensates are determined. For the scalar and twist-2 gluon condensates with mass dimension 4, we follow the approach proposed in [@Morita1; @Morita3], where, in the quenched approximation, the energy momentum tensor, expressed using gluonic operators, was matched with the corresponding quantity written down in form of the energy density $\epsilon$ and the pressure $p$, leading to $G_0 = G^{\mathrm{vac}}_0 - \frac{8}{11}\bigl[\epsilon(T) - 3p(T)\bigr]$ and $G_2 = -\frac{\alpha_s(T)}{\pi}\bigl[\epsilon(T) + p(T)\bigr]$ for the scalar and twist-2 gluon condensates. The functions $\epsilon(T)$, $p(T)$ and $\alpha_s(T)$ were then extracted from quenched lattice QCD data [@Boyd; @Kaczmarek]. We will in this study use the same numerical values for the $T$ dependent part of $G_0$ and $G_2$ as in [@Morita3]. As is shown there, both $G_0$ and $G_2$ exhibit a sudden decrease in the vicinity of $T_c$. It was suggested in previous studies that the OPE could break down at temperatures above $T_c$ as higher dimensional operators may become non-negligible [@Morita3]. To investigate this possibility, we include the scalar gluonic condensate with dimension 6, $\langle g^3 f^{abc} G^{a\nu}_{\mu} G^{b\lambda}_{\nu} G^{c\mu}_{\lambda} \rangle$, about which much less is known. To our knowledge, at $T=0$, there exists only an estimate based on the dilute instanton gas model, giving $ \langle g^3 f^{abc} G^{a\nu}_{\mu} G^{b\lambda}_{\nu} G^{c\mu}_{\lambda} \rangle = \frac{48 \pi^2}{5\rho_c^2} \langle \frac{\alpha_s}{\pi} G^2 \rangle, $ where $\rho_c$ is a representative value for the instanton radius, for which we use the established value of 0.33 fm [@Schafer]. Assuming that the relation above also holds at finite temperature, and taking into account the reduction of $\rho_c$ above $T_c$ [@Chu], we, however, found that the dimension 6 term does not influence the behavior of the spectral function much in the temperature region investigated in this Letter. Therefore, we conclude that even though the relation obtained from the dilute instanton gas model can only be considered to be a crude estimate, as long as it gives the correct order of magnitude, the contribution of the dimension 6 condensate is small and does not alter the results obtained in this study. ![image](Jpsi.etac.0.0.7.eps){width="7.0cm"} ![image](Jpsi.etac.1.0.7.eps){width="7.0cm"} Let us now turn to the MEM analysis of Eqs.(\[eq:OPE\]) and (\[eq:dispersion\]). First, we investigate the spectral function at $T = 0$ in the vector and pseudoscalar channel. To determine the upper boundary of the region of $\nu$ to be analyzed, we employ the criterion that the dimension 6 term should be smaller than 20% of the whole OPE expression of Eq.(\[eq:OPE\]), which gives $\nu^{\mathrm{V}}_{\mathrm{max}} = 8.03$ in the vector and $\nu^{\mathrm{PS}}_{\mathrm{max}} = 7.29$ in the pseudoscalar channel. We keep these values fixed when going to finite temperature. In fact, in the temperature region around $T_c$, the relative contribution of the dimension 6 term at $\nu^{\mathrm{V},\mathrm{PS}}_{\mathrm{max}}$ is even smaller, namely, around 10% or less. The lower boundary of $\nu$ is chosen to be $\nu^{\mathrm{V},\mathrm{PS}}_{\mathrm{min}} = 0.78$, corresponding to a Borel mass of $M = 3.0$ GeV. We have checked that the obtained spectral functions do not depend much on this choice. For the value of the charm quark mass $m_c$, we use a recent estimate giving $\overline{m}_c(m_c) = 1.277 \pm 0.026$ GeV [@Dehnadi], for $\alpha_s$ we employ the newest world average $\alpha_s(M_Z) = 0.1184 \pm 0.0007$ [@Bethke], while for the vacuum gluon condensate $G^{\mathrm{vac}}_0$ the standard value $G^{\mathrm{vac}}_0 = 0.012 \pm 0.0036$ $\mathrm{GeV}^4$ [@Shifman; @Colangelo] is applied. For the default model $m(\omega)$, we use a constant matched to the perturbative value of the spectral function at high energy, as was done in similar studies using lattice QCD [@Asakawa]. The resulting spectral functions are given on the left side of Fig. \[fig:zerofinite\]. We observe in both channels a clear ground state peak, corresponding to $\eta_c$ and $J/\psi$. The spectral functions also exhibit a second peak, which is, however, not statistically significant. These second peaks most likely reflect the existence of several excited states, which the MEM analysis is not able to resolve, quite similar to the situation encountered in lattice studies. Furthermore, it is seen that the spectral function of the vector channel approaches the default model in the region close to $\omega = 0$, which, however, should not be confused with a contribution of the scattering term. This behavior is an artifact caused by our usage of the derivative sum rule in this channel and should thus not be considered to be a physical effect. Numerically, the peak representing $\eta_c$ lies at $3.02$ GeV, while the one standing for $J/\psi$ is found at $3.06$ GeV. Thus, we see that the ground state in both channels reproduces the experimental value with a precision of the order of $50$ MeV. In the vector channel, the residue can be related to the electronic width of the corresponding resonance. We can obtain this residue from Fig. \[fig:zerofinite\] simply by integrating the spectral function in the region of the peak, which gives 0.162 $\mathrm{GeV}^2$, which is in good agreement with the experimental value of 0.173 $\mathrm{GeV}^2$. On the other hand, we observe that the hyperfine splitting between $\eta_c$ and $J/\psi$ is underestimated. All these findings are in qualitative agreement with the results obtained in the conventional analysis of the charmonium sum rules [@Marrow]. Next, we increase the temperature according to Eq.(\[eq:OPE\]). The resulting spectral functions are shown on the right side of Fig. \[fig:zerofinite\] at temperatures between $0.9$ $T_c$ and $1.2$ $T_c$. It is seen in the figure that the behavior of the spectral functions changes abruptly in the vicinity of $T_c$. First, both ground state peaks experience a shift to lower energies of the order of 50 MeV, before dissolving quickly into the continuum above the critical temperature. Investigating the spectral functions in more detail, one observes that $\eta_c$ disappears already at $T=1.0$ $T_c$, while $J/\psi$ survives a bit longer, but also appears to be melted to a large degree at $T=1.1$ $T_c$. This sudden qualitative change of the spectral function mainly originates from the changes of the third and fourth terms in Eq.(\[eq:OPE\]), which can be traced back to the rapid adjustment of the thermodynamic quantities $\epsilon(T)$ and $p(T)$ around $T_c$. It is reassuring to note that our results are consistent with the findings of [@Morita1; @Morita3] in the sense that both observe a negative energy shift of the peaks around $T_c$. In these earlier works, it was, however, not possible to discuss the possible melting of the peaks because a relativistic Breit-Wigner form for the spectral function was assumed at all investigated temperatures. For obtaining firm conclusions, one has to test the reliability of the MEM procedure at finite temperature, where systematic effects decrease the reproducibility and resolution of the spectral function obtained from MEM. In lattice studies, this reduced reliability is primarily caused by the reduction of the data points in the imaginary time correlator, due to periodicity and the reduction of the maximal time extent. In the case of QCD sum rules, this problem does not exist, as Eq.(\[eq:OPE\]) is given as a continuous function at any temperature and therefore the same number of data points can be used. Nevertheless, the reliability of the MEM technique is still reduced at finite temperature due to the uncertainties of the thermodynamic functions $\epsilon(T)$ and $p(T)$, whose contribution grows with temperature and therefore increases the error of Eq.(\[eq:OPE\]). In order to confirm that the change of the spectral function in Fig. \[fig:zerofinite\] is not an artifact, we reanalyze Eq.(\[eq:OPE\]) at $T=0$, but use the errors of $T\neq0$ in the analysis. The results are then compared to the ones given in Fig. \[fig:zerofinite\], to investigate the net temperature effect on the spectral function. We find from this analysis that while the height of the peaks of the spectral functions at $T=0$ is indeed reduced because of the increased error, this effect is much smaller than the actual reduction of the peaks around $T_c$, seen on the right side of Fig. \[fig:zerofinite\]. We therefore conclude that the disappearance of the peaks observed in Fig. \[fig:zerofinite\] is a physical effect and is not induced by an artifact of the MEM analysis. In summary, we have extracted the spectral functions of the pseudoscalar and vector channel at both zero and finite temperature using a combined analysis of QCD sum rules and MEM. At $T=0$, the MEM technique is able to clearly resolve the lowest energy peaks, corresponding to the $\eta_c$ and $J/\psi$ resonances. The positions of both peaks agree with the experimental values with a precision of about 50 MeV. At finite temperature, we find that $\eta_c$ and $J/\psi$ melt quickly after the temperature is raised above the deconfinement temperature $T_c$, caused by the sudden change of the dimension-4, scalar and twist-2 gluon condensates in this temperature region. We have checked that this effect is not an artifact of the systematics of the MEM analysis. These results quantitatively disagree with the earlier findings of lattice studies which suggest that both $\eta_c$ and $J/\psi$ can survive at temperatures of up to 1.5 $T_c$ or higher. It, however, has to be mentioned that our results are in fact consistent with the latest lattice results [@Ding], finding the peaks of $\eta_c$ and $J/\psi$ to be largely distorted between $0.73$ $T_c$ and $1.46$ $T_c$. It remains to be seen whether or not the two methods will converge to compatible conclusions when more accurate analyses will become available in the future. This work was supported by KAKENHI under Contracts No. 22105503, No. 19540275 and by YIPQS at the Yukawa Institute for Theoretical Physics. P.G. gratefully acknowledges the support by the Japan Society for the Promotion of Science for Young Scientists (Contract No. 21.8079). K.M. thanks FIAS for support. [99]{} V. A. Novikov et al., Phys. Rept. 41, 1 (1978). T. Matsui and H. Satz, Phys. Lett. B 178, 416 (1986). M. Asakawa and T. Hatsuda, Phys. Rev. Lett. 92, 012001 (2004). S. Datta et al., Phys. Rev. D 69, 094507 (2004). T. Umeda , K. Nomura and H. Matsufuru, Eur. Phys. J. C 39, 9 (2004). A. Jacov$\acute{a}$c et al., Phys. Rev. D 75, 014506 (2007). M.A. Shifman, A.I. Vainshtein and V.I. Zakharov, Nucl. Phys. B147, 385 (1979); B147, 448 (1979). A.I. Bochkarev and M.E. Shaposhnikov, Nucl. Phys. B268, 220 (1986). T. Hatsuda, Y. Koike and S.H. Lee, Nucl. Phys. B394, 221 (1993). K. Morita and S.H. Lee, Phys. Rev. Lett. 100, 022301 (2008); K. Morita and S.H. Lee, Phys. Rev. C 77, 064904 (2008); Y. Song, S.H. Lee and K. Morita, Phys. Rev. C 79, 014907 (2009). K. Morita and S.H. Lee, Phys. Rev. D 82, 054008 (2010). P. Gubler and M. Oka, Prog. Theor. Phys. 124, 995 (2010), arXiv:1005.2459 \[hep-ph\]. J. Marrow, J. Parker and G. Shaw, Z. Phys. C 37, 103 (1987). K. Morita and S.H. Lee, arXiv:1012.3110 \[hep-ph\]. M. Asakawa, T. Hatsuda, and Y. Nakahara, Prog. Part. Nucl. Phys. 46, 459 (2001). R.K. Bryan, Eur. Biophys. J. 18, 165 (1990). G. Boyd et al., Nucl. Phys. B469, 419 (1996). O. Kaczmarek et al., Phys. Rev. D 70, 074505 (2004). T. Schafer and E.V. Shuryak, Rev. Mod. Phys. 70, 323 (1998). M.-C. Chu and S. Schramm, Phys. Rev. D 51, 4580 (1995). B. Dehnadi et al., arXiv:1102.2264 \[hep-ph\]. S. Bethke, Eur. Phys. J. C 64, 689 (2009). P. Colangelo and A. Khodjamirian, “At the Frontier of Particle Physics/Handbook of QCD" (World Scientific, Singapore, 2001), Volume 3, 1495. H.-T. Ding et al., PoS LAT2010, 180 (2010).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We find the explicit coordinate transformation which links two exact cosmological solutions of the brane world which have been recently discovered. This means that both solutions are exactly the same with each other. One of two solutions is described by the motion of a domain wall in the well-known 5-dimensional Schwarzshild-AdS spacetime. Hence, we can easily understand the region covered by the coordinate used by another solution.' address: - \| Canadian Institute for Theoretical Astrophysics, University of Toronto\ Toronto, ON, M5S 3H8\ and\ Department of Physics and Astronomy, University of Victoria\ Victoria, BC, Canada V8W 3P6 - \| DAMTP, University of Cambridge\ Silver Street, Cambridge CB3 9EW, United Kingdom\ Department of Physics, The University of Tokyo, Tokyo 113-0033, Japan\ and\ Research Centre for the Early Universe(RESCEU),\ The University of Tokyo, Tokyo 113-0033, Japan - ' Department of Physics, Waseda University, Shinjuku, Tokyo 169-8555, Japan ' author: - Shinji Mukohyama - Tetsuya Shiromizu - 'Kei-ichi Maeda' title: Global structure of exact cosmological solutions in the brane world --- The Randall-Sundrum brane world[@RS1; @RS2] may give drastic changes to the conventional gravity theory or cosmology [^1]. Since the 3-brane world is motivated by the relation between the $11$-dimensional supergravity and $E_8 \times E_8$ superstring theory [@Horava], the drastic changes might be realistic. Hence we have to seriously think of this scenario. In this respect, the conventional 4-dimensional theory should be recovered in low energy limits. This recovery is easily confirmed on a brane with positive tension [@Tess; @Tama]. Hence for simplicity, hereafter, we will consider the single brane case with the positive tension. For the brane world cosmology with negative tension, the issue is still in debate [@Tess; @Tama; @Csaki1; @Radion; @Radion2; @Horowitz], because the fine-tuning and radius stabilization problem seem to be necessary to recover the conventional Einstein theory and the single brane is not apparently acceptable. Recently several authors have found exact cosmological solutions in the brane world [@Kraus; @Ida; @BDEL; @Mukohyama; @Vollic] [^2]. In one of them, the 3-brane is described as a ‘domain wall’ moving in 5-dimensional black-hole geometries [@Kraus; @Ida]. It may be worth noting that the radiation dominated Friedmann universe is also expected by the AdS/CFT correspondence [@Gubser; @ADSCFT]. Another one was given by exactly solving the Einstein equations in the Gaussian normal coordinate [@BDEL; @Mukohyama; @Vollic]. These global solutions will be important to discuss its stability as the full spacetime, the gravitational force between two bodies on the brane and the cosmological perturbation on the brane. Since each of the two solutions is general enough in each coordinate system, it is easily expected that both solutions represent the same spacetime in different coordinate systems. In this brief note, we give the explicit coordinate transformation from the coordinate used in Refs. [@Kraus; @Ida] to the Gaussian normal coordinate adopted in Refs. [@BDEL; @Mukohyama; @Vollic]. Moreover, we identify the region which the Gaussian normal coordinate covers. As stated above, some authors [@Kraus; @Ida] considered a domain-wall moving in the 5-dimensional ‘Schwarzshild-AdS’(Sch-AdS) spacetimes [@TBH] with the metric $$ds^2 = g_{\mu\nu}dx^{\mu}dx^{\nu} = -f(r) dT^2 + \frac{dr^2}{f(r)} + r^2 d\Sigma^2_K, \label{eqn:metric-Sch-AdS}$$ where $d\Sigma^2_K$ is a metric of a unit three-dimensional sphere, plane or hyperboloid for $K=+1,0$ or $-1$, respectively, and $$f(r) = K + \frac{r^2}{l^2} - \frac{\mu}{r^2}.$$ Here, $l$ ($>0$) and $\mu$ ($\ge 0$ for $K=+1$ or $K=0$, $\ge -l^2/4$ for $K=-1$) are constants. The constant $l$ gives the curvature scale of the bulk spacetime. Hereafter, we denote coordinates on the three-dimensional manifold (sphere, plane or hyperboloid) by $x^i$. Therein the domain wall is the 3-brane, whose orbit is given by $r=R(T)$, and the bulk spacetime is the 5-dimensional Sch-AdS spacetime. Assuming the $Z_2$-symmetry which is inspired by the reduction from M-theory to $E_8 \times E_8$ heterotic string theory [@Horava], Israel’s junction condition [@israel] gives the Friedmann equation on the brane: $$\Bigl( \frac{{\dot a}}{a}\Bigr)^2 = \frac{8\pi G_N}{3}\rho -\frac{K}{a^2}+ \frac{\Lambda_4}{3}+\frac{\kappa_5^4}{36}\rho^2+\frac{\mu}{a^4}, \label{eq:fried}$$ where the function $R(T)$ representing the orbit is reinterpreted as a scale factor $a(\tau)$ with the proper time $\tau$, and $\rho$ is energy density of matter on the domain wall. Here, $\kappa_5$ is the $5$-dimensional gravitational constant, $G_N=\kappa_5^4 \lambda /48\pi$, $\Lambda_4=\kappa_5^4\lambda^2/12-3l^{-2}$ and $\lambda$ is the vacuum energy on the brane. The deviation from the conventional Friedmann equation is expressed as the forth and fifth terms in the right-hand side of Eq. (\[eq:fried\]). The fifth term comes from the ‘electric’ part of the 5-dimensional Weyl tensor[@Tess]. The form of Eq. (\[eqn:metric-Sch-AdS\]) seems simple enough to handle while a metric derived by several authors [@BDEL; @Mukohyama; @Vollic] looks rather complicated. However, as already explained above, it is expected that both of metrics express the same spacetime. Yet, the Weyl tensor for the metric derived in Refs. [@BDEL; @Mukohyama; @Vollic] becomes zero in the limit of the infinite value of the fifth coordinate (or the affine parameter in the Gaussian normal coordinate). This implies that the limit of the infinite affine parameter does not correspond to the Cauchy horizon since the Weyl tensor cannot be zero except for the conformal infinity as long as the bulk spacetime is not exactly the anti-deSitter spacetime. In order to show these explicitly, we transform coordinates in the metric (\[eqn:metric-Sch-AdS\]) so that the transformed coordinate system becomes a Gaussian normal coordinate system based on a hypersurface given by $r=R(T)$. For the purpose of the coordinate transformation, let us consider geodesics which intersect with the hypersurface $r=R(T)$ perpendicularly. Note that these geodesics are spacelike and have zero $x^i$-components, provided that the hypersurface $r=R(T)$ is timelike. First, because of the existence of the Killing field $(\partial/\partial T)^{\mu}$ in the bulk spacetime, the unit tangent vector $u^{\mu}$ of a geodesic should satisfy $$\begin{aligned} g_{\mu\nu}u^{\mu} (\partial/\partial T)^{\nu} & = & -E,\nonumber\\ g_{\mu\nu}u^{\mu}u^{\nu} & = & 1,\end{aligned}$$ where $E$ is an integration constant. Hence, we obtain $u^{\mu}$ as $$u^{\mu}\partial_{\mu} = \frac{E}{f(r)}\partial_T \pm\sqrt{f(r)+E^2}\partial_r.$$ In the above, we assumed $u^i=0$ because we are interested only in geodesics whose tangent have zero $x^i$-components. The trajectory of the geodesic is given by $$\frac{dx^{\mu}}{d w} = u^{\mu}, \label{eqn:dxdw=u}$$ where $w$ is the affine parameter. For the case of $4\mu +l^2(E^2+K)^2>0$, the $r$-component can be integrated as $$2 r^2 + l^2(E^2+K) = \sqrt{4l^2\mu+l^4(E^2+K)^2} \cosh{[2l^{-1}(\pm w+w_0)]}, \label{eqn:r^2-integral}$$ where $w_0$ is a constant. For the cases of $4\mu +l^2(E^2+K)^2=0$ and $4\mu +l^2(E^2+K)^2<0$, the $r$-component of Eq. (\[eqn:dxdw=u\]) are integrated to give different expressions of $r$ in terms of $w$. However, the final form of the metric we shall obtain below is common for all cases. Hence, in the following arguments we show explicit calculation for the first case only. Let us determine the constants $E$ and $w_0$ so that the geodesic intersects with the hypersurface $r=R(T)$ perpendicularly at $T=T_0$ and that the affine parameter $w$ is zero on the hypersurface: $$\begin{aligned} u^{\mu} & \propto & g^{\mu\nu}\partial_{\nu}(r-R(T))\qquad \mbox{at}\quad T=T_0,r=R(T_0),\nonumber\\ 2 R^2(T_0) + l^2(E^2+K) & = &\sqrt{4l^2\mu+l^4(E^2+K)^2} \cosh{(2l^{-1}w_0)},\end{aligned}$$ These can be solved to give $$\begin{aligned} E & = & E(T_0) \equiv \pm R'(T_0) \sqrt{\frac{f(R(T_0))}{f^2(R(T_0))-{R'}^2(T_0))}}, \label{eqn:value-E}\\ w_0 & = & w_0(T_0)\equiv \frac{l}{2}\cosh^{-1}\left[ \frac{2R^2(T_0)+l^2(E^2(T_0)+K)}{\sqrt{4l^2\mu +l^4(E^2+K)^2}} \right], \label{eqn:value-w0}\end{aligned}$$ where we have taken a convention such that $\cosh^{-1}X>0$ for $X>1$. Note that the set $(T_0,w,x^i)$ can be considered as a coordinate system. Next, we can introduce a new coordinate $t$ so that $t$ becomes a proper time on the hypersurface $r=R(T)$: $$\begin{aligned} \left(\frac{\partial t}{\partial T_0}\right)_{w,x^i} & = & \frac{f(a(t))}{\sqrt{f(a(t))+\dot{a}^2(t)}},\\ \left(\frac{\partial t}{\partial w}\right)_{T_0,x^i} & = & \left(\frac{\partial t}{\partial x^i}\right)_{T_0,w} = 0,\end{aligned}$$ where $a(t)=R(T_0(t))$. Note that the function $a(t)$ is well-defined in a whole coordinate patch as well as on the hypersurface $r=r(T)$, or $w=0$, since both $t$ and $T_0$ are constant along the geodesic. In this coordinate, $E(T_0)$ given by Eq. (\[eqn:value-E\]) has a simple expression $$E(T_0) = \pm\dot{a}(t),\label{eqn:value-E2}$$ where a dot denotes $(\partial/\partial t)_{w,x^i}$. Hence, by substituting Eqs. (\[eqn:value-E2\]) and (\[eqn:value-w0\]) into Eq. (\[eqn:r\^2-integral\]), we can rewrite the original coordinate $r$ in terms of new coordinates $t$ and $w$ as $$\begin{aligned} r^2 = \varphi(t,w)a^2(t), \label{eqn:r^2}\end{aligned}$$ where $$\varphi(t,w) = \cosh{(2l^{-1}w)} + \frac{l^2}{2}(H^2+Ka^{-2}) (\cosh{(2l^{-1}w)} - 1) \pm \sqrt{1 + l^2(H^2+Ka^{-2}-\mu a^{-4})}\sinh{(2l^{-1}w)}. \label{eqn:varphi}$$ Here $H(t)$ is defined by $$H(t) = \frac{\dot{a}(t)}{a(t)}.$$ Now let us confirm that the new coordinate system $(t,w,x^i)$ is actually a Gaussian normal coordinate system. For this purpose, it is sufficient to show that $$dw = g_{\mu\nu}u^{\mu}dx^{\nu},$$ or $$\begin{aligned} \left(\frac{\partial w}{\partial T}\right)_{r,x^i} & = & \mp\dot{a}(t),\nonumber\\ \left(\frac{\partial w}{\partial r}\right)_{T,x^i} & = & \pm\frac{\sqrt{f(r)+\dot{a}^2(t)}}{f(r)}. \label{eqn:dw},\nonumber\\ \left(\frac{\partial w}{\partial x^i}\right)_{T,r} & = & 0\end{aligned}$$ is integrable. The integrability condition $ddw=0$ is equivalent to $$\left(\frac{\partial t}{\partial T}\right)_{r,x^i}/ \left(\frac{\partial t}{\partial r}\right)_{T,x^i} = -\frac{f(r)\sqrt{f(r)+\dot{a}^2}}{\dot{a}}.$$ This condition is easily confirmed by using the relation $$\left(\frac{\partial t}{\partial T}\right)_{r,x^i}/ \left(\frac{\partial t}{\partial r}\right)_{T,x^i} = -\left(\frac{\partial r}{\partial w}\right)_{t,x^i}/ \left(\frac{\partial T}{\partial w}\right)_{t,x^i},$$ and $$\begin{aligned} \left(\frac{\partial T}{\partial w}\right)_{t,x^i} & = & \left(\frac{\partial T}{\partial w}\right)_{T_0,x^i} = u^T = \pm\frac{\dot{a}(t)}{f(r)},\nonumber\\ \left(\frac{\partial r}{\partial w}\right)_{t,x^i} & = & \left(\frac{\partial r}{\partial w}\right)_{T_0,x^i} = u^r = \pm \sqrt{f(r)+\dot{a}^2}. \label{eqn:w-derivative}\end{aligned}$$ Finally, we can calculate $(\partial r/\partial t)_{w,x^i}$ from (\[eqn:r\^2\]) as well as $(\partial T/\partial w)_{t,x^i}$ and $(\partial r/\partial w)_{T,x^i}$ from (\[eqn:w-derivative\]). Moreover, from Eq. (\[eqn:dw\]), $$\left(\frac{\partial T}{\partial t}\right)_{w,x^i}/ \left(\frac{\partial r}{\partial t}\right)_{w,x^i} = - \left(\frac{\partial w}{\partial r}\right)_{T,x^i}/ \left(\frac{\partial w}{\partial T}\right)_{r,x^i} = \mp\frac{\sqrt{f(r)+\dot{a}^2}}{f(r)\dot{a}(t)}.$$ Thus, it is easily shown that $$ds^2 = -\frac{\psi^2(t,w)}{\varphi(t,w)}dt^2 + \varphi(t,w)a(t)^2d\Sigma^2_K + dw^2,\label{eqn:BDEL-metric}$$ where $$\begin{aligned} \psi(t,w) & \equiv & \varphi + \frac{1}{2H} \left(\frac{\partial\varphi}{\partial t}\right)_{w,x^i} \nonumber\\ & = & \cosh{(2l^{-1}w)} + \frac{l^2}{2}(H^2+\dot{H}) (\cosh{(2l^{-1}w)}-1) \pm \frac{1+\frac{l^2}{2}(2H^2+\dot{H}+Ka^{-2})} {\sqrt{1+l^2(H^2+Ka^{-2}-\mu a^{-4})}}\sinh{(2l^{-1}w)}.\end{aligned}$$ By defining $k$ and $C$ by $$\begin{aligned} k & = & 2l^{-1},\nonumber\\ C & = & -l^2\mu,\end{aligned}$$ the functions $\psi(t,w)$ and $\varphi(t,w)$ are rewritten as $$\begin{aligned} \psi(t,w) & = & \cosh{(kw)} + 2k^{-2}(H^2+\dot{H}) (\cosh{(kw)}-1) \pm \frac{1+2k^{-2}(2H^2+\dot{H}+Ka^{-2})} {\sqrt{1+4k^{-2}(H^2+Ka^{-2})+Ca^{-4})}}\sinh{(kw)},\nonumber\\ \varphi(t,w) & = & \cosh{(kw)} + 2k^{-2}(H^2+Ka^{-2}) (\cosh{(kw)} - 1) \pm \sqrt{1+4k^{-2}(H^2+Ka^{-2})+Ca^{-4})}\sinh{(kw)}. \label{eqn:psi-varphi}\end{aligned}$$ Using the Friedmann equation of Eq. (\[eq:fried\]) and setting $C={\cal C}\kappa_5^4 l^2$, we can see that the metric (\[eqn:BDEL-metric\]) with (\[eqn:psi-varphi\]) is equivalent to the expression obtained in Ref. [@BDEL]. As shown in Refs. [@BDEL; @Mukohyama; @Vollic] the lower signs should be taken in all equations if we glue two copies of the region $w\ge 0$ to obtain the brane-world with [positive]{} tension. Thus, hereafter, we take the lower signs. From now on we show where the Gaussian normal coordinate covers in the Sch-AdS spacetime. For simplicity, we will consider only $\Lambda_4=0$ cases. We concentrate on the $\mu\neq 0$ cases since the $\mu=0$ case is easier and can be understood in a similar way. Note that the metric (\[eqn:metric-Sch-AdS\]) has an event horizon at $r=r_h$, where $r_h$ is given by $$r_h^2 = \frac{l^2}{2}(\sqrt{K^2+4l^{-2}\mu}-K).$$ In the following arguments we shall show that the Gaussian normal coordinate system $(t,w,x^i)$ covers the region beyond the event horizon and the wormhole. First, let us consider the case when the condition $$(H^2a^2+K)^2 + 4l^{-2}\mu > 0$$ is satisfied. For $K=+1$, this condition is automatically satisfied since $\mu\ge 0$ for the bulk [black hole]{} spacetimes. This conditions is automatically satisfied also when $K=0$ and $Ha\ne 0$. In this case, it is easily shown that $$r^2 = \varphi(t,w)a(t)^2 \to \infty \qquad (w\to\infty)$$ and that $$r^2 = \varphi(t,w)a(t)^2 \ge r_{min}^2,\qquad (w\ge 0) \label{eqn:r-lowerlimit}$$ where $$r_{min}^2 = \frac{l^2}{2} \left[\sqrt{(H^2a^2+K)^2+4l^{-2}\mu}-(H^2a^2+K)\right].$$ Note that $r_{min}(t)$ has been determined by the variation of $w$ under a fixed $t$. The equality in Eq.(\[eqn:r-lowerlimit\]) holds at $w=w_{min}(t)$, where $w_{min}(t)$ ($>0$) is given by $$\cosh{(2l^{-1}w_{min}(t))} = \frac{2l^{-2}a^2+H^2a^2+K}{\sqrt{(H^2a^2+K)^2+4l^{-2}\mu}}.$$ We can easily show that $r_{min}\to 0$ and $w_{min}\to 0$ as $a\to 0$ by using the Friedmann equation on the brane. On the other hand, there is another minimum $r_{min}^(w)$ which is determined by the variation of $t$ under a fixed $w$. The orbit of $r_{min}^(w)$ is the same as $\psi (t,w)=0$ because $\psi$ is proportional to $\partial_t r^2$. These minimum will be important when we draw the conformal diagram. Now it is easily shown by using the lower-limit of $\mu$ ($\mu\ge 0$ for $K=+1$ and $K=0$, $\mu\ge -l^2/4$ for $K=-1$, and we are considering the $\mu\ne 0$ case) that $r_{min}^2\le r_h^2$. The equality holds if and only if $Ha=0$. Hence, the coordinate $(t,w,x^i)$ actually covers the region beyond the event horizon and the wormhole. Next, let us consider the case when $K=-1$ and the condition $$(H^2a^2-1)^2 + 4l^{-2}\mu = 0$$ is satisfied. In this case, $r^2$ approaches to a finite value in the limit $w\to\infty$: $$r^2 = \varphi(t,w)a(t)^2 \to \frac{l^2}{2}(1-H^2a^2) < \frac{l^2}{2} \qquad (w\to\infty).$$ On the other hand, $$r_h^2 > \frac{l^2}{2}.$$ Therefore, the coordinate $(t,w,x^i)$ reaches the region beyond the event horizon. Finally, let us consider the case of $K=-1$ and the time when condition $$(H^2a^2-1)^2 + 4l^{-2}\mu < 0$$ is satisfied. In this case, there exists a value of $w$ ($>0$) such that $r^2=0$. Note that $r=0$ corresponds to a singularity inside the horizon since tetrad components of the Weyl tensor for the metric (\[eqn:metric-Sch-AdS\]) diverge at $r=0$. Thus, the coordinate $(t,w,x^i)$ covers the region inside the event horizon and reaches the singularity inside the horizon. Now it is easy to see the global structure. Figure \[fig:Fig1\] is for $\mu>0$, $K=+1$. Figure \[fig:Fig2\] is for $\mu>0$, $K=0,-1$. Figures \[fig:Fig3\] and \[fig:Fig4\] are for $\mu=0$, $K=+1$ and for $\mu=0$, $K=0,-1$, respectively. Figure \[fig:Fig5\] is for $\mu<0$, $K=-1$. In these figures, the region covered by the Gaussian normal coordinate is expressed as the shaded region. Note that in Figures \[fig:Fig1\], \[fig:Fig2\] and \[fig:Fig5\], for a non-zero small value of $w$, $\partial_t$ should be past directed for small $t$, turns to the opposite direction at a time $t_$ when the orbit of $\partial_t$ reaches the $r_{min}^(w)$, and future directed for large $t$. To obtain these figures, we have used the following two facts. First, the hypersurface $t=t_0$ is always spacelike, and should come in contact with the hypersurface $r=r_{min}(t_0)$ at $w=w_{min}(t_0)$, where $t_0$ is a constant. Secondly, for a fixed $w$, $r\to\infty$ in the limits $a\to 0$ and $a\to\infty$, providing some reasonable assumptions (eg. $\Lambda_4=0$ and $\rho a^2\to 0$ in the limit $a\to\infty$). This means that the constant-$t$ hypersurface should become null in these limits. In this brief note, we have given the coordinate transformation between the metric of Eq. (\[eqn:metric-Sch-AdS\]) and of Eq. (\[eqn:BDEL-metric\]). As a result, we could see the region where the Gaussian normal coordinate covers. For general cases of single brane with the positive tension and $\mu \neq 0$, the coordinate $w$ labelling the extra dimension does not terminate at the ‘Cauchy horizon’ and goes beyond the event horizon. For some cases, the coordinate goes through the wormhole and reaches another domain of communication where we can define the total energy well [@AD]. The energy should be $\mu$ for cases which we considered and we can prove the positivity on the slice without naked singularities at least $K=0$ cases [@Bogo; @Wool]. This might argue us to consider general issues such as the positivity of the ADM energy [@PET], cosmic no-hair [@Wald], singularity theorem [@GR] and so on. According to Ref. [@Tess], the effective 4-dimensional Einstein equation is given by $${}^{(4)}G_{\mu\nu}= -\Lambda_4 q_{\mu\nu}+8\pi G_N T_{\mu\nu} +\kappa_5^4 \Pi_{\mu\nu}-E_{\mu\nu}, \label{eq:effective}$$ where $T_{\mu\nu}$ is the energy-momentum tensor on the brane, $\Pi_{\mu\nu}$ is a quadratic term of $T_{\mu\nu}$, and $E_{\mu\nu}={}^{(5)}C_{\mu\alpha\nu\beta}n^\alpha n^\beta$ is the ‘electric’ part of the 5-dimensional Weyl tensor. It is important to clarify whether the right-hand side of Eq. (\[eq:effective\]) satisfies the local energy condition [@GR] or not, because it is closely related to many interesting subjects such as stability of the Minkowski spacetime. For the Sch-AdS spacetime, $E_{00}\sim -\mu/a^4$ and then $-E_{00} \geq 0$ is guaranteed by the above argument. This indicates the tendency such that the right-hand side satisfies the local energy condition. However, we know that the perturbation analysis on the Randall & Sundrum brane world implies $E_{00} \geq 0$ [@Tess2]. Hence, we cannot quickly give the definite answer on the signature. This is crucial problem for general issues. SM would like to thank Professor W. Israel for continuing encouragement and Professor L. Kofman for his warm hospitality in Canadian Institute for Theoretical Astrophysics. TS would like to thank D. Langlois, M. Sasaki and T. Tanaka for discussions. SM’s work is supported by the CITA National Fellowship and the NSERC operating research grant. TS’s work is supported by JSPS Postdoctal Fellowship for Research Abroad. KM’s work is supported in part by Monbusho Grant-in Aid for Specially Promoted Research No. 08102010. L. Randall and R. Sundrum, Phys. Rev. Lett. [83]{}, 3370 (1999) \[hep-ph/9905221\]. L. Randall and R. Sundrum, Phys.Rev.Lett. [83]{}, 4690 (1999) \[hep-th/9906064\]. V. A. Rubakov and M. E. Shaposhnikov, Phys. Lett. [152B]{}, 136 (1983);\ M. Visser, Phys. Lett. [B159]{}, 22 (1985);\ K. Akama, Prog. Theor. Phys. [78]{}, 184 (1987). P. Horava and E. Witten, Nucl. Phys. [460]{}, 506(1996). W. Isarel, Nuovo Cim. [44B]{}, 1 (1966). T. Shiromizu, K. Maeda and M. Sasaki, gr-qc/9910076. J. Garriga and T. Tanaka, hep-th/9911055. C. Csaki, M. Graesser, C. Kold and J. Terning, hep-ph/9906513. C. Csaki, M. Graesser, L. Randal and J. Terning, hep-ph/9911406. P. Kanti, I. I. Kogan, K. A. Olive and M. Pospelov, hep-ph/9912266. R. Emparan, G. T. Horowitz and R. C. Myers, hep-th/9911043. P. Kraus, JHEP [9912]{}, 011 (1999) \[hep-th/9910149\]. D. Ida, gr-qc/9912002. P. Binétruy, C. Deffayet, U. Ellwanger and D. Langlois, hep-th/9910219. S. Mukohyama, hep-th/9911165, to appear in Phys. Lett. [B]{}. D. N. Vollick, hep-th/9911181. N. Kaloper, hep-th/9905210. T. Nihei, hep-ph/9905487. H. B. Kim and H. D. Kim, hep-th/9909053. P. Binetruy, C. Deffayet and D. Langlois, hep-th/9905012. E. E. Flanagan, S. H. H. Tye, I. Wasserman, hep-ph/9910498. J. M. Cline, C. Grojean and G. Servant, hep-ph/9906523. C. Csaki, M. Graesser, L. Randall and J. Terning, hep-ph/9911406. P. Kanti, I. I. Kogan, K. A. Olive and M. Prospelov, hep-ph/9909481. J. Garriga and M. Sasaki, hep-th/9912118. S. S. Gubser, hep-th/9912001. For AdS/CFT correspondece,\ J. Maldacena, Adv. Theor. Math. Phys., [2]{}, 231(1998);\ O. Aharony, S. S. Gubser, J. Maldacena, H. Ooguri and Y. Oz, hep-th/9905111 and reference therein D. Birmingham, Class. Quant. Grav. [16]{}, 1197(1999). L. F. Abbot and S. Deser, Nucl. Phys. [195]{}, 76(1982);\ A. Ashtekar and A. Magnon, Class. Quantum Grav. [1]{}, L39(1984);\ A. Ashtekar and S. Das, hep-th/9911230. G. W. Gibbons, S. W. Hawking, G. T. Horowitz and M. J. Perry, Commun. Math. Phys. [88]{}, 295(1983). E. Woolgar, Class. Quantum Grav. [11]{},1881(1994). R. Schoen and S. T. Yau, Commun. Math. Phys. [79]{}, 23(1981)\ E. Witten, Commun. Math. Phys. [80]{}, 381(1981). R. M. Wald, Phys. Rev. [D28]{}, 2118(1983). For example, R. M. Wald, [General Relativity]{}, (Univ. Chicago Press, Chicago, 1984). M. Sasaki, T. Shiromizu and K. Maeda, hep-th/9912233. [^1]: For earlier works on related scenarios, see Refs. [@earlier-works]. [^2]: In this field, a lot of works related have been done so far [@Csaki1; @Kaloper; @Nihei; @Kim; @BDL; @Flanagan; @Cline; @Csaki2; @Olive; @Misao].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We derive a bound on the magnetic dipole moment of the top quark in the context of the effective Lagrangian approach by using the ratios $R_b=\Gamma_b/\Gamma_{h}$, $R_l=\Gamma_{h}/\Gamma_l$ and the $Z$ width. We take into account the vertex and oblique corrections.' author: - \| R. Martínez$^1$ and J.-Alexis Rodríguez$^{1,2}$\ 1. Depto. de Física, Universidad Nacional, Bogotá, Colombia\ 2. Centro Internacional de Física, Bogotá, Colombia title: Constraint on the magnetic moment of the top quark --- The most recent analyses of precision measurements at the Large Electron Positron (LEP) collider lead to the conclusion that the predictions of the Standard Model (SM) of electroweak interactions, based on the gauge group $SU(2)_L\otimes U(1)_Y$ are in excellent agreement with the experimental results. Recently the discovery of the top quark has been announced by the Collider Detector at Fermilab (CDF) and D0 collaborations[@cdf]. The direct measurement of the top quark mass $m_t$ is in agreement with the indirect estimates derived by confronting the SM $m_t$ dependent higher order corrections with the LEP and other experimental results. The measurement of the top quark mass reduced the number of free parameters of the SM. A precise knowledge of the value of the top mass will improve the sensitivity of searches of new physics through small indirect effects. The precise measurements of the $g-2$ value of the electron provides a test of its point-like character. Similarly, measurements of the electric- and chromo- magnetic moments of the quarks can be important to study physics beyond the SM. In particular, the chromomagnetic moment of the top quark can affect its production in the $p\overline{p}$ and $e^{+}e^{-}$ reactions[@pro]. The SM predicts how the top quark should behave under these interactions, so any deviation from this behaviour would provide us with a probe of new physics beyond the SM. If new physics is found in this sector, it could probably originates from a non standard symmetry breaking mechanism. This is because the top mass is of the order of the electroweak (EW) breaking scale, and hence it is conceivable that the top-quark properties are sensitive to unsuppressed EW breaking effects[@peccei]. The aim of the present work is to extract indirect information on the magnetic dipole moment of the top quark from LEP data, specifically we use the ratios $R_b$ and $R_l$ defined by $$\begin{aligned} R_b&=&\frac{\Gamma (Z\rightarrow b\bar{b})}{\Gamma (Z\rightarrow hadron)} \; , \nonumber \\ R_l&=&\frac{\Gamma (Z\rightarrow hadron)}{\Gamma (Z\rightarrow l\bar{l})} \label{rbl} \end{aligned}$$ and the $Z$ width, in the context of an effective Lagrangian approach. The oblique and QCD corrections to the $b$ quark and hadronic $Z$ decay widths cancel off in the ratio $R_b$. This property makes $R_b$ very sensitive to direct corrections to the $Zb\overline{b}$ vertex, specially those involving the heavy top quark [@nardi], while $\Gamma_Z$ and $R_l$ are more sensitive to oblique corrections. The effective Lagrangian approach is a convenient model independent parametrization of the low-energy effects of the new physics that may show up at high energies[@wudka]. Effective Lagrangians, employed to study processes at a typical energy scale $E$ can be written as a power series in $1/\Lambda$, where the scale $\Lambda$ is associated with the heavy particles masses of the underlying theory[@burgues]. The coefficients of the different terms in the effective Lagrangian arise from integrating out the heavy degrees of freedom that are characteristic of a particular model for new physics. In order to define an effective Lagrangian it is necessary to specify the symmetry and the particle content of the low-energy theory. In our case, we require the effective Lagrangian to be $CP$-conserving, invariant under SM symmetry $SU(2)_L\otimes U(1)_Y$, and to have as fundamental fields the same ones appearing in the SM spectrum. Therefore we consider a Lagrangian in the form $${\cal L}_{eff}={\cal L}_{SM}+\sum_{n}\alpha_n {\cal O}^n$$ where the operators ${\cal O}^n$ are of dimension greater than four. In the present work, we consider the following dimension six and $CP$-conserving operators, $$\begin{aligned} O_{uW}^{ab} &=&\bar{Q}_L^a\sigma ^{\mu \nu }W_{\mu \nu }^i\tau ^i\tilde{\phi}% U_R^b \; ,\nonumber \\ O_{uB}^{ab} &=&\bar{Q}_L^a\sigma ^{\mu \nu }YB_{\mu \nu }\tilde{\phi}U_R^b \; , \label{operador}\end{aligned}$$ where $Q_L^a$ is the quark isodoublet, $U_R^b$ is the up quark isosinglet, $a$ , $b$ are the family indices, $B_{\mu \nu }$ and $W_{\mu \nu }$ are the $U(1)_Y$ and $SU(2)_L$ field strengths, respectively, and $\tilde{\phi}=i\tau_2\phi^$. We use the notation introduced by Buchmüller and Wyler [@wyler]. In the case of the operators $O_{uB}^{ab}$ and $O_{uW}^{ab}$, some degrees of family mixing is made explicit (corresponding to $a\neq b$) without breaking SM gauge invariance. After spontaneous symmetry breaking, these fermionic operators generate also effective vertices proportional to the anomalous magnetic moments of quarks. The above operators for the third family give rise to the anomalous $t\bar{t}\gamma $ vertex and the unknown coefficients $\epsilon _{uB^{}}^{ab}$ and $\epsilon_{uW}^{ab}$ are related respectively with the anomalous magnetic moment of the top quark through $$\delta \kappa _t=-\sqrt{2}{\frac{m_t}{m_W}}\frac g{eQ_t}(s_W\epsilon _{uW}^{33}+c_W\epsilon _{uB}^{33})\; .$$ where $s_W$ denotes the sine of the weak mixing angle. The expression for $R_b$ is given by $$R_b=R_b^{SM}(1+(1-R_b^{SM})\delta _b) \; ,$$ where $R_b^{SM}$ is the value predicted by the SM and $\delta _b$ is the factor which contains the new physics contribution, and it is defined as follows $$\delta _b=2{\frac{{g_V^{SM}g_V^{NP}+g_A^{SM}g_A^{NP}}}{{(g_V^{SM})^2+(g_A^{SM})^2}}}$$ and $g_{V}^{SM}$ and $g_A^{SM}$ are the vector and axial vector couplings of the $Zb\bar{b}$ vertex normalized as $g_V^{SM}=-1/2+2 s_W^2/3$ and $g_A^{SM}=-1/2$. The contributions from new physics, eq. (\[operador\]), to $R_l$ and $\Gamma_Z$ are of two classes. One from vertex correction to $Zb\bar{b}$ in the $\Gamma_{hadr}$ and the other from the oblique correction through $\Delta\kappa$ in the $\sin^2\theta_W$. These can be written as $$\begin{aligned} R_l&=&R_l^{SM} (1-0.1851\; \Delta\kappa +0.2157 \;\delta_b) \; ,\nonumber \\ \Gamma_Z&=&\Gamma_Z^{SM} (1-0.2351\;\Delta\kappa+0.1506 \;\delta_b)\end{aligned}$$ where $\Delta\rho$ is equal to zero for the operators that we are considering. The contribution of the above effective operators to the $Zb\overline{b}$ vertex is given by the Feynman diagrams shown in fig. 1, where a heavy dot denotes an effective vertex. After evaluating the Feynman diagrams, with insertions of the effective operators $O_{uB}^{ab}$ and $O_{uW}^{ab}$ we obtain $$\begin{aligned} g_V^{NP} &=&4 \sqrt{2} \epsilon _{uW}^{33} G_F m_W^3 m_t \big \{ 3 c_W (\tilde{C}_{12}- \tilde{C}_{11})-{\frac{m_t^2}{\sqrt{2} m_W^2}}(C_{12}-C_{11}+C_0) \nonumber \\ &&+{\frac{(1+a)}{8c_W}}(C_{11}+C_{12}+C_0)+\frac{1}{\sqrt{2}} (C_{12}-C_{11}-C_0) \nonumber \\ &&-\frac{3a}{4c_Wm_Z^2}(B_1-B_0)% \big \} \; ,\\ g_A^{NP} &=&4 \sqrt{2}\epsilon _{uW}^{33} G_f m_W^3 m_t \big \{ -{\frac a{2c_W}}% (C_0+C_{12}-C_{11})-\frac{1}{\sqrt{2}}(C_{12}-C_0-C_{11}) \nonumber \\ &&+\frac{m_t^2}{\sqrt{2} m_W^2}(C_{12}-C_{11}+C_0) -{\frac{2 m_ts_W^2}{m_W}}(\tilde{C}_0+\tilde{C}_{12}-% \tilde{C}_{11}) \nonumber \\ &&-\frac{3}{4c_Wm_Z2}(B_1-B_0)\big \} \; \end{aligned}$$ for the operator $O_{uW}$ and, $$\begin{aligned} g_V^{NP} &=&g_A^{NP}=\frac{4\sqrt{2}}{3}\epsilon _{uB}^{33} G_F m_W^3 m_t \frac{s_W}{c_W} [\frac{% m_t^2}{\sqrt{2}m_W^2}(C_{12}-C_{11}+C_0) \nonumber \\ &&-\frac{1}{\sqrt{2}}(-C_{11}+C_{12}-C_0)] \; ,\end{aligned}$$ for the operator $O_{uB}$. In the above equations $a=1-{\frac 83}s_W^2$ while $C_{ij}=C_{ij}(m_W,m_t,m_t)$, $\tilde{C}_{ij}=\tilde{C}_{ij}(m_t,m_W,m_W)$ and $B_i=B_i(0,m_t,m_W)$ are the Passarino-Veltman scalar integral functions [@veltman]. The combination $B_0-B_1$ has a pole in $d=4$ dimensions that is identified with the logarithmic dependence on the cutoff. Using the prescription given in ref. [@hariwara], the pole can be replaced by $\ln\Lambda^2/m_Z^2$. The operators (\[operador\]) contribute to the fermion processes at one loop level, giving oblique corrections to the gauge boson self energies. The contribution is essentially coming from the $\Sigma_{\gamma Z}(m_Z^2)$ self energy. Therefore these operators only contribute to $\Delta\kappa$ parameter [@oblicuas]. For $\Delta\kappa$ we have obtained the same results of the eqs. (50) and (51) of ref. [@renard]. To obtain the physical quantities $R_b$, $R_l$ and $\Gamma_Z$ as a function of $\delta\kappa_t$ instead of two parameters $\epsilon^{33}_{uW}$ and $\epsilon^{33}_{uB}$, we consider that only one coefficient at the time is different from zero at the scale $\Lambda$. Then we proceed to sum the value of the contribution of each operator in order to avoid cancellation between them. With this prescription we get an optimal bound. In Fig. 2 we display $R_b$ as a function of $\delta \kappa _t$. The horizontal lines represent the experimental measurements $R_b^{exp}=0.2178\pm 0.0011$ [@lep]; our bound can be expressed as $0.38\!\leq\! \delta\kappa_t\!\leq\! 1.21$. In Fig. 3 we show $R_l$ fraction versus $\delta\kappa_t$ with the horizontal lines representing the experimental result, $R_l^{exp}=20.778\pm 0.029$; for this case the bound can be expresses as $0.02\!\leq\!\delta\kappa_t\! \leq\! 0.48$. Finally in Fig. 4 we plot $\Gamma_Z$ versus $\delta\kappa_t$ with the experimental value $\Gamma_Z^{exp}=2.4946\pm 0.0027$ GeV; the limit is $0\!\leq\!\delta\kappa_t\!\leq \! 0.48$. The SM values for the parameters that we have used are $\Gamma_Z=2.4972$ GeV, $R_l=20.747$ GeV, $R_b=0.2157$, $\Gamma_{hadr}=1743.4$ MeV and $\Gamma_l=84.03$ MeV; with the input parameters: $m_t=175$ GeV, $\alpha_s(m_Z)=0.118$, $m_Z=91.1861$ GeV, $m_H=100$ GeV and $\Lambda=1$ TeV. In conclusion, the corrections through $R_l$ and $\Gamma_Z$ put a better bound on $\delta\kappa_t$ than the vertex correction from $R_b$. Our results from figs. 3 and 4 are of the same order of the results of the eqs. (56) and (57) of ref. [@renard], with the appropriate replacement between $f_{tW\phi}$ and $f_{tB\phi}$ and the magnetic dipole moment of the top quark. These bounds agrees also with the one obtained of the same effective operators in reference [@martinez], by using the CLEO result on $B(b\to s\gamma)$. We would like to thank M. Perez and E. Nardi for their comments. We thank COLCIENCIAS for financial support. [9]{} CDF Collaboration, F. Abe et. al., Phys. Rev. Lett. [74]{}, 2626 (1995); D0 Collaboration, S. Abachi et. al., Phys. Rev. Lett. [74]{}, 2632 (1995). G. Kane, G. Ladinsky, and C.P. Yuan, Phys. Rev. [D 45]{}, 1531 (1992); D. Atwood, A. Kagan, and T. Rizzo, Phys. Rev. [D 52]{}, 6264 (1995); T. Rizzo, Phys. Rev. [D 53]{}, 2326 (1996). R. D. Peccei and X. Zhang, Nucl. Phys. [B 337]{}, 269 (1990); Ehad Malkawi and C. -P. Yuan, Phys. Rev. [D 50]{}, 4462 (1994). S. Mrenna and C.- P. Yuan, Phys. Lett. [B 367]{}, 188 (1996); E. Ma, Phys. Rev. [D 53]{}, 2276 (1996); P. Bammert, et. al., Phys. Rev. [D 54]{}, 4275 (1996). H. Georgi, Nucl. Phys. [B 361]{}, 339 (1991); A. De Rújula et. al., Nucl. Phys. [B 369]{}, 3 (1992); J. Wudka, Int. J. Mod. Phys. [ A9]{}, 2301 (1994). C. P. Burgess and D. London, Phys. Rev. [D 48]{}, 4337 (1993); Phys. Rev. Lett. [69]{}, 3428 (1993). W. Buchmüller and D. Wyler, Nucl. Phys. [B 268]{}, 621 (1986); Phys. Lett. [B 197]{}, 379 (1987). G. Passarino and M. Veltman, Nucl. Phys. [B 160]{}, 151 (1979). K. Hagiwara et. al., Phys. Lett. [B 283]{}, 353 (1992); Phys. Rev. [D 48]{}, 2182 (1993). M. Peskin and T. Takeuchi, Phys. Rev. [D 46]{}, 381 (1992); V. A. Novikov, L. B. Okun and M. I. Vysotsky, Nucl. Phys. [B 397]{}, 35 (1993); P. Altarelli and F. Caravaglios, Nucl. Phys. [B 405]{}, 3 (1993); K. Hagiwara et. al., Z. Phys. [C 64]{}, 559 (1994). G. J. Gounaris, F. M. Renard and C. Verzegnassi, Phys. Rev. [D 52]{}, 451 (1995). G. Altarelli, hep-ph/9611239. R. Martinez, M. A. Perez and J. Toscano, Phys. Lett. [B 340*]{}, 91 (1994) Figure Captions Figure 1. Feynman diagrams contributing to the $Z\to b\bar{b}$ decay. The heavy dots denote an effective vertex. Figure 2. $R_b$ as a function of $\delta\kappa_t$ for $m_t=175$ GeV. The horizontal lines are the experimental results. Figure 3. Same as fig. 2 for $R_l$. Figure 4. $\Gamma_Z$ as a function of $\delta\kappa_t$ for $m_t=175$ GeV. The horizontal lines are the experimental results.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The triggering mechanism(s) and critical condition(s) of solar flares are still not completely clarified, although various studies have attempted to elucidate them. We have also proposed a theoretical flare-trigger model based on MHD simulations [@kusano12], in which two types of small-scale bipole field, the so-called Opposite Polarity (OP) and Reversed Shear (RS) types of field, can trigger flares. In this study, we evaluated the applicability of our flare-trigger model to observation of 32 flares that were observed by the Solar Dynamics Observatory (SDO), by focusing on geometrical structures. We classified the events into six types, including the OP and RS types, based on photospheric magnetic field configuration, presence of precursor brightenings, and shape of the initial flare ribbons. As a result, we found that approximately 30% of the flares were consistent with our flare-trigger model, and the number of RS type triggered flares is larger than that of the OP type. We found none of the sampled events contradicts our flare model, although we cannot clearly determine the trigger mechanism of 70% of the flares in this study. We carefully investigated the applicability of our flare-trigger model and the possibility that other models can explain the other 70% of the events. Consequently, we concluded that our flare-trigger model has certainly proposed important conditions for flare-triggering.' author: - Yumi Bamba - Kanya Kusano title: \| Evaluation of Applicability of a Flare Trigger Model\ based on Comparison of Geometric Structures --- Introduction {#sec:intro} ============ A solar flare is a sudden release of magnetic energy in the solar corona, and it is widely accepted that flare occurrence is related to topological changes of the magnetic field such as caused by magnetic reconnection (cf. CSHKP model, @carmichael64 [@sturrock66; @hirayama74; @kopppneuman76]). Two major physical models of solar eruptions including flares and coronal mass ejections (CMEs) are proposed so far: “ideal magnetohydrodynamics (MHD) models” and “magnetic reconnection models”. They emphasize different aspects of the mechanism of solar eruptions. The ideal MHD models point out that some ideal MHD instabilities cause the onset of solar eruptions. For instance, the torus instability [e.g. @bateman78; @kliemtorok06; @demoulin10; @kliem14] or helical kink instability [e.g. @gerrard01; @torok04; @fangibson03; @torokkliem05] have been proposed as the mechanism of solar eruptions. The torus instability is determined by the decay rate of the external poloidal field $B_{ex}$, and the critical condition is defined by the decay index ${n = -d({ln} B_{ex})/d({ln} R)}$. Recently, the critical range for onset of the torus instability is proposed as $n\sim1.3-1.5$ [@zuccarello15]. Some observational and computational studies using non-linear force-free field (NLFFF) [e.g. @cheng11; @kliem13; @savcheva15] modeling support the torus instability although the torus instability required an external agent that lifts the flux rope to an unstable height range. The helical kink instability model proposes that solar eruptions are caused by growth of the helical kink instability. The helical kink instability can grow when magnetic twist becomes strong enough, and it can grow even if a twisted flux rope does not reach the critical height of the torus instability. @hassanin16 showed a nice comparison between observations and an MHD simulation for a confined M-class flare. @liu16 also showed that there is consistency between the helical kink instability mechanisms and sequential confined flares from a combination of observations and NLFFF modeling. However, the helical kink instability requires higher twist in a flux rope and observations do not always show such a high twist. The ideal MHD models postulate that a flux rope exists prior to eruption as mentioned above. On the other hand, magnetic reconnection models note that some kind of magnetic reconnection causes the onset of a solar eruption. Two kinds of reconnection models have been proposed so far. One is the “magnetic breakout model” [@Antiochos99; @karpen12] and the other is the “tether-cutting model” [@moore01; @moore06]. The magnetic breakout model considers a multi-flux topology that consists of multiple flux systems with a coronal null point [e.g. @priest96; @sun12]. Sheared magnetic arcades can be created through many means such as slow photospheric shear. This leads to storage of free energy. Breakout reconnection can then remove the overlying magnetic field, and the core magnetic flux is free to erupt. @reva16 showed breakout reconnection associated with a CME using observations from TESIS EUV telescope. The tether-cutting model proposes that magnetic reconnection in the core of a sheared magnetic arcade creates a flux rope and causes the eruption. @liu13 showed evidence of tether-cutting reconnection using a NLFFF model using solar observational data of a flare productive AR. The breakout and tether-cutting models are clearly reviewed from an observational perspective in @schmieder15. There are various studies trying to find out what observable parameter(s) determine flare and solar eruption onset condition(s), while various models are proposed from computational studies and have been validated by observations, as mentioned above. One such model is flux cancellation [@vanBallegooijen89; @aulanier10], whose basic concept is that tether-cutting reconnection forms a flux rope and it partially cancels magnetic flux in the core of a sheared arcade causing the MHD instability to be triggered. @Green11 clearly showed flux cancelation and associated flux rope formation in the photoshere prior to an eruption. Another model is the emerging flux model proposed by @chenshibata00. They considered two cases based on the positional relationship between emerging flux and a flux rope. If emerging flux appears below a flux rope (i.e. on the polarity inversion line: PIL) and partial magnetic cancellation occurs, local magnetic pressure is decreased and the pressure gradient causes upflow that pushes the flux rope upward to erupt. On the other hand, if a flux rope appears beside the overlying field (i.e. away from the PIL), the overlying field that prohibits the flux rope from eruption is weakened by magnetic reconnection with the emerging flux, and the flux rope may erupt. Many other studies proposed various magnetic properties, that seem to relate to the onset of solar eruptions, such as length of highly sheared PILs [@hagyard84b], magnetic flux close to high gradient PILs [@schrijver07], free magnetic energy based on presence of strong gradient PILs [@falconer08], area and total magnetic flux in an active region (AR) [@higgins11], and total unsigned current helicity over an AR [@bobra15], etc. However, it is still unclear what triggers the MHD instability or magnetic reconnection that leads to an eruption. Although each model proposed a solar eruption scenario, it is likely that the eruption onset results from the feedback and interaction of different processes. For instance, @hagyard84a suggested that an overall instability driving energy release results from the positive feedback between reconnection and eruption of the sheared field. @kusano12 (henceforth “K12”) numerically demonstrated the feedback model, in which positive feedback between an ideal MHD instability and magnetic reconnection causes explosive growth of solar eruptions. Their basic idea is that “internal magnetic reconnection” between different spatial-scale magnetic fields (between a large-scale sheared field and a small-scale bipole field) forms a double-arc twisted loop and triggers reconnection among the large-scale sheared fields by destabilizing the double-arc twisted loop. In fact, @ishiguro17 recently showed that the double-arc twisted loop becomes unstable if the parameter $\kappa$, given by the product of magnetic twist and the fraction of reconnected flux, exceeds a threshold. Their model is consistent with the tether-cutting reconnection scenario and well explains the physical process from the theoretical point of view. K12 proposed that the feedback interaction can be triggered by reconnection between magnetic fields of different spatial-scale if two geometrical parameters; sheared angle $\theta_{0}$ of large-scale field and azimuth $\varphi_{e}$ of a small-scale bipole field, satisfy some condition. We briefly review the K12 model later in Section \[sec:KB12\] and Figure \[fig:setting\]. They surveyed combinations of $\theta_{0}$ and $\varphi_{e}$ and found that there are two specific cases; the Opposite Polarity (OP) and Reversed Shear (RS) fields, which can trigger a flare. The essence of the K12 model is that a positive feedback interaction of an ideal MHD instability and flare reconnection can be triggered by the two types of small-scale “trigger fields”. @bamba13 developed a way to measure the angles $\theta_{0}$ and $\varphi_{e}$ using the photospheric magnetic field and chromospheric brightenings observed by the Hinode satellite [@kosugi07]. They found that either the OP or RS type magnetic field condition was satisfied prior to several major flares. In this study, we aim to evaluate the consistency between the flare-trigger condition of K12 and observations, focusing on geometrical structures. For that purpose, we investigate more flares than @bamba13 since only a small number of flares were studied in their papers. Specifically, we determine what fraction of the flares are consistent with the OP or RS types and whether a flare occurred with some condition other than OP or RS field. We increased the number of events analyzed to 32. This was possible due to the large field-of-view (FOV) of the Solar Dynamics Observatory (SDO; @pesnell12). The paper is organized in the following manner. We first review the theoretical model of K12, which we focused on this study, and describe the data analysis methods and event selection criteria in Section \[sec:data\_analysis\]. Then we classified the events into six different types, including the OP and RS types, based on the analysis, and describe distinctive features for each type, in Section \[sec:results\]. In Section \[sec:discussion\], we discuss other possibilities than K12’s model for triggering some specific types of events which we classified in this study, and discuss caveats to the analysis identifying the flare-triggering site. We finally summarize the conclusions of the present study in Section \[sec:summary\]. Data Analysis and Event Classification {#sec:data_analysis} ====================================== Synopsis of the Theoretical Model {#sec:KB12} --------------------------------- We briefly review the theoretical model proposed by K12. The model simplifies the magnetic field structure of an AR and quantitatively surveyed magnetic configurations which can produce solar eruptions (flares). Figure \[fig:setting\](a) shows their simulation setup. They imposed a small bipole field onto the PIL of the large-scale bipole field representing an active region. The internal magnetic reconnection can occur between the sheared field (red arcade) and the small-bipole field (blue arcade), and it precedes flare reconnection among the large-scale sheared fields (red arcades) distributed along the PIL if some condition is satisfied. The surveyed parameters, which describe whether flares occur or not, are $\theta_{0}$ and $\varphi_{e}$; defined as illustrated in both panels (a, b). $\theta_{0}$ is the angle of the magnetic field relative to the potential field and it increases counter-clockwise in the range of 0$^{\circ}$-90$^{\circ}$. $\varphi_{e}$ is the counter-clockwise rotation angle of the small-scale bipole field relative to the large-scale potential field of in the range of 0$^{\circ}$-360$^{\circ}$. The parameters $\theta_{0}$ and $\varphi_{e}$ correspond to the magnetic twist and complexity of the magnetic field, respectively. The combination of ($\theta_{0}$, $\varphi_{e}$) characterizes the flare-trigger conditions, and either the OP or RS type bipole field can trigger the positive feedback process of an MHD instability and magnetic reconnection, leading to flares (and sometimes CMEs). The pre-existing sheared fields connect via the OP type field and form a double-arc twisted flux rope, then an eruptive MHD instability is triggered and the instability causes flare reconnection, i.e., “eruption-induced reconnection” starts. If the RS type of field exists, a part of the pre-existing sheared field is cancelled by reconnection with the RS field in the core region, and the sheared field collapses inwards because magnetic pressure is decreased. Then flare reconnection is triggered and the twisted flux rope erupts. It can be described as a “reconnection-induced eruption” process since the erupting flux rope is formed after the flare reconnection. Several observables which are relevant to the flare-trigger condition, have been derived from the simulations, and are summarized below. 1. The magnetic shear angle $\theta_{0}$ and the angle structure of the PIL disturbance at small scales $\varphi_{e}$ is observed to be consist with the conditions of either the OP or RS type. 2. Precursor brightenings should be seen near the PIL where the OP or RS configuration is satisfied. This represents the internal reconnection in the lower atmosphere. Hence it is inferred that precursor brightenings can be observed in any emission line which is formed in the lower atmosphere, for example, the chromosphere. 3. The flare ribbons, which appear in the initial flare phase, should have a clearly sheared configuration as illustrated by the thick yellow lines in Figure \[fig:setting\]. It results from a theoretical prediction that the flare reconnection can be reinforced by an instability which is driven by large-scale (non-potential) fields (such as the red arcades in Figure \[fig:setting\]). A sheared two-ribbon structure and precursor brightening on a highly sheared PIL is common in many other models such as flux cancellation, flux emergence, any tether-cutting reconnection, and also magnetic breakout models. The key feature of the K12 model is that the precursor brightening should appear near the local PIL where the OP or RS conditions are satisfied. It is due to the fact that precursor brightenings may represent local heating caused by the internal reconnection between the pre-existing sheared field and trigger field. Moreover, @bamba13 [@bamba14] reported that precursor brightenings were observed in the lower atmosphere from the upper photosphere to the transition region, rather than in the corona. This is consistent with the K12 simulation in which the OP or RS flux is injected from the bottom boundary into the corona. Therefore, in this study, we use the above three features (geometrical structure of magnetic field; precursor brightening in the chromosphere; and initial flare ribbons) to test the K12 model. Data Description {#sec:data} ---------------- We used data obtained by the Helioseismic and Magnetic Imager (HMI, @schou12) and the Atmospheric Imaging Assembly (AIA, @lemen12), both are onboard the SDO satellite. These instruments observe the full disk of the Sun with a large FOV of $2000^{\prime\prime} \times 2000^{\prime\prime}$. HMI observes the polarization states Stokes-I, Q, U, and V in the photospheric Fe I line (6173[Å]{}) with a spatial resolution of $1^{\prime\prime}$. In this study, we used the HMI level 1.5 line-of-sight (LOS) magnetograms ([hmi.M\_45s]{} series) and the Spaceweather HMI Active Region Patch (SHARP, [hmi.sharp\_cea\_720s]{}, @bobra14) series, that contains vertical and horizontal components of the photospheric magnetic field. The $180^{\circ}$ ambiguity is resolved and magnetic field vectors are remapped to the Lambert Cylindrical Equal-Area (CEA) projection for SHARP data. Data cadences are 45 sec. for the LOS magnetograms and 720 sec. for SHARP data, respectively. AIA observes solar atmosphere in ten EUV and UV channels, but in this study, we only used the AIA 1600 [Å]{} (continuum and C IV line) images that are sensitive to emission from the upper photosphere and transition region (log T $\sim 5.0$). The spatial resolution is $1.5^{\prime\prime}$ and the cadence is 24 sec. for the 1600 [Å]{} data. Analysis Procedure {#sec:method} ------------------ We followed the analysis method developed in @bamba13. The analysis method was developed for Hinode data, and @bamba14 examined the applicability of the techniques to SDO data. Here we summarize the essence of the procedure (see details in @bamba13 [@bamba14]). 1. We calibrated the HMI level 1.5 LOS magnetograms and AIA level 1.0 images to remove spatial fluctuations (spacecraft jitter) and to resample the HMI and AIA images to the same size. 2. We superimposed the PILs and strong brightening contours seen in AIA images onto the HMI LOS magnetograms. We define the flare-triggering site as a region located where the center of the initially sheared flare ribbons and precursor brightening were seen. In other words, if there is no precursor brightening between the initially sheared flare ribbons, we are not able to define the flare-triggering site for the event. 3. We measured the magnetic shear angle $\theta_{0}$ along the flaring PIL and azimuth $\varphi_{e}$ at the flare-triggering site. $\theta_{0}$ was measured as the angle of the initially sheared flare ribbons relative to direction $\bm{N}$, perpendicular to the averaged PIL and representing the direction of the potential field. The averaged PIL used to determine $\bm{N}$ was derived from a low-pass filtering of the LOS magnetic field through a Fourier transformation. The angle of the initially sheared flare ribbons was determined by the averaged angle of the sheared ribbons. We therefore are able to measure the angle only for the events that show two clear flare ribbons in the initial phase. $\varphi_{e}$ is the angle perpendicular to the local PIL in the flare-triggering site relative to $\bm{N}$. Event Selection and Classification Criteria {#sec:criteria} ------------------------------------------- Following to @bamba13, we selected events by the criteria summarized below. - GOES class was larger than M5.0. - Event occurred in the period of 2010 February 11 to 2014 February 28. - The SDO/HMI and AIA 1600 [Å]{} data covered a period of six hours before and after the flare onset for each event. - Flaring site was located within $\pm~750^{\prime\prime}$ from the solar disk center. Then we sampled 32 flare events, as summarized in Table \[table:event\_list\], and classified the events into six types by evaluating whether the important features I, II, and III of Section \[sec:KB12\] were observed. The classification procedure and criteria are summarized in Figure \[fig:chart\], and we labeled each type of event as below. The quantitative values of the angles ($\theta_{0}$, $\varphi_{e}$) stated below were uniquely defined in the present study based on the simulations of K12 for the event classification. Note that we only measured the angles for the events which showed two clear sheared ribbons and precursor brightenings at a single position, hence we did not measure the angles for any events classified as the Multiple Trigger Candidates type. Opposite Polarity (OP) type : \ The event satisfies the ($\theta_{0}$, $\varphi_{e}$) condition of the OP type in Figure 2 of K12. In particular, the shear angle and azimuth should satisfy $0^{\circ} \leq \theta_{0} \leq 90^{\circ}$ and $124^{\circ} \leq \varphi_{e} \leq 225^{\circ}$ at the region/timing where/when the last precursor brightening was seen. Qualitatively, the magnetic field in the small-bipole region has an opposite polarity pattern relative to that averaged over the whole AR. Reversed Shear (RS) type : \ The event satisfies the ($\theta_{0}$, $\varphi_{e}$) condition of the RS type in Figure 2 of K12. In particular, the shear angle and azimuth should satisfy $40^{\circ} \leq \theta_{0}$ and $225^{\circ} \leq \varphi_{e} \leq 335^{\circ}$ at the region/timing where/when the last precursor brightening was seen. The RS type has a $\theta_{0}$ range of approximately $40-90^{\circ}$. The local magnetic shear in the small-scale bipole region is towards the opposite direction of the global magnetic shear of the AR. Contradict-K12 type : \ The event occurs “no-flare” ($\theta_{0}$, $\varphi_{e}$) condition in Figure 2 of K12. In particular, $0^{\circ} \leq \varphi_{e} \leq 120^{\circ}$ and $250^{\circ} \leq \varphi_{e}$ with a small $\theta_{0}$ value at the region/timing where/when the last precursor brightening was seen. Multiple Trigger Candidates type : \ The event clearly shows two sheared ribbons (similar to the thick yellow lines in Figure \[fig:setting\]) in the initial flare phase. However, there are multiple small-scale bipole fields on which precursor brightenings are observed on the local PIL before the flare onset. In other words, there are multiple small-scale bipole fields which show the important features II. and III. of Section \[sec:KB12\], before the flare onset. No-precursor Brightening type : \ The event clearly shows two sheared ribbons in the initial flare phase. However, it does not show any precursor brightening over the local PIL of the small-scale bipole field located at the center of the two ribbons, that exists for at least two observation frames (approximately at least over a period of 1.5 min.) in AIA 1600 [Å]{} images from six hours before to the flare onset time. In brief, the event shows only the important structure III. of Section \[sec:KB12\]. Complicated Ribbon Type : \ The event shows complicated initial flare ribbons which are nothing like the thick yellow lines in Figure \[fig:setting\], i.e., they are much different from feature III suggested by the simulations. More specifically, the flare ribbons seen in the AIA 1600 [Å]{} images overplotted on the HMI LOS magnetogram are not able to distinguish its positive and negative polarities during the period from 5 minutes before flare onset to peak time. This type includes cases where flare ribbons appear in parallel (not sheared like Figure \[fig:setting\]). Results {#sec:results} ======= The selected flare events and classification results are listed in Table \[table:event\_list\], and Table \[table:list\_summary\] is a summary of the classification. First, we did not find any events of the Contradict-K12 type. It means that there was no flare event which was investigated in this study that occurred under “no-flare” combinations of $(\theta_{0}, \varphi_{e})$. It suggests that there was no event that was clearly inconsistent with the flare-trigger conditions proposed by K12, at least among the events investigated in the present study. Second, six of the 32 events satisfied the RS type condition. Events No. 1 and 2 are the same RS type events examined in previous studies (@bamba13 [@bamba14]). Figure \[fig:sample\](a-1)-(a-3) shows an example of an RS type event: the X5.4 flare that occurred on 2012 March 7 in AR NOAA 11429. The white/black background indicates the positive/negative polarity of the LOS magnetic field. The green lines and red contours outline the PILs (0 G lines) of the LOS magnetic field and strong emission (2000 DN) in AIA 1600 [Å]{}, respectively. The image shows flare ribbons that initially form two clearly sheared ribbon structures on both sides of the PIL, as outlined by black and white broken lines. The yellow arrow indicates the flare-triggering site, and precursor brightenings were intermittently seen in the region, as reported in @bamba13 [@bamba14]. We measured the magnetic shear angle $\theta_{0}$ and azimuth $\varphi_{e}$ around the flare-triggering site, and the angles were ($\theta_{0}, \varphi_{e})\sim(108^{\circ}, 313^{\circ}$) in the case of the X5.4 flare as summarized in Table \[table:event\_list\]. In contrast, there was no OP type event in our selected sample. Note that @bamba13 showed two examples of OP type events that occurred in 2006, that are out of the time range of this study because the events occurred before the SDO launch. Third, four of the 32 events were classified as Multiple Trigger Candidates type. As an example, the M6.0 flare that occurred on 2011 August 3 in AR NOAA 11261 is shown in Figure \[fig:sample\](b-1)-(b-3). These events showed the important features I. and II. However, there were multiple candidates for precursor brightenings at different locations, and a single flare-triggering site could not be identified. In the case of the RS type events, a precursor brightening seen at a single location between the initial two ribbons. However, in the Multiple Trigger Candidates type events, precursor brightenings were intermittently seen at several points on the PIL that was in between the two sheared ribbons. Hence, it was difficult to conclude that a single PIL was the flare-triggering site. Moreover, magnetic field structures in the precursor brightening sites were very small spatially in some cases, less than $2^{\prime\prime}$, that is close to the $1^{\prime\prime}$ spatial resolution of HMI. Especially, azimuth $\varphi_{e}$, which is measured at the point of the precursor brightening site, is highly sensitive to the spatial configuration (i.e. the shape of the PIL) of the site. In this study, we chose the flare-triggering sites to be the location where there were precursor brightenings, and for these cases, it was extremely difficult to measure $\varphi_{e}$ at the correct location. Therefore, in this study, we did not measure the angles for these four Multiple Trigger Candidates type events. Note that the lifetime and spatial size of the precursor brightenings seen in the RS and Multiple Trigger Candidates (and OP) types are different between different events. @bamba13 suggested that the critical magnetic flux differs among events because the critical perturbation amplitude needed to trigger the instability depends on how the system is unstable. Related to this, the spatial size and duration of the heating caused by internal reconnection within the trigger field represented by precursor brightenings is different among the events. In addition, empirically from this and other studies [@bamba13; @bamba14; @Palacios15; @bamba17a; @bamba17b; @Wang17; @Woods17], the duration varies from several hours to a few seconds and the spatial size were from a few to dozens of arcseconds. Another noticeable thing is that 22 events were classified as No-precursor Brightening or Complicated Ribbon type. In other words, approximately 70% of the selected events did not show any clear features that were predicted by K12’s MHD simulation such as sheared two-ribbon structure or precursor brightenings. Panels (c-1)-(c-3) show an example of a No-precursor Brightening type event: the M6.5 flare that occurred on 2013 April 11 in AR NOAA 118719. The 11 events that are classified as No-precursor Brightening type clearly showed a sheared two-ribbon structure in the initial flare phase as suggested by the simulation. However, no brightening was seen on the PIL between the two sheared ribbons, in at least two observation frames (approximately at least over a period of 1.5 min.), in AIA 1600 [Å]{} images, from six hours before to the flare onset time. Many of the No-precursor Brightening type events showed similar behavior: that the initial flare ribbons did not widely propagate. The flare ribbons suddenly appeared as bright points, and the intensity became enhanced at almost the same location within a few minutes. Panels (d-1)-(d-3) show an example of the Complicated Ribbon type: the M9.3 flare that occurred on 2013 October 24 in AR NOAA 11877. Initially the flare ribbons had a complex shape or sometimes looked almost like a single ribbon. Here, we briefly summarize the classification results. - There was no Contradict-K12 type event that satisfied the “no-flare” condition of K12. - 30% of the events, including six RS type events and four Multiple Trigger Candidates type, were consistent with K12. - Approximately 70% of the events (22 of the analyzed 32 events) were classified as either the No-precursor Brightening or Complicated Ribbon types that did not clearly show the key features suggested by K12. Discussion {#sec:discussion} ========== The result that 70% of the events did not show the key features suggested by the MHD simulation of K12 is interesting. Therefore, we here discuss whether the events classified as Complicated Ribbon and No-precursor Brightening type could be explained by K12’s model or other models. Complicated Ribbon type {#sec:dis_CR} ----------------------- It is not surprising that many of the flares showed complicated flare ribbons because the coronal magnetic fields are naturally intertwined. Figure \[fig:d\_11515\] shows a time series analysis of the M5.6 flare (No. 17 listed in Table \[table:event\_list\]) that occurred on 2012 July 2 in AR NOAA 11515, which was classified as a Complicated Ribbon type. The M5.6 flare occurred in a weak negative polarity region between the leading sunspot (LS) and satellite spot (SS). The red contour in panel (e) shows the initial flare ribbon of the M5.6 flare, but it is very different from the sheared two-ribbon structure which is suggested by K12’s simulation (illustrated by the yellow thick line in Figure \[fig:setting\]). The M5.6 flare followed a filament eruption and the preceding C2.9 flare whose flare ribbons are seen in Figure \[fig:d\_11515\](d) in the same region, as @louis14 reported. The filament was rooted in the rear part of the LS and weak negative region as illustrated in Figure \[fig:d\_11515\](a, b) by blue arcs. We found a tiny brightening near the southern root of the filament (the region surrounded by the yellow square in Figure \[fig:d\_11515\](a, b)) over a small wedge-like structure, as can be seen in the enlarged image in panel (a), and it started from 10:10 UT on 2012 July 2. The brightening region gradually extended (as seen in panel (b)), and the C2.9 flare occurred at 10:33 UT (panel (c)). The flare ribbons of the C2.9 event were also complicated as seen in panel (d). The M5.6 flare started at 10:43 UT (panel (e)) from a region slightly to the northwest of the C-class flaring region. Then two clear ribbons (PR and NR) appeared as sheared ribbons in the latter phase (panel (f)). We measured the angles $\theta_{0}$ and $\varphi_{e}$ before the C2.9 flare and eruption onset, assuming that the large-scale magnetic field structure around the SS had not drastically changed. The result of the measurement was $(\theta_{0}, \varphi_{e})\sim(126^{\circ}, 351^{\circ})$, and it is consistent with the RS type condition. From the result, we were able to conclude that the M5.6 flare was triggered by the RS type magnetic field structure in a three-step process, including the C2.9 flare and the filament eruption. @bamba17a also identified the flare-triggering site for an X1.0 flare, which showed three complicated flare ribbons in the initial phase and which can be classified as Complicated Ribbon type, through a detailed analysis. They suggested that precise analyses of spatial and temporal relationships between magnetic field structures and overlying structures such as filaments and precursor brightenings are effective for identifying the flare-triggering site for the Complicated Ribbon type event. Further, @bamba17b confirmed by spectroscopic observation that precursor brightenings can be a marker of magnetic reconnection in the lower atmosphere, such as the internal magnetic reconnection proposed by K12. We hence investigated whether precursor brightenings were seen on a PIL in the region where flare ribbons appeared with regards to the Complicated Ribbon type events. Table \[table:list\_typeD\] shows a list of the Complicated Ribbon type events and whether precursor brightenings exist or not. Brightenings on a PIL were observed in eight events out of the 11 Complicated Ribbon type events. Accordingly, we propose that it is possible that the flare-triggering site for the eight events may be found through detailed analysis, and we need precise analysis for each individual event. On the other hand, we further should consider the relationship between the Complicated Ribbon type events and a flare-trigger by magnetic reconnection at a coronal null point [@Longcope05; @Titov07; @Titov09]. The flare ribbons of the above M5.6 flare (seen in Figure \[fig:d\_11515\](f)) look like the circular-shaped ribbons such as studied by for example, @Masson09. Complex circular-shaped or scattered flare ribbons distributed from a coronal null point origin may be possible in this case. No-precursor Brightening type {#sec:dis_NB} ----------------------------- With regard to the events which did not cause precursor brightenings such as the No-precursor Brightening type and three of the Complicated Ribbon type events, the possibility that the flares were triggered by a different physical process(es) from that proposed by K12 cannot also be ruled out. K12 proposed that major flares should be preceded by internal magnetic reconnection between the flare-triggering flux of the OP or RS types and pre-existing sheared magnetic field. They treated precursor brightenings in the lower atmosphere on the flare-triggering site as a proxy of the internal reconnection. Therefore, we still can consider other physical process(es). For instance, if a long twisted flux rope was strapped down by overlying magnetic field and the trapping field was weakened by some cause such as emerging flux (case-B scenario of @chenshibata00), then in this case, precursor brightenings may be seen at a foot point of the trapping field, which may be away from the flaring site. We should confirm theexistence of a long twisted flux rope, and investigate how the flux rope was formed during the AR evolution, using observational data at high temperature, such as taken by the Hinode/X-ray telescope. Otherwise, gradual magnetic flux cancellation at a foot point of the overlying field may also weaken the trapping field [@vanBallegooijen89; @Zhang01; @Green11]. Precursor brightenings may be observed if the flux cancellation was caused by magnetic reconnection between opposite polarity fluxes. However, it is conceivable that no brightening may be observed in the case of flux cancellation by sinking of small magnetic fluxes in the photosphere. Another possibility for no-precursor brightening event is that a magnetic flux tube had formed in advance by some cause such as a preceding flare or photospheric motion and it became unstable to the ideal MHD instabilities of torus, kink and double-arc modes [@kliemtorok06; @torokkliem05; @ishiguro17]. We note that there may be a possibility that precursor brightenings can be seen in a different wavelength from AIA 1600 [Å]{}, which was used in the present study and which is sensitive to emission from the upper photosphere and lower transition region (log T $\sim 5.0$). Especially, in the case of the RS type, magnetic shear cancellation by internal reconnection could occur at any altitude while the internal reconnection should occur in the lower atmosphere in the OP type case. This is because the altitude of the internal reconnection in the RS type case depends on the spatial size of the pre-existing sheared field and flare-trigger field. Therefore, it might be likely that we cannot detect internal reconnection of the RS type by emission in 1600 [Å]{}, if the internal reconnection proceeds in the coronal region. We can also consider the breakout trigger scenario [@Antiochos99], in which magnetic reconnection occurs at much higher altitude than focused on in this study. Precursor brightenings may be observed in higher temperature lines than AIA 1600 [Å]{} as in the RS type case event. Another possibility is that the size of the internal reconnection region working as a trigger was too small to be observed. Recently, @ishiguro17 developed a theoretical model of the new Double-Arc Instability, and according to that model they predict that the size of the triggering reconnection depends on the strength of the magnetic twist. If magnetic twist is high enough, even tiny reconnection can trigger the Double-Arc Instability and work as a flare trigger. Therefore, it is necessary to distinguish which types of flares are caused by which physical process, i.e., which types of flares are explained by which model, by more precisely analyzing each of the events using multiple wavelengths data. Multiple Trigger Candidates type {#sec:dis_MTR} -------------------------------- The Multiple Trigger Candidates type events, in which there were several flare-triggering site candidates, is still understandable by the flare-trigger model of K12. It is not necessary to restrict the number of flare-triggering sites for some cases. @bamba13 [@bamba14] suggested that not only the geometrical conditions $\theta_{0}$ and $\varphi_{e}$ but also the total magnetic flux in the flare-triggering site and/or its temporal evolution contribute to the critical conditions for flare occurrence. They also suggested that the critical magnetic flux required to trigger an instability depends on the proximity of the system to an unstable state. Hence the combined contribution of more than one small-scale flux, that satisfies either the OP and/or RS type condition, is also conceivable as a flare-trigger. @bamba17b analyzed a region which includes two RS type bipole structures, and studied the sequential process from the internal reconnection, which was represented by precursor brightenings, to destabilization of the large-scale system in the AR, using spectroscopic data. However, they were not able to determine the detailed process by which the two flare-triggering sites contributed to the flare occurrence because the data they employed only covered the odd RS type region. Thus, we need more analyses to clarify the combination effect of multiple and different types of trigger regions. These No-precursor Brightening and Multiple Trigger Candidates type events are future topics of discussion. Difference of the Incidence rate of OP and RS types {#sec:incidence} --------------------------------------------------- Another interesting topic for discussion is the result that the OP type events were rare compared to the RS types, at least in this study. Some other studies found events which were suggested to be triggered by the OP type small magnetic structure [@Palacios15; @Wang17; @Woods17]. However, so far we found only two OP type flare events (already shown by @bamba13 by the same method of this study) whereas six events were classified as the RS type in this study. We consider that the difference in incidence between the two types results from a difference of flexibility in the geometrical conditions, such as displacement of the flare-trigger field from the highly sheared PIL and the height of the internal reconnection. For instance, the internal reconnection in the RS type case may be less sensitive to altitude, and it can occur more frequently than internal reconnection triggered by the OP type structure. Otherwise the RS type case may be sensitive to displacement of the flare-triggering site from the highly sheared (flaring) PIL. In the OP type case, pre-existing large-scale magnetic arcades (red arcades in Figure \[fig:setting\]) should directly reconnect with the flare-trigger field (blue arcade in Figure \[fig:setting\]), in order to form long twisted flux ropes. Therefore, the twisted flux rope, which will erupt and trigger flare reconnection, could not be formed if the small bipole is away from the PIL. Conversely, in the RS type case, magnetic shear cancellation, i.e. the internal magnetic reconnection between pre-existing magnetic field (red arcades in Figure \[fig:setting\]) and trigger field (the blue arcade in Figure \[fig:setting\]), can occur as long as the small bipole exists under the pre-existing sheared magnetic arcades. Thus, the sheared magnetic arcades can collapse and flare reconnection can be triggered by shear cancellation as long as the RS type field exists within the sheared arcade even if it is located away from the PIL. From observational results, such as those reported by @bamba17a, it is also suggested that the RS type flare-trigger could work even if it is located slightly away from the flaring PIL, even though the trigger field was injected just above the PIL in the K12 simulations. Therefore, the differences in sensitivity to the geometrical conditions and incidence are originally derived from differences in the physical processes for flare-triggering between the two types. Thus it is not surprising that the OP type flare-trigger is rare than the RS type flare-trigger. However, it is still unclear how a distant field that is away from the PIL can trigger flares in the OP and RS cases on the actual solar surface, and we need to statistically investigate the distance between flaring PILs and flare-trigger fields using observational data. Summary {#sec:summary} ======= In this study, we aimed to evaluate the consistency between the theoretical flare-trigger model of K12 and a variety of major flares that occurred on the actual solar surface, by focusing on the geometrical structure. We selected 32 major flares and tried to identify the flare-triggering site by applying the analysis method that was developed by @bamba13 [@bamba14] to SDO/HMI LOS magnetograms and AIA 1600 [Å]{} images. We classified the 32 events into six types of groups including K12’s OP and RS types. The most noteworthy result is that 30% of the events (including the RS type events and Multiple Trigger Candidates type) were consistent with K12. Moreover, there was no event that contradicted K12 and satisfied their “no-flare” condition at least in the events sampled in this study. Eight of the 11 Complicated Ribbon type events could be interpreted by K12’s flare-trigger model, even though we need more precise analysis of the spatial and temporal relationships between the magnetic field and overlying structures such as filament and precursor brightenings. Meanwhile, we found that there is a possibility that different physical process(es) from those proposed by K12 cannot also be ruled out for the events which did not show any precursor brightenings, such as the No-precursor Brightening type events and the three Complicated Ribbon type events. We still can consider other physical processes such as the emerging flux model proposed by @chenshibata00 and the magnetic flux cancellation model proposed by @vanBallegooijen89 [@Zhang01; @Green11], to trigger these events. Extended studies are required to reveal the physical process(es) which causes the different types of flares. Even so, our result that 30% of the events which were investigated in this study were consistent with the flare-trigger model of K12 leads to the conclusion that the observable features and ($\theta_{0}$, $\varphi_{e}$) parameters of K12 are important to understand the flare triggering. Moreover, we found that there was no OP type event in the events analyzed in this study while there were six RS type events. We hypothesized that this difference of incidence between the OP and RS types is likely owing to differences in the physical processes between the two-types: the RS type condition can be satisfied more easily because the geometrical conditions of the RS type are more flexible than those of the OP type. We need an extended study to clarify the occurrence difference between the OP and RS types, and the consistency/inconsistency between the Complicated Ribbon type, No-precursor Brightening type, and Multiple Trigger Candidates type events and the physical flare-trigger process of K12. Especially, it is important to comprehensively analyze the relationship between the photospheric magnetic field structures and chromospheric/coronal features, using not only AIA 1600 [Å]{} images but also multiple wavelengths. Moreover, we caution that the limitations of this study given the single UV wavelength used, and relying upon the LOS magnetic field data for the PIL location in AIA images, may introduce bias and errors in the results and their interpretation (see Appendix \[sec:projection\]). As such, it is suggested that further study be performed using additional data and more careful analysis. We should reveal the role of the small-scale OP- or RS type field for a flare in various magnetic field topologies, using both simulations and precise observations. The authors deeply appreciate the many instructive comments and encouragements from Dr. K. D. Leka, Dr. Satoshi Inoue, Dr. David H. Brooks, Dr. Shinsuke Imada, and researchers in ISAS/JAXA and NAOJ. The authors thank Dr. T. Hara for providing data analysis tool. The HMI and AIA data have been used courtesy of NASA/SDO and the AIA and HMI science teams. This work was partly carried out at the Hinode Science Center at NAOJ and Nagoya University, Japan, and the Solar Data Analysis System (SDAS) operated by the Astronomy Data Center in cooperation with the Hinode Science Center of the NAOJ. This work was supported by MEXT/JSPS KAKENHI Grant Numbers JP16H07478, JP15H05812, JP15K21709, JP15H05814, and JP15J10092. The authors thank the anonymous referee for valuable comments that improved the clarity of the manuscript. Caveat in Using the LOS Magnetic Field Data to Detect Flare-triggering Site {#sec:projection} =========================================================================== In this study, we mainly used the HMI LOS magnetograms to investigate the photospheric magnetic field structures. The LOS magnetic field contains a projection-effect, that is, the LOS component of the magnetic field $\it{B}_{LOS}$ can be treated as the radial component $\it{B}_{r}$ only when the observing angle $\theta$ equal 0 (or $\mu = cos(\theta) = 1.0$). In other words, $\it{B}_{r}$ can be different from $\it{B}_{LOS}$ depending on the location of the AR on the solar surface, even very close to solar disk center [@leka17]. In this study (also @bamba13 [@bamba14]), events were selected based on the criterion that the flaring site was located within $\pm~750^{\prime\prime}$ of the solar disk center ($\mu \sim 0.67$), in consideration of this projection effect. However, we should keep in mind that there is a possibility that we may get a slightly different result between the analysis with $\it{B}_{LOS}$ and with $\it{B}_{r}$. The flare-trigger field can be a very small-scale structure (empirically estimated as dozens of Mm) compared to the spatial size of an AR, and the spatial size of the trigger field is different between different AR. Moreover, the azimuth $\varphi_{e}$ is highly sensitive to the configuration of the local PIL and the location of the precursor brightenings in the flare-triggering site. We should have an extended analysis (a future work) using $\it{B}_{r}$ instead of $\it{B}_{LOS}$ to clarify the consistency between the flare-trigger model of K12 and the events classified as Multiple Trigger Candidates type, No-precursor Brightening type, and Complicated Ribbon type. We also used SHARP data, converted into Lambert CEA projection, to measure the shear angle $\theta_{0}$ in the present study. The magnetic field vectors in SHARP data have been transformed into components of the heliographic coordinates ($\it{B}_{r}$, $\it{B}_{\theta}$, $\it{B}_{\phi}$), which were originally in a spherical coordinate system, and these are transformed into a planar CEA coordinate system. These unit vectors in each coordinate system are not precisely aligned except at the center of the patch. In fact, an error is caused when the vectors are transformed from a spherical coordinate system to a planar CEA coordinate system (@bobra14). In this study, we determined the flaring PIL with co-aligned $\it{B}_{LOS}$ and AIA 1600 [Å]{} images, however, it may not mach the PIL identified using SHARP CEA data. We could have a few-pixel shift in the location of the PIL depending on the location of the AR on the solar surface. Therefore, we averaged the angle $\theta_{0}$ over a region 5-10 times the area of the flare-trigger field. Antiochos, S. K., DeVore, C. R., & Klimchuk, J. A. 1999, , 510, 485 Aulanier, G., T[ö]{}r[ö]{}k, T., D[é]{}moulin, P., & DeLuca, E. E. 2010, , 708, 314 Bamba, Y., Kusano, K., Yamamoto, T. T., & Okamoto, T. J. 2013, , 778, 48 Bamba, Y., Kusano, K., Imada, S., & Iida, Y. 2014, , 66, S16 Bamba, Y., Inoue, S., Kusano, K., & Shiota, D. 2017, , 838, 134 Bamba, Y., Lee, K.-S., Imada, S., & Kusano, K. 2017, , 840, 116 Bateman, G. 1978, Cambridge, Mass., MIT Press, 1978. 270 p., Bobra, M. G., Sun, X., Hoeksema, J. T., et al. 2014, , 289, 3549 Bobra, M. G., & Couvidat, S. 2015, , 798, 135 Carmichael, H. 1964, NASA Special Publication, 50, 451 Chen, P. F., & Shibata, K. 2000, [ApJ]{}, 545, 524 Cheng, X., Zhang, J., Ding, M. D., Guo, Y., & Su, J. T. 2011, , 732, 87 D[é]{}moulin, P., & Aulanier, G. 2010, , 718, 1388 Fan, Y., & Gibson, S. E. 2003, [ApJL]{}, 589, L10 Falconer, D. A., Moore, R. L., & Gary, G. A. 2008, , 689, 1433-1442 Gerrard, C. L., Arber, T. D., Hood, A. W., & Van der Linden, R. A. M. 2001, , 373, 1089 Green, L. M., Kliem, B., & Wallace, A. J. 2011, , 526, A2 Hagyard, M. J., Teuber, D., West, E. A., & Smith, J. B. 1984, , 91, 115 Hagyard, M. J., Moore, R. L., & Emslie, A. G. 1984, Advances in Space Research, 4, 71 Hassanin, A., & Kliem, B. 2016, , 832, 106 Higgins, P. A., Gallagher, P. T., McAteer, R. T. J., & Bloomfield, D. S. 2011, Advances in Space Research, 47, 2105 Hirayama, T. 1974, [Solar Phys.]{}, 34, 323 Inoue, S., Hayashi, K., & Kusano, K. 2016, , 818, 168 Ishiguro, N., & Kusano, K. 2017, , 843, 101 Karpen, J. T., Antiochos, S. K., & DeVore, C. R. 2012, , 760, 81 Kentischer, T. J., Bethge, C., Elmore, D. F., et al. 2008, [SPIE]{}, 7014, 701413 Kliem, B., T[ö]{}r[ö]{}k, T. 2006, [Physical Review Letters]{}, 96, 255002 Kliem, B., Su, Y. N., van Ballegooijen, A. A., & DeLuca, E. E. 2013, , 779, 129 Kliem, B., Lin, J., Forbes, T. G., Priest, E. R., & T[ö]{}r[ö]{}k, T. 2014, , 789, 46 Kopp, R. A., & Pneuman, G. W. 1976, [Solar Phys.]{}, 50, 85 Kosugi, T., Matsuzaki, K., Sakao, T., et al. 2007, [Solar Phys.]{}, 243, 3 Kusano, K., Bamba, Y., Yamamoto, T. T., et al. 2012, , 760, 31 Leka, K. D., Barnes, G., & Wagner, E. L. 2017, , 292, 36 Lemen, J. R., Title, A. M., Akin, D. J., et al. 2012, , 275, 17 Liu, C., Deng, N., Lee, J., et al. 2013, , 778, L36 Liu, R., Kliem, B., Titov, V. S., et al. 2016, , 818, 148 Longcope, D. W. 2005, Living Reviews in Solar Physics, 2, 7 Louis, R. E., Puschmann, K. G., Kliem, B., Balthasar, H., & Denker, C. 2014, [A&A]{}, 562, A110 Masson, S., Pariat, E., Aulanier, G., & Schrijver, C. J. 2009, , 700, 559 Moore, R. L., Sterling, A. C., Hudson, H. S., & Lemen, J. R. 2001, [ApJ]{}, 552, 833 Moore, R. L., & Sterling, A. C. 2006, Washington DC American Geophysical Union Geophysical Monograph Series, 165, 43 Palacios, J., Cid, C., Guerrero, A., Saiz, E., & Cerrato, Y. 2015, , 583, A47 Pesnell, W. D., Thompson, B. J., & Chamberlin, P. C. 2012, , 275, 3 Priest, E. R., & Titov, V. S. 1996, Proceedings of the Royal Society of London Series A, 354, 2951 Reva, A. A., Ulyanov, A. S., Shestov, S. V., & Kuzin, S. V. 2016, , 816, 90 Savcheva, A., Pariat, E., McKillop, S., et al. 2015, , 810, 96 Schou, J., Scherrer, P. H., Bush, R. I., et al. 2012, , 275, 229 Schrijver, C. J. 2007, , 655, L117 Schmieder, B., Aulanier, G., & Vr[š]{}nak, B. 2015, , 290, 3457 Sturrock, P. A. 1966, [Nature]{}, 211, 695 Sun, X., Hoeksema, J. T., Liu, Y., Chen, Q., & Hayashi, K. 2012, , 757, 149 Titov, V. S. 2007, , 660, 863 Titov, V. S., Forbes, T. G., Priest, E. R., Miki[ć]{}, Z., & Linker, J. A. 2009, , 693, 1029 T[ö]{}r[ö]{}k, T., Kliem, B., & Titov, V. S. 2004, , 413, L27 Török, T., & Kliem, B. 2005, [ApJ]{}, 630, L97 van Ballegooijen, A. A., & Martens, P. C. H. 1989, , 343, 971 Wang, H., Liu, C., Ahn, K., et al. 2017, Nature Astronomy, 1, 0085 Woods, M. M., Harra, L. K., Matthews, S. A., et al. 2017, , 292, 38 Zhang, J., Wang, J., Deng, Y., & Wu, D. 2001, , 548, L99 Zuccarello, F. P., Aulanier, G., & Gilchrist, S. A. 2015, , 814, 126 ----- ------------- ------------ ------------ -------- ----------------------------- --------------------------------------------------------- No. Date Onset Time GOES X-ray AR Type Angle(s) (UT) Class (NOAA) 1 13-Feb-2011 17:28 M6.6 11158 Reversed Shear $\theta_{0}\sim88^{\circ}, \varphi_{e}\sim344^{\circ}$ 2 15-Feb-2011 01:44 X2.2 11158 Reversed Shear $\theta_{0}\sim86^{\circ}, \varphi_{e}\sim331^{\circ}$ 3 09-Mar-2011 23:13 X1.5 11166 Complicated Ribbon 4 03-Aug-2011 13:17 M6.0 11261 Multiple Trigger Candidates $\theta_{0}\sim51^{\circ}$ 5 04-Aug-2011 03:41 M9.3 11261 Complicated Ribbon 6 06-Sep-2011 01:35 M5.3 11283 No-precursor Brightening $\theta_{0}\sim80^{\circ}$ 7 06-Sep-2011 22:12 X2.1 11283 No-precursor Brightening $\theta_{0}\sim79^{\circ}$ 8 07-Sep-2011 22:32 X1.8 11283 Complicated Ribbon 9 08-Sep-2011 15:32 M6.7 11283 Complicated Ribbon 10 23-Jan-2012 03:38 M8.7 11402 Multiple Trigger Candidates $\theta_{0}\sim20^{\circ}$ 11 05-Mar-2012 02:30 X1.1 11429 No-precursor Brightening $\theta_{0}\sim60^{\circ}$ 12 07-Mar-2012 00:02 X5.4 11429 Reversed Shear $\theta_{0}\sim108^{\circ}, \varphi_{e}\sim313^{\circ}$ 13 07-Mar-2012 01:05 X1.3 11429 Reversed Shear $\theta_{0}\sim68^{\circ}, \varphi_{e}\sim295^{\circ}$ 14 09-Mar-2012 03:22 M6.3 11429 Reversed Shear $\theta_{0}\sim74^{\circ}, \varphi_{e}\sim333^{\circ}$ 15 10-Mar-2012 17:15 M8.4 11429 Reversed Shear $\theta_{0}\sim98^{\circ}, \varphi_{e}\sim312^{\circ}$ 16 10-May-2012 04:11 M5.7 11476 Complicated Ribbon 17 02-Jul-2012 10:43 M5.6 11515 Complicated Ribbon 18 04-Jul-2012 09:47 M5.3 11515 Multiple Trigger Candidates $\theta_{0}\sim56^{\circ}$ 19 05-Jul-2012 11:39 M6.1 11515 Complicated Ribbon 20 12-Jul-2012 15:37 X1.4 11520 No-precursor Brightening $\theta_{0}\sim77^{\circ}$ 21 13-Nov-2012 01:58 M6.0 11613 Complicated Ribbon 22 11-Apl-2013 06:55 M6.5 11719 No-precursor Brightening $\theta_{0}\sim52^{\circ}$ 23 24-Oct-2013 00:21 M9.3 11877 Complicated Ribbon 24 01-Nov-2013 19:46 M6.3 11884 No-precursor Brightening $\theta_{0}\sim82^{\circ}$ 25 03-Nov-2013 05:16 M5.0 11884 No-precursor Brightening $\theta_{0}\sim79^{\circ}$ 26 05-Nov-2013 22:07 X3.3 11890 Complicated Ribbon 27 08-Nov-2013 04:20 X1.1 11890 No-precursor Brightening $\theta_{0}\sim80^{\circ}$ 28 31-Dec-2013 21:45 M6.4 11936 No-precursor Brightening $\theta_{0}\sim47^{\circ}$ 29 01-Jan-2014 18:40 M9.9 11936 Complicated Ribbon 30 07-Jan-2014 10:07 M7.2 11944 No-precursor Brightening $\theta_{0}\sim95^{\circ}$ 31 07-Jan-2014 18:04 X1.2 11944 No-precursor Brightening $\theta_{0}\sim15^{\circ}$ 32 04-Feb-2014 03:57 M5.2 11967 Multiple Trigger Candidates $\theta_{0}\sim65^{\circ}$ ----- ------------- ------------ ------------ -------- ----------------------------- --------------------------------------------------------- Type Number of Events ----------------------------- ------------------ Opposite Polarity 0 Reversed Shear 6 Contradict-K12 0 Multiple Trigger Candidates 4 No-precursor Brightening 11 Complicated Ribbon 11 No. Date Onset Time (UT) GOES X-ray Class NOAA AR Precursor Brightenings ----- ------------- ----------------- ------------------ --------- ------------------------ 3 09-Mar-2011 23:13 X1.5 11166 Yes 5 04-Aug-2011 03:41 M9.3 11261 Yes 8 07-Sep-2011 22:32 X1.8 11283 No 9 08-Sep-2011 15:32 M6.7 11283 Yes 16 10-May-2012 04:11 M5.7 11476 No 17 02-Jul-2012 10:43 M5.6 11515 Yes 19 05-Jul-2012 11:39 M6.1 11515 Yes 21 13-Nov-2012 01:58 M6.0 11613 Yes 23 24-Oct-2013 00:21 M9.3 11877 Yes 26 05-Nov-2013 22:07 X3.3 11890 Yes 29 01-Jan-2014 18:40 M9.9 11936 No \[table:list\_typeD\]	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent observations have shown that the majority of the Andromeda galaxy’s satellites are aligned in a thin plane. On the theoretical side it has been proposed that galaxies acquire their gas via cold streams. In addition, numerical simulations show that the same streams also deliver gas clumps which could potentially develop into satellite galaxies. Assuming that cold streams are a major source of satellite systems around galaxies we calculate the probabilities in two different models to find a certain fraction of satellites within a thin plane around the central galaxy of the host halo with and without having the same sense of rotation. Using simple geometrical considerations and adopting a random orientation of the streams we demonstrate that the vast thin disk of satellites detected around Andromeda can naturally be explained within this framework. In fact, without any satellite scattering, two streams would lead to too many satellites in the thin plane, compared with the observations. Three streams reproduce the observations very well. Natural implications from our model are that all massive galaxies should have a thin plane of satellites and that the satellites should naturally distribute themselves not only into a single plane but into several inclined ones. We estimate the effect of additional satellites accreted from random directions and find it to be of minor relevance for a mild inflow of satellites from random directions.' author: - \| \ \ $^1$Institut für Astrophysik, Universität Wien, Türkenschanzstra[ß]{}e 17, 1180 Wien, Österreich\ $^2$Universitäts-Sternwarte München, Scheinerstra[ß]{}e 1, 81679 München, Deutschland\ $^3$Max Planck Fellow, Max-Planck-Institut für extraterrestrische Physik, Gie[ß]{}enbachstra[ß]{}e, 85748 Garching, Deutschland\ $^4$Centro de Astrobiolog[í]{}a (INTA-CSIC), Ctra de Torrej[ó]{}n a Ajalvir, km 4, 28850 Torrej[ó]{}n de Ardoz, Madrid, España\ $^5$Astro-UAM, Universidad Autónoma de Madrid, Unidad Asociada CSIC, 28049 Madrid, España\ bibliography: - 'probab18.bib' date: Draft version title: 'The co-planarity of satellite galaxies delivered by randomly aligned cold mode accretion streams' --- \[firstpage\] cosmology: theory – galaxies: evolution – galaxies: formation – galaxies: high redshift – methods: numerical – Local Group Introduction {#sec:intro} ============ Our understanding of how galaxies form has changed substantially in recent years. A decade ago it was thought [@blumenthal; @rees; @silk; @white] that galaxies collect their baryons through diffuse gas, spherically symmetrically falling into dark matter haloes and being shock-heated as it hits the gas residing in the haloes, the so-called hot mode accretion. Whether the gas eventually settles into the equatorial plane, forming a galactic disk was depending on the mass of the dark halo. Below a critical mass, the gas could cool efficiently, forming a disk galaxy, while for larger masses the cooling time would be longer than the Hubble time, leading to structures that resemble galaxy clusters with a large baryon fraction in the hot, diffuse intergalactic gas component. Recent theoretical work and simulations [@fardal; @bd03; @keresa; @keresb; @db06; @ocvirk; @DekelA_09a; @dekel13] however seem to indicate that at high redshift $(z \gtrsim 2)$, galaxies acquire their baryons primarily via cold streams of relatively dense and pristine gas with temperatures around $10^4$ K that penetrate through the diffuse shock-heated medium, the so-called cold mode accretion. These streams peak in activity around redshift 3. Having reached the inner parts of the host halo they will eventually form a dense, unstable, turbulent disc where rapid star formation is triggered [@oscara; @oscarb; @DekelA_09b; @cd; @andi; @c12; @marcello; @genel12; @genzel11; @dekel13; @cip]. $N$-body simulations suggest that about half the mass in dark-matter haloes is built-up smoothly, suggesting that the baryons are also accreted semi-continuously as the galaxies grow [@genel10]. Hydrodynamical cosmological simulations also show rather smooth gas accretion, including mini-minor mergers with mass ratios smaller than 1:10, that brings in about two thirds of the mass [@DekelA_09a]. The massive, clumpy and star-forming discs observed at $z\sim 2$ [@genzel08; @genel; @foerster2; @foerster3] may have been formed primarily via the smooth and steady accretion provided by the cold streams, with a smaller contribution by major merger events [@oscara; @cd]. Attempts to directly observe cold accretion streams are ongoing: @mich used cosmological hydrodynamical <span style="font-variant:small-caps;">amr</span> simulations to predict the characteristics of Ly$\alpha$ emission from the cold gas streams. The Ly$\alpha$ luminosity in their simulations is powered by the release of gravitational energy as the gas is flowing inwards with a rather constant velocity. The simulated Ly$\alpha$-blobs (LABs) are similar in many ways to the observed LABs. Some of the observed LABs may thus be regarded as direct detections of the cold streams that drove galaxy evolution at high $z$. Observations [@rauch; @erb] and new <span style="font-variant:small-caps;">amr</span> simulations incorporating radiative transfer support this model [@joki]. @mich2 made theoretical predictions about the likelihood of observing these streams in absorption which have very recently really been observed [@bouche2]. @mich5 looked at the amount of inflow – the mass accretion rate – both as a function of radius, mass and redshift for the three constituents gas, stars and dark matter. @danovich find that at the few virial radii vicinity of the galaxy, the streams tend to be confined to a stream plane, and embedded in a flat pancake that carries $\sim 20\%$ of the influx. There are on average three significant streams, of which one typically carries more than half the mass inflow. Given the fact that galaxies grow by cold stream accretion, the question arises whether some observational signatures of this phase are still detectable in present-day galaxies. @mich4 analysed the velocity of the accretion along the streams. ![image](dendmulti03com.eps){width="17.73cm"} @noam1 [@noam2] looked at the preferred direction and at the distribution of satellite galaxies in the Local Group. They used $N$-body dark matter only simulations. They find that the satellites within the Local Group should preferentially lie in a plane. @ibata showed in a seminal observational paper the existence of a planar subgroup of satellites in the Andromeda galaxy (M 31), comprising about half of the population as observed and described by @conn. The structure is at least 200 kilo-parsec in radius, but also extremely thin, with a perpendicular scatter of less than 12.6 kilo-parsec. Radial velocity measurements reveal that the satellites in this structure have the same sense of rotation about their host. This shows conclusively that substantial numbers of dwarf satellite galaxies share the same dynamical orbital properties and direction of angular momentum. Intriguingly, the plane they identify is approximately aligned with the pole of the Milky Way’s disk and with the vector between the Milky Way and Andromeda. @ibata2 stressed the importance of reproducing all of the observed properties of the plane, especially the fact that most of the satellites are co-rotating, when comparing thy numerical simulations. @ibata3 found observational evidences that thin planes of satellites might be a common phenomenon amongst galaxies in the universe. Following up on this @bowden argued that a thin satellite disc can persist over cosmological times if and only if it lies in the planes perpendicular to the long or short axis of a triaxial halo, or in the equatorial or polar planes of a spheroidal halo. In any other orientation, the disc thickness would double on $\sim 5$ Gyr timescales and so must have been born with what they call ’an implausibly small vertical scale-height’. Several scenarios have been proposed in order to explain the origin of planar satellite systems. Dwarf galaxies might be accreted in groups [@donghia08; @li08]. Alternatively, the disk of satellites might be the tidal debris of a major merger with a gas-rich galaxy with the tidal arms condensing into tidal dwarf galaxies [@wetzstein07; @bournaud08; @pawlowski11; @hammer]. This scenario is however in conflict with the observational evidence of substantial amounts of dark matter, dominating the kinematics of dwarf spheroidals [see however @kroupa97; @pawlowski12]. Interestingly, coherently rotating, quasi-planar distributions of satellites have also been found in cold-dark-matter simulations [@lovell11; @keller12; @bb14; @gillet] [see @pawlowskiemail; @gerhard for an alternative view]. @andrea found evidence in numerical simulation for the creation of polar ring galaxies by cold stream accretion. @temple1 showed that galaxy pairs align with galactic filaments and @temple2 found alignement between the angular distribution of satellite galaxies around the isolated primary galaxies in filaments and the direction of the filaments where those primaries are located in SDSS observations. This indicates that satellite great planes might be a natural result of how galaxies accrete material and substructure from the cosmic web. It is this idea, that is the motivation for our paper. We demonstrate that thin planes of substructure, consistent with the observations of @ibata are in fact expected if galaxies are fed by cold, satellite loaded streams. In section \[sec:calc\] we present two simple geometrical models to analyse the probability of generating a satellite disk with given thickness and satellite fraction, for various numbers of randomly oriented cold streams. Section \[sec:conc\] presents the conclusions. Calculations {#sec:calc} ============ It has been proposed that gas enters a host halo via smooth accretion streams [@DekelA_09a]. For our purposes more interestingly, dark matter haloes filled with gas that could potentially form stars and end up as satellite galaxies are seen in simulations to lie on top of those smooth accretion streams (e.g. figure 2 of @DekelA_09a shows gas clumps on top of gas streams, which are identified by @mich3 to be gas clumps surrounded by dark matter haloes). In figure \[fig:denmap\] we show maps from hydrodynamical <span style="font-variant:small-caps;">amr</span> simulations of high redshift $(z \sim 2.5)$, high mass $(M_{\rm vir} > 10^{11}$ M$_\odot)$ galaxies from two suites of cosmological simulations: the Horizon-MareNostrum simulation [hereafter , @ocvirk] and the suite of simulations [Adaptive Refinement Tree, @cak; @cd; @c12; @dekel13; @c14a]. The circles indicate the virial radii. Clearly visible are three cold streams, funnelling gas into the centre of the galaxy. One can also see several gas clumps, embedded in the streams. Those are smaller galaxies which are entering the host halo via the streams. These smaller structures could merge with the central galaxy or, as we will assume below, end up as satellites, orbiting their host. Note that there are no clumps outside a stream, indicating that the build-up of a satellite system is stream driven [see e.g: @tormen; @onuora; @noam1; @DekelA_09a]. Note further that the clumps are carried in the innermost parts of the streams which are substantially narrower than 13 kpc (i.e. the width of the plane of satellites of M31). Suppose now that galaxies and their satellite systems form in the focal point of a number $m$ of randomly oriented cold streams. What is then the probability that a certain fraction of satellites lies within a thin plane around the central galaxy of the host halo? @ibata found through observations that 15 out of Andromeda galaxy’s 27 satellites are within a plane with 200 kpc radius and 12.6 kpc vertical rms scatter. As the radius (total extent) and vertical thickness (rms scatter) are defined in a different way these values cannot be used directly in order to determine the opening angle of the disk. We therefore analysed again the 3-D coordinates of the 15 satellite galaxies selected by @ibata [see figure 10 of Conn et al. 2012] and computed the rms radius $R$ instead of the maximum radius to be able to compare it to the rms thickness $D$ (12.6 kpc) of the disk. The rms radius turns out to be 187 kpc. From that we computed the opening angle atan$(D/R)$ which is 3.8 degree. @ibata excluded two out of those 15 subhaloes since those are counter-rotating. Since our first analysis using radial orbits does not take the orientation of the rotation into account we will compare this analysis to the ratio $15 / 27 \simeq 56\, \%$. Our second analysis, deploying planar orbits does take the orientation of the rotation into account therefore we will compare that analysis to the ratio $13 / 27 \simeq 48\, \%$. In order to work out the likelihood of having of order half of the satellites in a plane as thin as 3.8 degree we have performed two sets of Monte Carlo experiments, drawing randomly orientated streams and calculating the probabilities of having a given ratio of subhaloes within a plane of certain thickness. Underlying our calculations are the following assumptions: (1) A vast majority of incoming subhaloes enter the host halo through cold streams, as seen in the arrangement of the gas clumps in figure 2 of @DekelA_09a. The effect of having additional subhaloes entering from random directions[^1] will be discussed in depth in section \[sec:addsph\]. (2) Some of the gas clumps carried by the cold streams end up as satellite galaxies orbiting the central galaxy today. (3) Any host halo has between two and seven streams. (4) The streams themselves are randomly distributed over the whole sky as seen from the centre of the host halo. We are aware of studies which present evidence that already the streams seem to lie in a single plane [@DekelA_09a; @danovich]. It is therefore a conservative assumption to distribute the streams randomly with equal solid angles on the sky carrying equal probabilities of having a stream. (5) The orientation of the cold streams does not change with respect to the host halo during the period of accreting satellites. Both of our models are built upon those five assumptions. \[B\]\[B\]\[1\]\[0\][$\Psi$ \[degree\]]{} \[B\]\[B\]\[1\]\[0\][cumulative distribution $P$]{} \[Br\]\[Br\]\[1\]\[0\][Ibata et al.]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 2 / 2]{} \[Br\]\[Br\]\[1\]\[0\][67 % = 2 / 3]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 3 / 3]{} \[Br\]\[Br\]\[1\]\[0\][50 % = 2 / 4]{} \[Br\]\[Br\]\[1\]\[0\][75 % = 3 / 4]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 4 / 4]{} \[Br\]\[Br\]\[1\]\[0\][40 % = 2 / 5]{} \[Br\]\[Br\]\[1\]\[0\][60 % = 3 / 5]{} \[Br\]\[Br\]\[1\]\[0\][80 % = 4 / 5]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 5 / 5]{} \[Br\]\[Br\]\[1\]\[0\][33 % = 2 / 6]{} \[Br\]\[Br\]\[1\]\[0\][50 % = 3 / 6]{} \[Br\]\[Br\]\[1\]\[0\][67 % = 4 / 6]{} \[Br\]\[Br\]\[1\]\[0\][83 % = 5 / 6]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 6 / 6]{} \[Br\]\[Br\]\[1\]\[0\][29 % = 2 / 7]{} \[Br\]\[Br\]\[1\]\[0\][43 % = 3 / 7]{} \[Br\]\[Br\]\[1\]\[0\][57 % = 4 / 7]{} \[Br\]\[Br\]\[1\]\[0\][71 % = 5 / 7]{} \[Br\]\[Br\]\[1\]\[0\][86 % = 6 / 7]{} \[Br\]\[Br\]\[1\]\[0\][100 % = 7 / 7]{} \[Br\]\[Br\]\[1\]\[0\][2 streams]{} \[Br\]\[Br\]\[1\]\[0\][3 streams]{} \[Br\]\[Br\]\[1\]\[0\][4 streams]{} \[Br\]\[Br\]\[1\]\[0\][5 streams]{} \[Br\]\[Br\]\[1\]\[0\][6 streams]{} \[Br\]\[Br\]\[1\]\[0\][7 streams]{} \[Br\]\[Br\]\[1\]\[0\][radial orbits]{} \[B\]\[B\]\[1\]\[0\][$1$]{} \[B\]\[B\]\[1\]\[0\][$0.9$]{} \[B\]\[B\]\[1\]\[0\][$0.8$]{} \[B\]\[B\]\[1\]\[0\][$0.7$]{} \[B\]\[B\]\[1\]\[0\][$0.6$]{} \[B\]\[B\]\[1\]\[0\][$0.5$]{} \[B\]\[B\]\[1\]\[0\][$0.4$]{} \[B\]\[B\]\[1\]\[0\][$0.3$]{} \[B\]\[B\]\[1\]\[0\][$0.2$]{} \[B\]\[B\]\[1\]\[0\][$0.1$]{} \[B\]\[B\]\[1\]\[0\][$5$]{} \[B\]\[B\]\[1\]\[0\][$10$]{} \[B\]\[B\]\[1\]\[0\][$15$]{} \[B\]\[B\]\[1\]\[0\][$20$]{} \[B\]\[B\]\[1\]\[0\][$25$]{} \[B\]\[B\]\[1\]\[0\][$30$]{} \[B\]\[B\]\[1\]\[0\][$35$]{} \[B\]\[B\]\[1\]\[0\][$40$]{} \[B\]\[B\]\[1\]\[0\][$45$]{} ![image](geometry09multi4_08.eps){width="17.73cm"} Radial orbits model {#sec:rom} ------------------- For this model, which deploys radial orbits, an additional four assumptions are needed: (6) Each stream is loaded with an equal amount of subhaloes. (7) The streams will hit the centre of the host halo directly head on, i.e. they have no significant impact parameter with respect to the central galaxy. (8) The subhaloes will initially stay on the same orbits as defined by their stream, i.e. radial orbits parallel to the stream the subhalo is coming from. (9) These orbits will eventually be circularised within a plane also parallel to this stream. Using these assumptions we have performed a series of Monte Carlo calculations. We start a Monte Carlo calculation by fixing the total number $m$ of streams, with $m$ in the range of two to seven. We then draw a large number $(10^7)$ of sets of $m$ random stream orientations. As mentioned earlier, randomly oriented means, that an equal solid angle of the sky as seen from the centre of the host halo has an equal probability of carrying a stream. Also all streams must reach the centre of the host halo directly and the position of one of the streams does not have any influence on the position of the other streams. For each set of streams we conduct for each number $n$ ($n \le m$) of streams out of the total $m$ streams a grid search for the ’optimum’ plane. ’Optimum’ means that the largest of $n$ angles between $n$ streams and the plane is minimal. For this we rasterise the whole unit sphere centred on the host halo with points equally distributed and never further apart than 0.2 degree. Equally distributed means in this context that all points on the sphere are angularly equidistant. The vectors connecting these points with the centre of the host halo are the normal vectors to our plane candidates. A single subset of $n$ streams out of the total of $m$ streams gives a single optimal plane, but there are different optimal planes for different values of $n$ within the same set of $m$ total streams. In practice, the program runs through the following steps: Given one of the planes from our grid search mentioned above, we calculate the angles between each of the $m$ streams and the plane. The resulting $m$ angles are sorted according to their value. The highest value denotes the opening angle within which all $n = m$ streams lie in one plane. The second highest value indicates the opening angle within which a number $n = m - 1$ streams belong to one plane. We collect the values of the minimal angles $\Psi$ for each set of streams and for each value of $n$. We then proceed to the next test plane of our grid search. ### Results {#sec:romresu} In figure \[fig:streams\] we show the cumulative distribution functions of having $n$ out of $m$ streams close to a plane with a maximum angular deviation of $\Psi$ (opening angle). These are the cumulative distribution functions of the minimum angle $\Psi$ for a set of $n$ streams out of the total of $m$ streams. Each panel corresponds to a different total number $m$ and each of the coloured dashed curves shows a different value of $n$. The vertical black line at 3.8 degree indicates the observation of @ibata. The two triangles on that line indicate the intersections with those predicted $n / m$ curves that lie most closely to the fraction $15/27 \simeq 56\, \%$ of Andromeda’s satellites that have been found to belong to the thin plane. These triangles therefore show upper and a lower values for our predicted probabilities. In the first five columns of table \[tab:nstreams\] we explicitly state the positions of the triangles. $m$ $n_{\rm upper}$ $n_{\rm lower}$ $P_{\rm upper}$ $P_{\rm lower}$ $P_{\rm planar}$ ----- ----------------- ----------------- ----------------- ----------------- ------------------ 2 2 2 1.0 1.0 1.0 3 2 2 1.0 1.0 0.588 4 3 2 1.0 0.622 0.243 5 3 2 1.0 0.920 0.115 6 4 3 0.996 0.317 0.078 7 4 3 1.0 0.558 0.053 : Quoted are for all possible total numbers $m$ of streams $m$ the two values for $n$ whose ratios are just above $(n_{\rm upper})$ and just below $(n_{\rm lower})$ the observed ratio of $15 / 27 \simeq 56\, \%$ (radial orbits model). In the next two columns the cumulative probability $P$ that $n$ out of $m$ streams lie in a single plane with a deviation of less than the observed 3.8 degree is quoted (radial orbits model). The final column denotes the probability $P_{\rm planar}$ of having $13 / 27 \simeq 48\, \%$ satellites within a single thin plane and also having the same sense of rotation within the planar orbits model.[]{data-label="tab:nstreams"} ![image](geometry162com.eps){width="17.73cm"} Since we assumed that all the streams pass trough the galaxy centre, two streams will always lie on a perfect plane. The ’2 / $m$’ lines in figure \[fig:streams\] therefore are always at a probability of $P=1.0$. For a three stream scenario we would expect at least 67 % of the satellites to lie within one plane. Since two, three or four stream configurations are most commonly seen in cosmological simulations [@danovich], we can already say that there is nothing surprising about the configuration seen around the Andromeda galaxy. On the contrary, the fact that so little satellites lie on a thin plane rules out a two stream or a three stream scenario for the Andromeda galaxy, unless secular evolution or other processes move some of the satellites out of that plane. Since all heavy galaxies are fulled by cold streams it is a natural consequence of this model that not only the Andromeda galaxy, but all galaxies having at least Andromeda’s mass should have a thin plane of satellites as Andromeda does. Recent observational work [@ibata3; @ibata4] points in this direction [See @cautun however for an alternative view]. The beauty of this model is that it can also explain the two counter-rotating satellite galaxies which belong to the plane of satellites. Other models have severe difficulties with explaining those two. Planar orbits model {#sec:pom} ------------------- Since @ibata2 pointed out the importance of reproducing all of the observed properties of the plane we decided to refine our model further, allowing for planar, co-rotating orbits. For this second model we use the common five assumptions from the beginning of section \[sec:calc\], but we exchange the old assumptions 6 to 9 and replace them by three more realistic ones: (6) The streams are loaded with a varying amount of subhaloes. Each stream has the same probability of carrying many or only few satellites. (7) The streams will not hit the centre of the host halo head on, but with a significant impact parameter in a random direction perpendicular to the stream itself. (8) The subhaloes will stay on planar orbits, in a way that the cross product of the vector of the stream and the vector of the impact parameter defines the plane of the orbit and the sense of the rotation within this plane. Using these assumptions we have performed another large number $(10^7)$ of Monte Carlo calculations. This second set of Monte Carlo simulations is very similar to the first set mentioned in section \[sec:rom\]: First we fix the total number $m$ of streams, with $m$ being in the range of two to seven. We then draw a large number of sets of $m$ random stream orientations. In this model we also draw for each stream a random impact parameter direction. The impact parameter direction is a unit vector which has to be perpendicular to the stream orientation vector which is also a unit vector. The amplitude of the impact parameter vector is always unity (the actual physical size of the impact parameter is irrelevant to our problem, due to the expected circularisation of the orbits) and its remaining direction angle (the overall vector has to be perpendicular to the stream orientation vector) is uniformly distributed between 0 and 2 $\pi$. Also for each stream we draw a satellite loading factor in a way that each stream has the same satellite loading factor distribution and the sum of the satellite loading factors of all $m$ streams always adds up to one. This is done by drawing a random number uniformly distributed between 0 and 1 for each of the $m$ streams. These numbers are then renormalised so that their sum equals to one. This approach might lead to streams which carry only a single or even no clumps. There is nothing peculiar about such a stream carrying only smooth gas but no clumps, it is still a perfectly valid stream. Simulations show indeed that such streams exists. An example for such a stream is the one plotted in the right hand panel of figure \[fig:denmap\] (the stream in the lower right hand corner). We now conduct a grid search for the ’optimum’ plane the same way we did in our first model described in section \[sec:rom\]. Now we collect the values of the minimal angles $\Psi$ for each possible combination of satellite loading factors. ### Results {#sec:pomresu} In figure \[fig:postreams\] we plot the cumulative distribution functions of having the indicated fraction of satellites in a plane with an opening angle $\Psi$ and the same sense of rotation. Each panel corresponds to a different total number $m$ of streams. The black circles indicate the observation of Andromeda which has a fraction $13 / 27 \simeq 48\, \%$ of its satellites within a thin plane with an opening angle of 3.8 degree and the same sense of rotation [@ibata]. In the last column of table \[tab:nstreams\] we explicitly state the positions of the circles. \[B\]\[B\]\[1\]\[0\][cumulative distribution $P$]{} \[Br\]\[Br\]\[1\]\[0\][Ibata et al.]{} \[Br\]\[Br\]\[1\]\[0\][2 streams]{} \[Br\]\[Br\]\[1\]\[0\][3 streams]{} \[Br\]\[Br\]\[1\]\[0\][4 streams]{} \[Br\]\[Br\]\[1\]\[0\][5 streams]{} \[Br\]\[Br\]\[1\]\[0\][6 streams]{} \[Br\]\[Br\]\[1\]\[0\][7 streams]{} \[B\]\[B\]\[1\]\[0\][radial orbits]{} \[Bl\]\[Bl\]\[1\]\[0\][extended radial orbits]{} \[Bl\]\[Bl\]\[1\]\[0\][planar orbits]{} \[B\]\[B\]\[1\]\[0\][$f_{\rm random}$]{} \[B\]\[B\]\[1\]\[0\][1]{} \[B\]\[B\]\[1\]\[0\][0.9]{} \[B\]\[B\]\[1\]\[0\][0.8]{} \[B\]\[B\]\[1\]\[0\][0.7]{} \[B\]\[B\]\[1\]\[0\][0.6]{} \[B\]\[B\]\[1\]\[0\][0.5]{} \[B\]\[B\]\[1\]\[0\][0.4]{} \[B\]\[B\]\[1\]\[0\][0.3]{} \[B\]\[B\]\[1\]\[0\][0.2]{} \[B\]\[B\]\[1\]\[0\][0.1]{} \[B\]\[B\]\[1\]\[0\][0]{} \[B\]\[B\]\[1\]\[0\][5]{} \[B\]\[B\]\[1\]\[0\][10]{} \[B\]\[B\]\[1\]\[0\][15]{} \[B\]\[B\]\[1\]\[0\][20]{} \[B\]\[B\]\[1\]\[0\][25]{} \[B\]\[B\]\[1\]\[0\][30]{} \[B\]\[B\]\[1\]\[0\][35]{} \[B\]\[B\]\[1\]\[0\][40]{} \[B\]\[B\]\[1\]\[0\][45]{} ![image](spinfbmmulti05.eps){width="17.73cm"} It is a feature of our model that all satellites entering from the same stream always stay in the same plane. It is also obvious that the stream with the highest satellite loading factor carries always at least a fraction $1 / m$ of the total satellites. Therefore a satellite fraction of $1 / m$ has always a probability of $P=1.0$ regardless of the opening angle $\Psi$, as seen in figure \[fig:postreams\]. Generalising this one can say that the majority of the signal is usually coming from a large value of the varying satellite loading factor and only a minor contribution is coming from alignments of the planes. Since in the two and in the three stream scenario, which are the most common ones in cosmological simulations [@danovich] the observational data points lie well above 50% we can firmly state that there is nothing surprising about the configuration seen around the Andromeda galaxy. In this model even a four stream scenario can only just be ruled out on a one sigma level. @gerhard present a very similar model of planar orbits, but with fixed satellite loading fractions. They report relatively small probabilities to reproduce the observed situation. The differences to our results arise from their use of fixed instead of varying satellite loading fractions and from their negligence of the two stream scenario. Both might introduce inaccuracies since varying satellite loading fractions are clearly more physical then fixed satellite loading fractions (compare the right panel of figure \[fig:denmap\]) and the two stream scenario is one of the most likely scenarios seen in numerical simulations [@danovich]. As in the radial orbits model a natural consequence of this model is that all galaxies having at least Andromeda’s mass should have such a thin plane of satellites. Recent observations might point in this direction [@ibata3]. On top of that, this model has a second natural consequence: It predicts that high mass haloes should have several planes of satellites, namely one plane for each stream. Indeed recent observational work [@tully] reports the detection of four inclined planes in the Local Group. To make definitive statements about multiple planes however more extended analyses are needed: One should test the statistical significance of the planes, one should consider velocity data or the biases introduced by the satellite survey volumes or having associated satellites to a first plane when looking for a second one and finally one should keep in mind that one of the planes identified is actually offset from Andromeda’s barycentre. So much more theoretical as well as observational work is still needed in the field of multiple planes. Non-stream satellites {#sec:addsph} --------------------- In this section we discuss the impact of satellites entering from random directions not connected to the cold streams. Those satellites could either be accreted at cosmic times at which at which there is no cold stream activity, i.e. outside $z \sim 1-4$ or of sattelites not connected to any of the streams. Satellites entering the host halo from random directions outside of any of the streams will hardly lie in the same plane as the ones accreted as part of the streams. Therefore extending our models to allow for satellites to enter from random directions will significantly lower the probabilities of a certain fraction of all satellites lying in a thin plane. Assuming that no satellite accreted from a random direction will ever end up in the plane of the other satellites, the fraction $f_{\rm all}$ of all the satellites however accreted lying within a thin plane obviously decreases as $f_{\rm all} = f_{\rm stream} \ \left(1 - f_{\rm random}\right)$, where $f_{\rm stream}$ is the fraction of the satellites accreted only through cold streams lying within a thin plane and $f_{\rm random}$ is the fraction of spherically or randomly accreted satellites. To calculate the probability $P_{\rm all}$ of a certain fraction of all satellites lying in a thin plane we first fix $f_{\rm all}$ to the observed values of 15 out of 27 or 13 out of 27, depending on the underlying model and calculate the required fraction $f_{\rm stream}$ of the satellites accreted through the cold streams only lying within a thin plane according to the above formula. The probability $P_{\rm all}$ of a certain fraction of all satellites lying in a thin plane as a function of $f_{\rm stream}$ can now be taken from our two models. They are presented in figures \[fig:streams\] and \[fig:postreams\] with $f_{\rm stream}$ indicated by different line types or by colour, labelled as ’satellite fraction’. In figure \[fig:spinf\] we show how the different probabilities of having the observed thin plane of satellites in host galaxies with a varying number of cold streams in our two models decrease with increasing $f_{\rm random}$. In the left panel the radial orbits model is shown. Since we have only a few discrete data points for the satellite fraction in this model it can only show upper and lower bounds. For configurations having two to four streams the upper bound of the observed degree of co-planarity has a cumulative probability of more than 50% if one allows for up to 44.4% spherically accreted satellites. Since this panel is difficult to read and has only limited significance we also show in the middle panel an “extended radial orbits model” whose only difference to the pure radial orbits model is that we allow for a varying satellite loading factor. This gives narrow lines in this plot. One sees that we still get the observed degree of co-planarity with a cumulative probability of more than 50% if we allow for up to 35.4% spherically accreted satellites even in the seven stream configuration. Scenarios with less streams can have higher fraction of spherically accreted satellites. In the right panel we show the planar orbits model. One sees that one still gets a thin plane of co-rotating satellites as observed in Andromeda in 50% of all the cases if we allow for up to 28.6% spherically accreted satellites in the two stream configuration or up to 4.7% spherically accreted satellites in the three stream configuration. So a mild inflow of satellites from random direction does not alter the conclusions drawn from either of our models. Conclusions {#sec:conc} =========== Inspired by recent observations we investigated the probabilities of a certain fraction of subhaloes lying within a thin plane around the central galaxy of the host halo within the framework of the cold stream scenario. We performed two sets of Monte Carlo simulations that draw randomly orientated streams assuming that the satellites stay on circularised orbits whose planes are still parallel to the stream. We estimated the effect of additional satellites accreted from random directions, leading to the following results: - The configuration seen around the Andromeda galaxy is a natural result of cold stream accretion. A plane as thin as observed can be generated in both of our models with probabilities as high as 50% to 100% for two or three streams, which are the most common configurations seen in cosmological simulations. The radial orbits model however neglects the orientations of the rotation and only takes into account the positions of the satellites. Only the planar orbits model takes both the positions as well as the orientations of the rotation into account. - Without scattering of dwarf satellites, the Andromeda galaxy must have been fed by more than two streams otherwise we would expect more than only 48% of its satellite galaxies to lie within a very thin plane with the same sens of rotation. The most likely number of streams is three. If Andromeda’s satellite system was produced by two streams in fact a large number of objects must have been scattered out of the thin plane and it would be interesting to investigate whether and how this scattering process affects the internal structure of these satellites. - Since both of our models which are both based on the cold stream feeding mechanism guarantee an initial disc that is born with a remarkable small vertical scale-height, scattering processes as discussed by @bowden will most likely play only a minor role. - A natural implication from both our models is that all galaxies having at least the mass of Andromeda should have such a thin plane of satellites. Interestingly recent observations point in that direction [@ibata3]. The high probabilities found for the planar orbits model come about because of the high probability to assign $\ge 48\%$ of the satellites to a single stream which itself is a direct consequence of the fact that the most likely number of cold streams flowing into a single host halo is relatively low, namely between two and four [@danovich]. Also one should keep in mind that the radial orbits model neglects the orientations of the rotation and only takes into account the positions of the satellites. Only the planar orbits model takes both the positions as well as the orientations of the rotation into account. - An additional implication from our second, planar orbits model is that the satellites should naturally distribute themselves into several inclined planes. Indeed @tully reported that there are four inclined planes altogether in the Local Group[^2]. - Allowing satellites to accrete from random directions lowers the probabilities of having the required fraction of satellites in a thin disk. A mild inflow $(\le 25\%)$ of satellites from random directions however does not change our conclusions. We are aware that the current analyses has its limitations. So far we do not incorporate the evolution of the satellite’s orbits, the alignment of the Milky Way and its plane of satellites with respect to Andromeda and its respective plane of satellites nor the possibility of having multiple planes of satellites as observed by @tully. We conclude that the special spatial alignment of Andromeda’s satellite galaxies can naturally be explained by cold stream accretion and simple geometry. Acknowledgements {#acknowledgements .unnumbered} ================ Tobias Goerdt is a Lise Meitner fellow. We thank Rodrigo Ibata, Ben Moore, Doug Potter and Romain Teyssier for their kindness in sharing simulation and observational data with us and Oliver Czoske for the fruitful discussions. Tobias Goerdt would like to thank the University Observatory Munich for their hospitality, where parts of this work were carried out. Parts of the computational calculations were done at the Leibniz-Rechenzentrum under project number pr86ci and at the Vienna Scientific Cluster under project number 70522. This work was supported by FWF project number M 1590-N27 and by MINECO project number AYA 2012-32295. \[lastpage\] [^1]: e.g. accretion of satellites at cosmic times at which there is no cold stream activity, i.e. outside $z \sim 1-4$. [^2]: Concerning multiple planes one should remember the words of caution from the end of section \[sec:pomresu\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'Causal inference in multivariate time series is challenging due to the fact that the sampling rate may not be as fast as the timescale of the causal interactions. In this context, we can view our observed series as a subsampled version of the desired series. Furthermore, due to technological and other limitations, series may be observed at different sampling rates, representing a mixed frequency setting. To determine instantaneous and lagged effects between time series at the true causal scale, we take a model-based approach based on structural vector autoregressive (SVAR) models. In this context, we present a unifying framework for parameter identifiability and estimation under both subsampling and mixed frequencies when the noise, or shocks, are non-Gaussian. Importantly, by studying the SVAR case, we are able to both provide identifiability and estimation methods for the causal structure of both lagged and instantaneous effects at the desired time scale. We further derive an exact EM algorithm for inference in both subsampled and mixed frequency settings. We validate our approach in simulated scenarios and on two real world data sets.' author: - \| Alex Tank\ Department of Statistics\ University of Washington\ [email protected] - \| Emily Fox\ Department of Statistics\ University of Washington\ [email protected] - \| Ali Shojaie\ Department of Biostatistics\ University of Washington\ [email protected] bibliography: - 'ng\_subsamp.bib' title: Identifiability and Estimation of Structural Vector Autoregressive Models for Subsampled and Mixed Frequency Time Series ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'The accuracy of object detectors and trackers is most commonly evaluated by the Intersection over Union (IoU) criterion. To date, most approaches are restricted to axis-aligned or oriented boxes and, as a consequence, many datasets are only labeled with boxes. Nevertheless, axis-aligned or oriented boxes cannot accurately capture an object’s shape. To address this, a number of densely segmented datasets has started to emerge in both the object detection and the object tracking communities. However, evaluating the accuracy of object detectors and trackers that are restricted to boxes on densely segmented data is not straightforward. To close this gap, we introduce the relative Intersection over Union (rIoU) accuracy measure. The measure normalizes the IoU with the optimal box for the segmentation to generate an accuracy measure that ranges between 0 and 1 and allows a more precise measurement of accuracies. Furthermore, it enables an efficient and easy way to understand scenes and the strengths and weaknesses of an object detection or tracking approach. We display how the new measure can be efficiently calculated and present an easy-to-use evaluation framework. The framework is tested on the DAVIS and the VOT2016 segmentations and has been made available to the community.' author: - 'Tobias Böttger${}^{+}$' - 'Patrick Follmann${}^{+}$' - 'Michael Fauser${}^{+}$' bibliography: - 'trackingbib.bib' date: \| ${}^{+}$MVTec Software GmbH, Munich, Germany\ ${}^{}$Technical University of Munich (TUM)\ \ title: Measuring the Accuracy of Object Detectors and Trackers --- Introduction {#sec:introduction} ============ Visual object detection and tracking are two rapidly evolving research areas with dozens of new algorithms being published each year. To compare the performance of the many different approaches, a vast amount of evaluation datasets and schemes are available. They include large detection datasets with multiple object categories, such as PASCAL VOC [@everingham_2015_pascal_voc], smaller, more specific detection datasets with a single category, such as cars [@roman_2015_cod20k], and sequences with multiple frames that are commonly used to evaluate trackers such as VOT2016 [@vot_2016], OTB-2015 [@wu_otb_2015], or MOT16 [@mot_2016]. Although very different in their nature, all of the benchmarks use axis-aligned or oriented boxes as ground truth and estimate the accuracy with the Intersection over Union (IoU) criterion. Nevertheless, boxes are very crude approximations of many objects and may introduce an unwanted bias in the evaluation process, as is displayed in . Furthermore, approaches that are not restricted to oriented or axis-aligned boxes will not necessarily have higher accuracy scores in the benchmarks. To address these problems, a number of densely segmented ground truth datasets has started to emerge [@lin_2014_coco; @perazzi_2016_benchmark; @voji_2017_pixel_wise]. Unfortunately, evaluating the accuracy of object detectors and trackers that are restricted to boxes on densely segmented data is not straightforward. For example, the VOT2016 Benchmark [@vot_2016] generates plausible oriented boxes from densely segmented objects and the COCO 2014 Detection challenge [@lin_2014_coco] uses axis-aligned bounding boxes of the segmentations to simplify the evaluation protocol. Hence, approaches may have a relatively low IoU with the ground truth, although their IoU with the actual object segmentation is the same (or even better) than that of the ground truth box (see (c)). To enable a fair evaluation of algorithms restricted to axis-aligned or oriented boxes on densely segmented data we introduce the relative Intersection over Union accuracy (rIoU) measure. The rIoU uses the best possible axis-aligned or oriented box of the segmentation to normalize the IoU score. The normalized IoU ranges from 0 to 1 for an arbitrary segmentation and allows to determine the true accuracy of a scheme. For tracking scenarios, the optimal boxes have further advantages. By determining three different optimal boxes for each sequence, the optimal oriented box, the optimal axis-aligned box and, the optimal axis-aligned box for a fixed scale, it is possible to identify scale changes, rotations, and occlusion in a sequence without the need of by-frame labels. The optimal boxes are obtained in a fast and efficient optimization process. We validate the quality of the boxes in the experiments section by comparing them to a number of exhaustively determined best boxes for various scenes. The three main contributions of this paper are: 1. The introduction of the relative Intersection over Union accuracy (rIoU) measure, which allows an accurate measurement of object detector and tracker accuracies on densely segmented data. 2. The proposed evaluation removes the bias introduced by restricting the ground truth to boxes for densely segmented data (such as COCO 2014 Detection Challenge [@lin_2014_coco] or VOT2016 [@vot_2016]). 3. A compact, easy-to-use, and efficient evaluation scheme for evaluating object trackers that allows a good interpretability of a trackers strengths and weaknesses. The proposed measure and evaluation scheme is evaluated on a handful of state-of-the-art trackers for the DAVIS [@perazzi_2016_benchmark] and VOT2016 [@vot_2016] datasets and will be made available to the community. Related Work {#sec:relatedwork} ============ In the object detection community, the most commonly used accuracy measure is the Intersection over Union (IoU), also called Pascal overlap or bounding box overlap [@everingham_2015_pascal_voc]. It is commonly used as the standard requirement for a correct detection, when the IoU between the predicted detection and the ground truth is at least $0.5$ [@lin_2014_coco]. In the tracking community, many different accuracy measures have been proposed, most of them center-based and overlap-based measures [@cehovin_2016_visual; @vot_2016; @Kristan_2016_novel; @nawaz_2013_protocol; @smeulders_2014_visual; @wu_otb_2015]. To unify the evaluation of trackers, Čehovin [[et al.]{}]{}[@cehovin_2014_new_tracker; @cehovin_2016_visual] provide a highly detailed theoretical and experimental analysis of the most popular performance measures and show that many of the accuracy measures are highly correlated. Nevertheless, the appealing property of the IoU measure is that it accounts for both position and size of the prediction and ground truth simultaneously. This has lead to the fact that, in recent years, it has been the most commonly used accuracy measure in the tracking community [@vot_2016; @wu_otb_2015]. For example, the VOT2016 [@vot_2016] evaluation framework uses the IoU as the sole accuracy measure and identifies tracker failures when the IoU between the predicted detection and the ground truth is $0.0$ [@Kristan_2016_novel]. Since bounding boxes are very crude approximations of objects [@lin_2014_coco] and cannot accurately capture an object’s shape, location, or characteristics, numerous datasets with densely segmented ground truth have emerged. For example, the COCO 2014 dataset [@lin_2014_coco] includes more than 886,000 densely annotated instances of 80 categories of objects. Nevertheless, on the COCO detection challenge the segmentations are approximated by axis-aligned bounding boxes to simplify the evaluation. As stated earlier, this introduces an unwanted bias in the evaluation. A further dataset with excellent pixel accurate segmentations is the DAVIS dataset [@perazzi_2016_benchmark], which was released in 2017. It consists of 50 short sequences of manually segmented objects which, although originally for video object segmentation, can also be used for the evaluation of object trackers. Furthermore, the segmentations used to generate the VOT2016 ground truths have very recently been released [@voji_2017_pixel_wise]. In our work, we enable the evaluation of object detection and tracking algorithms that are restricted to output boxes on densely segmented ground truth data. The proposed approach is easy to add to existing evaluations and improves the precision of the standard IoU accuracy measure. Relative Intersection over Union (rIoU) ======================================= Using segmentations for evaluating the accuracy of detectors or trackers removes the bias a bounding-box abstraction induces. Nevertheless, the IoU of a box and an arbitrary segmentation generally does not range from $0$ to $1$, where the maximum value depends strongly on the objects’ shape. For example, in (b) the best possible axis-aligned box only has an IoU of 0.66 with the segmentation. To enable a more precise measurement of the accuracy, we introduce the relative Intersection over Union (rIoU) of a box $\mathcal{B}$ and a dense segmentation $\mathcal{S}$ as $$\label{eq:rIoU} \Phi_{rIoU} \left( \mathcal{S}, \mathcal{B}\right ) = \frac{\Phi_{IoU}(\mathcal{S},\mathcal{B})}{\Phi_{opt}(\mathcal{S})},$$ where $\Phi_{IoU}$ is the Intersection over Union (IoU), $$\label{eq:IoU} \Phi_{IoU} \left( \mathcal{S}, \mathcal{B}\right ) = \frac{\left \vert \mathcal{S} \cap \mathcal{B} \right \vert}{\left \vert \mathcal{S} \cup \mathcal{B} \right \vert},$$ and $\Phi_{opt}$ is the best possible IoU a box can achieve for the segmentation $\mathcal{S}$. In comparison to the usual IoU ($\Phi_{IoU}$), the rIoU measure ($\Phi_{rIoU}$) truly ranges from $0$ to $1$ for all possible segmentations. Furthermore, the measure makes it possible to interpret ground truth attributes such as scale change or occlusion, as is displayed later in section \[sec:theoretical\]. The calculation of $\Phi_{opt}$, required to obtain $\Phi_{rIoU}$, is described in the following section. Optimization ------------ An oriented box $\mathcal{B}$ can be parameterized with 5 parameters $$\label{eq:boxparam} b = \left (r_c,c_c,w,h,\phi\right ),$$ where $r_c$ and $c_c$ denote the row and column of the center, $w$ and $h$ denote the width and height, and $\phi$ the orientation of the box with respect to the column-axis. An axis-aligned box can equally be parameterized with the above parameters by fixing the orientation to $0^\circ$. For a given segmentation $\mathcal{S}$, the box with the best possible IoU is $$\label{eq:optimization} \Phi_{opt}(\mathcal{S}) = \max_b\,\,\, \Phi_{IoU} (\mathcal{S},\mathcal{B}(b)) \hspace{1cm} s.t.\,\, b \in \mathbb{R}_{>0}^4 \times [0^\circ,90^\circ).$$ For a convex segmentation, the above problem can efficiently be optimized with the method of steepest descent. To handle arbitrary, possibly unconnected, segmentations, we optimize with a multi-start gradient descent with a backtracking line search. The gradient is approximated numerically by the symmetric difference quotient. We use the diverse set of initial values for the optimization process displayed in . The largest axis-aligned inner box (black) and the inner box of the largest inner circle (magenta) are completely within the segmentation. Hence, in the optimization process, they will gradually grow and include background if it improves $\Phi_{IoU}$. On the other hand, the bounding boxes (green and blue) include the complete segmentation and will gradually shrink in the optimization to include less of the segmentation. The oriented box with the same second order moments as the segmentation (orange) serves as an intermediate starting point [@rosin_1999_measuring]. Hence, only if the initial values converge to different optima do we need to expend more effort. In these cases, we randomly sample further initial values from the interval spanned by the obtained optima with an added perturbation. In our experiements we used 50 random samples. Although this may lead to many different optimizations, the approach is still very efficient. A single evaluation of $\Phi_{IoU} (\mathcal{S},\mathcal{B})$ only requires an average of $0.04$ms for the segmentations within the DAVIS [@perazzi_2016_benchmark] dataset in HALCON[^1] on an IntelCore i7-4810 CPU @2.8GHz with 16GB of RAM with Windows 7 (x64). As a consequence, the optimization of $\Phi_{opt}$ requires an average of 1.3s for the DAVIS [@perazzi_2016_benchmark] and 0.7s for the VOT2016 [@vot_2016] segmentations. [![[blackswan]{} from DAVIS [@perazzi_2016_benchmark]. The initial values of the optimization process of are displayed. We use the axis-aligned bounding box (green), the oriented bounding box (blue), the inner square of the largest inner circle (magenta), the largest inner axis-aligned box (black) and the oriented box with the same second order moments as the segmentation (orange).[]{data-label="fig:initialvalues"}](initial_values_no_overlap.png "fig:"){width="40.00000%"}]{} Validation ---------- To validate the optimization process, we exhaustively searched for the best boxes in a collection of exemplary frames from each of the 50 sequences in the DAVIS dataset [@perazzi_2016_benchmark]. The validation set consists of frames that were challenging for the optimization process. In a first step, we validated the optimization for axis-aligned boxes. The results in indicate that the optimization is generally very close or identical to the exhaustively determined boxes. table [validate\_no\_rot.dat]{}; For the oriented boxes, one of the restrictions we can make is that the area must at least be as large as the smallest inner box of the segmentation and may not be larger than the bounding oriented box. Nevertheless, even with further heuristics, the number of candidates to test is in the number of billions for the sequences in the DAVIS dataset. Given a pixel-precise discretization for $r_c,c_c,w,h$ and a $0.5^\circ$ discretization of $\phi$, it was impossible to find boxes with a better IoU than the optimized oriented boxes in the validation set. This is mostly due to the fact that the sub-pixel precision of the parameterization (especially in the angle $\phi$) is of paramount importance for the IoU of oriented boxes. Theoretical Trackers {#sec:theoretical} ==================== The concept of theoretical trackers was first introduced by Čehovin [[et al.]{}]{}[@cehovin_2016_visual] as an “excellent interpretation guide in the graphical representation of results”. In their paper, they use perfectly robust or accurate theoretical trackers to create bounds for the comparison of the performance of different trackers. In our case, we use the boxes with an optimal IoU to create upper bounds for the accuracy of trackers that underlie the box-world assumption. We introduce three theoretical trackers that are obtained by optimizing for a complete sequence. Given the segmentation $\mathcal{S}$, the first tracker returns the best possible axis-aligned box ([box-axis-aligned]{}), the second tracker returns the optimal oriented box ([box-rot]{}) and the third tracker returns the optimal axis-aligned box with a fixed scale ([box-no-scale]{}). The scale is initialized in the first frame with the scale of the box determined by [box-axis-aligned]{}. [ ![[motorbike]{} from DAVIS [@perazzi_2016_benchmark]. The increasing gap between the [box-no-scale]{} and the other two theoretical trackers indicates a scale change of the motorbike. The drop in all three theoretical trackers around frame 25 indicates that the object is being occluded. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker"}](motorbike008.png "fig:"){width="19.00000%"} ![[motorbike]{} from DAVIS [@perazzi_2016_benchmark]. The increasing gap between the [box-no-scale]{} and the other two theoretical trackers indicates a scale change of the motorbike. The drop in all three theoretical trackers around frame 25 indicates that the object is being occluded. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker"}](motorbike016.png "fig:"){width="19.00000%"} ![[motorbike]{} from DAVIS [@perazzi_2016_benchmark]. The increasing gap between the [box-no-scale]{} and the other two theoretical trackers indicates a scale change of the motorbike. The drop in all three theoretical trackers around frame 25 indicates that the object is being occluded. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker"}](motorbike024.png "fig:"){width="19.00000%"} ![[motorbike]{} from DAVIS [@perazzi_2016_benchmark]. The increasing gap between the [box-no-scale]{} and the other two theoretical trackers indicates a scale change of the motorbike. The drop in all three theoretical trackers around frame 25 indicates that the object is being occluded. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker"}](motorbike034.png "fig:"){width="19.00000%"} ![[motorbike]{} from DAVIS [@perazzi_2016_benchmark]. The increasing gap between the [box-no-scale]{} and the other two theoretical trackers indicates a scale change of the motorbike. The drop in all three theoretical trackers around frame 25 indicates that the object is being occluded. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker"}](motorbike042.png "fig:"){width="19.00000%"} ]{} table [theoretical\_no\_scale.dat]{}; table [theoretical\_no\_rot.dat]{}; table [theoretical\_rect2.dat]{}; The theoretical tracker can be used to normalize a tracker’s IoU for a complete sequence, which enables a fair interpretation of a tracker’s accuracy and removes the bias from the box-world assumption. Furthermore, the three different theoretical trackers make it possible to interpret a tracking scene without the need of by-frame labels. As is displayed in , the difference between the [box-no-scale]{}, [box-axis-aligned]{}, and [box-rot]{} trackers indicates that the object is undergoing a scale change. Furthermore, the decreasing IoUs of all theoretical trackers indicate that the object is either being occluded or deforming to a shape that can be approximated less well by a box. For compact objects, the difference of the [box-rot]{} tracker and the [box-axis-aligned]{} tracker indicates a rotation or change of perspective, as displayed in . [ ![[dog]{} from DAVIS [@perazzi_2016_benchmark]. The gaps between the [box-axis-aligned]{} and [box-rot]{} tracker indicate a rotation of the otherwise relatively compact segmentation of the dog. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker_dog"}](dog000.png "fig:"){width="19.00000%"} ![[dog]{} from DAVIS [@perazzi_2016_benchmark]. The gaps between the [box-axis-aligned]{} and [box-rot]{} tracker indicate a rotation of the otherwise relatively compact segmentation of the dog. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker_dog"}](dog011.png "fig:"){width="19.00000%"} ![[dog]{} from DAVIS [@perazzi_2016_benchmark]. The gaps between the [box-axis-aligned]{} and [box-rot]{} tracker indicate a rotation of the otherwise relatively compact segmentation of the dog. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker_dog"}](dog023.png "fig:"){width="19.00000%"} ![[dog]{} from DAVIS [@perazzi_2016_benchmark]. The gaps between the [box-axis-aligned]{} and [box-rot]{} tracker indicate a rotation of the otherwise relatively compact segmentation of the dog. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker_dog"}](dog035.png "fig:"){width="19.00000%"} ![[dog]{} from DAVIS [@perazzi_2016_benchmark]. The gaps between the [box-axis-aligned]{} and [box-rot]{} tracker indicate a rotation of the otherwise relatively compact segmentation of the dog. The best possible IoU is never above 0.80 for the complete sequence.[]{data-label="fig:theoretical_tracker_dog"}](dog047.png "fig:"){width="19.00000%"} ]{} table [theoretical\_no\_rot\_dog.dat]{}; table [theoretical\_rect2\_dog.dat]{}; Experiments =========== We evaluate the accuracy of a handful of state-of-the art trackers on the DAVIS [@perazzi_2016_benchmark] and VOT2016 [@vot_2016] datasets with the new rIoU measure. We initialize the trackers with the best possible axis-aligned box for the given segmentation. Since we are primarily interested in the accuracy and not in the trackers robustness, we do not reinitialize the trackers when they move off target. Please note that the accuracy of the robustness measure is also improved when using segmentations; The failure cases (hence $\Phi_{IoU} = 0$) are identified earlier since $\Phi_{IoU}$ is zero when the tracker has no overlap with the segmentation and not with a bounding box abstraction of the object (which may contain a large amount of background, see, e.g., ). We restrict our evaluation to the handful of (open source) state-of-the-art trackers displayed in . A thorough evaluation and comparison of all top ranking trackers is beyond the scope of this paper. The evaluation framework is made available and constructed such that it is easy to add new trackers from MATLAB[^2], Python[^3] or HALCON. We include the Kernelized Correlation Filter (KCF) [@henriques_2015_high_speed] tracker since it was a top ranked tracker in the VOT2014 challenge even though it assumes the scale of the object to stay constant. The Discriminative Scale Space Tracker (DSST) [@danelljan_2014_accuracte] tracker is essentially an extension of KCF that can handle scale changes and outperformed the KCF by a small margin in the VOT2014 challenge. As further axis-aligned trackers, we include ANT [@cehovin_2016_ant], L1APG [@bao_2012_l1apg], and the best performing tracker from the VOT2016 challenge, the continuous convolution filters (CCOT) from Danelljan et al.* [@danelljan_2016_beyond]. We include the LGT [@cehovin_2013_tpami] as one of the few open source trackers that estimates the object position as an oriented box. \[t\] ![[bmx-trees]{} from DAVIS [@perazzi_2016_benchmark]. On the left, differences between [box-no-scale]{} and [box-axis-aligned]{} indicate that the object is changing scale and is occluded at frame 18 and around frames 60-70. In the middle plot, we compare the IoU of the axis-aligned box trackers and [box-axis-aligned]{}. The corresponding rIoU plot is shown on the right. It becomes evident that the ANT tracker fails when the object is occluded for the first time and the L1APG tracker at the second occlusion. The rIoU shows that DSST and CCOT perform very well, while the IoU would imply they are relatively weak.[]{data-label="fig:experiments"}](bmx000.png "fig:"){width="19.00000%"} ![[bmx-trees]{} from DAVIS [@perazzi_2016_benchmark]. On the left, differences between [box-no-scale]{} and [box-axis-aligned]{} indicate that the object is changing scale and is occluded at frame 18 and around frames 60-70. In the middle plot, we compare the IoU of the axis-aligned box trackers and [box-axis-aligned]{}. The corresponding rIoU plot is shown on the right. It becomes evident that the ANT tracker fails when the object is occluded for the first time and the L1APG tracker at the second occlusion. The rIoU shows that DSST and CCOT perform very well, while the IoU would imply they are relatively weak.[]{data-label="fig:experiments"}](bmx015.png "fig:"){width="19.00000%"} ![[bmx-trees]{} from DAVIS [@perazzi_2016_benchmark]. On the left, differences between [box-no-scale]{} and [box-axis-aligned]{} indicate that the object is changing scale and is occluded at frame 18 and around frames 60-70. In the middle plot, we compare the IoU of the axis-aligned box trackers and [box-axis-aligned]{}. The corresponding rIoU plot is shown on the right. It becomes evident that the ANT tracker fails when the object is occluded for the first time and the L1APG tracker at the second occlusion. The rIoU shows that DSST and CCOT perform very well, while the IoU would imply they are relatively weak.[]{data-label="fig:experiments"}](bmx030.png "fig:"){width="19.00000%"} ![[bmx-trees]{} from DAVIS [@perazzi_2016_benchmark]. On the left, differences between [box-no-scale]{} and [box-axis-aligned]{} indicate that the object is changing scale and is occluded at frame 18 and around frames 60-70. In the middle plot, we compare the IoU of the axis-aligned box trackers and [box-axis-aligned]{}. The corresponding rIoU plot is shown on the right. It becomes evident that the ANT tracker fails when the object is occluded for the first time and the L1APG tracker at the second occlusion. The rIoU shows that DSST and CCOT perform very well, while the IoU would imply they are relatively weak.[]{data-label="fig:experiments"}](bmx060.png "fig:"){width="19.00000%"} ![[bmx-trees]{} from DAVIS [@perazzi_2016_benchmark]. On the left, differences between [box-no-scale]{} and [box-axis-aligned]{} indicate that the object is changing scale and is occluded at frame 18 and around frames 60-70. In the middle plot, we compare the IoU of the axis-aligned box trackers and [box-axis-aligned]{}. The corresponding rIoU plot is shown on the right. It becomes evident that the ANT tracker fails when the object is occluded for the first time and the L1APG tracker at the second occlusion. The rIoU shows that DSST and CCOT perform very well, while the IoU would imply they are relatively weak.[]{data-label="fig:experiments"}](bmx079.png "fig:"){width="19.00000%"} In , we compare the average IoU with the average rIoU for the DAVIS and the VOT2016 datasets. Please note that we normalize each tracker with the IoU of the theoretical tracker that has the same abilities. Hence, the KCF tracker is normalized with the [box-no-scale]{} tracker, the LGT tracker with [box-rot]{}, and the others with [box-axis-aligned]{}. By these means, it is possible to observe how well each tracker is doing with respect to its abilities. For the DAVIS dataset, the KCF, ANT, L1APG, and LGT trackers all have the same absolute IoU, but when normalized by $\Phi_{opt}$, differences are visible. Hence, it is evident that the KCF is performing very well, given the fact that it does not estimate the scale. On the other hand, the LGT tracker, which has three more degrees of freedom, is relatively weak. A more detailed example analysis of the [bmx-trees]{} sequence from DAVIS [@perazzi_2016_benchmark] is displayed in . For the VOT2016 dataset, the overall accuracies are significantly worse than for DAVIS. On the one hand, this is due to the longer, more difficult sequences, and, on the other hand, due to the less accurate and noisier segmentations (see ). Nevertheless, the rIoU allows a more reliable comparison of different trackers. For example, ANT, LGT and DSST have almost equal average IoU value, while ANT clearly outperforms LGT and DSST with respect to rIoU. Again, we can see that the KCF tracker is quite strong regarding the fact that it cannot estimate the scale. Conclusion ========== In this paper, we have proposed a new accuracy measure that closes the gap between densely segmented ground truths and box detectors and trackers. We have presented an efficient optimization scheme to obtain the best possible detection boxes for arbitrary segmentations that are required for the new measure. The optimization was validated on a diverse set of segmentations from the DAVIS dataset [@perazzi_2016_benchmark]. The new accuracy measure can be used to generate three very expressive theoretical trackers, which can be used to obtain meaningful accuracies and help to interpret scenes without requiring by-frame labels. We have evaluated state-of-the-art trackers with the new accuracy measure on all segmentations within the DAVIS [@perazzi_2016_benchmark] and VOT2016 [@vot_2016] datasets to display its advantages. The complete code and evaluation system will be made available to the community to encourage its use and make it easy to reproduce our results. [^1]: MVTec Software GmbH, <https://www.mvtec.com/> [^2]: The MathWorks, Inc., <https://www.mathworks.com/> [^3]: Python Software Foundation, <https://www.python.org/>	{ "pile_set_name": "ArXiv" }
--- abstract: 'A new method of calculation of amplitudes of different processes in quantum electrodynamics is proposed. The method does not use the Feynman technique of trace of product of matrices calculation. The method strongly simplifies calculation of cross sections for different processes. The effectiveness of the method is shown on the cross-section calculation of Coulomb scattering, Compton scattering and electron-positron annihilation.' author: - 'K. S. Karplyuk' - 'O. O. Zhmudskyy' title: On Calculation of Amplitudes in Quantum Electrodynamics --- The most labor-intensive part of calculation of cross sections for different processes in quantum electrodynamics is the amplitude calculation for these processes. Such calculations for non-polarized electrons, that is, electrons with no definite incoming or outgoing spin states, can be simplified by the Feynman trace technology of the traces of products of $\gamma$-matrices calculation [@f]. In this paper we propose a method which strongly simplifies the amplitude calculation for any elementary processes. Let us start from the identity which is satisfied for matrices with arbitrary complex elements: $$\chi M\psi=\mathrm{Sp}\,\hat{\psi}\hat{\chi} M.$$ Here $M$ — arbitrary square matrix, $\psi$ — matrix-column, $\chi$ — matrix-row. The square matrix $\hat{\psi}$ has only one nonzero column (let it be the n-th) which is equal to $\psi$. The square matrix $\hat{\chi}$ has only one n-th nonzero row which is equal to $\chi$. In this paper we will use $\hat{\psi}$ matrix with the first non-zero column and the $\hat{\chi}$ matrix with the first non-zero row. According to (1) $$\bar{u}^fMu^i=\mathrm{Sp}\,\hat{u}^i\hat{\bar{u}}^f M=\mathrm{Sp}\,\hat{u}^i\hat{u}^{\dag f}\gamma^0 M.$$ Bispinors ${u}^i$ and ${u}^f$ represent the initial and final state of the fermions with momenta $\bm{p}^{i,f}$ and spins aligned along the unit vectors $\bm{s}^{i,f}$. They can be written as follows $$u^{i,f}=\sqrt{\frac{p_0^{i,f}+mc}{2p_0^{i,f}V}}\!\left(1+\frac{\bm{p}^{i,f}\bm{\varsigma}_1}{p_0^{i,f}+mc}\right)\!\! \frac{1+i\bm{s}^{i,f}\bm{\varsigma}_2}{\sqrt{2(1+s_z^{i,f})}} \left[\!\!\begin{array}{c} 1\\0\\ 0\\0 \end{array}\!\!\right].$$ These bispinors are normalized for one particle in a volume $V$, that is $\bar{u}^{i,f}\gamma^0u^{i,f}=u^{{i,f}\dag} u^{i,f}=\displaystyle\frac{1}{V}$. Here and below for brevity the following designations are used $$\begin{gathered} \bm{c}\bm{\gamma}=c_x\gamma^1+c_y\gamma^2+c_z\gamma^3, \end{gathered}$$ $$\begin{gathered} \bm{c}\bm{\varsigma}_1=c_x\gamma^0\gamma^1+c_y\gamma^0\gamma^2+c_z\gamma^0\gamma^3,\\ \bm{c}\bm{\varsigma}_2=c_x\gamma^2\gamma^3+c_y\gamma^3\gamma^1+c_z\gamma^1\gamma^2,\\ \bm{c}\bm{\pi}=c_x\gamma^0\gamma^2\gamma^3+c_y\gamma^0\gamma^3\gamma^1+c_z\gamma^0\gamma^1\gamma^2,\\ \pi^0=\gamma^1\gamma^2\gamma^3,\hspace{7mm}\hat{\iota}=\gamma^0\gamma^1\gamma^2\gamma^3,\end{gathered}$$ where matrices $\gamma$ are used in standard Dirac-Pauli representation. Let us use (3), and calculate the product $\hat{u}^i\hat{\bar{u}}^f$: $$\begin{gathered} \hat{u}^i\hat{\bar{u}}^f=\frac{1}{4V}\sqrt{\frac{(p_0^i+mc)(p_0^f+mc)}{p_0^ip_0^f(1+s_z^i)(1+s_z^f)}} \times\nonumber\\ \times\!{\Bigl(1\!+\!\frac{\bm{p}^i\bm{\varsigma}_1}{p_0^i+mc}\Bigr)\!\! \Bigl[a_0(1\!+\!\gamma^0)\!+\!\bm{a}(\bm{\pi}\!+\!\bm{\varsigma}_2)\Bigr]\!\! \Bigl(1\!+\!\frac{\bm{p}^f\bm{\varsigma}_1}{p_0^f+mc}\Bigr)\gamma_0}\!=\nonumber\\ =\frac{1}{4V}\sqrt{\frac{(p_0^i+mc)(p_0^f+mc)}{p_0^ip_0^f(1+s_z^i)(1+s_z^f)}} \Bigr\{a_0(1+\gamma^0)+\bm{a}(\bm{\pi}+\bm{\varsigma}_2)+\nonumber\\ +\frac{(\bm{a}\cdot\bm{p}^i)(\hat{\iota}-\pi^0)+ (a_0\bm{p}^i-\bm{a}\times\bm{p}^i) (\bm{\varsigma}_1-\bm{\gamma})}{p_0^i+mc}-\Bigl.\nonumber\\ -\frac{(\bm{a}\cdot\bm{p}^f)(\hat{\iota}+\pi^0)+ (a_0\bm{p}^f+\bm{a}\times\bm{p}^f) (\bm{\varsigma}_1+\bm{\gamma})}{p_0^f+mc}-\nonumber\\ -\frac{a_0(\bm{p}^i\cdot\bm{p}^f)-\bm{a}\cdot (\bm{p}^i\times\bm{p}^f)}{(p_0^i+mc)(p_0^f+mc)}(1-\gamma^0)-\nonumber\\ \Bigr.-\!\frac{a_0(\bm{p}^i\!\times\!\bm{p}^f)\!+\! (\bm{p}^i\!\cdot\!\bm{p}^f)\bm{a}\!-\!(\bm{a}\cdot\bm{p}^f)\bm{p}^i\!-\! (\bm{a}\!\cdot\!\bm{p}^i)\bm{p}^f} {(p_0^i+mc)(p_0^f+mc)}(\bm{\pi}\!-\!\bm{\varsigma}_2)\!\Bigl\}.\end{gathered}$$ Here $$\begin{gathered} a_0=\frac{i}{4}\bm{e}_z\!\cdot\!(\bm{s}^f\!\times\!\bm{s}^i)+\frac{1}{4}(1+\bm{s}^i\!\cdot\!\bm{s}^f+\bm{e}_z\!\cdot\!\bm{s}^i+\bm{e}_z\!\cdot\!\bm{s}^f),\\ \bm{a}\!=\!\frac{i}{4}\bigl[\bm{e}_z\!\!+\!\bm{s}^i\!\!+\!\bm{s}^f\!\!+\!(\bm{e}_z\!\cdot\!\bm{s}^f)\bm{s}^i\!\!+\!(\bm{e}_z\!\cdot\!\bm{s}^i)\bm{s}^f\!- \!\bm{e}_z(\bm{s}^i\!\cdot\!\bm{s}^f)\bigr]\!+\nonumber\\ +\frac{1}{4}\bigl(\bm{s}^f\!\!\times\bm{s}^i-\bm{e}_z\times\bm{s}^f+\bm{e}_z\times\bm{s}^i\bigr),\end{gathered}$$ where $\bm{e}_z$ is the unit vector along the $z$ axis. Note that $a_0$ and $\bm{a}$ depend on the direction of the spins $\bm{s}^i$ and $\bm{s}^f$ only, and do not depend on energies and momenta of initial and final fermions. As far as $\hat{u}^i\hat{\bar{u}}^f$is known, the trace of the matrix $\hat{u}^i\hat{\bar{u}}^fM$ can be calculated. In the general case matrix $M$ has a form: $$M=I+V_0\gamma^0+\bm{V}\bm{\gamma}+W_0\pi^0+\bm{W}\bm{\pi}+ \bm{E}\bm{\varsigma}_1+\bm{B}\bm{\varsigma}_2+J\hat{\iota}.$$ The unit matrix is the only one of the sixteen Dirac matrices which has nonzero trace. That is why in the matrix product of $\hat{u}^i\hat{\bar{u}}^f$ and $M$ it is enough to take into account only those terms which are proportional to the unit matrix. Such terms appear only for a multiplication of the same matrices. This remark simplifies the multiplication and trace calculation: $$\begin{gathered} \bar{u}^fMu^i=\mathrm{Sp}\,\hat{u}^i\hat{\bar{u}}^fM=\nonumber\\ =\sqrt{\frac{(p_0^i+mc)(p_0^f+mc)}{p_0^ip_0^f(1+s_z^i)(1+s_z^f)}} \frac{1}{V}\Bigl(a_0K_0-\bm{a}\cdot\bm{K}\Bigr).\end{gathered}$$ In three dimensions the scalar $K_0$ and the pseudo-vector $\bm{K}$ can be written as $$\begin{gathered} K_0=(I+V_0)+\frac{\bm{p}^i\cdot(\bm{E}+\bm{V})}{p_0^i+mc} -\frac{\bm{p}^f\cdot(\bm{E}-\bm{V})}{p_0^f+mc}+\nonumber\\ +\frac{(\bm{p}^i\times\bm{p}^f) \cdot(\bm{W}-\bm{B})-(\bm{p}^i\cdot\bm{p}^f)(I-V_0)}{(p_0^i+mc)(p_0^f+mc)},\\ \bm{K}\!=\!(\bm{W}\!+\!\bm{B})+\frac{\bm{p}^i(W_0\!+\!J)+\bm{p}^i\!\times\!(\bm{E}\!+\!\bm{V})}{p_0^i+mc}+\nonumber\\ +\frac{\bm{p}^f(W_0\!-\!J)+ \bm{p}^f\!\times\!(\bm{E}\!-\!\bm{V})}{p_0^f+mc}+\nonumber\\ +\frac{\bm{p}^i[\bm{p}^f\!\cdot\!(\bm{W}\!\!-\!\!\bm{B})]\!+\!\bm{p}^f[\bm{p}^i\!\cdot\!(\bm{W}\!\!-\!\!\bm{B})]\!-\! (\bm{p}^i\!\cdot\!\bm{p}^f)(\bm{W}\!\!-\!\!\bm{B})}{(p_0^i+mc)(p_0^f+mc)}-\nonumber\\ -\frac{(\bm{p}^i\times\bm{p}^f)(I-V_0)}{(p_0^i+mc)(p_0^f+mc)}.\end{gathered}$$ Note that $K_0$ and $\bm{K}$ depend on energies and momenta of initial and final fermion states only and do not depend on their polarization $\bm{s}^i$ and $\bm{s}^f$. In order to evaluate the probability and cross-section of the process the square of the amplitude of transaction must be calculated. Let us calculate $\|(a_0K_0-\bm{a}\bm{K})\|^2$ supposing that all coefficients in (7) are real. This is usually the case in quantum electrodynamics. Thus the modulus square $\|\bar{u}^fMu^i\|^2$ is $$\begin{gathered} \|\bar{u}^fMu^i\|^2=\nonumber\\ =\!{\frac{(p_0^i+mc)(p_0^f+mc)}{p_0^ip_0^f(1\!+s_z^i)(1\!+s_z^f)V^2}}(a_0K_0\!-\!\bm{a}\!\cdot\!\bm{K})(a_0^K_0\!-\!\bm{a}^\!\cdot\!\bm{K})\!=\nonumber\\ =\!{\frac{(p_0^i\!+\!mc)(p_0^f\!+\!mc)}{8V^2p_0^ip_0^f}}\Bigl[(1\!+\!\bm{s}^i\bm{s}^f)K_0^2\!+\!(1\!-\!\bm{s}^i\bm{s}^f)\bm{K}\!\cdot\!\bm{K}\!+\!\nonumber\\ +2(\bm{s}^i\bm{K})(\bm{s}^f\bm{K})+2(\bm{s}^i\times\bm{s}^f)\bm{K}K_0\Bigr].\end{gathered}$$ Expressions (8)-(11) are universal. They determine $\bar{u}^fMu^i$ and $\|\bar{u}^fMu^i\|^2$ for any processes in the quantum electrodynamics. Different processes differ by matrix $M$ only. The only thing we need to do in order to calculate $\bar{u}^fMu^i$ and $\|\bar{u}^fMu^i\|^2$ is to represent the interaction matrix $M$ in a form (7). Then substitute coefficients from matrix (7) into expressions (9)-(10). These expressions entirely determine $K_0$ and $\bm{K}$. Equations (8) and (11) give the algebraic expressions for $\bar{u}^fMu^i$ and $\|\bar{u}^fMu^i\|^2$. All that remains is the simplification of $\bar{u}^fMu^i$ and $\|\bar{u}^fMu^i\|^2$ as much as possible. Expression (11) determines the square of the amplitude of the process which corresponds to any desirable spin states of the incoming and outgoing fermions. This expression explicitly represents dependence on the fermion polarization that is why this dependence can be easily analyzed. If a detector is blind to polarization, i.e. for the final state both polarizations of fermions in $\bm{s}^f$ direction and in $-\bm{s}^f$ direction are registered, expression (11) must include the sum for both directions $\bm{s}^f$ and $-\bm{s}^f$: $$\|\bar{u}^fMu^i\|^2=\frac{(p_0^i+mc)(p_0^f+mc)}{4V^2p_0^ip_0^f}[K_0^2+\bm{K}\cdot\bm{K}].$$ Let us demonstrate the effectiveness of the above method on three examples: Coulomb scattering, Compton scattering and electron - positron annihilation. Coulomb Scattering ------------------ We will define the cross section of an electron of charge $e$ scattering on the Coulomb center of charge $Ze$ versus the square of the amplitude $\|\bar{u}^f\gamma^0u^i\|^2$ in a usual way: $$\begin{gathered} \frac{d\sigma}{d\Omega}=\frac{(2Zr_0mc^2)^2}{(2p\sin\frac{\theta}{2})^4}\left(\frac{Vp_o}{c}\right)^2\|\bar{u}^f\gamma^0u^i\|^2=\nonumber\\ =\left(\frac{Zr_0}{2}\right)^2\left(\frac{c}{v} \frac{1}{\sin\frac{\theta}{2}}\right)^4\left(1-\frac{v^2}{c^2}\right)V^2\|\bar{u}^f\gamma^0u^i\|^2.\end{gathered}$$ Here $r_0$ is the classical electron radius, $\theta$ is the scattering angle. The amplitude square $\|\bar{u}^f\gamma^0u^i\|^2$ is defined by the universal expression (11), in which $K_0$ and $\bm{K}$ must be calculated for the matrix $M=\gamma^0$. Hence, in expressions (9)-(10) we must set $V_0=1$. All other coefficients must be set to zero. We must also take into account that $p_0^i=p_0^f=p_0$ because for the Coulomb scattering energy is conserved: $$K_0=1+\frac{\bm{p}^i\cdot\bm{p}^f}{(p_0\!+\!mc)^2},\hspace{7mm} \bm{K}=\frac{\bm{p}^i\times\bm{p}^f}{(p_0\!+\!mc)^2}.$$ So, the cross section becomes: $$\begin{gathered} \frac{d\sigma}{d\Omega}=\frac{1}{2}\left(\frac{Z r_0}{2}\right)^2\left(\frac{c}{v} \frac{1}{\sin\frac{\theta}{2}}\right)^4\left(1-\frac{v^2}{c^2}\right) \left(\frac{p_0+mc}{2p_0}\right)^2 \times\nonumber\\ \times\Bigl[(1\!+\!\bm{s}^i\!\cdot\!\bm{s}^f)K_0^2\!+\!(1-\bm{s}^i\!\cdot\!\bm{s}^f)K^2\!+\! 2(\bm{s}^i\!\cdot\!\bm{K})(\bm{s}^f\!\cdot\!\bm{K})\!+\nonumber\\ +2(\bm{s}^f\times\bm{s}^i)\cdot\bm{K}K_0\Bigr].\end{gathered}$$ In expression (15) $K_0$ and $\bm{K}$ are determined according to equation (14). Expression (15) determines the differential cross section $d\sigma/d\Omega$ in the case with definite incoming and outgoing electron spin states. If the polarization of the final electron is not registered, expression (12) must be used: $$\frac{d\sigma}{d\Omega}=\left(\frac{Z r_0}{2}\right)^{\!\!2}\!\!\!\left(\frac{c}{v} \frac{1}{\sin\frac{\theta}{2}}\right)^{\!\!4}\!\!\!\left(1\!-\!\frac{v^2}{c^2}\right)\!\!\! \left(\frac{p_0\!+\!mc}{2p_0}\right)^{\!\!2}\!\!\!(K_0^2+K^2).$$ Expression $\displaystyle\left(\frac{p_0+mc}{2p_0}\right)^2 (K_0^2+\bm{K}\cdot\bm{K})$ can be simplified: $$\begin{gathered} \left(\frac{p_0+mc}{2p_0}\right)^2(K_0^2+\bm{K}\cdot\bm{K})=\nonumber\\ =\left(\frac{p_0+mc}{2p_0}\right)^2 \left\{ \left[1+\frac{p^2\cos\theta}{(p_0+mc)^2}\right]^2+ \frac{p^4\sin^2\theta}{(p_0+mc)^4}\right\}=\nonumber\\ =\frac{p_0^2+p_0^2-p^2(1-\cos\theta)}{2p_0^4}= \left(1-\frac{v^2}{c^2}\sin^2\frac{\theta}{2}\right)\end{gathered}$$ Recall that $\theta$ is an angle between $\bm{p}^i$ and $\bm{p}^f$. After substituting (17) into (16) we come up with unpolarized cross section for Coulomb scattering: $$\frac{d\sigma}{d\Omega}=\left(\frac{Z r_0}{2}\right)^2\left(\frac{c}{v} \frac{1}{\sin\frac{\theta}{2}}\right)^4\left(1-\frac{v^2}{c^2}\right) \left(1-\frac{v^2}{c^2}\sin^2\frac{\theta}{2}\right).$$ This is the well-known Mott scattering cross section [@m]. Note that for the polarized cross section calculation (15) and the unpolarized cross section calculation (18) instead of the Feynman technique (of trace of product of matrices calculation) we use expressions (9)-(12) which strongly simplify calculations. Compton Scattering ------------------ It is well-known that Compton scattering in the first order of probability theory is represented by two Feynman diagrams. Call them $a$ and $b$. According to the $a$ diagram an electron absorbs a photon of frequency $\omega_1$ first, and then emits a photon of frequency $\omega_2$. According to the $b$ diagram an electron emits a photon of frequency $\omega_2$ first, and then absorbs a photon of frequency $\omega_1$. The amplitudes for the two diagrams must be added and their sum squared. Assume that the incoming electron is at rest, hence $\bm{p}^i=0$, $p_0^i=mc$. Let us express the scattering cross section versus the square of the sum of the amplitudes $\|\bar{u}^fMu^i\|^2$ in a usual way: $$\begin{gathered} \frac{d\sigma}{d\Omega}=r_0^2mcp_0^fV^2\left(\frac{\omega_2}{\omega_1}\right)^2\|\bar{u}^fMu^i\|^2.\end{gathered}$$ Here $r_0$ is the classical electron radius. Matrix $M$ for the two diagrams is $$\begin{gathered} M=\\ =\bm{\mathfrak{e}}_2\bm{\gamma}\frac{p_{0a}\gamma^0\!-\!\bm{p}_a\bm{\gamma}\!+\!mc} {2mc\hbar k_1}\bm{\mathfrak{e}}_1\bm{\gamma}+\bm{\mathfrak{e}}_1\bm{\gamma}\frac{p_{0b}\gamma^0\!-\!\bm{p}_b\bm{\gamma}\!+\!mc} {(-2mc\hbar k_2)}\bm{\mathfrak{e}}_2\bm{\gamma}.\end{gathered}$$ Here $p_{0a}=mc+\hbar k_1$, $\bm{p}_a=\hbar\bm{k}_1$, $p_{0b}=mc-\hbar k_2$, $\bm{p}_b=-\hbar\bm{k}_2$, $\bm{k}_1$ and $\bm{k}_2$ are the wave vectors of photons $1$ and $2$, $k_1=\omega_1/c$, $k_2=\omega_2/c$, $\bm{\mathfrak{e}}_1$ and $\bm{\mathfrak{e}}_2$ are the unit vectors of polarization of photons $1$ and $2$, $\bm{\mathfrak{e}}_1\cdot\bm{k}_1=0$, $\bm{\mathfrak{e}}_2\cdot\bm{k}_2=0$. The polar angle in $d\Omega$ is measured from the $\bm{k}_1$ direction. Matrix multiplication in the expression for $M$ leads to the coefficients in equation (7): $$\begin{gathered} I=\frac{(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)}{2\hbar}\Bigl(\frac{1}{k_2}-\frac{1}{k_1}\Bigr),\hspace{3mm} V_{0}=\frac{(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)}{2mc\hbar}\Bigl(\frac{p_{0a}}{ k_1}-\frac{p_{0b}}{ k_2}\Bigr),\nonumber\\ \bm{V}=\frac{(\bm{p}_a\cdot\bm{\mathfrak{e}}_1)\bm{\mathfrak{e}}_2- \bm{\mathfrak{e}}_2\times(\bm{p}_a\times\bm{\mathfrak{e}}_1)}{2mc\hbar k_1}+\nonumber\\ +\frac{(\bm{p}_b\cdot\bm{\mathfrak{e}}_2)\bm{\mathfrak{e}}_1- \bm{\mathfrak{e}}_1\times(\bm{p}_b\times\bm{\mathfrak{e}}_2)}{(-2mc\hbar k_2)},\nonumber\\ \bm{W}\!=\!\frac{(\bm{\mathfrak{e}}_1\times\bm{\mathfrak{e}}_2)}{2mc\hbar}\Bigl(\frac{ p_{0a}}{ k_1}+\frac{ p_{0b}}{ k_2}\Bigr), \bm{B}=-\frac{(\bm{\mathfrak{e}}_1\times\bm{\mathfrak{e}}_2)}{2\hbar}\Bigl(\frac{1}{k_1}+\frac{1}{k_2}\Bigr),\nonumber\\ W_{0}=\frac{-\bm{\mathfrak{e}}_2\cdot(\bm{p}_a\times\bm{\mathfrak{e}}_1)}{2mc\hbar k_1}+\frac{\bm{\mathfrak{e}}_1\cdot(\bm{p}_b\times\bm{\mathfrak{e}}_2)}{2mc\hbar k_2},\hspace{3mm}\bm{E}=0,\hspace{3mm}J=0.\end{gathered}$$ Substitution of these coefficients into (9)-(10) gives us expressions for $K_0$ and $\bm{K}$: $$\begin{gathered} K_0=(I+V_0)+\frac{\bm{p}^f\cdot\bm{V}}{p_0^f+mc}=2\frac{\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2}{2mc}-\nonumber\\ -\frac{1}{2mc} \frac{\bm{p}^f\cdot[\bm{\mathfrak{e}}_2\times(\bm{\mathfrak{k}}_1\times\bm{\mathfrak{e}}_1)]} {p_0^f+mc}-\frac{1}{2mc} \frac{\bm{p}^f\cdot[\bm{\mathfrak{e}}_1\times(\bm{\mathfrak{k}}_2\times\bm{\mathfrak{e}}_2)]}{p_0^f+mc},\\ \bm{K}=(\bm{W}+\bm{B})+\frac{\bm{p}^fW_0-\bm{p}^f\times \bm{V}}{p_0^f+mc}=\nonumber\\ =\frac{1}{2mc} \frac{\bm{p}^f\times[\bm{\mathfrak{e}}_2\times(\bm{\mathfrak{k}}_1\times\bm{\mathfrak{e}}_1)]- \bm{p}^f[\bm{\mathfrak{e}}_2\cdot(\bm{\mathfrak{k}}_1\times\bm{\mathfrak{e}}_1)]}{p_0^f+mc}+\nonumber\\ +\frac{1}{2mc}\frac{\bm{p}^f\times[\bm{\mathfrak{e}}_1\times(\bm{\mathfrak{k}}_2\times\bm{\mathfrak{e}}_2)]- \bm{p}^f[\bm{\mathfrak{e}}_1\cdot(\bm{\mathfrak{k}}_2\times\bm{\mathfrak{e}}_2)]} {p_0^f+mc}.\end{gathered}$$ In expressions (20) and (21) $\displaystyle\bm{\mathfrak{k}}_1=\frac{\bm{k}_1}{k_1}$, $\displaystyle\bm{\mathfrak{k}}_2=\frac{\bm{k}_2}{k_2}$. Expressions for $K_0$ and $\bm{K}$ together with (11) and (19) determine the polarized cross section. In order to get the unpolarized cross section, expressions (12) and (19) must be used. The sum $K_0^2+K^2$ can be simplified and expressed as $$K_0^2+K^2=\frac{1}{2mc}\frac{1}{p_0^f+mc}\Bigl[4(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)^2+\frac{(\omega_1-\omega_2)^2}{\omega_1\omega_2}\Bigr].$$ Thus, the unpolarized cross section is $$\frac{d\sigma}{d\Omega}=\frac{r_0^2}{4}\left(\frac{\omega_2}{\omega_1}\right)^2 \Bigl[4(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)^2+\frac{(\omega_1-\omega_2)^2}{\omega_1\omega_2}\Bigr].$$ This is the well-known Klein-Nishina scattering cross section [@kn]. As in the previous case we don’t use the Feynman technique of trace of product of matrices calculation. Using the universal expressions (11) and (12) instead strongly simplified calculations. Annihilation ------------ Two Feynman diagrams represent the annihilation process in the first order of perturbation theory. The first diagram, call it a, corresponds to the process in which an incoming electron emits a photon $\gamma_1$ of frequency $\omega_1$, then a photon $\gamma_2$ of frequency $\omega_2$, and transfers to the state with negative energy. The second diagram, call it b, corresponds to the process in which $\gamma_1$ and $\gamma_2$ interchange. In order to calculate the annihilation probability we must add the amplitudes of these processes and then square it. For simplicity, assume that the electron is at rest, so $\bm{p}^i=0$, $p_0^i=mc$. Following Feynman, we treat the positron as an electron with negative energy moving backward in time. This electron has linear momentum and spin opposite in direction to the positron’s momentum and spin. It allows us to describe the positron by the same bispinor (3) if we set up $p_0^f=-p^+$, $\bm{p}^f=-\bm{p}^+$, $\bm{s}^f=-\bm{s}^+$, where the index + designates positron quantities. The same substitutions have to be done in expressions (9) and (10). Let us express the annihilation cross section versus the square of the sum of the amplitudes $\|\bar{u}^fMu^i\|^2$ in a usual way: $$\frac{d\sigma}{d\Omega}=r_0^2\frac{m\hbar^2\omega_1^2p_0^+V^2}{\|p^+\|(p_0^++mc)c}\|\bar{u}^fMu^i\|^2.$$ Matrix $M$, which corresponds to the sum of two diagrams is: $$\begin{gathered} M=\\ =\bm{\mathfrak{e}}_2\bm{\gamma}\frac{p_{0a}\gamma^0\!-\!\bm{p}_a\bm{\gamma}\!+\!mc} {(-2mc\hbar k_1)}\bm{\mathfrak{e}}_1\bm{\gamma}+\bm{\mathfrak{e}}_1\bm{\gamma}\frac{p_{0b}\gamma^0\!-\!\bm{p}_b\bm{\gamma}\!+\!mc} {(-2mc\hbar k_2)}\bm{\mathfrak{e}}_2\bm{\gamma}.\end{gathered}$$ Here $p_{0a}=mc-\hbar k_1$, $\bm{p}_a=-\hbar\bm{k}_1$, $p_{0b}=mc-\hbar k_2$, $\bm{p}_b=-\hbar\bm{k}_2$, $\bm{k}_1$ and $\bm{k}_2$ are the wave vectors of the photons $1$ and $2$, $k_1=\omega_1/c$, $k_2=\omega_2/c$, $\bm{\mathfrak{e}}_1$ and $\bm{\mathfrak{e}}_2$ are the unit vectors of the polarization of the photons $1$ and $2$, $\bm{\mathfrak{e}}_1\cdot\bm{k}_1=0$, and $\bm{\mathfrak{e}}_2\cdot\bm{k}_2=0$. The polar angle in the $d\Omega$ is measured from the $\bm{k}_1$ direction. After transformation of the matrix $M$ to the form (7) we can find coefficients in (7) in the reference frame in which incoming electron is at rest $\bm{p}^i=0$: $$\begin{gathered} I+V_0=2\frac{\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2}{2mc},\hspace{4mm} \bm{W}+\bm{B}=0,\hspace{4mm}\bm{E}=0,\hspace{4mm}J=0,\\ \bm{V}=-\frac{\bm{\mathfrak{e}}_1\times(\bm{\mathfrak{k}}_2\times\bm{\mathfrak{e}}_2)+ \bm{\mathfrak{e}}_2\times(\bm{\mathfrak{k}}_1\times\bm{\mathfrak{e}}_1)}{2mc}=\\ =-\frac{(\bm{\mathfrak{k}}_1+\bm{\mathfrak{k}}_2)(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)- \bm{\mathfrak{e}}_1(\bm{\mathfrak{e}}_2\cdot\bm{\mathfrak{k}}_1) -\bm{\mathfrak{e}}_2(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{k}}_2)}{2mc},\\ W_0=-\frac{\bm{\mathfrak{e}}_1\cdot(\bm{\mathfrak{k}}_2\times\bm{\mathfrak{e}}_2) +\bm{\mathfrak{e}}_2\cdot(\bm{\mathfrak{k}}_1\times\bm{\mathfrak{e}}_1)}{2mc}.\end{gathered}$$ Here $\displaystyle\bm{\mathfrak{k}}_{1,2}=\frac{\bm{k}_{1,2}}{k_{1,2}}$, and $\displaystyle k_{1,2}=\frac{\omega_{1,2}}{c}$. Using (9) and (10) scalar $K_0$ and pseudo-vector $\bm{K}$ can be calculated. $$\begin{gathered} K_0=\frac{\bigl[2(p_0^f+mc)+\hbar(k_1\!+\!k_2)(1+\cos\theta)\bigr](\bm{\mathfrak{e}}_1\!\cdot\!\bm{\mathfrak{e}}_2)}{2mc(p_0^f+mc)}-\nonumber\\ -\frac{\hbar(k_1\!+\!k_2)(\bm{\mathfrak{k}}_1\!\cdot\!\bm{\mathfrak{e}}_2)(\bm{\mathfrak{k}}_2\!\cdot\!\bm{\mathfrak{e}}_1)}{2mc(p_0^f+mc)}.\\ \bm{K}=\frac{\bm{p}^f\!\times\bigl[\bm{\mathfrak{e}}_1\times(\bm{\mathfrak{k}}_2\!\times\!\bm{\mathfrak{e}}_2)\bigr]+ \bm{p}^f\!\times\bigl[\bm{\mathfrak{e}}_2\times(\bm{\mathfrak{k}}_1\!\times\!\bm{\mathfrak{e}}_1)\bigr]}{2mc(p_0^f+mc)}+\nonumber\\ +\frac{\bm{p}^f\bigl[\bm{\mathfrak{e}}_1\cdot(\bm{\mathfrak{e}}_2\!\times\!\bm{\mathfrak{k}}_2)\bigr]+ \bm{p}^f\bigl[\bm{\mathfrak{e}}_2\cdot(\bm{\mathfrak{e}}_1\!\times\!\bm{\mathfrak{k}}_1)\bigr]}{2mc(p_0^f+mc)}.\end{gathered}$$ These expression for $K_0$ and $\bm{K}$ together with (11) and (24) determine the polarized cross section (both the electron and the positron have a given direction of spin). In order to calculate the unpolarized cross section expressions (12) and (24) must be used. The sum $K_0^2+K^2$ can be reduced to $$K_0^2+K^2=\frac{1}{2mc}\frac{1}{p_0^f+mc}\Bigl[4(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)^2-\frac{(\omega_1+\omega_2)^2}{\omega_1\omega_2}\Bigr].$$ Thus the unpolarized cross section is: $$\frac{d\sigma}{d\Omega}=\frac{r_0^2}{4}\frac{\hbar^2k_1^2}{\|p^+\|(p_0^++mc)} \Bigl[\frac{(\omega_1+\omega_2)^2}{\omega_1\omega_2}-4(\bm{\mathfrak{e}}_1\cdot\bm{\mathfrak{e}}_2)^2\Bigr].$$ This result entirely coincides with the one calculated by the technique of trace of product of matrices calculation [@f] and with the one calculated by Dirac [@d]. As it was shown in the above examples, the method proposed in this paper allows us to strongly simplify calculation of the polarized and unpolarized fermion cross sections in quantum electrodynamics. It is free from the necessity of calculation of trace of product of great amount of matrices. [12]{} 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: 'First, I give an elementary proof for the fact that the minimal crossing number is additive under composition of torus links. This result is generalized to the composition of homogeneous braids with alternating fibered links. Then there follow estimates for the crossing number of satellite knots. In the last chapter, I discuss a conjecture concerning the HOMFLY and the Kauffman polynomial.' author: - Hermann Gruber title: Estimates for the minimal crossing number --- Introduction. ============= A link is called composite iff there is a 2-sphere in $S^{3}$ which meets the link in exactly two points and decomposes it into two sublinks K and L, neither of which is an unknotted arc. Write K\#L if the link is the composition of K and L. Whether the minimal crossing number of a composite link is simply the sum of the minimal crossing numbers of the factor links, is a very natural and very old question already posed by Tait some hundred years ago. It is trivial to see that it cannot be greater than the sum, but the question if equality holds in general remains unsolved up to now. However, the question is answered if the factor links are all from a certain family of knots, called adequate knots[@Kirby], problem 1.65. Here, another such family is exposed, which contains the torus links and the alternating fibered links. Some of the arguments exposed here are closely following the ideas of Murasugi’s paper [@MuBraid], and some results are re-expressed and proved more shortly here. It turns out to be astonishingly easy to show that the crossing number of torus links is additive under composition. But no one has written up these things up to now, and it is one purpose of this paper to do this. Another aim is to turn around the classical points of view, gaining beautiful formulas estimating the minimal crossing number via the geometric concepts of braid index and (canonical) genus. For example, you will see that, in this application here, the canonical genus can be a considerably more powerful concept than the Seifert genus. Preliminaries. ============== Throughout this paper, it is immaterial if the knots are chiral or not. For this cause, don’t regard chiral pairs as distinct. For a (nontrivial and non-split) link K and a regular diagram D(K) (or simply D), let c(D) be the number of crossings in D. Write c(K) for the minimal crossing number, and b(K) for the braid index of K. Give the knot K an arbitrary orientation. By cutting out each crossing, respecting the orientation, convert the diagram D into a number of oriented closed curves in the (extended) plane, called Seifert circles of D. Write s(D) for the number of Seifert circles of D. From these circles, construct a spanning surface for K. The genus of this surface depends of the diagram: $g(D)=\frac{1}{2}(c(D)-s(D)-\|K\|+2)$, where \|K\| is the number of components of K. This can easily be seen by calculating the Euler characteristics of the constructed surface. The (Seifert) genus of a link K is defined as the minimal genus among all orientable surfaces bounded by K, the canonical (or weak) genus as $$\widetilde{g}(K)=\min _{D=D(K)}\{g(D)\}$$ and the free genus $g_{f}(K)$ is the minimum genus of all Seifert surfaces for K whose complement in $S^{3}$ is a handlebody. We mention $g(K)\leq g_{f}(K)\leq \widetilde{g}(K)$. Torus links, alternating fibered links and homogeneous braids. ============================================================== I shall note first the following self-evident lemma, which plays a central role in our discussion: For any link K, $$c(K)=\min _{D=D(K)}\left(2g(D)+s(D)\right)+\|K\|-2$$. By minimizing both the number of Seifert circles and the genus independently, we get the following estimate: \[bound1\] For every link K, $$c(K)\geq 2\widetilde{g}(K)+b(K)+\|K\|-2\geq 2g(K)+b(K)+\|K\|-2$$ What remains to show is $$\min _{D=D(K)}\{s(D)\}=b(K)$$ See Yamada’s proof [@Yamada]. Both of the above inequalities are sharp for torus links, so we can give a short and elementary proof of a result of Murasugi: [@MuBraid]Let K be a (p,q) torus link with $p\geq q\geq 2$. Then $$c(K)=2g(K)+b(K)+\|K\|-2=pq-p$$ Consider the standard representation D of K as a closed braid with q strands. Here, $c(D)=pq-p$ and $2g(D)=pq-p-q+\|K\|-2$. g(D) is minimal over all Seifert surfaces, see [@Cro]. As the diagram D shows, the braid index can be at most q, and since it cannot be lower than the bridge number, it must be equal to q. This result motivates the following definition: Say that a link K is in the family $\mathcal{F}$ (or, shortly, $\mathcal{F}$-link) if $c(K)= 2g(K)+b(K)+\|K\|-2$. Next, see that the weaker of the lower bounds is additive under composition: Let K be a composite link with the factor links $K_{1},...,K_{n}$. Then $$c(K)\geq \sum _{i=1}^{n}[2g(K_{i})+b(K_{i})+\|K_{i}\|-2]$$ Use the facts that the genus is additive under composition (see e. g. [@Ad]), and for the braid index it holds $b(K_{1}\#K_{2})=b(K_{1})+b(K_{2})-1$[@BirBraid]. The free genus is also additive under knot composition, see [@freegenus]. Thus for knots, we may also write here $g_{f}(K)$ instead of $g(K)$, sometimes leading to better results. If K is a composite link with factor links $K_{1},...,K_{n}$, and all factor links are in $\mathcal{F}$, then $$c(K)=\sum _{i=1}^{n}c(K_{i})$$ At this stage, we already see that the crossing number is additive for torus links; a problem repeatedly posed by Adams (see [@Ad], chapter 5.1); we see that the answer is almost as simple as the question in this special case. With two other results by Murasugi, we have that $\mathcal{F}$ also contains other classes of links: We repeat here almost exactly corollary 2 in [@MuBraid]: [@MuBraid] If K is an alternating fibered link, then K is in $\mathcal{F}$. Murasugi’s statement is $c(K)=deg\Delta _{K}+b(K)-1$, and with $deg\Delta _{K}-\|K\|+1\leq 2g(K)$, the rest follows. The following proposition is very closely related to proposition 7.4 in [@MuBraid]. However, I redo it here to point out a different aspect: to fit the definition of the family $\mathcal{F}$. Let $\gamma $ be a homogeneous n-braid, and L be the closure of $\gamma $. If $b(L)=n$, then L is in $\mathcal{F}$. Draw a diagram D by simply closing $\gamma $. You have for the degree of the reduced Alexander polynomial $deg\Delta _{L}=c(D)-s(D)+1$, since $\gamma $ is homogeneous. With $deg\Delta _{L}+1-\|K\|\leq 2g(L)$ , $s(D)=b(L)$, this leads to $2g(L)\geq c(D)-b(L)-\|K\|+2$, and because of $g(D)\geq g(L)$, $c(D)=2g(L)+b(L)+\|K\|-2$. By Lemma \[bound1\], you see that c(D) is minimal. Examples. ========= In the previous chapter, we have seen some families of $\mathcal{F}$-links. You may have observed that all links presented there were homogeneous (for a definition see[@Cro]), and fibered. However, it is neither true that every fibered homogeneous link is an $\mathcal{F}$-link nor that every $\mathcal{F}$-link is fibered. We take a look at an example: The Perko knot (denoted by $10_{161}$ in Rolfsen’s table [@Ro]) is fibered and homogeneous, and its (canonical) genus is 3, as well as its braid index. But its crossing number is 10. Therefore, the Perko knot is not in $\mathcal{F}$. I proceed with a discussion of alternating $\mathcal{F}$-links. A number $n(D)$ associated with an alternating link diagram D is called an alternating link invariant if $n(D_1)=n(D_2)$ for any two reduced alternating diagrams $D_1,D_2$ of the same link. Consider again Seifert circles: Let $D(L)$ be an alternating link diagram. Let $s_a(D)$ be the number of Seifert circles in $D$. Then $s_a(D)=:s_a(L)$ is an alternating link invariant. Let $D$ be a reduced alternating link diagram of L. Then every other reduced alternating diagram of L can be reached from $D$ by a finite number of flyping moves [@MuFlype]. The figure shows the flyping move. Hence, it is enough to show that $s(D)$ is not altered by the flyping move. When considering only the underlying Seifert circles in the plane, then flyping does not affect the connectedness of the circles inside the disk. If the strands on the right-hand side of the above figure have the same orientation, the flyping move does not change the number of Seifert circles. For the remaining case, we use the idemposition algebra for arrangements of (unoriented) circles in the plane (see[@GSB]). In this case, the effect of flyping outside the disk looks like: When counting Seifert circles, we must consider several cases for the above disk diagram on the left: The outgoing arcs of the diagram can be connected outside the diagram in two ways, and there are two possibilities how the ingoing arcs into the disk can be connected inside the disk: 1. the upper left arc is connected to the lower left arc outside the diagram (called numerator closure) 1. the upper left arc is connected inside the disk to the lower left arc 2. the upper left arc is connected inside the disk to the right ingoing arc 2. the upper left arc is connected to the upper right arc outside the diagram (called denominator closure) 1. the upper left arc is connected inside the disk to the lower left arc inside the disk 2. the upper left arc is connected to the right ingoing arc inside the disk. Now the important thing is to see that the left arcs are in all four cases connected to each other. We further need two idemposition types, see the following picture: To imitate flyping, do the following: Apply first a type 1 idemposition $\Phi$ on the left side. In all cases, theorem 14 in [@GSB] gives $s(\Phi(D))=s(D)+1$. The disc is now disconnected from the rest of the link. Turn it in the plane by $\pi$. Next, apply a type 2 idemposition $\Theta$ on the right side. Since the arcs on the right side of the image cannot be connected before applying $\Theta$, again by theorem 14, we have $s(\Theta(\Phi(D)))=(s(D)+1)-1$, and the proof is complete. We obtain immediately the following condition for alternating $\mathcal{F}$-links: \[sab\] An alternating link is in $\mathcal{F}$ iff $s_a(L)=b(L)$. Assume first $s_a(L)=b(L)$. It is a long known fact that the Seifert surface from an alternating diagram $D$ is minimal, i. e. $g(D)=g(L)$. Since $s_a(L)=s(D)=b(L)$, and a reduced alternating diagram has minimal number of crossings [@Ad], $c(L)=2g(L)+b(L)-1$. On the other hand, if $b(L)<s_a(L)$, for a reduced alternating diagram we have $c(D)>2g(D)+b(L)-1$. We proceed with a statement on planar Murasugi sums. For the definition of the terms used here, see [@MuBraid]. Let L be an alternating link, and $D(L)$ be the planar Murasugi sum of special alternating link diagrams $D(L_1),..., D(L_n)$. then $$s_a(L)-1 = \sum_{i=1}^n (s_a(L_i)-1)$$ Follows immmediately by the definition of planar Murasugi sum. It is now natural to ask whether a similar equality holds for the braid index. This is conjectured in [@MuBraid]: \[MurasugiSum\] Let $L$ be the planar Murasugi sum of special alternating links $L_1,..., L_n$. then $$b(L)-1 = \sum_{i=1}^n (b(L_i)-1)$$ I note that the special alternating $\mathcal{F}$-links are exactly those satisfying theorem 8.1 in [@MuBraid], called there “special alternating links of the nonmultiple type”. This is an important point for the following discussion. The pretzel links $\mathcal{P}(a_1,a_2,...a_{2n})$ with all $a_i\geq 2$ are special $\mathcal{F}$-links (which are not fibered). Let $D(L)$ be the planar Murasugi sum of special alternating link diagrams $D(L_1),..., D(L_n)$. Assume at least one of $L_1,..,L_n$ is a non-$\mathcal{F}$-link. Then $L$ is a non-$\mathcal{F}$-link. In the proof of theorem 8.1 in [@MuBraid], a local diagrammatic move upon a special alternating diagram of a non-$\mathcal{F}$ link reduces the number of Seifert circles. Use this move in $D(L)$ to obtain a diagram $D'$ with fewer Seifert circles. But now, $b(L)\leq s(D') < s_a(L)$. Recall \[sab\] to conclude that L is not in $\mathcal{F}$. Let $L$ be a $\mathcal{F}$-link. Then conjecture \[MurasugiSum\] holds. Since $L$ is in $\mathcal{F}$, $b(L)=s_a(L)=\sum_{i=1}^n (s_a(L_i)-1)$. On the other hand, for all $L_i$, $s_a(L_i)=b(L_i)$ because they must also be in $\mathcal{F}$ (which follows from the above proposition). I believe furthermore the alternating links constructed from special alternating $\mathcal{F}$-links are exactly the alternating $\mathcal{F}$-links. This would follow immediately from conjecture \[MurasugiSum\] and complete the classification of alternating $\mathcal{F}$-links. Application to satellite knots. =============================== We can apply Lemma \[bound1\] to give a “good” estimate for the crossing number of (p,q)-cable knots about a knot in $\mathcal{F}$. (Here, p means the linking number of the knot with a meridian of the essential knotted torus in which L lies.) Let p,q be positive integers with $\gcd (p,q)=1$. Let K be the (p,q)-cable knot about a $\mathcal{F}$-knot C. Then $c(K)\geq q(p-1) + p \cdot c(C)$. Schubert showed in ([@Schubert], pp. 247 seq.) that $$2 g(K)= (p-1)(q-1)+2p \cdot g(C)$$ Furthermore, the braid index of K is $p \cdot b(C)$ [@BirSat; @WillBraid]. Since $\gcd (p,q)=1$, $\|K\|=1$. Use \[bound1\] to obtain $c(K)\geq q(p-1) + p \cdot c(C)$. However, this bound is far from being sharp, as we conjecture that the crossing number is $q(p-1) + p^2 \cdot c(C)$. This may illustrate the difficulties in proving statements about the crossing number of satellite knots. Similar estimates can be easily established in the same spirit, which cover other special cases, as you may figure out yourself. Nevertheless, we state here another example. (The definitions for pattern types and the weights $w_i$ used in the following are the same as in [@Nutts; @BirSat].) Let K be a satellite knot with a $\mathcal{F}$-knot companion C and a closed alternating braid B as type 0 pattern. Then, $c(K) \geq c(B)+ b(B)\cdot c(C)$. Another result of Schubert ([@Schubert],p.192) states that in our case $$2g(K) \geq 2b(B) \cdot g(C) + 2g(B)$$ Since B is an $\mathcal{F}$-knot, $2g(B) = c(B)-b(B)+1$; the same holds for C. Here again $b(K)= b(B)\cdot b(C)$ holds, for B is a type 0 pattern [@BirSat]. Now, \[bound1\] together with both B and C being $\mathcal{F}$-knots: $$c(K) \geq c(B)+ b(B)\cdot c(C)$$ We can also show an unconditional estimate for satellite knots, as follows: \[sat1\] Let K be a satellite knot with companion C. Let B be the knot from the pattern. Then $$c(K) \geq \left\{ \begin{array}{l} w_0 \cdot(2g(C) + b(C) - 1) + g(B) + w_0 \textnormal{ for pattern type 0} \\ w_0 \cdot(2g(C) + b(C) - 1) + g(B) + w_0 + w_1 \textnormal{ for pattern type 1}\\ 2\alpha(C)-2 \textnormal{ else}\\ \end{array} \right.$$ There, $\alpha(K)$ is the arc index of K as defined in [@CroArc]. First consider type 0 and type 1 patterns: Schubert’s theorem on the genus of satellite knots reads for the general case as follows: $g(K) \geq n \cdot g(C) + g(B)$, where B is the knot from the pattern n is the linking number of a meridian of the essential torus with B. Here, $n=w_0$. The braid index of K is $w_0 \cdot b(C)$ for a type 0 pattern and $w_0 \cdot b(C)+w_1$ for a type 1 pattern respectively [@BirSat]. Using Lemma \[bound1\], we get $c(K)\geq w_0 \cdot(2g(C) + b(C) - 1) + g(B) + w_0$ for a type 0 pattern, resp. $c(K)\geq w_0 \cdot(2g(C) + b(C) - 1) + g(B) + w_0 + w_1$ for a type 1 pattern. For the other case, we use a different estimate given by Ohyama [@Ohyama]: $c(K)\geq 2b(K)-2$. Since we have a type k pattern, $b(K)\geq \alpha(C)$ ([@Nutts], prop.3.3.5). We can conclude $c(K)\geq 2 \alpha(C)-2$ in this case. We can furthermore solve here the problem whether the crossing number of a satellite link is greater than that of the companion ([@Kirby], problem 1.67) if the latter is an alternating fibered knot: Let K be a satellite knot and let its companion C be an alternating fibered knot. Then $c(K)\geq 2c(C)+2$. Cromwell showed that $\alpha(K)\geq c(K)+2$ if K is an alternating knot [@CroArc]. Since $K \in \mathcal{F}$, and for a proper satellite knot, the weight $w_0$ is greater than 1, the rest follows from the above proposition. The result can be generalized to more companion types and sharpened at the same time if we restrict the pattern type: Let K be a satellite knot and let its companion C be a $\mathcal{F}$-knot. Furthermore, assume K has a pattern of type 0 or 1. Then $c(K)> w_0 \cdot c(C)$. Obvious from \[sat1\]. The results worked out here are all dealing with the Seifert genus. But the canonical genus gives us more power; we deal with this in the next section. However, I don’t believe that any satellite link is in $\mathcal{F}$. Crossing number, canonical genus and knot polynomials ===================================================== Of course, the question of determining the minimal crossing number remains open in general. But we can state another beautiful lower bound in terms of the HOMFLY polynomial, which is again additive under composition. Define the two-variable Laurent polynomial in the variables $v,z$ $P[v,z]=P(K)$ as in [@HOMFLY], and write shortly $e$ (resp. $m$) and $E$ (resp. $M$) for the minimal and maximal non-zero exponent in $v$ (resp. $z$). For every link $K$, $$c(K)\geq M+\frac{1}{2}(E-e)$$ Use Lemma \[bound1\] together with $M\leq \min _{D=D(K)}{c(D)-s(D)+1}$ [@MoSeif] and the MWF- inequality [@MoSeif; @FranksWilliams]: $b(K)\geq \frac{1}{2}(E-e)+1$ Furthermore, $\|K\|=m$. Notice that the right-hand side is again additive under composition, since for a composite link $L\#M$ the polynomial is $P(L)\cdot P(M)$. Sometimes, this bound can lead to better results than the inequality using the genus. For instance, $M=6$ for the untwisted double of the trefoil, whereas the free genus is 2, and the genus is only 1. On the other hand, Stoimenow showed an example where $2g>M$ [@Sto]; and there are also some examples for which the MWF- inequality is not sharp. We continue with showing up a connection between the crossing number of a knot and the canonical genus of its double. (I noticed this when reading in [@MoSeif] that the canonical genus of the Whitehead double of the trefoil is 3, which is also the minimal crossing number) \[bound2\] Let K be a knot and $W_K$ be a Whitehead double of K. Then $$c(K)\geq \widetilde{g}(W_K)$$ First, assume that $W_K$ is untwisted. The proof is constructive: Draw a diagram with minimal crossing number of K. Then choose a parallel to run along the curve in the inverse direction of the original curve (diagrams of this kind are often called blackboard diagrams). Then, in a region which is not near a crossing, replace the antiparallel by a (say) positive clasp to obtain a diagram D of $W_K$ with $4c(K)+2$ crossings. Cut out the crossings to obtain a bunch of Seifert circles with a pattern as shown in the figure. ![image](./Seifert2.eps){width="100.00000%"} Near each underlying crossing of K, you have four emerging arcs. Since the underlying diagram of K is minimal (and hence reduced), every one of these four arcs belongs to a distinct Seifert circle. In the region with the clasp from the Whitehead doubling, you see a closed Seifert circle and two lines above and below it, belonging to the same Seifert circle. The following picture should illustrate this. ![image](./Seifert4.eps){width="100.00000%"} Since every arc closes up to a Seifert circle with another arc, we count $s(D)=2c(K)+1$ Seifert circles in the diagram, and clearly, this leads for the genus of this diagram to $g(D)=c(K)\geq \widetilde{g}(W_K)$ . Next, consider the case $W_K$ is twisted. Let $D_n$ (resp. $D_{n+1}$) be a diagram with n (resp. $n+1$) half-twists. The following picture illustrates that one more half-twist produces splits a Seifert circle in $D_n$ into two, and at the same time we have more crossing in $D_{n+1}$, thus every pair of diagrams $D_n$, $D_{n+1}$ has the same genus. See the figure. ![image](./Seifert5.eps){width="100.00000%"} Complete the proof by induction on n. I was informed that this proposition was found independently by Kidwell and Stoimenow [@StoKid]. Together with Morton’s canonical genus inequality, we can estimate the minimal crossing number from the polynomial of a satellite about it. A similar technique has already successfully been used in [@MoCable] for the braid index in special cases. Does an infinite family of knots exist with $\widetilde{g}(W_K)=c(K)$? Very recently, Stoimenow and Kidwell asked a question (see [@suki], Problem 1.15) which turns out to be closely related to this problem: Let K be a knot and $W_K$ be its Whitehead double. Let $F(K)$ be the Kauffman polynomial of K in the variables $a$ and $z$. Then $$2(\max\deg_z F(K) +1)=\max\deg_z P(W_K)$$ The truth of the conjecture would imply that there exists such an infinite family, namely the prime alternating knots. With the inequality found here, I can give further “evidence” for the truth of the above equality (both sides are bounded above by $2c(K)$), and attack a part of it: Let K be a prime alternating knot and Wh(K) a Whitehead double of K. Then $$2(\max\deg_z F(K) +1)\geq \max\deg_z P(W_K)$$ Kidwell showed that $c(K) \geq \max\deg_z F(K) +1$ with equality if K is a prime alternating knot [@Kidwell]. Morton’s inequality and proposition \[bound2\] give us $2c(K)\geq \max\deg_z P(W_K)$, and the proof is complete. Since we have also found now $$2c(K)\geq \max\deg_z P(W_K)$$ and $$2c(K)\geq 2(\max\deg_z F(K) +1)$$ I feel free to mention the following, which is quite similar: Let $G(K)$ be the Kauffman polynomial with coefficients reduced modulo 2. Then $b(W_K)\geq \alpha(K)$ (see [@Nutts], prop.3.3.5) together with $\alpha(K) \geq \max\deg_a G(K) - \min\deg_a G(K)+2$ (see [@Nutts], prop.4.4.1) imply: $$2b(W_K)-2\geq 2(\max\deg_a G(K) - \min\deg_a G(K)+1)$$ whereas $$2b(W_K)-2 \geq \max\deg_v P(W_K) -\min\deg_v P(W_K)$$ is the MWF-inequality. (Note that the degrees in $v$ of the HOMFLY polynomial depend heavily on how $W_K$ is twisted, whereas $G(K)$ is of course not affected.) I would like to thank Joan Birman, and Alexander Stoimenow for answering my questions, and providing useful hints. [1]{} Adams, Colin C. : “The knot book” (german), Berlin 1995 Birman, J.S. : “Studying links via closed braids IV: composite links and split links.” Invent. Math. 102 (1990), 115–139. Birman, J.S. ; Menasco, W. :“Special Positions for essential tori in link complements”, Topology 33, No. 3 (1994) 525–556 Cromwell, P. R. : “Homogeneous Links”, J. London Math. Soc. (series 2) 39 (1989), 535–552 Cromwell, P. : “Arc presentations of knots and links”, in: Knot Theory (Proc. Conference, Warsaw 1995), eds. V. F. R. Jones et al, Banach Center Publications 42, Warsaw, 1998, 57–64 Franks, J. ; Williams, R. F. : “Braids and the Jones-Conway polynomial” Trans. Amer. Math. Soc. 303 (1987), 97–108. Freyd, P. ; Hoste, J. ; Lickorish, W. B. R. ; Millet, K. ; Ocneanu, A. ; Yetter, D. : “A new polynomial invariant for knots and links”, Bull. Amer. Math. Soc. 12 (1985), 239–246. Kidwell, M. :“On the degree of the Brandt-Lickorish-Millet-Ho polynomial of a link”, Proc. Amer. Math. Soc. 100 (1987), 755–761. Kidwell, M. ; Stoimenow, A. :“Examples Relating to the Crossing Number, Writhe, and Maximal Bridge Length of Knot Diagrams”, to appear in Mich. Math. J. Kirby, R. :“Problems in low-dimensional topology”, on the web. Menasco, W. ; Thistlethwaite, M. : “The classification of alternating links”, Ann. Math. 138 (1993), 113–171 AMS Bulletin, Vol. 25, No. 2, October 1991, 403–412. Murasugi, K. : “On the braid index of alternating links”, Trans. Amer. Math. Soc. 326, No. 1 (1991), 237–260. Morton, H. : “Seifert circles and knot polynomials”, Proc. Cambridge Phil. Soc. 99 (1986), 107–109. Morton, H. : “The 2-variable polynomial of cable knots”, Proc. Cambridge Phil. Soc. 101 (1987), 267–278. Nutt, I. : “Braid index of satellite links”, PhD thesis, University of Liverpool (1995) Ohtsuki, T. (ed.): “Problems on invariants of knots and 3-manifolds”, Geometry and Topology Monographs, Vol. 4 Ohyama, Y. :“On the minimal crossing number and the braid index of links”, Can. J. Math. 45, No. 1 (1993), 117–131. Ozawa, M. : “Additivity of free genus of knots”, Topology 40 (2001), no. 4, 659–665. Rolfsen, D. :“Knots and links”, Berkeley 1976. Schubert, H. :“Knoten und Vollringe”, Acta Mathematica 90 (1953), 131–286 Spencer-Brown, G. :“Laws of form”, international edition, Lübeck 1997, “Appendix 5: Two proofs of the four-color map theorem” Stoimenow, A. : “On the crossing number of positive knots and braids and braid index criteria of Jones and Morton-Williams-Franks”, Trans. Amer. Math. Soc. 354 (2002), no. 10, 3927–3954 R. F. Williams:“The braid index of generalized cables”, Pacific J. Math. 155 (1992), no. 2, 369–375. Yamada, S. : “The minimal nunmber of Seifert circles equals the braid index” Invent. Math. 88 (1987), 347–356.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider a nonlocal functional $J_K$ that may be regarded as a nonlocal version of the total variation. More precisely, for any measurable function $u\colon \Rd \to \R$, we define $J_K(u)$ as the integral of weighted differences of $u$. The weight is encoded by a positive kernel $K$, possibly singular in the origin. We study the minimisation of this energy under prescribed boundary conditions, and we introduce a notion of calibration suited for this nonlocal problem. Our first result shows that the existence of a calibration is a sufficient condition for a function to be a minimiser. As an application of this criterion, we prove that halfspaces are the unique minimisers of $J_K$ in a ball, provided they are admissible competitors. Finally, we outline how to exploit the optimality of hyperplanes to recover a $\Gamma$-convergence result concerning the scaling limit of $J_K$.' address: 'Dipartimento di Matematica, Università di Pisa, Largo B. Pontecorvo 5, 56127 Pisa, Italy. E-mail: <[email protected]>' author: - Valerio Pagliari title: 'Halfspaces minimise nonlocal perimeter: a proof via calibrations' --- Introduction ============ We consider the $d$-dimensional vector space $\Rd$ equipped with the Euclidean inner product $\cdot$. In this note, we show that halfspaces are the unique local minimisers of the nonlocal functional $$\begin{aligned} \label{eq:JK}\begin{split} J_K(u;\Omega) & \coloneqq \frac{1}{2}\int_{\Omega}\int_{\Omega}K(y-x){\left\lvertu(y)-u(x)\right\rvert}\de y \de x \\ & \quad +\int_{\Omega}\int_{\Omega^c}K(y-x){\left\lvertu(y)-u(x)\right\rvert}\de y \de x, \end{split} \end{aligned}$$ where $\Omega \subset \Rd$ is a Lebesgue measurable set and $\Omega^c$ is its complement, while $u$ and $K$ are positive Lebesgue measurable functions on $\Rd$. Further hypotheses on the reference set $\Omega$ and on the kernel $K$ are stated below, see Subsection \[sec:setup\]. We recall that when $u=\chi_E$ is the characteristic function of the Lebesgue measurable set $E\subset\Rd$, that is $\chi_{E}(x)=1$ if $x\in E$ and $\chi_{E}(x)=0$ otherwise, then $J_K$ can be understood as a nonlocal perimeter of the set $E$ in $\Omega$. More generally, $J_K(u;\Omega)$ may be regarded as a nonlocal total variation of $u$ in $\Omega$. Nonlocal perimeters were firstly introduced by Caffarelli, Roquejoffre, and Savin [@CRS] to the purpose of describing phase field models that feature long-range space interactions. In their work, $K(x) = \left\| x \right\|^{-d-s}$, with $s\in(0,1)$. Subsequently, many authors have extended the analysis in several directions, and by now the literature has become vast; as a narrow list of papers that are more closely related to ours, we suggest that the interested reader may consult [@CV; @ADM; @CSV; @MRT; @CN] and the references therein. Let $B$ be the open unit ball in $\Rd$ with centre in the origin, put $\mathbb{S}^{d-1}\coloneqq \partial B$, and let ${\mathscr{L}}^d$ be the $d$-dimensional Lebesgue measure. Our aim is proving the following: \[stm:piani\] For all $\hat{n}\in\mathbb{S}^{d-1}$, we define $H\coloneqq {\left\{x\in\Rd : x\cdot \hat{n} > 0\right\}}$. Then, $$J_K( \chi_H; B ) \leq J_K( v; B )$$ for all ${\mathscr{L}}^d$-measurable $v\colon \Rd \to [0,1]$ such that $v(x)=\chi_H(x)$ for ${\mathscr{L}}^d$-a.e. $x\in B^c$. Moreover, for any other minimiser $u$ satisfying the same constraint, it holds $u(x)=\chi_H(x)$ ${\mathscr{L}}^d$-a.e. $x\in \Rd$. The proof that we propose relies on a general criterion for minimality, see Theorem \[stm:plateau\], which in turn involves a notion of calibration fitted to nonlocal problem at stake, see Definition \[stm:calib\]. Let us outline the structure of this note. In the next Subsection, we precise the mathematical framework of this paper and we set the notations in use. Section \[sec:criterion\] contains the definition of nonlocal calibration and the proof of Theorem \[stm:piani\]. Lastly, in Section \[sec:Gammaconv\], as a possible application of our main result, we discuss its role in the analysis of the scaling limit of the functional $J_K$. Set-up and notations {#sec:setup} -------------------- We remind that we work in $\Rd$, the $d$-dimensional Euclidean space, endowed with the inner product $\cdot$ and the associated norm $\left\| \, \cdot \, \right\|$. We let ${\mathscr{L}}^d$ and ${\mathscr{H}}^{d-1}$ be respectively the $d$-dimensional Lebesgue and the $(d-1)$-dimensional Hausdorff measure on $\Rd$. We shall henceforth omit to specify the measure w.r.t. which a set or a function is measurable, when the measure is $\Ld$ or the product $\Ld \otimes \Ld$ on $\Rd \times \Rd$; analogously, we shall use the expression “a.e.” in place of “$\Ld$-a.e.” and of “$\Ld \otimes \Ld$-a.e.”. If $u$ and $v$ are measurable functions, we shall also write “$u=v$ in $E$” as a shorthand for “$u(x)=v(x)$ for a.e. $x\in E$”. In this note, $\Omega\subset \Rd$ is an open and connected reference set such that ${\mathscr{L}}^d(\Omega)\in(0,+\infty)$. Later on, in Section \[sec:Gammaconv\], some regularity on the boundary $\partial \Omega$ will be required. For what concerns the kernel $K\colon \Rd \to [0,+\infty]$, it is not restrictive to assume that is even, i.e. $$K(x) = K(-x) \qquad \text{a.e. } x\in\Rd.$$ Besides, we suppose that $$\label{eq:summK} \int_{\Rd} \left( 1 \wedge \left\| x \right\| \right) K(x) \de x < +\infty,$$ where, if $t,s\in\R$, $t \wedge s$ equals the minimum between $t$ and $s$. This condition entails that $K \in L^1(B(0,r)^c)$ for all balls $B(0,r)$ with centre in the origin and radius $r>0$; in particular, $K$ might have a non-$L^1$ singularity in $0$. The main example of functions that fulfil is given by fractional kernels [@CRS; @L], i.e. kernels of the form $$K(x) = \frac{a(x)}{ {\left\lvertx\right\rvert}^{d+s} },$$ where $a\colon \Rd \to \R$ is an even function such that $0<\lambda \leq a(x) \leq \Lambda$ for some $\lambda,\Lambda\in \R$ and $s\in(0,1)$. A faster decay at infinity for $K$ will be needed in Section \[sec:Gammaconv\], see . We are interested in a variational problem concerning $J_K$, to which we shall informally refer as Plateau’s problem. Precisely, given a Lebesgue measurable set $E_0 \subset \Rd$ such that $J_K( \chi_{E_0}; \Omega)<+\infty$, we define the family $$\label{eq:F} {\mathscr{F}}\coloneqq{\left\{v\colon \Rd \to [0,1] : v \text{ is measurable and } v = \chi_{E_0} \text{ in } \Omega^c\right\}},$$ and we address the minimisation of $J_K(\,\cdot\,;\Omega)$ in the class ${\mathscr{F}}$; namely, we consider $$\label{eq:plateau} \inf{\left\{J_K(v;\Omega) : v\in{\mathscr{F}}\right\}}.$$ For $s\in\R$, let us set $T(s) \coloneqq \big( (0 \vee s) \wedge 1\big)$ ($t \vee s$ is the maximum between the real numbers $t$ and $s$). Observe that $T \circ \chi_{E_0} = \chi_{E_0}$ and $J_K( T \circ u; \Omega ) \leq J_K( u; \Omega )$, so the infimum in equals $$\inf{\left\{J_K(v;\Omega) : v\colon \Rd \to \R \text{ is measurable and } v = \chi_{E_0} \text{ in } \Omega^c \right\}}.$$ We therefore see that choice of ${\mathscr{F}}$ as the class of competitors is not restrictive. Standing our assumptions on $\Omega$, any set $E$ that has finite perimeter in $\Omega$ satisfies $J_K( \chi_E; \Omega ) <+\infty$, see [@MRT; @BP] We shall recall the definition of finite perimeter set later in this Subsection. As the functional $J_K(\,\cdot\,; \Omega)$ is convex, when $\Omega$ has finite measure, existence of solutions to can be established by the direct method of calculus of variations (see [@BP]; see also [@CSV] for an approach via approximation by smooth sets). In particular, as consequence of the following coarea-type formula: $$\label{eq:coarea} J_K(u;\Omega) = \int_{0}^{1}{\operatorname{Per}_K}({\left\{u>t\right\}};\Omega)\de t,$$ there always exists a minimiser which is a characteristic function. Indeed, for any $u\colon \Rd \to [0,1]$, there exists $t^\ast\in\R$ such that ${\operatorname{Per}_K}({\left\{u>t^\ast\right\}}; \Omega)\leq J_K(u; \Omega)$, otherwise would be contradicted. Thus, if $u$ is a minimiser of , then $\chi_{{\left\{u>t^\ast\right\}}}$ is minimising as well. Formula can be easily validated, see for instance [@CSV; @CN]. The family of functionals on $L^1(\Omega)$ such that a generalised Coarea Formula holds was firstly introduced by Visintin [@V]. It is well-known that existence of solutions to the classical counterpart of may be proved in the framework of geometric measure theory. We remind here some basic facts, while we refer to the monographs [@Ma; @AFP] for a thorough treatment of the subject. We say that $u\colon \Omega \to \R$ is a function of bounded variation in $\Omega$, and we write $u\in\mathrm{BV}(\Omega)$, if $u\in L^1(\Omega)$ and $${\left\lvert\D u\right\rvert}(\Omega) \coloneqq \sup{\left\{ \int_{\Omega} u(x)\mathrm{div}\zeta(x) \de x : \zeta\in C^{\infty}_c (\Rd;\Rd), {\left\Vert\zeta\right\Vert}_{L^\infty}\leq 1\right\}} < + \infty.$$ We dub ${\left\lvert\D u\right\rvert}(\Omega)$ the total variation of $u$ in $\Omega$. We also say that a measurable set $E$ is a set of finite perimeter in $\Omega$ when its characteristic function $\chi_E$ is a function of bounded variation in $\Omega$, and, in this case, we refer to $\operatorname{Per}(E;\Omega) \coloneqq {\left\lvert\D \chi_E\right\rvert}(\Omega)$ as perimeter of $E$ in $\Omega$. In this framework, the result that parallels the existence of solutions to reads as follows: there is a set $E$ with finite perimeter in $\Omega$ such that $\operatorname{Per}(E;\Omega)$ attains $$\label{eq:cl-plateau} \inf{\left\{{\left\lvert\D u\right\rvert}(\Omega) : u\colon \Rd \to [0,1] \text{ is measurable and } u = \chi_{E_0} \text{ in } \Omega^c\right\}}.$$ Finite perimeter sets stand as measure-theoretic counterparts of smooth hypersurfaces. For example, we may equip them with an inner normal: for any $x\in \mathrm{supp}{\left\lvert\D\chi_E\right\rvert}$, we define $$\label{eq:in-norm} \hat n(x)\coloneqq \lim_{r\to 0^+}\frac{\D\chi_E(B(x,r))}{{\left\lvert\D\chi_E\right\rvert}(B(x,r))},$$ where $\D \chi_E$ is the distributional gradient of $\chi_E$ and $B(x,r)$ is the open ball of centre $x$ and radius $r>0$. A fundamental result by De Giorgi [@D] states that $$\label{eq:per-Haus} \operatorname{Per}(E;\Omega)={\mathscr{H}}^{d-1}(\partial^\ast E\cap\Omega),$$ where $$\partial^\ast E \coloneqq {\left\{x\in \Rd : \hat n(x) \mbox{ exists and } {\left\lvert\hat n(x)\right\rvert}=1\right\}}$$ is the so-called reduced boundary of $E$. In addition, for any $x\in\partial^\ast E$, $$\label{eq:blowup} \frac{E-x}{r} \to {\left\{y\in\Rd : y\cdot \hat n(x) > 0\right\}} \quad\mbox{as } r\to 0^+ \mbox{ in } L^1_\mathrm{loc}(\Rd) .$$ Once existence of solutions to is on hand, a useful criterion to substantiate the minimality of a given competitor is provided by means of calibrations. The notion of calibration may be expressed in very general terms (see [@Mo; @HL] and references therein); as far as we are concerned, we say that a (classical) calibration for the finite perimeter set $E$ is a divergence-free vector field $\zeta\colon \Rd \to \Rd$ such that $\left\| \zeta(x) \right\| \leq 1$ a.e. and $\zeta(x) = \hat{n}(x)$ for ${\mathscr{H}}^{d-1}$-a.e. $x\in\partial^\ast E$. It can be shown that if the set $E$ admits a calibration, then its perimeter equals the infimum in . The goal of the next Section is validating a nonlocal analogue of this principle. A similar analysis has been independently carried out by Cabré in [@C]. In that work, the author proposes a notion of calibration akin to ours. He exploits it to establish optimality of nonlocal minimal graphs and to give as well a simplified proof of a result in [@CRS] stating that minimisers of the Plateau’s problem satisfy a zero nonlocal mean curvature equation in the viscosity sense. Minimality via calibrations {#sec:criterion} ============================= In this Section, we propose a notion of calibration adapted to the current nonlocal setting, and we show that the existence of a calibration is a sufficient condition for a function $u$ to minimise the energy $J_K$ w.r.t compact perturbations. Then, we show that halfspaces admit calibrations, and thus we infer their minimality. We remind that we assume that $\Rd\times\Rd$ is equipped with the product measure $\Ld \otimes \Ld$. \[stm:calib\] Let $u\colon \Rd \to [0,1]$ and $\zeta\colon \Rd\times\Rd \to \R$ be measurable functions. We say that $\zeta$ is a nonlocal calibration for $u$ if the following hold: 1. ${\left\lvert\zeta(x,y)\right\rvert}\leq 1$ for a.e. $(x,y)\in\Rd\times\Rd$; 2. for a.e. $x\in\Rd$, $$\label{eq:div=0} \lim_{r\to 0^+}\int_{B(x,r)^c} K(y-x)\left(\zeta(y,x)-\zeta(x,y)\right)\de y = 0;$$ 3. for a.e. $(x,y)\in\Rd\times\Rd$ such that $u(x)\neq u(y)$, $$\label{eq:nlnormal} \zeta(x,y)(u(y)-u(x))={\left\lvertu(y)-u(x)\right\rvert}.$$ The next remark collects some comments about the definition above. Let $\zeta\colon \Rd\times\Rd \to \R$ be a calibration for $u\colon \Rd \to [0,1]$. 1. It is not restrictive to assume that $\zeta$ is antisymmetric: indeed, $\tilde\zeta(x,y)\coloneqq ( \zeta(x,y)-\zeta(y,x) )/2$ is a calibration for $u$ as well. 2. In view of , the integral in is convergent for each $r>0$. We can regard as a nonlocal counterpart of the vanishing divergence condition that is prescribed for classical calibrations. Such nonlocal gradient and divergence operators where introduced in [@GO], and they have already been exploited to study nonlocal perimeters by Mazón, Rossi, and Toledo in [@MRT], where the authors propose a notion of $K$-calibrable set in relation to a nonlocal Cheeger energy. 3. Suppose that $u=\chi_E$ for some measurable $E\subset \Rd$. By , $\zeta$ must satisfy $$\zeta(x,y) = \begin{cases} -1 & \text{if } x\in E, y\in E^c \\ 1 & \text{if } x\in E^c, y\in E. \end{cases}$$ Heuristically, this means that the calibration gives the sign of the inner product between the vector $y-x$ and the inner normal to $E$ at the “crossing point”, provided the boundary of $E$ is sufficiently regular (see Figure \[fig:zeta\]). Indeed, if we imagine to displace a particle from to $x$ and $y$, $\zeta$ equals $-1$ when the particle exits $E$, and it equals $1$ if the particles enters $E$. (-0.1,-0.4) node[$E$]{}; (-1,0) .. controls (0,1) and (0,0) .. (1,0); (1,-1) arc (90:270:-0.5); (-0.5,-2) .. controls (0,-1) and (0,-1) .. (1,-1); (-0.5,-2) .. controls (-1.5,-2) and (-0.5,-1) .. (-1,0); (0.4,-0.6) node[$x$]{}; (0.5,-0.5) circle (0.5pt); (1.6,0.6) node[$y$]{}; (1.15,-0.26) node[$\hat n$]{}; (0.5,-0.5) – (1.5,0.5); (1,0) – (1,-0.5); (0.3,0) – (1.7,0); Our criterion reads as follows: \[stm:plateau\] Let $E_0\subset \Rd$ be a measurable set such that $J_K(\chi_{E_0};\Omega) <+\infty$, and let $\F$ be the family in . If for some $u\in\F$ there exists a calibration $\zeta$, then $$J_K(u;\Omega)\leq J_K(v;\Omega)\quad \text{for all } v\in{\mathscr{F}}.$$ Moreover, if $\tilde u\in{\mathscr{F}}$ is another minimiser, then $\zeta$ is a calibration for $\tilde u$ as well. By the definitions of $J_K(\,\cdot\,;\Omega)$, $\zeta$, and ${\mathscr{F}}$, for any $v\in{\mathscr{F}}$, $$\label{eq:lowbound0} J_K(v;\Omega)\geq a(v) + b_1(v) + b_0,$$ where $$\begin{gathered} a(v)\coloneqq \frac{1}{2}\int_{\Omega}\int_{\Omega}K(y-x)\zeta(x,y)(v(y)-v(x))\de y\de x,\\ b_1(v)\coloneqq - \int_{\Omega}\int_{\Omega^c} K(y-x) \zeta(x,y) v(x)\de y \de x, \\ b_0\coloneqq \int_{\Omega}\int_{\Omega^c} K(y-x) \zeta(x,y) \chi_{E_0}(y)\de y \de x. \end{gathered}$$ Since it is not restrictive to assume that $J_K(v;\Omega)$ is finite, we can suppose that $a(v)$, $b_1(v)$, and $b_0$ are finite as well. We claim that it suffices to prove that $a(v)=-b_1(v)$ to grant the minimality of $u$. Indeed, $a(v)=-b_1(v)$ yields $$\label{eq:lowbound} J_K(v;\Omega)\geq b_0 \quad\text{for all } v\in{\mathscr{F}},$$ and we remark that the lower bound $b_0$ is attained by $u$, because equality holds in for this function. Therefore, $u$ is a minimiser. Now, we prove that $a(v)=-b_1(v)$ for all $v\in{\mathscr{F}}$. Recalling that we can assume $\zeta$ to be antisymmetric, we have $$a(v)=-\int_{\Omega}\int_{\Omega}K(y-x)\zeta(x,y)v(x)\de y\de x.$$ Also, gets $$\begin{split} 0 & = -2\lim_{r\to 0^+}\int_{B(x,r)^c} K(y-x) \zeta(x,y)\de y \\ & = -2\lim_{r\to 0^+}\int_{B(x,r)^c \cap \Omega} K(y-x) \zeta(x,y)\de y -2\int_{\Omega^c} K(y-x) \zeta(x,y)\de y, \end{split}$$ whence $$a(v) = -\lim_{r\to 0^+} \int_\Omega\int_{B(x,r)^c \cap \Omega} K(y-x) \zeta(x,y) v(x)\de y \de x = -b_1(v).$$ Next, let $\tilde u\in{\mathscr{F}}$ be another minimiser of $J_K(\,\cdot\,;\Omega)$, that is $J_K(\tilde u;\Omega)=b_0$. Our purpose is proving that for a.e. $(x,y)\in\Rd\times\Rd$ such that $ \tilde u(x)\neq\tilde u(y) $ it holds $$\label{eq:optcond} \zeta(x,y)\left(\tilde u(y)-\tilde u(x)\right) = {\left\lvert\tilde u(y)-\tilde u(x)\right\rvert}.$$ First of all, note the equality holds for a.e. $(x,y)\in\Omega^c \times \Omega^c$, because $u=\tilde u$ in $\Omega^c$. Furthermore, from we have $$b_0=J_K(\tilde u;\Omega)\geq a(\tilde u)+b_1(\tilde u)+b_0 =b_0,$$ thus $$\begin{split} \frac{1}{2}\int_{\Omega}\int_{\Omega}K(y-x) \left[{\left\lvert\tilde u(y)-\tilde u(x)\right\rvert}-\zeta(x,y)(\tilde u(y)-\tilde u(x))\right]\de y \de x \\ +\int_{\Omega}\int_{\Omega^c}K(y-x) \left[{\left\lvert\tilde u(y)-\tilde u(x)\right\rvert}-\zeta(x,y)(\tilde u(y)-\tilde u(x))\right]\de y \de x = 0. \end{split}$$ The integrand appearing in the previous identity is positive, therefore we deduce that is satisfied for a.e. $(x,y)\in\Omega\times\Rd$. Eventually, in the case $x\in\Omega^c$ and $y\in\Omega$, we achieve the conclusion by exploiting the antisymmetry of $\zeta$. We take advantage of the previous theorem to prove that halfspaces are the unique local minimisers of $J_K(\,\cdot\,;B)$. This property has already been shown for fractional kernels in [@CRS; @ADM] by means of a reflection argument, which in fact turns out to be effective whenever $K$ is radial and strictly decreasing [@BP]. Here, we are able to deal with the case when the kernel is neither monotone nor radial. We start with the following lemma, whose proof is a simple verification: Given $\hat{n}\in\mathbb{S}^{d-1}$, let us set $$\zeta(x,y)\coloneqq \sign((y-x)\cdot \hat{n}) \qquad\text{and}\qquad H\coloneqq {\left\{x\in\Rd : x\cdot \hat{n} > 0\right\}}.$$ Then, $\zeta$ is a calibration for $\chi_H$. Now, we prove Theorem \[stm:piani\]. In view of Theorem \[stm:calib\] and of the Lemma above, we deduce that $\chi_H$ is a minimiser of the problem under consideration. Hence, we are left to prove uniqueness. Let $u\colon \Rd \to [0,1]$ be another minimiser. The second assertion in Theorem \[stm:calib\] grants that $\zeta(x,y)\coloneqq \sign((y-x)\cdot \hat{n})$ is a calibration for $u$ as well, so we get $$\label{eq:optcond-H} \sign((y-x)\cdot \hat{n})(u(y)-u(x))={\left\lvertu(y)-u(x)\right\rvert}\qquad \text{a.e. $(x,y)\in \Rd\times\Rd$.}$$ Let $N\subset \Rd\times\Rd$ be the negligible set of couples $(x,y)$ such that the previous equation does not hold; then, for all $(x,y) \in N^c \coloneqq \Rd \times \Rd \setminus N$ such that $x\cdot\hat{n}<y\cdot\hat{n}$, we have $u(x) \leq u(y)$. We assert that, in fact, the implication that we have just obtained holds true everywhere in $\Rd\times\Rd$. To see this, let $\rho\in L^1(\Rd)$ be a positive, radial function such that $\int\rho =1$ and whose support is contained in $B$. For ${\varepsilon}\in (0,1)$, we consider the family $\rho_{{\varepsilon}}(x)\coloneqq {\varepsilon}^{-d}\rho\left( {\varepsilon}^{-1} x \right)$ and the convolutions $u_{{\varepsilon}}\coloneqq\rho_{\varepsilon}\ast u\colon \Rd \to [0,1]$. Since $\int\rho =1$, for any couple $(x,y)\in \Rd\times\Rd$ we have $$\begin{split} u_{{\varepsilon}}(y) - u_{{\varepsilon}}(x) = & \int_{\Rd\times\Rd} \rho_{{\varepsilon}}(\xi) \rho_{{\varepsilon}}(\eta) \left[ u(y+\eta) - u(x+\xi) \right] \de \xi \de \eta \\ = & \int_{B(0,{\varepsilon}) \times B(0,{\varepsilon})} \rho_{{\varepsilon}}(\xi) \rho_{{\varepsilon}}(\eta) \left[ u(y+\eta) - u(x+\xi) \right] \de \xi \de \eta. \end{split}$$ Let us suppose that $\delta \coloneqq (y-x)\cdot \hat{n} > 0$. If we choose ${{\varepsilon}}\in(0,\delta/2)$, we see that $ (x+\xi) \cdot \hat{n} < (y+\eta) \cdot \hat{n} $ for all $ \xi,\eta \in B(0,{\varepsilon}) $, hence $u(x+\xi) \leq u(y+\eta)$ for a.e. $(\xi,\eta) \in B(0,{\varepsilon}) \times B(0,{\varepsilon})$. Consequently, for all $(x,y)\in \Rd\times \Rd$, it holds $u_{{\varepsilon}}(x) \leq u_{{\varepsilon}}(y)$ provided ${{\varepsilon}}$ is small enough. Letting ${\varepsilon}\to 0^+$, we find that $$\label{eq:mon} u(x) \leq u(y) \qquad\text{if } x\cdot\hat{n}<y\cdot\hat{n}$$ Next, we focus on the superlevel sets of $u$: for $t\in(0,1)$, we define $$E_t \coloneqq {\left\{ x : u(x) > t\right\}},$$ and we observe that if $(x,y)\in E_t \times E_t^c$, it must be $x\cdot \hat n \geq y \cdot \hat n$, otherwise, by we would have $u(x) \leq u(y)$. Therefore, there exists $\lambda_t \in \R$ such that $E_t \subset {\left\{x : x\cdot \hat{n} \geq \lambda_t\right\}}$ and $E_t^c \subset {\left\{y : y\cdot \hat{n} \leq \lambda_t\right\}}$, whence $\Ld( E_t\, {\!\bigtriangleup\!}\, {\left\{x : x\cdot \hat{n} \geq \lambda_t\right\}} ) = 0$ for all $t\in(0,1)$. Recalling that it holds $u = \chi_H$ in $B^c$, we infer that $\lambda_t = 0$ and this gets $$\Ld( E_t {\!\bigtriangleup\!}H ) = 0 \qquad \text{for all $t\in(0,1)$}.$$ Summing up, we proved that $u\colon \Rd \to [0,1]$ is a function such that, for all $t\in(0,1)$, the superlevel set $E_t$ coincides with the halfspace $H$, up to a negligible set. To reach the conclusion, we let ${\left\{t_k\right\}}_{k\in \N}\subset (0,1)$ be a sequence that converges to $0$ when $k\to +\infty$. Because it holds $${\left\{ x : u(x) = 0\right\}} = \bigcap_{k\in\N} E_{t_k}^c \quad\text{and}\quad {\left\{ x : u(x) = 1\right\}} = \bigcap_{k\in\N} E_{1 - t_k},$$ we see that $\Ld( {\left\{ x : u(x) = 0\right\}} {\!\bigtriangleup\!}H^c ) = 0$ and $\Ld( {\left\{ x : u(x) = 1\right\}} {\!\bigtriangleup\!}H) = 0$. Thus, $u = \chi_H$ in $\Rd$. $\Gamma$-limit of the rescaled energy {#sec:Gammaconv} ===================================== In this Section, we outline how to exploit Theorem \[stm:piani\] to study the limiting behaviour of certain rescalings of the energy $J_K$. In precise terms, we are interested in the $\Gamma$-convergence as ${\varepsilon}\to 0^+$ of ${\left\{ J_{K_{{\varepsilon}}}(\,\cdot\,;\Omega) \right\}}$ with respect to the $L^1_{\mathrm{loc}}(\Rd)$-convergence, where, for ${{\varepsilon}}>0$, we let $$K_{{\varepsilon}}(x) \coloneqq \frac{1}{{{\varepsilon}}^d} K\left( \frac{x}{{\varepsilon}} \right).$$ In [@BP], the analysis has already been carried out by Berendsen and the author of this note when $K$ is radial and strictly decreasing, but, as we concisely explain in the remainder of this note, the same arguments may be conveniently adapted to the current more general setting. We shall not deal with all the computations in depth, because our main interest here is how to take advantage of the minimality of halfspaces. This will be apparent in Lemma \[stm:sigmaK\]. We refer to the works in the bibliography for the technical details. For the sake of completeness, we recall the following definition: \[stm:defGammac\] Let $X$ be a set endowed with a notion of convergence and, for ${{\varepsilon}}>0$, let $f_{{\varepsilon}}\colon X\to[-\infty,+\infty]$ be a function. We say that the family ${\left\{f_{\varepsilon}\right\}}$ $\Gamma$-converges as ${{\varepsilon}}\to0^+$ to the function $f_0\colon X\to [-\infty,+\infty]$ w.r.t. the convergence in $X$ if 1. for any $x_0\in X$ and for any ${\left\{x_{\varepsilon}\right\}}\subset X$ that converges to $x_0$, it holds $$f_0(x_0)\leq\liminf_{{{\varepsilon}}\to 0}f_{\varepsilon}(x_{\varepsilon});$$ 2. for any $x_0\in X$ there exists ${\left\{x_{\varepsilon}\right\}}\subset X$ that converges to $x_0$ with the property that $$\limsup_{{{\varepsilon}}\to 0}f_{\varepsilon}(x_{\varepsilon})\leq f_0(x_0).$$ When $u\colon \Rd \to [0,1]$ is a measurable function, let us define $$\begin{gathered} J_{\varepsilon}^1(u; \Omega) \coloneqq\frac{1}{2}\int_\Omega\int_\Omega K_{\varepsilon}(y-x){\left\lvertu(y)-u(x)\right\rvert}\de y\de x, \\ J_{\varepsilon}^2(u; \Omega) \coloneqq\int_{\Omega}\int_{\Omega^c}K_{\varepsilon}(y-x){\left\lvertu(y)-u(x)\right\rvert}\de y \de x,\\ J_{\varepsilon}(u; \Omega) \coloneqq J_{\varepsilon}^1(u; \Omega) + J_{\varepsilon}^2(u; \Omega). \end{gathered}$$ Observe that, according to the notation in , $J_{\varepsilon}= J_{K_{\varepsilon}}$. We also introduce the limit functional $$J_0(u;\Omega) \coloneqq \begin{cases} \displaystyle{ \frac{1}{2}\int_{\Rd} K(z) \left( \int_\Omega {\left\lvert z \cdot \D u\right\rvert} \right) \de z } & \text{if } u\in\mathrm{BV}(\Omega), \\ +\infty & \text{otherwise}. \end{cases}$$ Our goal is proving the following: \[stm:Gconv\] Let $\Omega \subset \Rd$ be an open, connected, and bounded set with Lipschitz boundary. Let also $K\colon \Rd \to [0,+\infty)$ be an even function such that $$\label{eq:fastK} \int_{\Rd} K(x){\left\lvertx\right\rvert} dx < +\infty.$$ Then, for any measurable $u\colon \Rd \to [0,1]$ the following hold: 1. \[stm:lower\] For any family ${\left\{u_{\varepsilon}\right\}}$ that converges to $u$ in $L^1_\mathrm{loc}(\Rd)$, we have $$J_0(u;\Omega) \leq \liminf_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} J^1_{{\varepsilon}}(u_{{\varepsilon}};\Omega).$$ 2. \[stm:upper\] There exists a family ${\left\{u_{\varepsilon}\right\}}$ that converges to $u$ in $L^1_\mathrm{loc}(\Rd)$ such that $$\limsup_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} J_{{\varepsilon}}(u_{{\varepsilon}};\Omega) \leq J_0(u;\Omega).$$ We remark that, being $J_{\varepsilon}^2(\,\cdot\,;\Omega)$ positive, Theorem \[stm:Gconv\] entails the $\Gamma$-convergence of ${\left\{J_{{\varepsilon}}(\,\cdot\,;\Omega)\right\}}$ to $J_0(\,\cdot\,;\Omega)$ w.r.t. the $L^1_\mathrm{loc}(\Rd)$-convergence. Also, note that prescribes a condition that is more stringent than . Several results about the asymptotics of functionals akin to $J_{{\varepsilon}}$ have been considered in the literature [@BBM; @D; @P; @MRT; @AB]; in particular, we wish to mention the following one by Ponce: \[stm:ponce\] Let $\Omega\subset \Rd$ be an open bounded set with Lipschitz boundary and let $u\in \mathrm{BV}(\Omega)$. If holds, then $$\label{eq:pointlim} \lim_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} J^1_{\varepsilon}(u;\Omega) = J_0(u;\Omega)$$ Moreover, $J_0(\,\cdot\,;\Omega)$ is the $\Gamma$-limit as ${{\varepsilon}}\to 0^+$ of ${\left\{{{\varepsilon}}^{-1} J_{{\varepsilon}}(\,\cdot\,;\Omega)\right\}}$ w.r.t. the $L^1(\Omega)$-topology. We discuss separately the proofs of statements \[stm:lower\] and \[stm:upper\] in Theorem \[stm:Gconv\]. Preliminarly, we remark that we only need to study the $\Gamma$-convergence of $J_{{\varepsilon}}$ regarded as a functional on measurable sets, namely, for $E\subset \Rd$ measurable, we consider $$\begin{gathered} J^i_{{\varepsilon}}(E;\Omega) \coloneqq J^i_{{\varepsilon}}(\chi_E;\Omega) \qquad\text{for } i=1,2, \\ J_{\varepsilon}(E;\Omega) \coloneqq J_{{\varepsilon}}(\chi_E;\Omega), \end{gathered}$$ and the limit functional $$J_0(E;\Omega) \coloneqq J_0(\chi_E;\Omega).$$ Indeed, by appealing to results by Chambolle, Giacomini, and Lussardi [@CGL Propositions 3.4 and 3.5], it is possible to recover the $\Gamma$-convergence of $J_{{\varepsilon}}$ as a functional on measurable functions from the analysis of the restrictions; this is mainly due to convexity and to the validity of Coarea Formulas. So, as for the $\Gamma$-upper limit inequality, we need to show that, for any given measurable $E\subset \Rd$, there exists a family ${\left\{E_{\varepsilon}\right\}}$ that converges to $E$ in $L^1_\mathrm{loc}(\Rd)$ as ${{\varepsilon}}\to0^+$ such that $$\limsup_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} J_{{\varepsilon}}^1(E_{{\varepsilon}};\Omega) \leq J_0(E;\Omega).$$ Hereafter, by saying that the family of sets ${\left\{E_{\varepsilon}\right\}}$ converges to $E$ in $L^1_\mathrm{loc}(\Rd)$, we mean that $\chi_{E_{\varepsilon}} \to \chi_E$ in $L^1_\mathrm{loc}(\Rd)$. The desired inequality may be achieved as in [@BP] by reasoning on a class of sets ${\mathscr{D}}$ which is dense w.r.t. the energy $J_0$ among all measurable sets. We omit the details, since Theorem \[stm:piani\] plays no role in this step. Now we turn to the proof of the $\Gamma$-lower limit inequality. Our task is proving that, for any given measurable $E\subset \Rd$ and for any family ${\left\{E_{\varepsilon}\right\}}$ that converges to $E$ in $L^1_\mathrm{loc}(\Rd)$ as ${{\varepsilon}}\to0^+$, it holds $$\label{eq:lowerlim} J_0(E;\Omega) \leq \liminf_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} J_{{\varepsilon}}^1(E_{{\varepsilon}};\Omega).$$ In [@P], the approach to the $\Gamma$-lower limit inequality relies on representation formulas for the relaxations of a certain class of integral functionals. Here, following [@BP], we propose a strategy which combines the pointwise limit and Theorem \[stm:piani\]. Observe that we can write $$J_0(E;\Omega) \coloneqq \begin{cases} \displaystyle{ \int_{\partial^\ast E \cap \Omega} \sigma_K(\hat{n}(x)) \de {\mathscr{H}}^{d-1}(x) } & \text{if $E$ is a finite perimeter set in $\Omega$,} \\ +\infty & \text{otherwise}, \end{cases}$$ where $\hat{n}\colon \partial^\ast E \to \mathbb{S}^{d-1}$ is the measure-theoretic inner normal of $E$ (recall ) and $\sigma_K\colon \Rd \to [0,+\infty)$ is the anisotropic norm $$\sigma_K(p) \coloneqq \frac{1}{2}\int_{\Rd} K(z){\left\lvertz\cdot p\right\rvert}\de z, \qquad\text{for } p\in\Rd.$$ When $K$ is radial, $J_0$ coincides with De Giorgi’s perimeter, up to a multiplicative constant that depends on $K$ and on $d$. Indeed, if $K(x) = \bar K ({\left\lvertx\right\rvert})$ for some $\bar K\colon [0,+\infty) \to [0,+\infty)$, for any $\hat p\in\mathbb{S}^{d-1}$, we have that $$\begin{aligned} \sigma_K(\hat p) & = \frac{1}{2} \left(\int_{0}^{+\infty} \bar K(r) r^d \de r \right) \int_{\mathbb{S}^{d-1}} {\left\lverte\cdot \hat p\right\rvert} \de {\mathscr{H}}^{d-1}(e) \\ & = \frac{1}{2} \left(\int_{\Rd} K(x){\left\lvertx\right\rvert} \de x\right) \fint_{\mathbb{S}^{d-1}} {\left\lverte\cdot e_d\right\rvert} \de {\mathscr{H}}^{d-1}(e), \end{aligned}$$ where $e_d \coloneqq (0,\dots,0,1)$ is the last element of the canonical basis. By a blow-up argument à la Fonseca-Müller [@FM] that has already been applied to similar problems [@AB; @ADM], it turns out that the $\Gamma$-lower limit inequality holds as soon as one characterises the norm $\sigma_K$ in terms of the evaluation on halfspaces of the $\Gamma$-inferior limit of ${{\varepsilon}}^{-1} J_{{\varepsilon}}(\,\cdot\,;B)$. Precisely, we need to validate the following: \[stm:sigmaK\] For any $\hat{p}\in\mathbb{S}^{d-1}$, $$\sigma_K (\hat{p}) = \inf{\left\{ \liminf_{{{\varepsilon}}\to 0^+} \frac{1}{\omega_{d-1} {\varepsilon}} J^1_{\varepsilon}(E_{{\varepsilon}};B) : E_{\varepsilon}\to H_{\hat{p}} \text{ in } L^1(B) \right\}},$$ where $\omega_{d-1}$ is the $(d-1)$-dimensional Lebesgue measure of the unit ball in $\R^{d-1}$, and $H_{\hat{p}} \coloneqq {\left\{ x\in\Rd : x\cdot \hat{p} > 0 \right\}}$. It is in the proof of this Lemma that Theorem \[stm:piani\] comes into play. For $\hat{p}\in \mathbb{S}^{d-1}$, let us set $$\label{eq:sigma'K} \sigma'_K(\hat{p}) \coloneqq \inf{\left\{ \liminf_{{{\varepsilon}}\to 0^+} \frac{1}{\omega_{d-1} {\varepsilon}} J^1_{\varepsilon}(E_{{\varepsilon}};B) : E_{\varepsilon}\to H_{\hat{p}} \text{ in } L^1(B) \right\}}.$$ By , we now that $$\label{eq:sigmaK-P} \sigma_K(\hat{p}) = \lim_{{{\varepsilon}}\to 0^+} \frac{1}{\omega_{d-1}{\varepsilon}} J^1_{\varepsilon}(H_{\hat{p}};B),$$ hence $\sigma_K (\hat{p}) \geq \sigma'_K(\hat{p})$. To the purpose of proving the reverse inequality, we introduce a third function $\sigma''_K$ and we show that $\sigma_K \leq \sigma''_K \leq \sigma'_K$. So, for $\hat{p}\in \mathbb{S}^{d-1}$ and $\delta\in(0,1)$, we let $$\sigma''_K(\hat{p}) \coloneqq \inf{\left\{ \liminf_{{{\varepsilon}}\to 0^+} \frac{1}{ \omega_{d-1} {\varepsilon}} J^1_{\varepsilon}(E_{{\varepsilon}};B) : E_{\varepsilon}\to H_{\hat{p}} \text{ in } L^1(B) \text{ and } E_{\varepsilon}{\!\bigtriangleup\!}H_{\hat{p}} \subset B_{1-\delta} \right\}},$$ where $B_{1-\delta} \coloneqq B(0,1-\delta)$ and $E_{\varepsilon}{\!\bigtriangleup\!}H_{\hat{p}}$ is the symmetric difference between $E_{\varepsilon}$ and $H_{\hat{p}}$. We decide not use a notation that exhibits the dependence of $\sigma''_K$ on the parameter $\delta$ because a posteriori the values of $\sigma''_K$ are not influenced by it. We firstly show that $\sigma_K \leq \sigma''_K$. Let ${E_{\varepsilon}}$ be a family of measurable subsets of $\Rd$ such that $E_{\varepsilon}\cap B^c = H_{\hat{p}}\cap B^c$ and that $E_{\varepsilon}\to H_{\hat{p}}$ in $L^1(B)$. By Theorem \[stm:piani\], we have that $$\begin{aligned} 0 & \leq J_{{\varepsilon}}(E_{\varepsilon}; B) - J_{{\varepsilon}}(H_{\hat{p}}; B) \\ & = J^1_{{\varepsilon}}(E_{\varepsilon}; B) - J^1_{{\varepsilon}}(H_{\hat{p}}; B) - \left[ J^2_{{\varepsilon}}(E_{\varepsilon}; B) - J^2_{{\varepsilon}}(H_{\hat{p}}; B) \right]. \end{aligned}$$ If we also assume that $E_{\varepsilon}{\!\bigtriangleup\!}H_{\hat{p}} \subset B_{1-\delta}$, we see that $$\begin{gathered} J^2_{{\varepsilon}}(E_{\varepsilon}; B) - J^2_{{\varepsilon}}(H_{\hat{p}}; B) \\ = \int_{E_{\varepsilon}\cap B_{1-\delta}}\int_{H_{\hat{p}} \cap B^c} K_{\varepsilon}(y-x) \de y \de x - \int_{H_{\hat{p}} \cap B_{1-\delta}}\int_{H_{\hat{p}} \cap B^c} K_{\varepsilon}(y-x) \de y \de x \end{gathered}$$ and hence, noticing that ${\left\lverty-x\right\rvert}\geq\delta$ if $x\in B^c$ and $y\in B_{1-\delta}$, $$\begin{aligned} \frac{1}{{\varepsilon}}{\left\lvert J^2_{{\varepsilon}}(E_{\varepsilon}; B) - J^2_{{\varepsilon}}(H_{\hat{p}}; B) \right\rvert} & \leq \frac{2}{\delta} \int_{ E_{\varepsilon}{\!\bigtriangleup\!}H _{\hat{p}}} \int_{ B^c } K_{\varepsilon}(y-x) \frac{{\left\lverty-x\right\rvert}}{{\varepsilon}}\de y \de x \\ & \leq \frac{2}{\delta} \Ld (E_{\varepsilon}{\!\bigtriangleup\!}H _{\hat{p}}) \int_{ \Rd } K(z) {\left\lvertz\right\rvert} \de z. \end{aligned}$$ By our choice of ${\left\{E_{\varepsilon}\right\}}$ and , this yields $$\lim_{{{\varepsilon}}\to 0^+} \frac{1}{{\varepsilon}} {\left\lvert J^2_{{\varepsilon}}(E_{\varepsilon}; B) - J^2_{{\varepsilon}}(H_{\hat{p}}; B) \right\rvert} = 0,$$ whence $$\begin{aligned} 0 & \leq \liminf_{{\varepsilon}\to 0^+} \frac{1}{{\varepsilon}} \left[ J_{{\varepsilon}}(E_{\varepsilon}; B) - J_{{\varepsilon}}(H_{\hat{p}}; B) \right] \\ & = \liminf_{{\varepsilon}\to 0^+} \frac{1}{{\varepsilon}} \left[ J^1_{{\varepsilon}}(E_{\varepsilon}; B) - J^1_{{\varepsilon}}(H_{\hat{p}}; B) \right]. \end{aligned}$$ Recalling and the definition of $\sigma''_K$, we deduce $\sigma_K(\hat{p}) \leq \sigma''_K(\hat{p})$. To conclude, we are left to show that $\sigma''_K \leq \sigma'_K$. This may be done as in the proof of [@BP Lemma 3.11] by means of a suitable “gluing” lemma (see also [@ADM]). Acknowledgements {#acknowledgements .unnumbered} ---------------- The author warmly thanks Matteo Novaga for suggesting the problem and for providing insights about it. Part of this paper was written during a stay of the author at the Centre de Mathématiques Appliquées (CMAP) of the École Polytechnique in Paris. The author thanks the institution for hospitality and the Unione Matematica Italiana (UMI) for partially funding the stay via the UMI Grants for Ph.D. students.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the minimality of an isometric immersion of a Riemannian manifold into a strictly pseudoconvex CR manifold $M$ endowed with the Webster metric (associated to a fixed contact form on $M$), hence formulate a version of the CR Yamabe problem for CR manifolds-with-boundary. This is shown to be a nonlinear subelliptic problem of variational origin.' --- Minimality in CR geometry and the CR Yamabe 0.1in problem on CR manifolds with boundary Sorin Dragomir[^1] Introduction ============ Minimal surfaces $N^2$ in the lowest dimensional Heisenberg group ${\mathbb H}_1$, or more generally in a $3$-dimensional nondegenerate CR manifold, have been recently considered by a number of people (cf. N. Arcozzi & F. Ferrari, [@kn:ArFe], I. Birindelli & E. Lanconelli, [@kn:BiLa], J-H. Cheng et alt., [@kn:CHMY], N. Garofalo & S.D. Pauls, [@kn:GaPa], and S.D. Pauls, [@kn:Pau]) motivated by the interest in a Heisenberg version of the Bernstein problem, or by anticipating an appropriate formulation of the CR Yamabe problem on a CR manifold-with-boundary and a CR analog to the positive mass theorem. All the notions of minimality dealt with are but ordinary minimality of $N^2$ with respect to the ambient Webster metric. This is demonstrated by our Theorem \[c:1\] (though confined to the case where the characteristic direction $T = \partial /\partial t$ of ${\mathbb H}_1$ is tangent to $N^2$). We also study minimality of a given isometric immersion $\Psi : N^m \to {\mathbb H}_n$ of a $m$-dimensional Riemannian manifold $(N^m , g)$ into $({\mathbb H}_n , g_{\theta_0} )$ (the Heisenberg group carrying the Webster metric $g_{\theta_0}$ associated with the contact form $\theta_0 = d t + i \sum_{j=1}^n (z_j d \overline{z}^j - \overline{z}_j d z^j )$), cf. our Theorem \[t:min\]. A first step towards a Weierstrass type representation of minimal surfaces in ${\mathbb H}_n$ is taken in Theorem \[t:7\]. The Yamabe problem on a compact $n$-dimensional ($n \geq 3$) Riemannian manifold $(M , g)$ with boundary $\partial M$ is to deform conformally the given metric $\hat{g} = u^{4/(n-2)} g$ ($u > 0$) such that $(M , \hat{g})$ has constant scalar curvature and $\partial M$ is minimal in $(M , \hat{g})$. This is equivalent to solving the boundary value problem $$\Delta u - \frac{n-2}{4(n-1)} \rho_g u + C u^{(n+2)/(n-2)} = 0 \;\; {\rm in} \;\; M, \label{e:1}$$ $$\frac{\partial u}{\partial \eta} + \frac{n-2}{2} h_g u = 0 \;\; {\rm on} \;\; \partial M, \label{e:2}$$ where $\Delta$ and $\rho_g$ are respectively the Laplace-Beltrami operator and the scalar curvature of $(M , g)$, $h_g$ is the mean curvature of $\partial M \hookrightarrow (M , g)$, and $\eta$ is a unit outward normal on $\partial M$ with respect to $g$. When $M$ is closed ($\partial M = \emptyset$) the full solution to (\[e:1\]) is described in [@kn:LePa]. When $\partial M \neq \emptyset$ the problem (\[e:1\])-(\[e:2\]) was solved by J.F. Escobar, [@kn:Esc], under the assumptions that 1) $n \in \{ 3,4,5\}$, or 2) $n \geq 3$ and $\partial M$ has some nonumbilic point, or 3) $n \geq 6$, $\partial M$ is totally umbilical, and either $M$ is locally conformally flat or the Weyl tensor doesn’t vanish identically on $\partial M$. A CR analog of the Yamabe problem was formulated by D. Jerison & J.M. Lee, [@kn:JL1], though only on closed CR manifolds. Precisely, if $M$ is a $(2n+1)$-dimensional closed strictly pseudoconvex CR manifold on which a contact form $\theta$ has been fixed then the CR Yamabe problem is to look for a contact form $\hat{\theta} = u^{p-2} \theta$ ($p = 2 + 2/n$) such that the Tanaka-Webster connection of $(M , \hat{\theta})$ has constant pseudohermitian scalar curvature $\hat{\rho} = \lambda$. This is equivalent to solving $$- (2 + 2/n) \, \Delta_b u + \rho \, u = \lambda \, u^{p-1} \label{e:3}$$ (the [CR Yamabe equation]{}) where $\Delta_b$ and $\rho$ are respectively the sublaplacian[^2] and the pseudohermitian scalar curvature of $(M , \theta )$. D. Jerison & J.M. Lee solved (cf. [@kn:JL2]-[@kn:JL3]) the problem (\[e:3\]) under the assumption that[^3] $\lambda (M) < \lambda (S^{2n+1})$, where $\lambda (M)$ is the CR invariant $$\inf \{ \int_M ( b_n \\| \pi_H \nabla u \\|^2 + \rho u^2 ) \theta \wedge (d \theta )^n : \int_M \|u\|^p \theta \wedge (d \theta )^n = 1 \} .$$ Moreover, the inequality $\lambda (M) \leq \lambda (S^{2n+1})$ holds true. The remaining case $\lambda (M) = \lambda (S^{2n+1})$ was settled by N. Gamara & R. Yacoub, [@kn:GaYa]. It is noteworthy that the proof in [@kn:GaYa] doesn’t rely on a CR analog to the positive mass theorem, but rather on techniques within the theory of critical points at infinity (by analogy with A. Bahri & H. Brezis, [@kn:BaBr]). When $\partial M \neq \emptyset$ no formulation of the CR Yamabe problem is available as yet, perhaps due to the previous lack of a natural CR analog to minimality. Our approach (as well as in [@kn:JL2]) is to formulate the CR Yamabe problem as the Yamabe problem for the Fefferman metric $F_\theta$, a Lorentz metric on the total space $C(M)$ of the canonical circle bundle $S^1 \to C(M) \stackrel{\pi}{\rightarrow} M$ (cf. [@kn:Lee]). That is, to look for a positive function $u \in C^\infty (M)$ such that the Fefferman metric $F_{\hat{\theta}}$ corresponding to the contact form $\hat{\theta} = u^{p-2} \theta$ has constant scalar curvature. What is the appropriate boundary condition? When $\partial M$ is nonempty $C(M)$ is a manifold-with-boundary as well, and (by Theorem \[p:5\]) the tangent space $T_z (\partial C(M))$ is nondegenerate in $(T_z (C(M)), F_{\theta , z})$ at all points $z$, except for those projecting on ${\rm Sing}(T^T )$, the singular points of the tangential component (with respect to $\partial M$) of the characteristic direction $T$ of $d \theta$. It also turns out that $\partial C(M) \setminus \pi^{-1}({\rm Sing}(T^T ))$ is a Lorentz manifold (with the metric induced by $F_\theta$). Therefore, when ${\rm Sing}(T^T ) = \emptyset$ we may request that $\partial C(M)$ be minimal in $(C(M), F_{\hat{\theta}})$. By Theorem \[t:1\] this projects to the natural boundary condition (\[e:34\]) on $\partial M$, thus leading to the [CR Yamabe problem]{} (\[e:33\])-(\[e:34\]) on a CR manifold-with-boundary. This is shown (cf. Theorem \[p:11\]) to be a nonlinear subelliptic problem of variational origin. 0.1in [. The Author is grateful to E. Lanconelli for stimulating conversations on the arguments in this paper and for introducing him to the results in the preprint [@kn:BiLa]. Also, the Author wishes to express his gratitude for the hospitality and excellent working atmosphere in the Department of Mathematics of the University of Bologna and for discussions with N. Arcozzi and F. Ferrari (who kindly provided the preprint [@kn:ArFe]).]{} CR manifolds with boundary ========================== Let $M$ be an oriented $m$-dimensional $C^\infty$ manifold-with-boundary $\partial M$. A [CR structure]{} is a complex subbundle $T_{1,0}(M)$ of the complexified tangent bundle $T(M) \otimes {\mathbb C}$, of complex rank $n$ ($0 < n \leq [m/2]$), such that $$T_{1,0}(M) \cap T_{0,1}(M) = (0),$$ $$Z,W \in \Gamma^\infty (T_{1,0}(M)) \Longrightarrow [Z,W] \in \Gamma^\infty (T_{1,0}(M)).$$ Here $T_{0,1}(M) = \overline{T_{1,0}(M)}$ (complex conjugation). The pair $(M , T_{1,0}(M))$ is a [CR manifold]{} (with boundary) and the integer $n$ is its [CR dimension]{}. Also $k = m - 2n$ is its [CR codimension]{} and the pair $(n,k)$ is its [type]{}. There is a natural first order differential operator $\overline{\partial}_b$ (the [tangential Cauchy-Riemann operator]{}) given by $(\overline{\partial}_b u) \overline{Z} = \overline{Z}(u)$, for any $C^1$ function $u : M \to {\mathbb C}$ and any $Z \in T_{1,0}(M)$. Then $\overline{\partial}_b u = 0$ are the [tangential Cauchy-Riemann equations]{}. A solution to the tangential Cauchy-Riemann equations is a [CR function]{} on $M$. Let ${\rm CR}^r (M)$ denote the space of all CR functions on $M$ of class $C^r$. The boundary $\partial M$ is [noncharacteristic]{} for $T_{1,0}(M)$ if for any local frame $\{ T_\alpha : 1 \leq \alpha \leq n \}$ of $T_{1,0}(M)$ defined on the open subset $U \subseteq M$ one has $T_\alpha \not\in T(\partial M) \otimes {\mathbb C}$ (i.e. $T_{\alpha , x} \not\in T_x (\partial M) \otimes {\mathbb C}$, for some $x \in U \cap \partial M$) for some $1 \leq \alpha \leq n$. The [Levi distribution]{} of the CR manifold $(M , T_{1,0}(M))$ is $$H(M) = {\rm Re}\{ T_{1,0}(M) \oplus T_{0,1}(M) \} .$$ It carries the complex structure $$J : H(M) \to H(M), \;\; J(Z + \overline{Z}) = i (Z - \overline{Z}), \;\; Z \in T_{1,0}(M).$$ Assume from now on that $M$ is a CR manifold of type $(n,1)$ (of [hypersurface type]{}). $H(M)$ is oriented by $J$, hence the conormal bundle $$H(M)^\bot_x = \{ \omega \in T^_x (M) : {\rm Ker}(\omega ) \supseteq H(M)_x \} , \;\;\; x \in M,$$ is an oriented real line bundle, hence trivial. Let then $\theta$ be a global nowhere vanishing section in $H(M)^\bot$ (a [pseudohermitian structure]{} on $M$). The [Levi form]{} is $$L_\theta (Z, \overline{W}) = - i (d \theta )(Z, \overline{W}), \;\;\; Z, W \in T_{1,0}(M),$$ and $M$ is [nondegenerate]{} (respectively [strictly pseudoconvex]{}) if $L_\theta$ is nondegenerate (respectively positive definite) for some $\theta$. Also $M$ is [Levi flat]{} if $L_\theta = 0$ (equivalently, if $H(M)$ is integrable). An alternative definition of the Levi form is $$G_\theta (X,Y) = (d \theta )(X, J Y), \;\;\; X,Y \in H(M).$$ Note that $L_\theta$ and the ${\mathbb C}$-linear extension of $G_\theta$ coincide on $T_{1,0}(M) \otimes T_{0,1}(M)$. If $M$ is nondegenerate then any pseudohermitian structure $\theta$ is a [contact form]{}, i.e. $\theta \wedge (d \theta )^n$ is a volume form on $M$. Let $M$ be a nondegenerate CR manifold and $\theta$ a fixed contact form (the pair $(M, \theta )$ is commonly referred to as a [pseudohermitian manifold]{}). There is a unique vector field $T$ on $M$ such that $\theta (T) = 1$ and $(d \theta )(T , X) = 0$, for any $X \in T(M)$ ($T$ is the [characteristic direction]{} of $d \theta$). The [Webster metric]{} of $(M , \theta )$ is given by $$g_\theta (X,Y) = G_\theta (X , J Y), \;\; g_\theta (X,T) = 0, \;\; g_\theta (T,T) = 1,$$ for any $X,Y \in H(M)$. $g_\theta$ is a semi-Riemannian (Riemannian, if $M$ is strictly pseudoconvex and $L_\theta$ is positive definite) metric on $M$. Let $M$ be a nondegenerate CR manifold-with-boundary. Then the boundary $\partial M$ is noncharacteristic for $T_{1,0}(M)$. \[p:1\] The proof is by contradiction. Assume that there is a local frame $\{ T_\alpha \}$ of $T_{1,0}(M)$ on $U \subseteq M$ such that $T_\alpha \in T(\partial M) \otimes {\mathbb C}$, for all $1 \leq \alpha \leq n$. Then $T_{1,0}(M)_x \subset T_x (\partial M) \otimes {\mathbb C}$ for any $x \in U \cap \partial M$. Then, by taking complex conjugates, $T_{0,1}(M)_x \subset T_x (\partial M) \otimes {\mathbb C}$ hence, by looking at dimensions, $H(M)_x = T_x (\partial M)$, i.e. $L_{\theta , x} = 0$, a contradiction. $\square$ 0.1in From now on we assume that $M$ is nondegenerate. For each boundary point $x \in \partial M$ we set $$T_{1,0}(\partial M)_x = T_{1,0}(M)_x \cap [T_x (\partial M) \otimes {\mathbb C}].$$ Let $\{ T_\alpha : 1 \leq \alpha \leq n \}$ be a local frame of $T_{1,0}(M)$, defined on the local coordinate neighborhood $(U , \varphi = (x^1 , \cdots , x^{2n+1}))$. $U \cap \partial M$ consists of the points $x \in U$ such that $\varphi (x) \in \partial {\mathbb R}^{2n+1}_+ = {\mathbb R}^{2n} \times \{ 0 \}$. We may write $T_\alpha = f^A_\alpha \; \partial /\partial x^A$, for some $C^\infty$ functions $f^A_\alpha : U \to {\mathbb C}$. By Proposition \[p:1\] there is $\alpha$, say $\alpha = n$, such that $T_\alpha \not\in T(\partial M) \otimes {\mathbb C}$. Then $f_n^{2n+1}(x_0 ) \neq 0$ for some $x_0 \in U \cap \partial M$, and then $f^{2n+1}_n \neq 0$ on a whole neighborhood of $x_0$, which we may denote again by $U$. Then $$\{ T_j - \left( \lambda^{2n+1}_j /\lambda^{2n+1}_n \right) \; T_n : 1 \leq j \leq n-1 \}$$ is a local frame of $T_{1,0}(\partial M)$ on $U \cap \partial M$, hence $T_{1,0}(\partial M)$ has rank $n-1$. We got Let $M$ be a nondegenerate CR manifold-with-boundary, of CR dimension $n$. Then its boundary $\partial M$ is a CR manifold of type $(n-1,2)$, i.e. $T_{1,0}(\partial M) = T_{1,0}(M) \cap [T(\partial M) \otimes {\mathbb C}]$ is a CR structure of CR codimension $2$. \[p:2\] Let us look at a few examples. For instance, let ${\mathbb H}_n = {\mathbb C}^n \times {\mathbb R}$ be the Heisenberg group, with the CR structure spanned by $$Z_j = \frac{\partial}{\partial z_j} + i \, \overline{z}_j \, \frac{\partial}{\partial t}\, , \;\;\; 1 \leq j \leq n,$$ (if $n=1$ then $\overline{Z}_1$ is the [Lewy operator]{}, cf. [@kn:Lew]). ${\mathbb H}_n$ is a Lie group with the group law $$(z, t) \cdot (w , s) = (z + w , t + s + 2 \; {\rm Im}(z \cdot \overline{w})),$$ for $(z,t), \; (w, t) \in {\mathbb H}_n$, where $z \cdot \overline{w} = \delta_{jk} z^j \overline{w}^k$ (with the convention $z^j = z_j$), and $Z_j$ are left invariant. 0.1in[Example 1.]{} ${\mathbb H}_n^+ = \{ (z,t) \in {\mathbb H}_n : t \geq 0 \}$ is a CR manifold-with-boundary $\partial {\mathbb H}_n^+ = {\mathbb C}^n \times \{ 0 \}$. Let $U = \{ (z,t) \in {\mathbb H}_n^+ : z_n \neq 0 \}$. Then $$\{ \frac{\partial}{\partial z^a} - \frac{\overline{z}_a}{\overline{z}_n} \; \frac{\partial}{\partial z_n} : 1 \leq a \leq n-1 \} \label{e:4}$$ is a local frame of $T_{1,0}(\partial {\mathbb H}_n^+ )$ on $U \cap \partial {\mathbb H}_n^+$. In particular, the tangential Cauchy-Riemann equations on $\partial {\mathbb H}^+_n$ are $$z_n \; \frac{\partial u}{\partial \overline{z}_a} - z_a \; \frac{\partial u}{\partial \overline{z}_n} = 0, \;\;\; 1 \leq a \leq n-1.$$ $\square$ 0.1inThe [Heisenberg norm]{} is $\|x\| = (\|z\|^4 + t^2 )^{1/4}$, for any $x = (z,t) \in {\mathbb H}_n$, where $\|z\|^2 = z \cdot \overline{z}$. 0.1in[Example 2.]{} $\Omega_r = \{ x \in {\mathbb H}_n : \|x\| \leq r \}$ ($r > 0$) is a CR manifold-with-boundary $\partial \Omega_r = \Sigma_r = \{ x \in {\mathbb H}_n : \|x\| = r \}$ (the [Heisenberg sphere]{}, cf. [@kn:GaLa]). Let us set $\phi (z,t) = \|z\|^2 - i \; t$. Note that $\overline{\partial}_b \phi = 0$, i.e. $\phi \in {\rm CR}^\infty ({\mathbb H}_n )$. Taking into account that $$Z_j (\|x\|) = \frac{\overline{\phi}}{2\, \|x\|^3} \; \overline{z}_j \, , \;\;\; 1 \leq j \leq n,$$ it follows that (\[e:4\]) is a local frame of $T_{1,0}(\Sigma_r )$ on $\Sigma_r \cap \{ z \in {\mathbb H}_n : z_n \neq 0 \}$. The [Folland-Stein operators]{} are $${\mathcal L}_\alpha = - \frac{1}{2} \sum_{j=1}^n (Z_j \overline{Z}_j + \overline{Z}_j Z_j ) + i \; \alpha \; T, \;\;\; \alpha \in {\mathbb C}, \label{e:folste}$$ where $T = \partial /\partial t$. Let us consider the function $$\varphi_\alpha (z,t) = \phi (z,t)^{- (n+\alpha )/2} \; \overline{\phi (z,t)}^{\; - (n-\alpha )/2} \, ,$$ and the constant $c_\alpha = 2^{2-2n} \pi^{n+1}/\left( \Gamma (\frac{n+\alpha}{2}) \; \Gamma (\frac{n-\alpha}{2})\right)$. $\alpha \in {\mathbb C}$ is [admissible]{} if $c_\alpha \neq 0$ (equivalently if $\pm \alpha \in \{ n , \, n+2, \, n+4 , \, \cdots \}$). The Folland-Stein operators (\[e:folste\]) form a family of operators of the form $A + \alpha B$ (where $A$ is a second order hypoelliptic operator and $B$ is a first order operator) which are hypoelliptic for any admissible $\alpha$ (cf. [@kn:FoSt], p. 444). This is by now classical, and as well known the key ingredient in the proof is to build a fundamental solution to (\[e:folste\]) i.e. to show that ${\mathcal L}_\alpha ( \varphi_\alpha /c_\alpha ) = \delta$, for any admissible $\alpha$. It is noteworthy that the Heisenberg spheres $\Sigma_r$ are the level sets of $$\varphi_0 (z,t) = \|\phi (z,t)\|^{-n} = \left( \|z\|^4 + t^2 \right)^{-n/2}\, .$$ Let $\theta_0$ be the canonical pseudohermitian structure on ${\mathbb H}_n$ i.e. $$\theta_0 = d t + i \sum_{j=1}^n (z_j d \overline{z}^j - \overline{z}_j d z^j ).$$ ${\mathbb H}_n$ is strictly pseudoconvex and $L_{\theta_0}$ is positive definite. Moreover, the Webster metric of $({\mathbb H}_n , \theta_0 )$ is expressed by $$g_{\theta_0}(X_j , X_k ) = g_{\theta_0}(Y_j , Y_k ) = \delta_{jk} \, , \;\;\; g_{\theta_0}(X_j , Y_k ) = 0,$$ $$g_{\theta_0}(X_j , T) = g_{\theta_0} (Y_j , T) = 0, \;\;\; g_{\theta_0}(T,T) = 1,$$ where $$X_j = \frac{1}{\sqrt{2}} \, (Z_j + \overline{Z}_j ), \;\;\; Y_j = \frac{i}{\sqrt{2}} \, (Z_j - \overline{Z}_j ).$$ The Heisenberg spheres form a foliation of $({\mathbb H}_n , g_{\theta_0})$ whose normal bundle is the span of $$V = T + (\phi /t) z^j Z_j + (\overline{\phi}/t) \overline{z}^j \overline{Z}_j \, . \label{e:6}$$ \[p:fol\] Then perhaps (\[e:6\]) is the Heisenberg analog to the radial vector field in ${\mathbb R}^{2n+1}$ (see [@kn:GaLa], p. 331-332). [Proof of Proposition]{} \[p:fol\]. Let us set $$E_j = Z_j + \overline{Z}_j - \frac{1}{t} (\phi z_j + \overline{\phi} \overline{z}_j ) T, \;\; F_j = i(Z_j - \overline{Z}_j ) + \frac{i}{t} (\phi z_j - \overline{\phi} \overline{z}_j ) T.$$ Then $\{ E_j , F_j \}$ is a local frame of the tangent bundle of the foliation and a calculation shows that (\[e:6\]) satisfies $g_{\theta_0} (E_j , V) = g_{\theta_0}(F_j , V) = 0$. $\square$ 0.1in Let $M$ and $N$ be two CR manifolds with boundary. A [CR map]{} is a $C^\infty$ map $f : M \to N$ such that $(d_x f) T_{1,0}(M)_x \subseteq T_{1,0}(N)_{f(x)}$, for any $x \in M$. A [CR immersion]{} is an immersion and a CR map. A CR immersion $f : M \to N$ is [neat]{} if i) $f(M) \cap \partial N = f(\partial M)$ and ii) for each point $x \in \partial M$ there is a local chart $\psi : V \to {\mathbb R}^{m+p}_+$ of $N$ such that $f(x) \in V$ and $\psi^{-1}({\mathbb R}^m_+ ) = V \cap f(M)$ ($m = \dim (M)$). 0.1in[Example 3.]{} $\Sigma_r^+ = \Sigma_r \cap {\mathbb H}^+_n$ is a CR manifold-with-boundary $\partial \Sigma_r^+ = S^{2n-1}(r) \times \{ 0 \}$ and the inclusion $\Sigma^+_r \to {\mathbb H}^+_n$ is a neat CR immersion. $\square$ 0.1in[Example 4.]{} $S^{2n+1}_+ = S^{2n+1} \cap {\mathbb R}^{2n+2}_+$ is a CR manifold-with-boundary $\partial S^{2n+1}_+ = S^{2n} \times \{ 0 \}$. Let ${\mathcal C}$ be the Cayely transform $${\mathcal C}(\zeta ) = (\frac{\zeta^\prime}{1 + \zeta^{n+1}} \, , \, i \, \frac{1 - \zeta^{n+1}}{1 + \zeta^{n+1}}), \;\; \zeta = (\zeta^\prime , \zeta^{n+1}), \;\; 1 + \zeta^{n+1} \neq 0,$$ and $f : {\mathbb H}_n \to \partial \Omega_{n+1}$ the CR isomorphism $f(z,t) = (z, t+ i \|z\|^2 )$ with the obvious inverse $f^{-1}(z,w) = (z, {\rm Re}(w))$. Here $\Omega_{n+1}$ is the Siegel domain $$\Omega_{n+1} = \{ (z, w) \in {\mathbb C}^{n+1} : {\rm Im}(w) > \|z\|^2 \} .$$ Then $F = f^{-1} \circ {\mathcal C}$ is a neat CR diffeomorphism $$F : S^{2n+1}_+ \setminus \{ (0, \cdots , 0, -1)\} \to {\mathbb H}_n^+ .$$ Indeed if $\zeta \in S^{2n+1}_+$ and $\zeta^{n+1} = u + iv$ ($v \geq 0$) and $(z,t) = F(\zeta )$ then $t = 2v/[(1+u)^2 + v^2 ] \geq 0$. In particular $F$ descends to a CR diffeomorphism $(S^{2n} \times \{ 0 \}) \setminus \{ (0, \cdots , 0 , -1) \} \approx {\mathbb C}^n \times \{ 0 \}$. $\square$ 0.1in Let $M$ be a nondegenerate CR manifold-with-boundary. A complex $p$-form $\eta$ on $M$ is a $(p,0)$-[form]{} if $T_{0,1}(M) \, \rfloor \, \eta = 0$. Let $\Lambda^{p,0}(M) \to M$ be the bundle of all $(p,0)$-forms. If $M$ has CR dimension $n$ then the top degree $(p,0)$-forms are the $(n+1,0)$-forms. $K(M) = \Lambda^{n+1,0}(M)$ is the [canonical bundle]{} over $M$. There is a natural action of ${\mathbb R}_+ = (0, + \infty )$ on $K(M) \setminus \{ 0 \}$. Let $C(M)$ be the quotient space and $\pi : C(M) \to M$ the projection. This construction leads to a principal bundle $S^1 \to C(M) \to M$ (the [canonical circle bundle]{} over $M$). Let $\theta$ be a pseudohermitian structure on $M$ and $T$ the characteristic direction of $d \theta$. Given a local frame $\{ T_\alpha \}$ of $T_{1,0}(M)$ on a local coordinate neighborhood $(U, x^A )$, let $\theta^\alpha$ be the locally defined complex $1$-forms determined by $$\theta^\alpha (T_\beta ) = \delta_\beta^\alpha \, , \;\; \theta^\alpha (T_{\overline{\beta}} ) = 0, \;\; \theta^\alpha (T) = 0.$$ Here $T_{\overline{\alpha}} = \overline{T}_\alpha$. Then $$\pi^{-1}(U) \to U \times S^1 \, , \;\; [z] \mapsto (x \, , \, \lambda /\|\lambda \| ),$$ $$z = \lambda \, (\theta \wedge \theta^1 \wedge \cdots \wedge \theta^n )_x \, , \;\; \lambda \in {\mathbb C} \setminus \{ 0 \} , \;\; x \in M,$$ is a local trivialization chart of the canonical circle bundle. Let us set $\gamma : \pi^{-1} (U) \to {\mathbb R}$, $\gamma ([z]) = \arg (\lambda )$ (where $\arg : {\mathbb C} \to [0, 2 \pi )$). Then $(\pi^{-1}(U) , \, \tilde{x}^A = x^A \circ \pi , \, \gamma )$ are naturally induced local coordinates on $C(M)$ and $\pi^{-1}(U \cap \partial M)$ consists of all $c \in \pi^{-1}(U)$ with $\tilde{x}^{2n+1}(c) = 0$, i.e. $C(M)$ is a manifold-with-boundary modelled on ${\mathbb R}^{2n+1}_+ \times {\mathbb R}$. We obtained Let $M$ be a nondegenerate CR manifold-with-boundary. Then the total space $C(M)$ of the canonical circle bundle is a manifold-with-boundary $\partial C(M) = \pi^{-1}(\partial M)$. In particular $\partial C(M)$ is a principal $S^1$-bundle over $\partial M$. Let $\nabla$ be the unique linear connection on $M$ (the [Tanaka-Webster connection]{}) satisfying the axioms 1) $H(M)$ is parallel with respect to $\nabla$, 2) $\nabla J = 0$, $\nabla g_\theta = 0$, and 3) the torsion $T_\nabla$ of $\nabla$ is [pure]{}, i.e. $T_\nabla (Z, W) = 0$, $T_\nabla (Z, \overline{W}) = 2 i G_\theta (Z , \overline{W}) T$, and $\tau \circ J + J \circ \tau = 0$. Here $\tau (X) = T_\nabla (T ,X)$ is the [pseudohermitian torsion]{}. We set $A(X,Y) = g_\theta (\tau X, Y)$, for any $X,Y \in T(M)$. By a result of S. Webster, [@kn:Web], $A$ is symmetric. With respect to a local frame $\{ T_\alpha : 1 \leq \alpha \leq n \}$ of $T_{1,0}(M)$, defined on an open set $U \subseteq M$, it is customary to set $g_{\alpha\overline{\beta}} = L_\theta (T_\alpha , T_{\overline{\beta}})$ (the local coefficients of the Levi form), $\nabla T_\beta = \omega_\beta^\alpha \otimes T_\alpha$ (the connection $1$-forms) and $R^\nabla (T_A , T_B ) T_C = {{R_C}^D}_{AB} T_D$ (the curvature components). The range of the indices $A,B,C, \cdots$ is $\{ 0, 1, \cdots , n , \overline{1}, \cdots , \overline{n} \}$ (with the convention $T_0 = T$). Next, the [pseudohermitian Ricci tensor]{} is $R_{\lambda\overline{\mu}} = {{R_\lambda}^\alpha}_{\alpha\overline{\mu}}$ and the [pseudohermitian scalar curvature]{} is $\rho = g^{\alpha\overline{\beta}} R_{\alpha\overline{\beta}}$. When $M$ is strictly pseudoconvex and $\theta$ is a pseudohermitian structure such that $L_\theta$ is positive definite $C(M)$ carries a Lorentz metric $F_\theta$ such that $F_{\hat{\theta}} = e^{u \circ \pi} F_\theta$, where $\hat{\theta} = e^u \theta$, $u \in C^\infty (M)$ (in particular the [restricted conformal class]{} $[F_\theta ] = \{ e^{u \circ \pi} F_\theta : u \in C^\infty (M) \}$ is a CR invariant). Cf. J.M. Lee, [@kn:Lee], $F_\theta$ is given by $$F_\theta = \pi^ \tilde{G}_\theta + 2 (\pi^* \theta ) \odot \sigma , \label{e:ad2}$$ $$\label{e:ad1} \sigma = \frac{1}{n+2} \{ d \gamma + \pi^* (i \, \omega^\alpha_\alpha - \frac{i}{2} \, g^{\alpha\overline{\beta}} d g_{\alpha\overline{\beta}} - \frac{\rho}{4(n+1)} \, \theta ) \} .$$ $F_\theta$ is the [Fefferman metric]{} of $(M , \theta )$. Here $\tilde{G}_\theta$ is the degenerate $(0,2)$-tensor field on $M$ given by $$\tilde{G}_\theta (X,Y) = (d \theta )(X, J Y), \;\;\; \tilde{G}_\theta (T , Z) = 0,$$ for any $X, Y \in H(M)$ and any $Z \in T(M)$. Also $\odot$ denotes the symmetric tensor product. Let $S = \partial /\partial \gamma$ be the tangent to the $S^1$-action. $\sigma$ is a connection $1$-form in $S^1 \to C(M) \to M$. If $X \in T(M)$ is a tangent vector field on $M$ then $X^\uparrow \in T(C(M))$ will denote the horizontal lift of $X$ with respect to the connection ${\mathcal H} = {\rm Ker}(\sigma )$. Although the submersion $\pi : C(M) \to M$ is not semi-Riemannian (its fibres are degenerate) a technique similar to that in [@kn:One] leads to For any $X,Y \in H(M)$ $$\nabla^{C(M)}_{X^\uparrow} Y^\uparrow = (\nabla_X Y)^\uparrow - (d \theta )(X,Y) T^\uparrow - (A(X,Y) + (d \sigma )(X^\uparrow , Y^\uparrow )) \hat{S},$$ $$\nabla^{C(M)}_{X^\uparrow} T^\uparrow = (\tau X + \phi X)^\uparrow ,$$ $$\nabla^{C(M)}_{T^\uparrow} X^\uparrow = (\nabla_T X + \phi X)^\uparrow + 2(d \sigma )(X^\uparrow , T^\uparrow ) \hat{S},$$ $$\nabla^{C(M)}_{X^\uparrow} \hat{S} = \nabla^{C(M)}_{\hat{S}} X^\uparrow = (J X)^\uparrow ,$$ $$\nabla^{C(M)}_{T^\uparrow} T^\uparrow = V^\uparrow , \;\; \nabla^{C(M)}_{\hat{S}} \hat{S} = 0,$$ $$\nabla^{C(M)}_{\hat{S}} T^\uparrow = \nabla^{C(M)}_{T^\uparrow} \hat{S} = 0,$$ where $\phi : H(M) \to H(M)$ is given by $G_\theta (\phi X , Y) = (d \sigma )(X^\uparrow , Y^\uparrow )$, and $V \in H(M)$ is given by $G_\theta (V , Y) = 2 (d \sigma )(T^\uparrow , Y^\uparrow )$. Also $\hat{S} = ((n+2)/2) S$. \[l:1\] Lemma \[l:1\] relates the Levi-Civita connection $\nabla^{C(M)}$ of $(C(M), F_\theta )$ to the Tanaka-Webster connection of $(M, \theta )$. Cf. [@kn:BaDrUr] for the proof of Lemma \[l:1\]. The geometry of the first fundamental form of the boundaries ============================================================ Let $M$ be a strictly pseudoconvex CR manifold and $\theta$ a contact form on $M$ such that $G_\theta$ is positive definite. Let $T(\partial M)^\bot \to \partial M$ be the normal bundle of $\partial M \hookrightarrow (M , g_\theta )$. Let ${\rm tan}_x : T_x (M) \to T_x (\partial M)$ and ${\rm nor}_x : T_x (M) \to T(\partial M)^\bot_x$ be the projections associated with the direct sum decomposition $$T_x (M) = T_x (\partial M) \oplus T(\partial M)^\bot_x \, , \;\;\; x \in \partial M.$$ If $T$ is the characteristic direction of $d \theta$ then we set $T^\bot = {\rm nor} (T)$ and $T^T = {\rm tan}(T)$. Let ${\rm Null}(j^* F_\theta )$ consist of all $V \in T(\partial C(M))$ such that $F_\theta (V,W) = 0$, for any $W \in T(\partial C(M))$. Let us consider the closed set ${\rm Sing}(T^T ) = \{ x \in \partial M : T^T_x = 0 \}$ and set $\Omega = \partial M \setminus {\rm Sing}(T^T )$. Then $${\rm Null}(j^* F_\theta )_z = \begin{cases} 0, & z \in \pi^{-1}(\Omega ), \cr {\rm Ker}(d_z \pi ), & z \in \pi^{-1}({\rm Sing}(T^T )), \cr \end{cases}$$ for any $z \in \partial C(M)$. Moreover $(\pi^{-1}(\Omega )\, , \, j^* F_\theta )$ is a Lorentz manifold. \[p:5\] Here $j : \partial C(M) \hookrightarrow C(M)$ is the inclusion. Hence $\partial C(M)$ is degenerate at each point $z \in \pi^{-1}({\rm Sing}(T^T ))$. In particular, if $\partial M$ is tangent to $T$ then the boundary $(\partial C(M) \, , \, j^* F_\theta )$ is a Lorentz manifold. 0.1in[Example 1.]{} ([continued]{}) $T(\partial {\mathbb H}_n^+ )$ is the span of $\{ X_j - \sqrt{2}\, y_j T , \; Y_j + \sqrt{2} \, x_j T : 1 \leq j \leq n \}$ hence $\xi = T + \sqrt{2} \, y^j X_j - \sqrt{2} \, x^j Y_j$ is normal to $\partial {\mathbb H}_n^+$ (with $z^j = x^j + i y^j$). Then $T$ decomposes as $$T = a^j (X_j - \sqrt{2} \, y_j T) + b^j (Y_j + \sqrt{2} \, x_j T) + c \xi ,$$ $$a^j = - \frac{\sqrt{2}\, y^j}{1 + 2 \|z\|^2} \, , \;\; b^j = \frac{\sqrt{2} \, x^j}{1 + 2 \|z\|^2} \, , \;\; c = \frac{1}{1 + 2 \|z\|^2} \, .$$ Then $T^\bot = c \xi$ and (with the conventions in Theorem \[p:5\]) ${\rm Sing}(T^T ) = \{ 0 \}$ hence $(\partial C({\mathbb H}_n^+ ) \setminus \pi^{-1} (0) \, , \, j^* F_{\theta_0})$ is a Lorentz manifold. $\square$ 0.1in [Proof of Theorem]{} \[p:5\]. Let $V \in T(\partial C(M))$ such that $F_\theta (V , W) = 0$ for any $W \in T(\partial C(M))$ i.e. $$(\pi^* \tilde{G}_\theta )(V , W) + (\pi^* \theta )(V) \sigma (W) + (\pi^* \theta )(W) \sigma (V) = 0.$$ By taking into account $$T(C(M)) = {\rm Ker}(\sigma ) \oplus {\rm Ker}(d \pi ) \label{e:old1}$$ we may decompose $V = V_H + V_V$, with $V_H \in {\rm Ker}(\sigma )$. Then $$\tilde{G}((d \pi ) V_H , (d \pi )W_H ) + \theta ((d \pi ) V_H ) \sigma (W_V ) + \theta ((d \pi ) W_H ) \sigma (V_V ) = 0. \label{e:old2}$$ As $\partial C(M)$ is a saturated set, it is tangent to the $S^1$-action. Hence we may apply (\[e:old2\]) for $W = S \in {\rm Ker}(d \pi ) \subset T(\partial C(M))$. As $\sigma (S) = 1/(n+2)$ we obtain $$\theta ((d \pi ) V_H ) = 0,$$ i.e. $(d \pi )V_H \in H(M)$, hence (\[e:old2\]) becomes $$\tilde{G}_\theta ((d \pi ) V_H , (d \pi )W_H ) + \theta ((d \pi ) W_H ) \sigma (V_V ) = 0. \label{e:old3}$$ Applying (\[e:old3\]) for $W = V$ gives $$G_\theta ((d \pi )V_H , (d \pi )V_H ) = 0$$ hence $(d \pi ) V_H = 0$, and then $V_H = 0$ (due to ${\rm Ker}(\sigma ) \cap {\rm Ker}(d \pi ) = (0)$). Therefore, on one hand $$\label{e:incl} {\rm Null}(j^* F_\theta ) \subseteq {\rm Ker}(d \pi )$$ and on the other (\[e:old3\]) becomes $$\theta ((d \pi ) W_H ) \sigma (V_V ) = 0. \label{e:old4}$$ Let $x_0 \in \Omega$ (so that $T^T_{x_0} \neq 0$) and $z_0 \in \pi^{-1}(x_0 )$. We may apply (\[e:old4\]) for $W = (T^T )^\uparrow$, at the point $z_0$. Yet $$(\pi^* \theta )(W_H )_{z_0} = \theta (T^T )_{x_0} = \\| T^T \\|^2_{x_0} \neq 0$$ hence (by (\[e:old4\])) $\sigma (V_V )_{z_0} = 0$, or $(V_V )_{z_0} = 0$ and we may conclude that ${\rm Null}(j^* F_\theta )_{z_0} = (0)$. To complete the proof of Theorem \[p:5\] it suffices to show that ${\rm Null}(j^* F_\theta )_{z}$ is $1$-dimensional, for any $z \in \pi^{-1} (C)$. Let us set $x = \pi (z)$. Then, for any $W \in T (\partial C(M))$ $$F_\theta (S , W)_z = (\pi^* \theta )(W)_z \sigma (S)_z = \frac{1}{n+2} \, \theta_x ((d_z \pi ) W_z ) =$$ $$= g_{\theta , x} (T^\bot_{x} , (d_z \pi )W_z ) = 0$$ as $(d_z \pi ) W_z$ is tangent to $\partial M$. Hence $S_z \in {\rm Null}(j^* F_\theta )_z$ (and we may apply (\[e:incl\])). Since $F_\theta (S,S) = 0$ and $S$ is tangent to $\partial C(M)$, $F_\theta$ is indefinite on $T(\partial C(M))$. However (by the first part of Theorem \[p:5\]) $F_\theta$ is nondegenerate on $T(\pi^{-1}(\Omega ))$ hence $(j^* F_\theta )_z$ has signature $(2n, 1)$ at each $z \in \pi^{-1}(\Omega )$. $\square$ Let $M$ be a strictly pseudoconvex CR manifold-with-boundary and $\theta$ a contact form with $G_\theta$ positive definite. Let $T$ be the characteristic direction of $d \theta$. The property that $T \in T(\partial M)$ is not CR invariant. If $T \in T(\partial M)$ and $\hat{T}$ is the characteristic direction of $d \hat{\theta}$, where $\hat{\theta} = e^{2u} \theta$ [(]{}$u \in C^\infty (M)$[)]{}, then ${\rm Sing}(\hat{T}^T ) = \emptyset$. [Proof]{}. Let us consider a local orthonormal (with respect to $g_\theta$) frame of $T(\partial M)$ of the form $\{ E_1 , \cdots , E_{2n-1} , T \}$, so that $E_a \in H(M)$, $1 \leq a \leq 2n-1$. Next, let us complete $\{ E_a \}$ to a local orthonormal frame $\{ E_1 , \cdots , E_{2n} \}$ of $H(M)$ and set $T_\alpha = (1/\sqrt{2}) (E_\alpha + i E_{\alpha + n})$, $1 \leq \alpha \leq n$. Given another contact form $\hat{\theta} = e^{2u} \theta$ ($u \in C^\infty (M)$) the characteristic direction of $d \hat{\theta}$ is expressed by $$\hat{T} = e^{-2u} (T + i u^{\overline{\alpha}} T_{\overline{\alpha}} - i u^\alpha T_\alpha ) =$$ $$= e^{-2u} \{ T + \frac{i}{\sqrt{2}} (u^{\overline{\alpha}} - u^\alpha ) E_\alpha + \frac{1}{\sqrt{2}} (u^{\overline{\alpha}} + u^\alpha ) E_{\alpha + n } \}$$ dove $u^\alpha = u_\alpha = T_\alpha (u)$ (as $L_\theta (T_\alpha , T_{\overline{\beta}}) = \delta_{\alpha\beta}$). Let $\xi$ be a unit normal on $\partial M$. Then $\xi \in \{ \pm E_{2n} \}$ hence ${\rm Sing}(\hat{T}^T ) = {\rm Sing}(T) = \emptyset$. $\square$ 0.1in If $z \in C(M)$ we denote by $\beta_z : T_{\pi (z)}(M) \to {\rm Ker}(\sigma_z )$ the inverse of the ${\mathbb R}$-linear isomorphism $d_z \pi : {\rm Ker}(\sigma_z ) \to T_{\pi (z)} (M)$. It is an elementary matter that Given $v \in T_x (\partial M)$ its horizontal lift $\beta_z v$, $z \in \pi^{-1} (x)$, is tangent to $\partial C(M)$. Indeed, let $a : (-\epsilon , \epsilon ) \to \partial M$ be a smooth curve such that $a(0) = x$ and $\dot{a}(0) = v$. Let $X \in T(\partial M)$ be a tangent vector field such that $X_x = v$. Let $a^\uparrow : (-\epsilon , \epsilon ) \to C(M)$ be the unique horizontal lift of $a$, issuing at $z$. As $\pi (a^\uparrow (t)) = a(t)$ one has $a^\uparrow (t) \in \partial C(M)$, $\|t\| < \epsilon$. On the other hand $\dot{a}^\uparrow (0) \in {\rm Ker}(\sigma_z )$ and it projects on $v$ hence $$T_z (\partial C(M)) \ni \dot{a}^\uparrow (0) = X^\uparrow_z = \beta_z v.$$ $\square$ 0.1in We set $T(\partial M)^\uparrow = \{ \beta X : X \in T(\partial M) \}$ and ${\mathcal V}_z = {\rm Ker}(d_z \pi )$, for $z \in \partial C(M)$. As observed above, $\partial C(M)$ is tangent to the $S^1$-action hence $\mathcal V$ is a smooth distribution on $\partial C(M)$. Let $M$ be a strictly pseudoconvex CR manifold-with-boundary. One has the decomposition $$T(\partial C(M)) = T(\partial M)^\uparrow \oplus {\mathcal V} \label{e:old6}$$ Moreover, if $\partial M$ is tangent to the characteristic direction $T$ of $d \theta$ then $$T(\partial C(M))^\bot \subseteq {\rm Ker}(\sigma ), \;\;\; (d \pi ) T(\partial C(M))^\bot \subseteq H(M), \label{e:old7}$$ $${\rm Ker}(\sigma ) = T(\partial M)^\uparrow \oplus T(\partial C(M))^\bot . \label{e:old8}$$ Here $T(\partial C(M))^\bot \to \partial C(M)$ is the normal bundle of $j : \partial C(M) \hookrightarrow (C(M), F_\theta )$. \[p:6\] [Proof of Lemma]{} \[p:6\]. Note that $$T(\partial M)^\uparrow \cap {\mathcal V} \subseteq {\rm Ker}(\sigma ) \cap {\rm Ker}(d \pi ) = (0),$$ hence the sum $T(\partial M)^\uparrow + {\mathcal V}$ is direct. The arguments preceding Lemma \[p:6\] show that $T(\partial M)^\uparrow \oplus {\mathcal V} \subseteq T(\partial C(M))$. Viceversa, let $V \in T(\partial C(M)) \subset T(C(M))$. Then (by the decomposition (\[e:old1\])) $$V = X^\uparrow + f \, S, \label{e:old9}$$ for some $X \in T(M)$ and $f \in C^\infty (C(M))$. Then $$X_{\pi (z)} = (d_z \pi ) V_z \in T_{\pi (z)}(\partial M), \;\;\; z \in \partial C(M),$$ i.e. $X \in T(\partial M)$ and then $T(\partial C(M)) \subseteq T(\partial M)^\uparrow \oplus {\mathcal V}$. To check (\[e:old7\]) let $V \in T(\partial C(M))^\bot \subset T(C(M))$ and use (\[e:old1\]) to decompose as in (\[e:old9\]). By assumption $T \in T(\partial M)$ hence $T^\uparrow \in T(\partial C(M))$ and then $$0 = F_\theta (V , T^\uparrow ) = \tilde{G}((d \pi )V , (d \pi )T^\uparrow ) + \theta ((d \pi ) T^\uparrow ) \sigma (V) =$$ $$= \tilde{G}_\theta (X , T) + \frac{f}{n+2} = \frac{f}{n+2}$$ i.e. $f = 0$, or $V = X^\uparrow \in {\rm Ker}(\sigma )$. To check the second statement in (\[e:old7\]) let $$V \in T(\partial C(M))^\bot \subseteq {\rm Ker}(\sigma ) = T(M)^\uparrow = H(M)^\uparrow \oplus ({\mathbb R} T)^\uparrow$$ i.e. $V = Y^\uparrow + f T^\uparrow$, for some $Y \in H(M)$. Moreover $S \in {\rm Ker}(d \pi ) \subset T(\partial C(M))$, hence $S$ and $V$ are orthogonal $$0 = F_\theta (S , V) = \theta ((d \pi ) V) \sigma (S) = \frac{f}{n+2}$$ i.e. $f = 0$, or $V \in H(M)^\uparrow$. (\[e:old7\]) is proved and may be equivalently written $$T(\partial C(M))^\bot \subseteq H(M)^\uparrow .$$ When $T^\bot = 0$ the space $T(\partial C(M))$ is nondegenerate in $(T(C(M)) , F_\theta )$ hence so does the perp space $T(\partial C(M))^\bot$. Also $$T(C(M)) = T(\partial C(M)) \oplus T(\partial C(M))^\bot .$$ Let us prove (\[e:old8\]). First $$T(\partial M)^\uparrow \cap T(\partial C(M))^\bot \subseteq T(\partial C(M)) \cap T(\partial C(M))^\bot = (0)$$ hence the sum $T(\partial M)^\uparrow + T(\partial C(M))^\bot$ is direct and (by (\[e:old7\])) $$T(\partial M)^\uparrow \oplus T(\partial C(M))^\bot \subseteq {\rm Ker}(\sigma ). \label{e:old10}$$ Finally (by (\[e:old6\])) $${\rm Ker}(\sigma ) \oplus {\rm Ker}(d \pi ) = T(C(M)) = T(\partial C(M)) \oplus T(\partial C(M))^\bot =$$ $$= T(\partial M)^\uparrow \oplus {\rm Ker}(d \pi ) \oplus T(\partial C(M))^\bot$$ and (\[e:old10\]) yields (\[e:old8\]). $\square$ 0.1in From now on we assume that $\partial M$ is tangent to $T$. Then let us consider a local orthonormal frame $\{ E_1 , \cdots , E_{2n-1} , T \}$ of $T(\partial M)$, with respect to $i^* g_\theta$ (the first fundamental form of $i : \partial M \hookrightarrow M$), defined on some open set $U \subseteq \partial M$. In particular $E_a \in H(M)$, $1 \leq a \leq 2n-1$. Let $M$ be a strictly pseudoconvex CR manifold-with-boundary. Let $\theta$ be a contact form on $M$ such that $G_\theta$ is positive definite and let $T$ be the characteristic direction of $d \theta$. Assume that $\partial M$ is tangent to $T$. Then $$\{ E_1^\uparrow , \cdots , E_{2n-1}^\uparrow , T^\uparrow \pm \frac{n+2}{2} \, S \}$$ is a local orthonormal frame of $T(\partial C(M))$, with respect to $j^* F_\theta$, defined on the open set $\pi^{-1} (U) \subseteq \partial C(M)$. In particular $T^\uparrow - ((n+2)/2) S$ is a global timelike vector field on $\partial C(M)$, i.e. $(\partial C(M) , j^* F_\theta )$ is a spacetime. \[p:7\] See also [@kn:BeEh]. The proof is straightforward. The geometry of the second fundamental form of the boundaries ============================================================= As $(\partial C(M), j^* F_\theta )$ is a Lorentz submanifold of $(C(M), F_\theta )$ we may write the Gauss equation $$\nabla^{C(M)}_X Y = \nabla^{\partial C(M)}_X Y + {\mathbb B}(X,Y),$$ for any $X, Y \in T(\partial C(M))$. Here $\nabla^{\partial C(M)}$ is the induced connection and ${\mathbb B}$ is the second fundamental form of $j : \partial C(M) \hookrightarrow C(M)$. Cf. e.g. [@kn:O'Ne], p. 100. At this point, we wish to compute the mean curvature vector of $j$ $${\mathbb H} = \frac{1}{2n+1} \; {\rm trace}_{j^* F_\theta} ({\mathbb B}).$$ To this end it is convenient to use the local frame in Proposition \[p:7\]. Let $M$ be a strictly pseudoconvex CR manifold-with-boundary, of CR dimension $n$, and $\theta$ a contact form on $M$ such that $G_\theta$ is positive definite. Assume that $\partial M$ is tangent to the characteristic direction $T$ of $d \theta$. Let $\{ E_1 , \cdots , E_{2n-1} , T \}$ be a local $g_\theta$-orthonormal frame of $T(\partial M)$ and $\xi$ a unit normal vector field on $\partial M$, both defined on the open set $U \subseteq \partial M$. Then the mean curvature vector ${\mathbb H}$ of the immersion $j : \partial C(M) \hookrightarrow C(M)$ is given by $$\label{e:old11} {\mathbb H}_z = \frac{1}{2n+1} \, \sum_{a=1}^{2n-1} g_\theta (\nabla_{E_a} E_a \, , \, \xi )_{\pi (z)} \, \xi_{z}^\uparrow$$ for any $z \in \pi^{-1}(U)$. Here $\nabla$ is the Tanaka-Webster connection of $(M , \theta )$. In particular ${\mathbb H} = (2n/(2n+1)) \, H^\uparrow$, where $H$ is the mean curvature vector of the immersion $i : \partial M \hookrightarrow M$. Therefore, $\partial C(M)$ is minimal in $(C(M), F_\theta )$ if and only if $\partial M$ is minimal in $(M , g_\theta )$. \[t:1\] [Example 5]{}. ${\mathbb R}^{2n}_+ \times {\mathbb R}$ is a strictly pseudoconvex CR manifold (with the CR structure induced from ${\mathbb H}_n$) whose boundary $N = \partial ({\mathbb R}^{2n}_+ \times {\mathbb R})$ is tangent to $T = \partial /\partial t$. The normal bundle of the boundary is the span of $\xi = \partial /\partial y^n - 2 x_n T$. By the Gauss formula, the second fundamental form of the boundary is given by $$B(\frac{\partial}{\partial x^n} , \frac{\partial}{\partial x^n} ) = - 4 y_n \xi , \;\; B(\frac{\partial}{\partial x^\alpha} , \frac{\partial}{\partial x^n} ) = - 2 y_\alpha \xi , \;\; B(\frac{\partial}{\partial x^\alpha} , \frac{\partial}{\partial x^\beta} ) = 0,$$ $$B(\frac{\partial}{\partial x^\alpha} , \frac{\partial}{\partial y^\beta}) = 0, \;\; B(\frac{\partial}{\partial x^n} , \frac{\partial}{\partial y^\beta} ) = 2 x_\beta \xi , \;\; B(\frac{\partial}{\partial y^\alpha} , \frac{\partial}{\partial y^\beta}) = 0,$$ $$B(\frac{\partial}{\partial x^\alpha} , T) = 0, \;\; B(\frac{\partial}{\partial x^n} , T) = \xi , \;\; B(\frac{\partial}{\partial y^\alpha} , T ) = 0, \;\; B(T , T) = 0.$$ Here $1 \leq \alpha , \beta \leq 2n-1$. On the other hand, the induced metric on $N$ is given by $$g : \left( \begin{array}{ccc} 2(\delta_{ij} + 2 y_i y_j ) & - 4 y_i x_\beta & - 2 y_i \\ - 4 x_\alpha y_j & 2(\delta_{\alpha\beta} + 2 x_\alpha x_\beta ) & 2 x_\alpha \\ - 2 y_j & 2 x_\beta & 1 \end{array} \right)$$ hence (by an argument similar to the proof of Lemma \[l:ex\]) the corresponding cometric on $T^* (N)$ is given by $$\label{e:com} g^{-1} : \left( \begin{array}{ccc} \frac{1}{2} \delta^{ij} & 0 & y^i \\ 0 & \frac{1}{2} \delta^{\alpha\beta} & - x^\alpha \\ y^j & - x^\beta & 1 + 2 \|x^\prime \|^2 + 2 \|y\|^2 \end{array} \right)$$ where $x^\prime = (x_1 , \cdots , x_{2n-1})$, $\|x^\prime \|^2 = x_\alpha x^\alpha$ and $\|y\|^2 = y_j y^j$. Finally a calculation (based on (\[e:com\])) shows that $2n H = g^{ab} B(\partial_a , \partial_b ) = 0$, i.e. $N$ is minimal in $({\mathbb R}^{2n}_+ \times {\mathbb R} , g_{\theta_0})$. In particular (by Theorem \[t:1\]) $\partial C({\mathbb R}^{2n}_+ \times {\mathbb R})$ is minimal in $(C({\mathbb R}^{2n}_+ \times {\mathbb R}) , F_{\theta_0})$. $\square$ 0.1in Let $\{ X_A : 1 \leq A \leq 2n+1 \}$ be a local $F_\theta$-orthonormal frame of $T(\partial C(M))$, i.e. $F_\theta (X_A , X_B ) = \epsilon_A \delta_{AB}$, with $\epsilon_1 = \cdots = \epsilon_{2n} = 1 = -\epsilon_{2n+1}$. Then ${\mathbb H}$ is locally given by $${\mathbb H} = \frac{1}{2n+1} \, \sum_A \epsilon_A {\mathbb B}(X_A , X_A ).$$ [Proof of Theorem]{} \[t:1\]. Using the local frame furnished by Lemma \[p:7\] we obtain $$(2n+1) {\mathbb H} = \sum_{a=1}^{2n-1} {\mathbb B}(E_a^\uparrow , E_a^\uparrow ) + 2 (n+2) {\mathbb B}(T^\uparrow , S). \label{e:old12}$$ As a consequence of Lemma \[l:1\] we have $$\nabla^{C(M)}_{E_a^\uparrow} E_a^\uparrow = (\nabla_{E_a} E_a )^\uparrow - \frac{n+2}{2} \; A(E_a , E_a ) S, \label{e:old13}$$ $$\nabla^{C(M)}_{T^\uparrow} S = 0. \label{e:old14}$$ The equation (\[e:old14\]) implies ${\mathbb B}(T^\uparrow , S) = 0$ (with the corresponding simplification of (\[e:old12\])). As $T \in T(\partial M)$ we have $$T(\partial M)^\bot \subseteq H(M).$$ We need the following $\,$ Assume that $\partial M$ is tangent to $T$. Let $T(\partial M)^\bot \to \partial M$ be the normal bundle of the immersion $i : \partial M \hookrightarrow M$. Then $$[T(\partial M)^\bot ]^\uparrow = T(\partial C(M))^\bot . \label{e:old15}$$ \[l:5\] [Proof of Lemma]{} \[l:5\]. Let $\xi \in T(\partial M)^\bot$ and $V \in T(\partial C(M)) = T(\partial M)^\uparrow \oplus {\rm Ker}(d \pi )$, i.e. $V = X^\uparrow + f \, S$. Let us set $X_H := X - \theta (X) T \in H(M)$. Then $$F_\theta (V , \xi^\uparrow ) = \tilde{G}_\theta (X , \xi ) + f \; F_\theta (S, \xi^\uparrow ) =$$ $$= G_\theta (X_H , \xi ) + f\; \theta (\xi ) \sigma (S) = g_\theta (X_H , \xi ) = 0$$ because $X,T \in T(\partial M)$ implies $X_H \in T(\partial M)$. It follows that $$[T(\partial M)^\bot ]^\uparrow \subseteq T(\partial C(M))^\bot .$$ The desired equality follows by inspecting dimensions. $\square$ Let $\xi$ be a unit normal vector field on $\partial M$, defined on the open set $U \subseteq N$. Then (by Lemma \[l:5\]) $\xi^\uparrow$ is a unit normal vector field on $\partial C(M)$. Then (by the Gauss equation and by (\[e:old13\])) $$F_\theta ({\mathbb B}(E_a^\uparrow , E_a^\uparrow )\, , \, \xi^\uparrow ) = F_\theta (\nabla^{C(M)}_{E_a^\uparrow} E_a^\uparrow \, , \, \xi^\uparrow ) = F_\theta ((\nabla_{E_a} E_a )^\uparrow \, , \, \xi^\uparrow ) =$$ $$= \tilde{G}_\theta (\nabla_{E_a} E_a \, , \, \xi ) = g_\theta (\nabla_{E_a} E_a \, , \, \xi )$$ which yields (\[e:old11\]). $\square$ 0.1in The Levi-Civita connection $\nabla^{g_\theta}$ of $(M , g_\theta )$ is related to the Tanaka-Webster connection $\nabla$ of $(M , \theta )$ by $$\nabla^{g_\theta}_X Y = \nabla_X Y + (\Omega (X,Y) - A(X,Y)) T + \label{e:old16}$$ $$+ \tau (X) \theta (Y) + \theta (X) J Y + \theta (Y) J X ,$$ for any $X,Y \in T(M)$. Here $\Omega = - d \theta$. Cf. e.g. [@kn:BaDr], p. 238. Thus, for any $X, Y \in H(M)$ $$\nabla^{g_\theta}_X Y = \nabla_X Y + (\Omega (X,Y) - A(X,Y)) T$$ and then $$\nabla^{g_\theta}_{E_a} E_a = \nabla_{E_a} E_a - A(E_a , E_a ) T$$ implies (as $g_\theta (T , \xi ) = 0$) $$(2n+1) {\mathbb H} = \sum_{a} g_\theta (\nabla^{g_\theta}_{E_a} E_a \, , \, \xi ) \xi^\uparrow =$$ $$= \sum_{a} g_\theta (B(E_a , E_a ) \, , \, \xi ) \xi^\uparrow = 2n g_\theta (H \, , \, \xi ) \xi^\uparrow$$ because $\nabla^{g_\theta}_T T = 0$ implies $B(T,T) = 0$. Here $B$ is the second fundamental form of $i : \partial M \hookrightarrow M$ and $H = (1/(2n)) \, {\rm trace}_{g_\theta} (B)$ is its mean curvature vector. Then ${\mathbb H} = (2n/(2n+1)) \, H^\uparrow$. $\square$ Let $M$ be a strictly pseudoconvex CR manifold-with-boundary and $\theta$ such that $T \in T(\partial M)$. Then $\partial C(M)$ has nonumbilic points in $(C(M), F_\theta )$. Moreover $\partial M$ is totally umbilical in $(M, g_\theta )$ if and only if $${\mathbb B}(X^\uparrow , Y^\uparrow ) = \frac{2n+1}{2n} \, F_\theta (X^\uparrow , Y^\uparrow ) {\mathbb H},$$ $${\mathbb B}(X^\uparrow , T^\uparrow ) = \{ (d \sigma )(X^\uparrow , \xi^\uparrow ) + g_\theta (X , J \xi ) \} \xi^\uparrow ,$$ for any $X, Y \in T(\partial M) \cap H(M)$. \[t:3\] [Proof]{}. By (\[e:old16\]) and the Gauss formula for the immersion $\partial M \hookrightarrow (M, g_\theta )$ $$B(X,Y) = g_\theta (\nabla_X Y , \xi ) \xi , \;\;\; B(X,T) = g_\theta (\tau X + J X , \xi ) \xi ,$$ for any $X,Y \in T(\partial M) \cap H(M)$. Next, by Lemma \[l:1\] and the Gauss formula for the immersion $\partial C(M) \hookrightarrow (C(M), F_\theta )$ $$\label{e:sff1} {\mathbb B}(X^\uparrow , Y^\uparrow ) = B(X,Y)^\uparrow ,$$ $$\label{e:sff2} {\mathbb B}(X^\uparrow , T^\uparrow ) = B(X,T)^\uparrow + \{ (d \sigma )(X^\uparrow , \xi^\uparrow ) + g_\theta (X , J \xi ) \} \xi^\uparrow ,$$ $$\label{e:sff3} {\mathbb B}(X^\uparrow , \hat{S}) = - g_\theta (X , J \xi ) \xi^\uparrow , \;\;\; {\mathbb B}(T^\uparrow , \hat{S}) = 0.$$ Note that $J \xi$ is tangent to $\partial M$. Assume that ${\mathbb B} = F_\theta \otimes {\mathbb H}$. Then (by (\[e:sff3\])) $J \xi$ is orthogonal to $\partial M$, hence $\xi = 0$, a contradiction. The last statement in Theorem \[t:3\] follows from $B = g_\theta \otimes H$ and (\[e:sff1\])-(\[e:sff2\]). $\square$ Minimal submanifolds ==================== The purpose of this section to investigate minimal submanifolds in the Heisenberg group ${\mathbb H}_n$. First, we establish the relationship between the notion of $X$-minimality of N. Arcozzi & F. Ferrari, cf. (3) in [@kn:ArFe], I. Birindelli & E. Lanconelli, cf. (3.23) in [@kn:BiLa], and N. Garofalo & S.D. Pauls, cf. (2.5) in [@kn:GaPa] (see also [@kn:Pau]) and minimality of an isometric immersion (between Riemannian manifolds). Second, we prove the following Let $\Psi : N \to {\mathbb H}_n$ be an isometric immersion of a $m$-dimensional Riemannian manifold $(N , g)$ into $({\mathbb H}_n , g_{\theta_0})$. Then $\Psi$ is minimal if and only if $$\Delta \Psi = 2 J T^\bot \label{e:delta}$$ where $\Delta$ is the Laplace-Beltrami operator of $(N , g)$. In particular, there are no minimal isometric immersions $\Psi$ of a compact Riemannian manifold $N$ into the Heisenberg group such that $T$ is tangent to $\Psi (N)$. \[t:min\] Compare to Theorem 6.2 and Corollaries 6.1 and 6.2 in [@kn:DW], p. 45-48. Let $M = {\mathbb H}_1$ be the lowest dimensional Heisenberg group and $\varphi : {\mathbb H}_1 \to {\mathbb R}$ a $C^2$ function. Let us set $$N = \{ x \in {\mathbb H}_1 : \varphi (x) = 0 \}$$ and assume there is an open neighborhood $O \supset N$ such that $$\|\nabla \varphi (x)\| \geq \alpha > 0, \;\;\; x \in O. \label{e:old17}$$ Here $\nabla \varphi$ is the Euclidean gradient of $\varphi$. Let $(z, t)$ be the natural coordinates on ${\mathbb H}_1 = {\mathbb C} \times {\mathbb R}$ and set $Z = Z_1 = \partial /\partial z + i \, \overline{z} \, \partial /\partial t$ (the generator of $T_{1,0}({\mathbb H}_1 )$). Let $\theta_0 = d t + i (z \, d \overline{z} - \overline{z} \, d z)$ be the canonical contact form on ${\mathbb H}_1$. Note that $L_{\theta_0} (Z , \overline{Z}) = 1$. The Tanaka-Webster connection of $({\mathbb H}_1 , \theta_0 )$ is given by $$\Gamma^A_{BC} = 0, \;\;\; A,B,C \in \{ 1 , \overline{1} , 0 \} .$$ Let us set $X_1 = \frac{1}{\sqrt{2}} \, (Z + \overline{Z})$ and $X_2 = Y_1 = \frac{i}{\sqrt{2}} \, (Z - \overline{Z})$. We shall prove the following Let $N = \{ x \in {\mathbb H}_1 : \varphi (x) = 0 \}$ be a surface in ${\mathbb H}_1$ such that $(\ref{e:old17})$ holds. Assume that $N$ is tangent to the characteristic direction $T = \partial /\partial t$ of $({\mathbb H}_1 , \theta_0 )$. Let $\xi$ be a unit normal vector field on $N$. Then the mean curvature vector of $N$ in $({\mathbb H}_1 , g_{\theta_0} )$ is given by $$H = - \frac{1}{2} \, \sum_{j=1}^2 X_j \left( \frac{X_j \varphi}{\|X \varphi \|} \right) \; \xi . \label{e:old18}$$ Here $\| X \varphi \|^2 = (X_1 \varphi )^2 + (X_2 \varphi )^2$ is the $X$-gradient of $\varphi$. \[c:1\] [Proof of Theorem]{} \[c:1\]. $T(N)$ is the span of $\{ E \, , \, T \}$ while $T(N)^\bot$ is the span of $\xi$, where $$E = \frac{1}{\|X \varphi \|} \{ (X_2 \varphi ) X_1 - (X_1 \varphi ) X_2 \} , \;\; \xi = \frac{1}{\|X \varphi \|} \{ (X_1 \varphi ) X_1 + (X_2 \varphi ) X_2 \} ,$$ so that $g_{\theta_0} (E, E) = 1$ and $g_{\theta_0} (\xi , \xi ) = 1$. A calculation (based on $\nabla_{X_j} X_k = 0$) leads to $$\nabla_{X_1} E = X_1 \left( \frac{X_2 \varphi}{\|X \varphi \|} \right) X_1 - X_1 \left( \frac{X_1 \varphi}{\|X \varphi \|} \right) X_2 \, ,$$ $$\nabla_{X_2} E = X_2 \left( \frac{X_2 \varphi}{\|X \varphi \|} \right) X_1 - X_2 \left( \frac{X_1 \varphi}{\|X \varphi \|} \right) X_2 \, ,$$ hence $$\nabla_E E = \frac{1}{\|X \varphi \|} \left\{ \left[ (X_2 \varphi ) X_1 \left( \frac{X_2 \varphi}{\|X \varphi \|} \right) - (X_1 \varphi ) X_2 \left( \frac{X_2 \varphi}{\| X \varphi \|} \right) \right] X_1 + \right. \label{e:old19}$$ $$+ \left. \left[ (X_1 \varphi ) X_2 \left( \frac{X_1 \varphi}{\|X \varphi \|} \right) - (X_2 \varphi ) X_1 \left( \frac{X_1 \varphi}{\|X \varphi \|} \right) \right] X_2 \right\} .$$ Then (by (\[e:old19\])) $$g_{\theta_0} (\nabla_E E , \xi ) = - \sum_{j=1}^2 X_j \left( \frac{X_j \varphi}{\|X \varphi \|} \right) + \label{e:old20}$$ $$+ \frac{1}{\|X \varphi \|^2} \left\{ (X_1 \varphi )^2 \; X_2 \left( \frac{X_2 \varphi}{\|X \varphi \|} \right) + (X_1 \varphi )^2 \; X_1 \left( \frac{X_1 \varphi}{\| X \varphi \|} \right) + \right.$$ $$+ \left. (X_1 \varphi )(X_2 \varphi ) \; X_1 \left( \frac{X_2 \varphi}{\|X \varphi \|} \right) + (X_1 \varphi )(X_2 \varphi ) \; X_2 \left( \frac{X_1 \varphi}{\|X \varphi \|} \right) \right\} .$$ Using the identity $$\|X \varphi \| X_j (\|X \varphi \|) = (X_1 \varphi ) \; X_j X_1 \varphi + (X_2 \varphi ) \; X_j X_2 \varphi$$ one may show that the second term in the right hand member of (\[e:old20\]) is $\|X \varphi \|^{-4}$ times $$(X_2 \varphi )^2 \{ (X_2 X_2 \varphi ) \|X \varphi \| - (X_2 \varphi ) X_2 (\|X \varphi \|) \} +$$ $$+ (X_1 \varphi )^2 \{ (X_1 X_1 \varphi )\|X \varphi \| - (X_1 \varphi ) X_1 (\|X \varphi \|) \} +$$ $$+ (X_1 \varphi )(X_2 \varphi ) \{ (X_1 X_2 \varphi ) \|X \varphi \| - (X_2 \varphi ) X_1 (\|X \varphi \|) \} +$$ $$+ (X_1 \varphi )(X_2 \varphi ) \{ (X_2 X_1 \varphi ) \|X \varphi \| - (X_1 \varphi ) X_2 (\|X \varphi \|) \} =$$ $$= - \{ (X_1 \varphi ) X_1 (\|X \varphi \|) + (X_2 \varphi ) X_2 (\|X \varphi \|) \} \{ (X_1 \varphi )^2 + (X_2 \varphi )^2 \} +$$ $$+ \|X \varphi \| \{ (X_1 \varphi )^2 X_1 X_1 \varphi + 2 (X_1 \varphi )(X_2 \varphi ) X_1 X_2 \varphi +(X_2 \varphi )^2 X_2 X_2 \varphi \}$$ (as $[X_1 , X_2 ] = - 2 \; T$ and $T(\varphi ) = 0$) or $$- (X_2 \varphi ) \|X \varphi \| \{ (X_1 \varphi ) X_2 X_1 \varphi + (X_2 \varphi ) X_2 X_2 \varphi \} -$$ $$- (X_1 \varphi ) \|X \varphi \| \{ (X_1 \varphi ) X_1 X_1 \varphi + (X_2 \varphi ) X_1 X_2 \varphi \} +$$ $$+ \|X \varphi \| \{ (X_1 \varphi )^2 X_1 X_1 \varphi + 2 (X_1 \varphi ) (X_2 \varphi ) X_1 X_2 \varphi + (X_2 \varphi )^2 X_2 X_2 \varphi \} = 0$$ hence (\[e:old20\]) leads to (\[e:old18\]). $\square$ 0.1in Let us prove Theorem \[t:min\]. Let $(x^1 , \cdots , x^{2n} , x^0 )$ be the Cartesian coordinates on ${\mathbb R}^{2n+1}$ and $(U, u^1 , \cdots , u^m )$ a local coordinate system on $N$. Let $H(\Psi )$ be the mean curvature vector of $\Psi : N \to {\mathbb H}_n$. Then $H(\Psi ) = H^A \partial_A$, where $\partial_A$ is short for $\partial /\partial x^A$. Let $g_0 = g_{\theta_0}$ be the Webster metric of $({\mathbb H}_n , \theta_0 )$ and $D^0$ the Levi-Civita connection of $({\mathbb H}_n , g_0 )$. We set $B_\alpha^A = \partial \Psi^A /\partial u^\alpha$, so that $\Psi_* (\partial /\partial u^\alpha ) = B^A_\alpha \partial_A$. Let $\{ E_1 , \cdots , E_m \}$ be a local orthonormal (with respect to $g$) frame of $T(N)$, defined on $U$. Then $E_\alpha = E_\alpha^\beta \partial /\partial u^\beta$. Taking into account that $E_\alpha^\beta B^A_\beta = E_\alpha (\Psi^A )$, the Gauss formula of $\Psi$ $$D^0_{E_\alpha} E_\beta = \Psi_* D_{E_\alpha} E_\beta + B(E_\alpha , E_\beta )$$ may be written $$\{ E_\alpha (E_\beta \Psi^A ) - (D_{E_\alpha} E_\beta )(\Psi^A ) \} \partial_A =$$ $$= B(E_\alpha , E_\beta ) - E_\alpha (\Psi^A ) E_\beta (\Psi^B ) D^0_{\partial_A} \partial_B \, .$$ Here $D$ is the Levi-Civita connection of $(N , g)$ and $B$ is the second fundamental form of $\Psi$. Contraction of $\alpha$ and $\beta$ gives $$(\Delta \Psi^A )\partial_A = m H(\Psi ) - \sum_{\alpha = 1}^m E_\alpha (\Psi^A ) E_\alpha (\Psi^B ) D^0_{\partial_A} \partial_B \label{e:min33}$$ Since $$\partial_j = \frac{\partial}{\partial x^j} = Z_j + \overline{Z}_j - 2 y^j T, \;\; \partial_{j+n} = \frac{\partial}{\partial y^j} = i(Z_j - \overline{Z}_j ) + 2 x^j T, \label{e:partials}$$ it follows that the Tanaka-Webster connection of $({\mathbb H}_n , \theta_0 )$ satisfies $$\nabla_{\partial_j} \partial_k = \nabla_{\partial_{j+n}} \partial_{k+n} = 0,$$ $$\nabla_{\partial_j} \partial_{k+n} = - \nabla_{\partial_{j+n}} \partial_k = 2 \delta_{jk} T, \label{e:min34}$$ $$\nabla_{\partial_A} T = \nabla_T \partial_B = 0.$$ Let $J$ be the complex structure in $H({\mathbb H}_n )$, extended to a $(1,1)$-tensor field on ${\mathbb H}_n$ by requesting that $J T = 0$. Using $D^0 = \nabla - (d \theta_0 ) \otimes T + 2 (\theta_0 \odot J)$ it follows that $$E_\alpha (\Psi^A ) E_\beta (\Psi^B ) D^0_{\partial_A} \partial_B = E_\alpha (\Psi^A ) E_\beta (\Psi^B ) \nabla_{\partial_A} \partial_B -$$ $$- (d \theta )(E_\alpha , E_\beta )T + \theta (E_\alpha ) J \Psi_* E_\beta + \theta (E_\beta ) J \Psi_* E_\alpha$$ where $\theta = \Psi^* \theta_0$. On the other hand, by (\[e:min34\]) $$E_\alpha (\Psi^A ) E_\beta (\Psi^B ) \nabla_{\partial_A} \partial_B =$$ $$= 2 \sum_{j=1}^n \{ E_\alpha (\Psi^j ) E_\beta (\Psi^{j+n}) - E_\alpha (\Psi^{j+n}) E_\beta (\Psi^{j}) \} T .$$ Also $\sum_\alpha \theta (E_\alpha ) = \sum_\alpha g_0 (T , \Psi_* E_\alpha ) = \sum_\alpha g(T^T , E_\alpha )$ hence $$\sum_\alpha \theta (E_\alpha ) J \Psi_* E_\alpha = J \Psi_* T^T = - J T^\bot ,$$ so that (\[e:min33\]) becomes $m H(\Psi ) = \Delta \Psi - 2 J T^\bot$ (yielding (\[e:delta\])). $\square$ 0.1in Our Theorem \[c:1\] demonstrates that the Webster metric is the “correct” choice of ambient metric. Nevertheless, even the geometry of a hyperplane in $({\mathbb H}_n \, , \, g_0 )$ turns out to be rather involved. In the sequel, we work out explicitly the case of $\{ z \in {\mathbb H}_n : t = 0 \}$. 0.1in [Example 1.]{} ([continued]{}) Let $\Psi : \partial {\mathbb H}_n^+ \to {\mathbb H}_n^+$ be the inclusion and $g = \Psi^* g_0$ (the first fundamental form of $\Psi$). Let $\Delta$ be the Laplace-Beltrami operator of $(\partial {\mathbb H}^n_+ , g)$. We may state The coordinate functions $z^j$ on $\partial {\mathbb H}_n^+ \approx {\mathbb C}^n$ satisfy $\Delta z^j = 2\, z^j /(1 + 2 \|z\|^2 )$. Consequently the boundary of $({\mathbb H}^+_n , g_0 )$ is minimal. \[p:ex1\] Note that $$\theta_0 (\partial_i ) = - 2 y_i \, , \;\;\; \theta_0 (\partial_{i+n}) = 2 x_i \, ,$$ $$(d \theta_0 )(\partial_i , \partial_j ) = (d \theta_0 )(\partial_{i+n} , \partial_{j+n}) = 0, \;\; (d \theta_0 )(\partial_i , \partial_{j+n}) = 2 \delta_{ij} \, ,$$ $$J \partial_j = \partial_{j+n} - 2 x_j T , \;\;\; J \partial_{j+n} = - \partial_j - 2 y_j T.$$ Then by (\[e:old16\]) (with $\tau = 0$) and by (\[e:min34\]) it follows that $$D^0_{\partial_i} \partial_j = - 2 (y_i \delta^k_j + y_j \delta_i^k ) \frac{\partial}{\partial y^k} + 4 (y_i x_j + y_j x_i ) T,$$ $$D^0_{\partial_i} \partial_{j+n} = 2 (y_i \frac{\partial}{\partial x^j} + x_j \frac{\partial}{\partial y^i} ) + 4(y_i y_j - x_i x_j ) T,$$ $$D^0_{\partial_{i+n}} \partial_{j+n} = -2(x_i \delta_j^k + x_j \delta_i^k ) \frac{\partial}{\partial x^k} - 4 (x_i y_j + x_j y_i ) T.$$ Next, we shall need the Gauss formula $$D^0_{\partial_a} \partial_b = D_{\partial_a} \partial_b + B(\partial_a , \partial_b ) ,$$ where $D$ is the Levi-Civita connection of $(\partial {\mathbb H}_n^+ , g)$. We obtain $$B(\partial_i , \partial_j ) = 4c(y_i x_j + y_j x_i ) \xi ,$$ $$B(\partial_i , \partial_{j+n}) = 4c(y_i y_j - x_i x_j ) \xi , \label{e:B}$$ $$B(\partial_{i+n} , \partial_{j+n}) = - 4 c (x_i y_j + x_j y_i ) \xi ,$$ hence $\Psi$ is not totally geodesic, and if $D_{\partial_a} \partial_b = \Gamma^c_{ab} \partial_c$ then $$\Gamma^k_{ij} = - 4 c (y_i x_j + y_j x_i ) y^k \, , \;\; \Gamma^{k + n}_{i + n \; j + n} = - 4 c (x_i y_j + x_j y_i ) x^k \, ,$$ $$\Gamma^{k+n}_{ij} = 4 c (y_i x_j + y_j x_i ) x^k - 2 (y_i \delta_j^k + y_j \delta_i^k ) ,$$ $$\Gamma^k_{i \; j+n} = 2 y_i \delta_j^k - 4 c (y_i y_j - x_i x_j ) y^k \, , \label{e:G}$$ $$\Gamma^{k+n}_{i \; j+n} = 2 x_j \delta_i^k + 4 c (y_i y_j - x_i x_j ) x^k \, ,$$ $$\Gamma^k_{i+n \; j+n} = - 2(x_i \delta_j^k + x_j \delta_i^k ) + 4 c (x_i y_j + x_j y_i ) y^k \, .$$ We need the following The local coefficients of the cometric $g^{-1}$ on $T^* (\partial {\mathbb H}_n^+ )$ are given by $$\label{e:gmeno} g^{-1} : \left( \begin{array}{cc} \frac{1}{2} \delta^{ij} - c y^i y^j & c y^i x^j \\ c x^i y^j & \frac{1}{2} \delta^{ij} - c x^i x^j \end{array} \right) .$$ Consequently $$\Delta u = \frac{1}{2} \, \Delta_0 u + 2 c \frac{\partial u}{\partial r} - c \{ y^i y^j \frac{\partial^2 u}{\partial x^i \partial x^j} - 2 y^i x^j \frac{\partial^2 u}{\partial x^i \partial y^j} + x^i x^j \frac{\partial^2 u}{\partial y^i \partial y^j} \}$$ for any $u \in C^2 (\partial {\mathbb H}_n^+ )$, where $\Delta_0$ is the ordinary Laplacian on ${\mathbb R}^{2n}$ and $\partial /\partial r$ is the radial vector field $x^j (\partial /\partial x^j ) + y^j (\partial /\partial y^j )$. \[l:ex\] By Lemma \[l:ex\] it follows that $\Delta x^j = 2 c x^j$ and $\Delta y^j = 2 c y^j$, hence the first statement in Proposition \[p:ex1\]. On the other hand $T^\bot = c \xi$ implies $J T^\bot = c \; \partial /\partial r$ hence (by Theorem \[t:min\]) $H(\Psi ) = 0$. Note that the mean curvature vector may be also computed from $2n H(\Psi ) = g^{ab} B(\partial_a , \partial_b )$ by (\[e:B\]) and (\[e:gmeno\]). It remains that we prove Lemma \[l:ex\]. The first statement is elementary yet rather involved. The identities $g_{ac} g^{cb} = \delta_a^b$ may be written $$\begin{cases} 2 (\delta_{ij} + 2 y_i y_j ) g^{jk} - 4 y_i x_j g^{j+n,k} = \delta^k_i \, , \cr 2(\delta_{ij} + 2 y_i y_j ) g^{j,k+n} - 4 y_i x_j g^{j+n,k+n} = 0, \cr - 4 x_i y_j g^{jk} + 2(\delta_{ij} + 2 x_i x_j ) g^{j+n,k} = 0, \cr - 4 x_i y_j g^{j,k+n} + 2 (\delta_{ij} + 2 x_i x_j ) g^{j+n,k+n}= \delta^k_i \, . \cr \end{cases} \label{e:expl}$$ Contraction of the first two equations (respectively of the last two equations) by $y^i$ (respectively by $x^i$) gives $$(1 + 2 \|y\|^2 ) y_j g^{jk} - 2 \|y\|^2 x_j g^{j+n,k} = y^k \, ,$$ $$(1+ 2\|y\|^2 ) y_j g^{j,k+n} - 2 \|y\|^2 x_j g^{j+n,k+n} = 0,$$ $$2 \|x\|^2 y_j g^{jk} - (1 + 2 \|x\|^2 ) x_j g^{j+n,k} = 0,$$ $$- 2 \|x\|^2 y_j g^{j,k+n} + (1 + 2 \|x\|^2 ) x_j g^{j+n,k+n} = x^k \, ,$$ where from $$y_j g^{jk} = \frac{c}{2} (1 + 2 \|x\|^2 ) y^k \, , \;\; x_j g^{j+n,k} = c \|x\|^2 y^k \, ,$$ $$y_j g^{j,k+n} = c \|y\|^2 x^k \, , \;\; x_j g^{j+n,k+n} = \frac{c}{2} (1 + 2 \|y\|^2 ) x^k \, ,$$ and substitution back into (\[e:expl\]) yields (\[e:gmeno\]). To compute the Laplacian $$\Delta u = \frac{\partial}{\partial x^a} \left( g^{ab} \frac{\partial u}{\partial x^b}\right) + g^{ab} \frac{\partial}{\partial x^a} \left( \log \sqrt{G} \right) \frac{\partial u}{\partial x^b}$$ (with $G = \det [g_{ab}]$) we recall that $\partial (\log \sqrt{G} )/\partial x^a = \Gamma^b_{ba}$ hence (by (\[e:G\])) $$\frac{\partial}{\partial x^a} \left( \log \sqrt{G} \right) = 2 c x_a \, , \;\;\; 1 \leq a \leq 2n .$$ Then (\[e:gmeno\]) yields the result. $\square$ The CR Yamabe problem ===================== Let $M$ be a compact strictly pseudoconvex CR manifold-with-boundary, of CR dimension $n$, and $\theta$ a contact form on $M$ with $G_\theta$ positive definite. Let us assume that $\partial M$ is tangent to the characteristic direction $T$ of $d \theta$. Let us set $p = 2 + 2/n$ and $f = (p-2) \log u$, with $u \in C^\infty (M)$, $u > 0$. If $\hat{\theta} = e^f \theta$ then $\partial C(M)$ is minimal in $(C(M), F_{\hat{\theta}})$ if and only if $$\label{e:new29} \frac{\partial (u \circ \pi )}{\partial \eta} - n\; F_\theta ({\mathbb H}, \eta) \; u \circ \pi = 0 \;\;\; {\rm on} \;\; \partial C(M),$$ where $\eta$ and ${\mathbb H}$ are respectively an outward unit normal and the mean curvature vector of the immersion $\partial C(M) \hookrightarrow (C(M), F_\theta )$. In particular, if $\xi$ and $H$ are an outward unit normal and the mean curvature vector of the immersion $\partial M \hookrightarrow (M , g_\theta )$ then [(\[e:new29\])]{} projects to $$\label{e:new30} \frac{\partial u}{\partial \xi} - \frac{2n^2}{2n+1} \; g_\theta (H , \xi ) \; u = 0 \;\;\; {\rm on} \;\; \partial M.$$ \[l:3\] The first statement in Lemma \[l:3\] is of course well known in conformal geometry. We give a brief proof for the convenience of the reader. If $\hat{\theta} = e^f \theta$ the corresponding Fefferman metric is $F_{\hat{\theta}} = e^{f \circ \pi} F_\theta$ hence the Levi-Civita connections $\hat{D}$ and $D$ (of $F_{\hat{\theta}}$ and $F_\theta$, respectively) are related by $$\hat{D}_V W = D_V W + \frac{1}{2} \{ V(f) W + W(f) V - F_\theta (V,W) D (f\circ \pi ) \} , \label{e:29}$$ for any $V, W \in T(C(M))$, where $D (f\circ \pi )$ is the gradient of $f\circ \pi$ with respect to $F_\theta$. Our assumption $T \in T(\partial M)$ and Proposition \[p:5\] imply that $T(\partial C(M))$ is nondegenerate in $T(C(M))$ with respect to $F_\theta$, hence with respect to $F_{\hat{\theta}}$ as well. Let $\mathbb B$ and $\hat{\mathbb B}$ be the second fundamental forms of the immersions $\partial C(M) \hookrightarrow (C(M), F_\theta )$ and $\partial C(M) \hookrightarrow (C(M), F_{\hat{\theta}})$. Then (by (\[e:29\]) and the Gauss formula) $$\hat{\mathbb B} = {\mathbb B} - \frac{1}{2} \; F_\theta \otimes (D(f \circ \pi ))^\bot \, . \label{e:30}$$ Taking traces in (\[e:30\]) shows that the mean curvature vectors of the two immersions are related by $\hat{\mathbb H} = e^{-f} \{ {\mathbb H} - \frac{1}{2} (D(f \circ \pi ))^\bot \}$ hence $\partial C(M)$ is minimal in $(C(M) , F_{\hat{\theta}})$ if and only if ${\mathbb H} = (1/2) (D(f \circ \pi ))^\bot$ and (\[e:new29\]) is proved. Let $\xi$ be an outward unit normal on $\partial M$ in $(M , g_\theta )$. Then $\eta = \xi^\uparrow$ is an outward unit normal on $\partial C(M)$ in $(C(M), F_{\theta})$. Then (by Theorem \[t:1\]) the mean curvatures of $\partial M \hookrightarrow (M , g_\theta )$ and $\partial C(M) \hookrightarrow (C(M), F_\theta )$ are related by $$F_\theta ({\mathbb H}, \eta ) = \frac{2n}{2n+1} \; g_\theta (H , \xi ) \circ \pi$$ hence (\[e:new29\]) projects on $M$ to give (\[e:new30\]). $\square$ 0.1in We may consider the problem $$- b_n \, \Delta_b u + \rho \; u = \lambda \; u^{p-1} \;\;\; {\rm in} \;\; M, \label{e:33}$$ $$\frac{\partial u}{\partial \xi} - \frac{2n^2}{2n+1} \; \mu_\theta \; u = 0 \;\;\; {\rm on} \;\; \partial M, \label{e:34}$$ (the [CR Yamabe problem]{} on a CR manifold-with-boundary) where $$\Delta_b u = {\rm div}(\nabla^H u), \;\;\; u \in C^2 (M),$$ is the [sublaplacian]{} of $(M , \theta )$, $b_n = 2 + 2/n$, $\lambda$ is a constant, and $\mu_\theta = g_\theta (H , \xi ) \in \{ \pm \\| H \\| \}$. Also $\nabla u$ is the gradient of $u$ with respect to $g_\theta$ and $\nabla^H u = \pi_H \nabla u$ (the [horizontal gradient]{}) where $\pi_H : T(M) \to H(M)$ is the projection associated with the direct sum decomposition $T(M) = H(M) \oplus {\mathbb R} T$. The divergence operator is meant with respect to the volume form $\omega = \theta \wedge (d \theta )^n$. The problem (\[e:33\])-(\[e:34\]) is a nonlinear subelliptic problem of variational origin. Indeed, we may state Let us set $$A_\theta (u) = \int_M \{ b_n \\| \nabla^H u \\|^2 + \rho \; u^2 \} \omega - a_n \int_{\partial M} \mu_\theta \; u^2 \; d \sigma \, ,$$ $$B_\theta (u) = \int_M \|u\|^p \omega \, ,$$ where $\sigma = {\rm vol}(i^* g_\theta )$, the canonical volume form associated with the induced metric $i^* g_\theta$ on $\partial M$, and $a_n = 2^{n+2} \, (n + 1)! \, n/(2n+1)$. Moreover, let $$Q_\theta (u) = \frac{A_\theta (u)}{B_\theta (u)}\, , \;\;\; Q(M) = \inf \{ Q_\theta (u) : u \in C^\infty (M), \;\; u > 0 \} .$$ If $u \in C^\infty (M)$ is a positive function such that $Q_\theta (u) = Q(M)$ then $u$ is a solution to [(\[e:33\])-(\[e:34\])]{} with $\lambda = (p/2) Q(M)$, a CR invariant of $M$. \[p:11\] [Proof]{}. If $\{ T_\alpha \}$ is a local frame of $T_{1,0}(M)$ then the horizontal gradient is expressed by $\nabla^H u = u^\alpha T_\alpha + u^{\overline{\alpha}} T_{\overline{\alpha}}$, where $u^\alpha = g^{\alpha\overline{\beta}} u_{\overline{\beta}}$ and $u_{\overline{\beta}} = T_{\overline{\beta}}(u)$, hence $\\| \nabla^H u \\|^2 = 2 u_\alpha u^\alpha$. Then $$\frac{d}{d t} \{ A_\theta (u + t h) \}_{t=0} = 2 \int_M \{ b_n (u^\alpha h_\alpha + u_\alpha h^\alpha ) + \rho u h \} \omega - 2 a_n \int_{\partial M} \mu_\theta \, u \, h \; d \sigma ,$$ for any $h \in C^2 ({\rm Int}(M)) \cap C^1 (M)$ (where ${\rm Int}(M) = M \setminus \partial M$). On the other hand $$\int_M u^\alpha h_\alpha \, \omega = \int_M \{ T_\alpha (u^\alpha h) - h T_\alpha (u^\alpha ) \} \omega =$$ $$= \int_M {\rm div}(h u^\alpha T_\alpha ) \omega - \int_M \{ T_\alpha (u^\alpha ) + u^\alpha {\rm div}(T_\alpha ) \} h \, \omega .$$ Note that ${\rm div}(T_\alpha ) = \Gamma^\beta_{\beta\alpha}$ hence $T_\alpha (u^\alpha ) + u^\alpha {\rm div}(T_\alpha ) = {u^\alpha}_\alpha$, where ${u^\alpha}_\beta = g^{\alpha\overline{\gamma}} u_{\overline{\gamma}\beta}$ and $u_{\alpha\overline{\beta}} = (\nabla^2 u)(T_\alpha , T_{\overline{\beta}})$. The complex Hessian is meant with respect to the Tanaka-Webster connection i.e. $$(\nabla^2 u)(X,Y) = (\nabla_X d u )Y = X(Y(u)) - (\nabla_X Y)(u),$$ for any $X, Y \in {\mathcal X}(M)$. Note that $\omega = c_n \; d \, {\rm vol}(g_\theta )$ (with $c_n = 2^n n!$). Then (by Green’s lemma) $$\int_M u^\alpha h_\alpha \, \omega = c_n \int_{\partial M} h\, u^\alpha g_\theta (T_\alpha , \xi ) d \sigma - \int_M {u^\alpha}_\alpha \, h \, \omega .$$ As the sublaplacian is locally given by $$\Delta_b u = {u^\alpha}_\alpha + {u^{\overline{\alpha}}}_{\overline{\alpha}}$$ we may conclude that $$\frac{d}{d t} \{ A_\theta (u + t h) \}_{t=0} = 2 \int_{M} (- b_n \, \Delta_b u + \rho \, u) h \, \omega + \label{e:35}$$ $$+ 2 \int_{\partial M} [b_n c_n g_\theta (\nabla^H u , \xi ) - a_n \mu_\theta u ] h \, d \sigma .$$ Also $$\frac{d}{d t} \{B_\theta (u + t h)\}_{t=0} = p \int_M u^{1 + 2/n} h \, \omega . \label{e:36}$$ As $T \in T(\partial M)$ one has $\xi \in H(M)$ hence $g_\theta (\nabla^H u , \xi ) = \xi (u)$ (also denoted by $\partial u/\partial \xi$). If $u$ achieves $Q(M)$ $$\frac{d}{d t} \{ Q_\theta (u + t h)\}_{t=0} = 0$$ hence $$2 \int_M (-b_n \, \Delta_b u + \rho \, u) h \, \omega + 2 \int_{\partial M} [b_n c_n \, \xi (u) - a_n \mu_\theta u ] h \, d \sigma -$$ $$- p \; Q_\theta (u) \, \int_M u^{1 + 2/n} h \, \omega = 0.$$ In particular this holds for $\left. h\right\|_{\partial M} = 0$ hence $$-b_n \, \Delta_b u + \rho u = (p/2) Q(M) u^{1 + 2/n}$$ and going back to arbitrary $h$ $$\frac{\partial u}{\partial \xi} - \frac{a_n}{b_n c_n} \, \mu_\theta \, u = 0 \;\;\; {\rm on} \;\; \partial M$$ which is (\[e:34\]) because $a_n /(b_n c_n ) = 2n^2 /(2n+1)$. The proof that $Q(M)$ is a CR invariant is similar to the arguments in [@kn:JL2], p. 174-175. Let $E^+ \to M$ be the ${\mathbb R}_+$-bundle spanned by $\theta$ and let us set $$E^\alpha_x = \{ \nu : E^+_x \to {\mathbb R} : \nu (t \theta_x ) = t^{-\alpha} \nu (\theta_x ), \; {\rm for \; all} \; t > 0\} , \;\; (\alpha > 0)$$ for any $x \in M$. Then $(\nu_\theta )_x (t \theta_x ) = 1/t$ defines a global frame $\{ \nu_\theta \}$ of $E^1 \to M$ (and of course $\{ \nu_\theta^\alpha \}$ is a global frame of $E^\alpha \to M$). We need the [CR invariant sublaplacian]{} $$L : \Gamma^\infty (E^{n/2}) \to \Gamma^\infty (E^{1 + n/2}), \;\; L (u \, \nu_\theta^{n/2}) = (- b_n \, \Delta_b u + \rho \, u) \nu_\theta^{1 + n/2} .$$ By definition $\int_M u \, \nu_\theta^{n+1} = \int_M u \, \omega$. A section $s = u \nu^\alpha_\theta$ in $E^\alpha$ is [positive]{} if $u > 0$. Finally, the fact that $Q(M)$ is a CR invariant follows from $$\label{e:37} Q(M) = \inf \{ \int_M (L s) \otimes s \; : \; s \in \Gamma^\infty (E^{n/2})$$ $${\rm a \; positive \; section \; such \; that} \;\; \int_M s^p = 1 \} .$$ The identity (\[e:37\]) follows from the fact that the sets $\{ A_\theta (u) : B_\theta (u) = 1 , u > 0 \}$ and $\{ A_\theta (u)/B_\theta (u) : u > 0 \}$ coincide and from the calculation $$\int_M (L s) \otimes s = \int_M (-b_n \, u \Delta_b u + \rho u^2 ) \omega ,$$ $$\int_M u (\Delta_b u) \; \omega = \int_M \{ {\rm div}(u \nabla^H u) - \\| \nabla^H u \\|^2 \} \omega =$$ $$= c_n \int_{\partial M} u \, \frac{\partial u}{\partial \xi} \, d \sigma - \int_M \\| \nabla^H u \\|^2 \omega ,$$ hence (by (\[e:34\])) $\int_M (L s) \otimes s = A_\theta (u)$, for any $s = u \nu_\theta^{n/2} \in \Gamma^\infty (E^{n/2})$. Minimal surfaces in ${\mathbb H}_n$ =================================== Let $(N, g)$ be a $2$-dimensional Riemannian manifold and $\Psi : N \to {\mathbb H}_n$ a minimal isometric immersion of $(N , g)$ into $({\mathbb H}_n , g_0 )$. Let $(U , z = x + i y)$ be isothermal local coordinates on $N$, i.e. locally $$g = 2 E (d x^2 + d y^2 ),$$ for some $E \in C^\infty (U)$, $E > 0$. As well known the Laplace-Beltrami operator of $(N , g)$ is locally given by $$\Delta u = \frac{2}{E} \; \frac{\partial^2 u}{\partial z \partial \overline{z}} \, , \;\;\; u \in C^2 (N).$$ Let us set $F^j = \Psi^j + i \Psi^{j+n}$, $1 \leq j \leq n$, and $f = \Psi^0$. Also, we consider $K : U \to {\mathbb C}$ given by $$K = \frac{\partial f}{\partial z} + i \sum_{j=1}^n (F^j \frac{\partial \overline{F}^j}{\partial z} - \overline{F}^j \frac{\partial F^j}{\partial z}) .$$ The normal component of the characteristic vector field $T = \partial /\partial t$ of $d \theta_0$ is locally given by $$T^\bot = ( 1 - \frac{2}{E} \|K\|^2 ) T - \label{e:tbot}$$ $$- \frac{1}{E} \{ ( \overline{K} \frac{\partial F^j}{\partial z} + K \frac{\partial F^j}{\partial \overline{z}} ) Z_j + ( \overline{K} \frac{\partial \overline{F}^j}{\partial z} + K \frac{\partial \overline{F}^j}{\partial \overline{z}} ) \overline{Z}_j \} .$$ \[l:8\] [Proof]{}. The characteristic direction decomposes as $T = \Psi_* T^T + T^\bot$, where $T^T = \lambda \partial /\partial z + \overline{\lambda} \partial /\partial \overline{z}$, for some $\lambda \in C^\infty (U)$. Taking the inner product with $\Psi_* \partial /\partial \overline{z}$ yields $\lambda = \overline{K}/E$ hence (\[e:partials\]) yields (\[e:tbot\]). $\square$ Let $\Psi : N \to {\mathbb H}_n$ be an isometric immersion of $(N , g)$ into $({\mathbb H}_n , g_0 )$. Then $$2 \sum_{j=1}^n \frac{\partial F^j}{\partial z} \frac{\partial F^j}{\partial \overline{z}} + K^2 = 0, \label{e:46}$$ $$\sum_{j=1}^n \left( \left\| \frac{\partial F^j}{\partial z} \right\|^2 + \left\| \frac{\partial F^j}{\partial \overline{z}} \right\|^2 \right) + \|K\|^2 \neq 0. \label{e:47}$$ \[l:9\] [Proof]{}. A calculation based on (\[e:partials\]) shows that the Webster metric of $({\mathbb H}_n , \theta_0 )$ is given (with respect to the frame $\{ \partial /\partial x^j , \; \partial /\partial y^j , \; \partial /\partial t \}$) by $$g_0 : \left( \begin{array}{ccc} 2 (\delta_{jk} + 2 y_j y_k ) & - 4 y_j x_k & - 2 y_j \\ - 4 x_j y_k & 2 (\delta_{jk} + 2 x_j x_k ) & 2 x_j \\ - 2 y_k & 2 x_k & 1 \end{array} \right)$$ hence $$g_\theta (\Psi_* \frac{\partial}{\partial z} \, , \, \Psi_* \frac{\partial}{\partial \overline{z}}) = \Psi^A_z \Psi^B_{\overline{z}} g_{AB} = \|K\|^2 + \sum_j (\|F^j_z \|^2 + \|F^j_{\overline{z}}\|^2 ),$$ $$g_\theta (\Psi_* \frac{\partial}{\partial z} \, , \, \Psi_* \frac{\partial}{\partial z}) = \Psi^A_z \Psi^B_z g_{AB} = K^2 + \sum_j F^j_z F^j_{\overline{z}} ,$$ (where $g_{AB} = g_0 (\partial_A , \partial_B )$). Since $\Psi$ is an isometric immersion $$g_0 (\Psi_* \frac{\partial}{\partial x} \; , \; \Psi_* \frac{\partial}{\partial y}) = 0, \label{e:48}$$ $$g_0 (\Psi_* \frac{\partial}{\partial x} \; , \; \Psi_* \frac{\partial}{\partial x}) = g_0 (\Psi_* \frac{\partial}{\partial y} \; , \; \Psi_* \frac{\partial}{\partial y}), \label{e:49}$$ and then (\[e:48\])-(\[e:49\]) yield (\[e:46\])-(\[e:47\]), respectively. $\square$ 0.1in Note that (again by (\[e:partials\])) $$\Delta \Psi = (\Delta \psi^A ) \partial_A = (\Delta F^j ) Z_j + (\Delta \overline{F}^j ) \overline{Z}_j +$$ $$+ \{ \Delta f + 2 \sum_{j=1}^n (\Psi^{j} \Delta \Psi^{j+n} - \Psi^{j+n} \Delta \Psi^{j}) \} T$$ and (by Lemma \[l:8\]) $$i E \, J T^\bot = (\overline{K} F^j_z + K F^j_{\overline{z}}) Z_j - (\overline{K} \; \overline{F}^j_z + K \overline{F}^j_{\overline{z}}) \overline{Z}_j$$ hence the minimality condition (\[e:delta\]) becomes $$\Delta F^j = - \frac{2i}{E} (\overline{K} F^j_z + K F^j_{\overline{z}}), \;\;\; 1 \leq j \leq n, \label{e:delf}$$ and $\Delta f = \frac{i}{2} \sum_j (\overline{F}^j \Delta F^j - F^j \Delta \overline{F}^j )$ or (by (\[e:delf\])) $$\Delta f = \frac{1}{E} \{ \overline{K} (\|F\|^2 )_z + K (\|F\|^2 )_{\overline{z}} \} . \label{e:delfp}$$ Let $N$ be a Riemann surface. An immersion $\Psi : N \to {\mathbb H}_n$ is [conformal]{} if (\[e:48\])-(\[e:49\]) hold, for any local complex coordinate system $(U , z = x + i y)$ on $N$. Moreover (\[e:delf\])-(\[e:delfp\]) lead to the following definition. A [minimal surface]{} in ${\mathbb H}_n$ is a Riemann surface $N$ together with a conformal immersion $\Psi : N \to {\mathbb H}_n$ such that $$F^j_{z\overline{z}} + i (\overline{K} F^j_z + K F^j_{\overline{z}}) = 0, \;\;\; 1 \leq j \leq n, \label{e:delf1}$$ $$f_{z\overline{z}} - \frac{1}{2} \{ \overline{K} (\|F\|^2 )_z + K (\|F\|^2 )_{\overline{z}} \} = 0. \label{e:delfp1}$$ Here $\|F\|^2 = \sum_j F^j \overline{F}^j$. We may state the following Let $\Omega \subset {\mathbb C}$ be a simply connected domain and $\Psi : \Omega \to {\mathbb H}_n$ a minimal surface such that $J T^\bot = 0$ [(]{}e.g. $\Psi (\Omega )$ is tangent to the characteristic direction of $d \theta_0$[)]{}. Let us set $\Phi = \partial \Psi /\partial z$. Then $\Phi$ is holomorphic and [(\[e:46\])-(\[e:47\])]{} hold in $\Omega$. Viceversa, let $\Phi : \Omega \to {\mathbb C}^{2n+1}$ be a holomorphic map and let us set $$\label{e:psi} \Psi^A (z) = {\rm Re} \int_o^z \Phi^j (\zeta ) d \zeta , \;\;\; A \in \{ 0, 1, \cdots , 2n \} ,$$ for any $z \in \Omega$, where $o \in \Omega$ is a fixed base point. Let $K : \Omega \to {\mathbb C}$ be given by $$K = \Phi^0 - 2 \sum_{j=1}^n \{ \Phi^j \; {\rm Re} \int_o^z \Phi^{j+n}(\zeta ) d \zeta + \Phi^{j+n} \; {\rm Re} \int_o^z \Phi^j (\zeta ) d \zeta \} .$$ If the following identities hold in $\Omega$ $$2 \sum_{j=1}^n \{ \|\Phi^j \|^2 - \|\Phi^{j+n}\|^2 + i(\Phi^{j+n} \overline{\Phi}^j + \Phi^j \overline{\Phi}^{j+n} ) \} + K^2 = 0, \label{e:new55}$$ $$2 \sum_{j=1}^n (\|\Phi^j \|^2 + \|\Phi^{j+n}\|^2 ) + \|K\|^2 \neq 0, \label{e:new56}$$ $$\overline{K} (\Phi^j + i \Phi^{j+n}) + K (\overline{\Phi}^j + i \overline{\Phi}^{j+n}) = 0, \;\;\; 1 \leq j \leq n, \label{e:54}$$ then $\Psi : \Omega \to {\mathbb H}_n$ is a minimal immersion such that $J T^\bot = 0$. \[t:7\] Compare to Theorem 8.1 in [@kn:DW], p. 58. [Proof of Theorem]{} \[t:7\]. (\[e:46\])-(\[e:47\]) follow from Lemma \[l:9\]. Next $J T^\bot = 0$ and (\[e:delf\])-(\[e:delfp\]) yield $\partial \Phi /\partial \overline{z} = 0$ in $\Omega$. Viceversa, given a holomorphic map $\Phi : \Omega \to {\mathbb C}^{2n+1}$ the function $\Psi^A$ given by (\[e:psi\]) is well defined (by the classical theorem of Cauchy the integral doesn’t depend upon the choice of path from $o$ to $z$) and $\partial \Psi /\partial z = \Phi$ hence (\[e:new55\])-(\[e:new56\]) yield (\[e:46\])-(\[e:47\]) so that (\[e:48\])-(\[e:49\]) are satisfied and $g_0 (\Psi_* \partial /\partial x \, , \, \Psi_* \partial /\partial x) \neq 0$, i.e. $\Psi$ is a conformal immersion. Finally (\[e:54\]) may be written $$\overline{K} F^j_z + K F^j_{\overline{z}} = 0, \;\;\; 1 \leq j \leq n,$$ which is equivalent (by Lemma \[l:8\]) to $J T^\bot = 0$ and (\[e:delf1\])-(\[e:delfp1\]) imply minimality. $\square$ [12345]{} N. Arcozzi & F. Ferrari, [Metric normal and distance function in the Heisnberg group]{}, preprint, 2003. A. Bahri & H. Brezis, [Nonlinear elliptic equations]{}, in “Topics in Geometry in memory of J. D’Atri”, Ed. by S. Gindikin, Birkhäuser, Boston-Basel-Berlin, 1996, pp. 1-100. E. Barletta & S. Dragomir, [On the CR structure of the tangent sphere bundle]{}, Le Matematiche, Catania, (2)L(1995), 237-249. E. Barletta & S. Dragomir & H. Urakawa, [Yang-Mills fields on CR manifolds]{}, preprint, 2004. J.K. Beem & P.E. Ehrlich, [Global Lorentzian geometry]{}, Marcel Dekker, Inc., New York - Basel, 1981. I. Birindelli & E. Lanconelli, [A negative answer to a one-dimensional symmetry problem in the Heisenberg group]{}, preprint, 2003. J-H. Cheng & J-F. Hwang & A. Malchiodi & P. Yang, [Minimal surfaces in pseudohermitian geometry]{}, preprint, 2004. S. Dragomir & J.C. Wood, [Sottovarietà minimali ed applicazioni armoniche]{}, Quaderni dell’Unione Matematica Italiana, Vol. 35, Pitagora Editrice, Bologna, 1989. S. Dragomir, [On a conjecture of J.M. Lee]{}, Hokkaido Math. J., (1)23(1994), 35-49. J.F. Escobar, [The Yamabe problem on manifolds with boundary]{}, J. Diff. Geometry, 35(1992), 21-84. G.B. Folland & E.M. Stein, [Estimates for the $\overline{\partial}_{b}$-complex and analysis on the Heisenberg group]{}, Comm. Pure Appl. Math., 27(1974), 429-522. N. Gamara & R. Yacoub, [CR Yamabe conjecture - the conformally flat case]{}, Pacific J. Math., (1)201(2001), 121-175. N. Garofalo & E. Lanconelli, [Frequency functions on the Heisenberg group, the uncertanty principle and unique continuation]{}, Ann. de l’Inst. Fourier, (2)40(1990), 313-356. N. Garofalo & S.D. Pauls, [The Bernstein problem in the Heisenberg group]{}, preprint, 2002. D. Jerison & J.M. Lee, [A subelliptic, nonlinear eigenvalue problem and scalar curvature on CR manifolds]{}, Contemp. Math., 27(1984), 57-63. D. Jerison & J.M. Lee, [The Yamabe problem on CR manifolds]{}, J. Diff. Geometry, 25(1987), 167-197. D. Jerison & J.M. Lee, [CR normal coordinates and the Yamabe problem]{}, J. Diff. Geometry, 29(1989), 303-344. J.M. Lee, [The Fefferman metric and pseudohermitian invariants]{}, Trans. A.M.S., (1)296(1986), 411-429. J.M. Lee & T. Parker, [The Yamabe problem]{}, Bull. Amer. Math. Soc., 17(1987), 37-91. H. Lewy, [On the local character of the solution of an atypical linear differential equation in three variables and a related theorem for regular functions of two complex variables]{}, Ann. of Math., 64(1956), 514-522. B. O’Neill, [The fundamental equations of a submersion]{}, Michigan Math. J., 13(1966), 459-469. B. O’Neill, [Semi-Riemannian geometry]{}, Academic Press, New York-LOndon-Paris-San Diego-San Francisco-Sao Paulo-Sydney-Tokyo-Toronto, 1983. S.D. Pauls, [Minimal surfaces in the Heisenberg group]{}, Geometriae Dedicata, 104(2004), 201-231. S.M. Webster, [Pseudohermitian structures on a real hypersurface]{}, J. Diff. Geometry, 13(1978), 25-41. [^1]: Author’s address: Università degli Studi della Basilicata, Dipartimento di Matematica, Campus Macchia Romana, 85100 Potenza, Italy, e-mail: [[email protected]]{}. The Author acknowledges support from INdAM, Italy, within the interdisciplinary project [Nonlinear subelliptic equations of variational origin in contact geometry]{}. [^2]: As to the sign convention the sublaplacian in [@kn:JL2] is $-\Delta_b$. [^3]: If $n \geq 2$ and $M$ is not locally CR equivalent to $S^{2n+1}$ then $\lambda (M) < \lambda (S^{2n+1})$, cf. [@kn:JL2].	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper is concerned with the analysis of correlation between two high-dimensional data sets when there are only few correlated signal components but the number of samples is very small, possibly much smaller than the dimensions of the data. In such a scenario, a principal component analysis (PCA) rank-reduction preprocessing step is commonly performed before applying canonical correlation analysis (CCA). We present simple, yet very effective approaches to the [joint]{} model-order selection of the number of dimensions that should be retained through the PCA step [and]{} the number of correlated signals. These approaches are based on reduced-rank versions of the Bartlett-Lawley hypothesis test and the minimum description length information-theoretic criterion. Simulation results show that the techniques perform well for very small sample sizes even in colored noise.' author: - 'Yang Song, Peter J. Schreier, David Ramírez, and Tanuj Hasija [^1] [^2]' bibliography: - 'refs.bib' title: 'Canonical correlation analysis of high-dimensional data with very small sample support' --- Bartlett-Lawley statistic, canonical correlation analysis, model-order selection, principal component analysis, small sample support. Introduction {#intro} ============ Correlation analysis based on only small sample support is a challenging task yet with important applications in areas as diverse as biomedicine (e.g. [@Lin2006; @Correa2010]), climate science (e.g. [@Wallace1992; @Shabbar2004]), array processing (e.g. [@GeICASSP09]), and others. In this paper, we look at the scenario where the data sets have large dimensions but there are only few correlated signal components. Probably the most common way of analyzing correlation between two data sets is canonical correlation analysis (CCA) [@HotellingBio36]. In CCA, the observed data $\xB \in \Cn^n$ and $\yB \in \Cn^m$ are transformed into $p$-dimensional internal (latent) representations ${\bf a} = {\bf S}\xB$ and ${\bf b} = {\bf T}\yB$, where $p = \min(n,m)$, using linear transformations described by the matrices ${\bf S} \in \Cn^{p \times n}$ and ${\bf T} \in \Cn^{p \times m}$. The key idea is to determine ${\bf S}$ and ${\bf T}$ such that most of the correlation between $\xB$ and $\yB$ is captured in a low-dimensional subspace. CCA proceeds as follows. First two vectors (“projectors”) ${\bf s}_1 \in \Cn^n$ and ${\bf t}_1 \in \Cn^m$ are determined such that the absolute value of the scalar correlation coefficient $k_1$ between the internal variables $a_1 = {\bf s}_1^T \xB$ and $b_1 = {\bf t}_1^T \yB$ is maximized. The internal variables $(a_1, b_1)$ constitute the first pair of [canonical variables]{}, and $k_1$ is called the first [canonical correlation (coefficient)]{}. The next pair of canonical variables $(a_2, b_2)$ maximizes the absolute value of the scalar correlation coefficient $k_2$ (the second canonical correlation) between $a_2 = {\bf s}_2^T \xB$ and $b_2 = {\bf t}_2^T \yB$, subject to the constraint that they are to be uncorrelated with the first pair. A total of $p$ correlations is determined in this manner, and ${\bf S} = [{\bf s}_1, ..., {\bf s}_p]^T$, ${\bf T} = [{\bf t}_1, ..., {\bf t}_p]^T$. CCA can be performed via the singular value decomposition of the coherence matrix [@ScharfSPT0300] $$\RB_{xx}^{-1/2} \RB_{xy} \RB_{yy}^{-1/2} = {\bf F} {\bf K} {\bf G}^H,\label{equ_CCA}$$ where $\RB_{xy}$ is the cross-covariance matrix between $\xB$ and $\yB$, and $\RB_{xx}$ and $\RB_{yy}$ are the auto-covariance matrices of $\xB$ and $\yB$. The canonical correlations $0 \leq k_i \leq 1$ are the singular values, which are the diagonal elements of the diagonal matrix ${\bf K}$. The transformations that generate the latent representations ${\bf a}$ and ${\bf b}$ are then described by ${\bf S} = {\bf F}^H \RB_{xx}^{-1/2}$ and ${\bf T} = {\bf G}^H {\bf R}_{yy}^{-1/2}$. In practice, we do not know the covariance matrices and must estimate them from samples. If CCA is performed based on sample covariance matrices, it leads to sample canonical correlations $\hat{k}_i$. If the number of samples $M$ is not significantly larger than the dimensions $m$ and $n$, these $\hat{k}_i$’s can be extremely misleading as they are generally substantially overestimated. Indeed, if $M < m+n$ then $m+n - M$ sample canonical correlations are always identically one, which means that they do not carry any information at all about the true population canonical correlations [@PezeshkiACSSC04]. In order to avoid this, we perform a dimension-reduction preprocessing step before applying CCA. The most common type of preprocessing is principal component analysis (PCA). That is, instead of applying (\[equ\_CCA\]) directly to the sample covariance matrices, we first extract a reduced number $r_x$ of components from $\xB$ that account for a large fraction of the total variance in $\xB$. Similarly, we extract $r_y$ components from $\yB$ that account for a large fraction of the total variance in $\yB$. CCA is then performed on the components extracted from $\xB$ and $\yB$. The necessity of a PCA step preceding CCA for small sample sizes was shown in [@NadakuditiSSP11] using random matrix theory tools. The paper [@NadakuditiSSP11], however, did not answer the critical question of how to determine $r_x$ and $r_y$ such that the estimated $\hat{k}_i$’s best reflect the true population canonical correlations $k_i$. At the same time, a key question in any correlation analysis is how many correlated signals there are. If we had access to the population canonical correlations, we could simply count the number of nonzero $k_i$’s. Since we don’t, we need to estimate the number $d$ of correlated signals from the estimated $\hat{k}_i$’s. This is a model-order selection problem. In this paper, we present approaches to [jointly]{} determine, for a PCA-CCA setup, the ranks $r_x$ and $r_y$ of the PCA step and the number $d$ of correlated signals based on extremely small sample support, with $M$ possibly less or even substantially less than $m + n$. These approaches rely on the fact that, while $m$ and $n$ may be very large, the number of correlated signals $d$ is often small. However, a complicating factor of the PCA-CCA setup is that PCA is designed to extract components that account for most of the variance [within one]{} data set, but these components are not necessarily the ones that account for most of the correlation [between two]{} data sets. In the literature, most of the work on model-order selection deals with either (i) determining the number of signals in a single data set [@WaxASSPT0485; @NadakuditiSPT0708; @Lu2015] or (ii) the number of correlated signals between two data sets, but without a PCA step [@BartlettBio41; @LawleyBio59; @FujikoshiBiometrika79; @ZhangSPT0493; @ChenIET-RSN1096; @GundersonJMA97; @StoicaSPT0496]. There is only little work on the [joint]{} model-order selection in a PCA-CCA setup, most of which is rather ad hoc [@Zwick1986a; @Hwang2013] and only [@Roseveare2015] presents a systematic approach. However, none of these joint PCA-CCA techniques works in the sample-poor case. In the absence of any methodical approach in the sample-poor regime, it is common to use very simple rules of thumb such as “choose the PCA ranks such that a certain percentage (e.g., 70%) of the total variance/energy in each data set is retained” (see, e.g., [@Wallace1992]). Needless to say, such rules based on experience only work for specific scenarios. In general, there are two main approaches to model-order selection: hypothesis tests and information-theoretic criteria. [Hypothesis tests]{} [@BartlettBio41; @LawleyBio59] are usually series of binary generalized likelihood ratio tests (GLRTs). Starting at $s = 0$, they test whether the model has order $s$ (the null hypothesis) or order greater than $s$ (the alternative). If the null hypothesis is rejected, $s$ is incremented and a new test is run. This proceeds until the null hypothesis is not rejected or the maximum model order is reached. The disadvantage of hypothesis tests is that they require the subjective selection of a probability of false alarm. This can be avoided by using [information-theoretic criteria]{} (ICs) (e.g., [@WaxASSPT0485]), which compute a score as a function of model order. This score is the difference between the likelihood for the observed data, which measures how well the model fits the observed data, and a penalty function. With increasing order there is an increasing number of free parameters, and so the model fit becomes better. In order to avoid overfitting, complex models are penalized by the penalty function, which increases with model order. The best trade-off is achieved when the difference of likelihood and penalty function is maximized. It should be noted that the GLRT and IC methods for model-order selection are actually closely linked [@StoicaSPL1110]—a fact that we will exploit, as well. In this paper, we present approaches to the joint model-order selection in a PCA-CCA setup based on reduced-rank versions of both the Bartlett-Lawley hypothesis test and the minimum description length (MDL) IC [@WaxASSPT0485]. As far as we know, these are currently the only techniques capable of handling the combined PCA-CCA approach in the sample-poor regime. An early version of this paper was presented at ICASSP 2015 [@SongICASSP15]. [ We would also like to contrast our work with so-called [sparse CCA]{} (e.g., [@SparseCCA1; @SparseCCA2]). In sparse CCA, a sparsity constraint is placed on the projectors $\sB_i$ and ${\bf t}_i$, which means that each canonical variable $a_i$ or $b_i$ is a linear combination of only a few components in $\xB$ and $\yB$, respectively. While sparse CCA was not proposed to deal with the sample-poor scenario, in principle it can be used as an alternative to PCA-CCA [if]{} there is a priori information that the projectors are sparse. However, in many scenarios of interest (e.g., the applications in biomedicine, climate science, and array processing cited above) there is no justification to assume sparse projectors. When applied to non-sparse problems, sparse CCA will not work well.]{} Our program for this paper is as follows. In Section \[sec:probform\], we formulate the problem and illustrate the issues that arise when performing CCA based on very small sample sizes and how a combined PCA-CCA approach can address these. We present our approaches based on the hypothesis test in Section \[sec:hyptest\] and based on the MDL-IC in Section \[sec:mdl\]. Extensive simulation results are shown in Section \[sec:sim\]. Problem Formulation {#sec:probform} =================== We observe $M$ independent and identically distributed (i.i.d.) sample pairs $\xB_i \in \Cn^n$, $\yB_i \in \Cn^m$ that are drawn from the two-channel measurement model $$\begin{aligned} {\bf x}&= {\bf A}_x {\bf s}_x + {\bf n}_x, \nonumber \\ {\bf y}&= {\bf A}_y {\bf s}_y + {\bf n}_y. \label{eq:model}\end{aligned}$$ The signals $\sB_x \in \Cn^{d + f_x}$ and $\sB_y \in \Cn^{d + f_y}$ are jointly Gaussian with zero means and cross-covariance matrix $$\RB_{s_xs_y} = \left[\begin{array}{cc} {\bf diag}(\rho_1\sigma_{x,1}\sigma_{y,1},\ldots,\rho_d\sigma_{x,d}\sigma_{y,d}) & {\bf 0}_{d\times f_y} \\ {\bf 0}_{f_x\times d} & {\bf 0}_{f_x\times f_y} \end{array}\right],$$ where $\sigma_{x,i}$ is the unknown standard deviation of signal component $s_{x,i}$, $\sigma_{y,i}$ the unknown standard deviation of signal component $s_{y,i}$, and $\rho_i$ the unknown correlation coefficient between $s_{x,i}$ and $s_{y,i}$. Hence, the first $d$ components of $\sB_x$ and $\sB_y$ are correlated, whereas the next ($f_x$, $f_y$) components are independent between $\sB_x$ and $\sB_y$. The correlated components may be stronger or weaker than the independent components. Without loss of generality, we assume the auto-covariance matrices $\RB_{s_xs_x}$ and $\RB_{s_ys_y}$ to be diagonal. The matrices $\AB_x \in \Cn^{n \times (d+f_x)}$ and $\AB_y \in \Cn^{m \times (d+f_y)}$ as well as the dimensions $d$, $f_x$, and $f_y$ are deterministic but unknown. Without loss of generality, $\AB_x$ and $\AB_y$ are assumed to have full column-rank. With all these assumptions, $\|\rho_i\|$ is the $i$th canonical correlation coefficient $k_i$ between $\sB_x$ and $\sB_y$. The noise vectors ${\bf n}_x \in \Cn^n$ and ${\bf n}_y \in \Cn^m$ are independent of each other, independent of the signals, zero-mean Gaussian, and with [unknown]{} (arbitrary) covariance matrices. Compared to the dimensions $m$ and $n$ (which may be very large), we assume that there are only few correlated signals and only few independent signals with variance larger than the correlated signals (but there can be many independent signals with variance smaller than the correlated signals). [However, because we [do not]{} assume that the mixing matrices $\AB_x$ and $\AB_y$ in (\[eq:model\]) are sparse, the cross-covariance matrix $\RB_{xy}$ between the observed vectors $\xB$ and $\yB$ is [not sparse]{} and sparse CCA is generally not suitable for this scenario.]{} ![Sample canonical correlation coefficients $\hat{k}_i$ for different sample sizes $M$, averaged over $1000$ runs. The are three nonzero population canonical correlations, which are $0.9$, $0.7$, and $0.5$, depicted as $$. In all cases shown, the true $k_i$’s are significantly overestimated.[]{data-label="fig:wrongCCs"}](can_corr_vs_M.pdf) We collect the $M$ sample pairs in data matrices $\XB = [\xB_1, ..., \xB_M]$ and $\YB = [\yB_1, ..., \yB_M]$, from which we compute the sample covariance matrices $\hat{\RB}_{xx} = \XB\XB^H/M$, $\hat{\RB}_{yy} = \YB\YB^H/M$, and $\hat{\RB}_{xy} = \XB\YB^H/M$. In the case of small sample support, the sample canonical correlations $\hat{k}_i$, $i = 1, ..., p$, $p = \min(n,m)$, computed from the sample covariance matrices can be extremely misleading. It has been shown in [@PezeshkiACSSC04] that when $M < m+n$, at least $m+n - M$ sample canonical correlations will be identically one regardless of the two-channel model that generates the data samples. In such a small sample scenario, the $\hat{k}_i$’s cannot be used to infer the number of correlated signals. But even in the case with $M$ greater (but not substantially greater) than $m+n$ the sample $\hat{k}_i$’s are generally significantly overestimated. This is shown in Fig. \[fig:wrongCCs\], which displays the sample canonical correlations for a model of dimension $m = n = 20$, with $d = 3$ correlated components and $f_x = f_y = 0$ independent components for different sample sizes $M$. Even for $M = 200$, where the number of samples is ten times the dimension of the system, the $\hat{k}_i$’s for $i \geq 4$ are quite wrong, and it is impossible from visual inspection to determine the number of correlated components. This motivates the use of a rank-reduction preprocessing step. The most common type of preprocessing is PCA, and a combined PCA-CCA approach is the setup that we consider in our paper. So let us investigate what effect rank reduction has on the estimated canonical correlations. The PCA step retains those $r_x$ and $r_y$ components in $\XB$ and $\YB$, respectively, that account for most of their total variance. These components can be computed as follows. We first determine the singular value decompositions (SVDs) of the data matrices $\XB = \UB_x \SigmaB_x \VB_x^H$ and $\YB = \UB_y \SigmaB_y \VB_y^H$. Then the reduced-rank PCA descriptions of $\XB$ and $\YB$ are $$\begin{aligned} \XB_{r_x} & = \UB_x^H(:,1:r_x) \XB \in \Cn^{r_x \times M}, \nonumber \\ \YB_{r_y} & = \UB_y^H(:,1:r_y) \YB \in \Cn^{r_y \times M}, \label{equ:XrYr}\end{aligned}$$ where ${\bf U}_x(:,1:r_x)$ denotes the $n\times r_x$ matrix containing the first $r_x$ columns of $\UB_x$, which are associated with the largest $r_x$ singular values of ${\bf X}$, and ${\bf U}_y(:,1:r_y)$ denotes the $m\times r_y$ matrix containing the first $r_y$ columns of $\UB_y$, which are associated with the largest $r_y$ singular values of ${\bf Y}$. Now let $\widetilde{\RB}_{xx} = \XB_{r_x}\XB^H_{r_x}/M$, $\widetilde{\RB}_{yy} = \YB_{r_y}\YB_{r_y}^H/M$, and $\widetilde{\RB}_{xy} = \XB_{r_x}\YB_{r_y}^H/M$ be the sample covariance matrices from the reduced-dimensional PCA descriptions. The corresponding estimated canonical correlations $\hat{k}_i(r_x,r_y)$ may be computed as the singular values of the reduced-dimensional sample coherence matrix, which is [@PezeshkiACSSC04] $$\begin{aligned} & {\bf{\widetilde R}}_{xx}^{-1/2} {\bf{\widetilde R}}_{xy} {\bf{\widetilde R}}_{yy}^{-1/2} \nonumber \\ =&\UB^H_x(:,1:r_x) \UB_x \left(\SigmaB_x\SigmaB_x^H\right)^{-1/2} \UB_x^H \UB_x(:,1:r_x) \nonumber \\ &\times \UB^H_x(:,1:r_x) \UB_x \SigmaB_x \VB_x^H \VB_y \SigmaB_y^H \UB_y^H \UB_y(:,1:r_y) \nonumber \\ &\times \UB^H_y(:,1:r_y) \UB_y \left(\SigmaB_y\SigmaB_y^H\right)^{-1/2} \UB_y^H \UB_y(:,1:r_y) \nonumber \\ =&\left[{\bf I}_{r_x},{\bf 0}_{r_x\times(m-r_x)}\right] \left(\SigmaB_x\SigmaB_x^H\right)^{-1/2} \left[{\bf I}_{r_x},{\bf 0}_{r_x\times(m-r_x)}\right]^H \nonumber \\ &\times \left[{\bf I}_{r_x},{\bf 0}_{r_x\times(m-r_x)}\right] \SigmaB_x \VB_x^H \VB_y \SigmaB_y^H \left[{\bf I}_{r_y},{\bf 0}_{r_y\times(m-r_y)}\right]^H \nonumber \\ &\times \left[{\bf I}_{r_y},{\bf 0}_{r_y\times(m-r_y)}\right] \left(\SigmaB_y\SigmaB_y^H\right)^{-1/2} \left[{\bf I}_{r_y},{\bf 0}_{r_y\times(m-r_y)}\right]^H \nonumber \\ =& \SigmaB_x^{-1}(1:r_x,1:r_x)\left[\SigmaB_x(1:r_x,1:r_x),{\bf 0}_{r_x\times(M-r_x)}\right]\VB_x^H \VB_y \nonumber \\ & \times\left[\SigmaB_y(1:r_y,1:r_y),{\bf 0}_{r_y\times(M-r_y)}\right]^H\SigmaB_y^{-1}(1:r_y,1:r_y) \nonumber\\ =& \left[{\bf I}_{r_x},{\bf 0}_{r_x\times(M-r_x)}\right] \VB_x^H \VB_y \left[{\bf I}_{r_y},{\bf 0}_{r_y\times(M-r_y)}\right]^H \nonumber\\ =& \VB_x^H(:,1:r_x) \VB_y(:,1:r_y).\label{equ:VxVy}\end{aligned}$$ The thus computed canonical correlations $\hat{k}_i(r_x,r_y)$, $i = 1, ..., r$, $r = \min(r_x,r_y)$, depend on the ranks $r_x$ and $r_y$. As seen in (\[equ:VxVy\]), the $i$th estimated canonical correlation $\hat{k}_i(r_x,r_y)$ can be found as the $i$th largest singular value of ${\bf V}^H_x(:,1:r_x) {\bf V}_y(:,1:r_y)$, where $\VB_x$ and $\VB_y$ are the matrices of right singular vectors of $\XB$ and $\YB$, respectively. To avoid defective unit canonical correlations, we must choose $r_x+r_y \leq M$ and $\max(r_x,r_y) \leq p$. This, however, does not tell us what the optimum choices for $r_x$ and $r_y$ are such that the $\hat{k}_i(r_x,r_y)$’s are as close to the true canonical correlations as possible. Intuitively, it seems that $r_x$ and $r_y$ should be chosen large enough to capture as much of the correlated signal components as possible without including too much noise. If the correlated components are weaker than some of the independent components, this will inevitably mean that the PCA preprocessing step must also keep those stronger independent components. On the other hand, if the correlated components are also the strongest components, it would be better if the PCA step got rid of the independent components. Hence, without noise, $r_x$ would ideally be chosen between $d$ and $d+f_x$, and $r_y$ between $d$ and $d+f_y$. With noise, the ranks $r_x$ and $r_y$ may also fall outside of these ranges, depending on the properties of the noise and the relative strengths of the signals. It can be shown using Cauchy’s interlacing theorem (the result is presented as Lemma \[lemma\_interlace\] in Appendix \[app:interlace\]) that increasing the ranks of the PCA steps increases [every]{} estimated canonical correlation coefficient. Hence, choosing too large an $r_x$ or $r_y$ will lead to estimated canonical correlations that are greater, possibly significantly greater, than the true canonical correlations. On the other hand, if $r_x$ and $r_y$ are not large enough, then the rank-reduced representation does not contain all of the correlated components, and thus the estimated canonical correlations can be too small. ![Effect of rank reduction on the estimated canonical correlations $\hat{k}_i(r)$, averaged over $1000$ runs. The are $d = 3$ correlated signal components with population canonical correlation coefficients $0.9$, $0.8$, and $0.7$ (depicted as $$), and $f_x = f_y = 2$ stronger independent signal components. For $r > 5$, the canonical correlation coefficients are all overestimated. For $r < 5$, the nonzero coefficients are underestimated. The ranks of the PCA steps for $\xB$ and $\yB$ are the same: $r = r_x = r_y$.[]{data-label="fig_cck"}](can_corr_vs_r.pdf) These considerations can be illustrated by the following example, where $M = 30$ and $m = n = 20$. There are $d = 3$ correlated signals (each with variance 1.5) and $f = f_x = f_y = 2$ independent signals (each with variance 5). Since the independent signals are stronger than the correlated signals (and the numbers $f_x$ and $f_y$ of independent signals in $\sB_x$ and $\sB_y$ are identical), we would expect $r_x = r_y = d + f = 5$ to be the optimum rank for the PCA step. Indeed, Fig. \[fig\_cck\] shows that choosing $r = r_x = r_y$ greater than 5 leads to $\hat{k}_i$’s that are too large, whereas $r$ less than 5 leads to $\hat{k}_i$’s that are too small. While the exact relationships depend on the variances of signal and noise components and the correlation coefficients, the principle observed here generalizes to other settings. Order selection based on hypothesis test {#sec:hyptest} ======================================== Traditional test ---------------- In the case of sufficient number of samples, the traditional hypothesis test [@BartlettBio41; @LawleyBio59] for determining the number $d$ of correlated components between $\xB$ and $\yB$ is a series of binary hypothesis tests. Starting with $s = 0$, it tests the null hypothesis $H_0$: $d = s$ versus the alternative hypothesis $H_1$: $d > s$. If $H_0$ is rejected, $s$ is incremented and a new test is run. This proceeds until $H_0$ is not rejected or $s = p = \min(n,m)$ is reached. The binary test in [@BartlettBio41; @LawleyBio59] is a generalized likelihood ratio test (GLRT) of $H_0$ vs. $H_1$. For a given number $s$ of correlated signals, let $\Omega_s$ denote the parameter space of the model, which consists of the auto- and cross-covariance matrices. The maximum value of the log-likelihood function for a given number $s$ of correlated signals, maximized over the parameter space $\Omega_s$, is [@StoicaSPT0496] $$\ell_{\max}(\XB,\YB\|\Omega_s) = -M \ln \prod_{i=1}^s \left(1 - \hat{k}_i^2 \right).$$ Canonical correlation coefficients close to 1 are strong evidence of correlation between $\xB$ and $\yB$ and thus lead to large $\ell_{\max}$. Now let $\Omega_{d > s}$ denote the parameter space of all models where the assumed number of correlated signals $d$ is greater than $s$. The generalized log-likelihood ratio for testing $H_0$ vs. $H_1$ is [@ChenIET-RSN1096] $$\begin{aligned} \Lambda(n,m,s) & = \ell_{\rm max}({\bf X},{\bf Y}\|\Omega_{d=s})-\ell_{\rm max}({\bf X},{\bf Y}\|\Omega_{d>s}) \nonumber \\ & = \ell_{\rm max}({\bf X},{\bf Y}\|\Omega_{s})-\ell_{\rm max}({\bf X},{\bf Y}\|\Omega_{p}) \nonumber \\ & = M \ln \prod_{i=s+1}^p \left(1 - \hat{k}_i^2\right), \label{equ:GLRT}\end{aligned}$$ where the second identity follows from the fact that the maximum of the likelihood function, under the constraint $d > s$, occurs when the model has the most degrees of freedom, i.e., for $d = p = \min(n,m)$. The cross-covariance matrix has $N_\Omega(n,m,s) = 2s(m+n-s)$ degrees of freedom [@StoicaSPT0496]. Wilks’ theorem [@Wilks1938] says that $-2 \Lambda(n,m,s)$ is asymptotically (as $M \rightarrow \infty$) $\chi^2$-distributed with degrees of freedom equal to the difference of the dimensions of the parameter spaces $\Omega_p$ and $\Omega_s$:[^3] $$\begin{aligned} N_\Lambda(n,m,s) & = N_\Omega(n,m,p) - N_\Omega(n,m,s) \nonumber \\ & = 2p(m + n - p) - 2s(m + n - s) \nonumber \\ & = 2(m - s)(n - s)\end{aligned}$$ For finite $M$, the closeness of the $\chi^2$-approximation may be improved by replacing $-2\Lambda(n,m,s)$ with the Bartlett-Lawley statistic [@BartlettBio41; @LawleyBio59] $$\begin{aligned} C(n,m,s) = & - \! 2 \left( M-s-\frac{m+n+1}{2}+\sum_{i=1}^{s} \hat{k}_i^{-2} \right) \nonumber \\ & \times \ln\prod_{i=s+1}^{p}\left(1-\hat{k}_i^2\right).\label{cs1}\end{aligned}$$ This correction makes the moments of the test statistic equal to the moments of the $\chi^2$-distribution. As long as $M$ is large compared to $m$ and $n$, the statistic $C(n,m,s)$ is generally very close to a $\chi^2$-distribution. [ Note that this is independent of the covariance matrix of the noise, since it is not used anywhere in the derivation.]{} This allows computation of a test threshold $T(n,m,s)$ for a given probability of false alarm. Test with PCA preprocessing --------------------------- Instead of running the test directly on $\XB$ and $\YB$, we would like to apply the test to the reduced-rank PCA descriptions $\XB_{r_x}$ and $\YB_{r_y}$ obtained in (\[equ:XrYr\]). By performing PCA on $\xB$ and $\yB$, we create a new reduced-rank two-channel model: $$\begin{aligned} {\bf x}_{r_x}& ={\bf U}^H_x(:,1:r_x){\bf x} \nonumber \\ &={\bf U}^H_x(:,1:r_x){\bf A}_x {\bf s}_x + {\bf U}^H_x(:,1:r_x){\bf n}_x \nonumber \\ &=\widetilde{{\bf A}}_x {\bf s}_x + \widetilde{{\bf n}}_x, \nonumber\\ {\bf y}_{r_y}& ={\bf U}^H_y(:,1:r_y){\bf y} \nonumber \\ &={\bf U}^H_y(:,1:r_y){\bf A}_y {\bf s}_y + {\bf U}^H_y(:,1:r_y){\bf n}_y \nonumber \\ &=\widetilde{{\bf A}}_y {\bf s}_y + \widetilde{{\bf n}}_y. \label{eq:model2}\end{aligned}$$ In this model, the new matrices $\widetilde{\AB}_x$ and $\widetilde{\AB}_y$ have full rank because $\AB_x$ and $\AB_y$ are assumed to have full rank. With the PCA preprocessing the GLRT statistic is $$\Lambda(r_x,r_y,s) = M \ln \prod_{i=s+1}^r \left(1 - \hat{k}_i^2(r_x,r_y) \right), \label{equ:Lambda}$$ and the Bartlett-Lawley statistic is $$\begin{aligned} C(r_x,r_y,s) = & - \! 2\left( M-s-\frac{r_x+r_y+1}{2}+\sum_{i=1}^{s} \hat{k}_i^{-2}(r_x,r_y) \right) \nonumber \\ & \times \ln\prod_{i=s+1}^{r}\left(1-\hat{k}^2_i(r_x,r_y)\right) \label{cs2}\end{aligned}$$ for $s=0,\ldots,r-1$ with $r = \min(r_x,r_y)$. The challenge in the reduced-rank version of the hypothesis test is thus to [jointly]{} determine the best ranks $r_x,r_y$ of the PCA steps and the number $d$ of correlated signals. As long as the number of samples $M$ is large compared to the minimum PCA dimension $r = \min(r_x,r_y)$ but $r_x$ and $r_y$ are not too small (which we will explain in the next paragraph), the new test statistic $C(r_x,r_y,d)$ under $H_0: d = s$ is still approximately $\chi^2$-distributed with $2(r_x - d)(r_y - d)$ degrees of freedom. We denote by $r_{\max}$ the largest $r$ for which the $\chi^2$-distribution holds well enough. Of course, requiring $M$ to be large with respect to $r$ is a much more relaxed condition than requiring $M$ to be large with respect to the dimensions $n$ and $m$. This is because $r_x$ and $r_y$ do not have to be chosen greater (unless there are strong noise components) than $d + f_x$ and $d + f_y$, respectively, which are usually much smaller than $n$ and $m$. There is, however, a complication. By applying PCA to $\xB$ and $\yB$, we might eliminate some of the correlated components if the PCA ranks $r_x$ and $r_y$ are not chosen large enough. If this is the case, then the number of correlated components $\tilde{d}$ in the [reduced-rank descriptions]{} $\xB_{r_x}$ and $\yB_{r_y}$ will be [smaller]{} than the number of correlated components $d$ between $\xB$ and $\yB$. As a consequence, $C(r_x,r_y,d)$ will no longer resemble a $\chi^2$-distribution. Instead, $C(r_x,r_y,\tilde{d})$ with $\tilde{d} < d$ will now be approximately $\chi^2$. By choosing $r_x$ and $r_y$ not large enough it thus becomes likely that the null hypothesis “there are $\tilde{d}$ correlated signals” is not rejected, thus deciding for a smaller number $\tilde{d}$ than the true $d$. We are now getting closer to writing down a rule for jointly selecting $r_x$, $r_y$, and $d$. In order to motivate this rule, we summarize the preceding discussion: Provided the PCA ranks $r_x$ and $r_y$ are chosen sufficiently large to capture all correlated components while $r$ is still small compared to $M$, i.e., $r \leq r_{\max}$, the statistic $C(r_x,r_y,d)$ in (\[cs2\]) is approximately $\chi^2$ [(again irrespective of the noise covariance matrix)]{}. This means that in a series of binary tests of $H_0 : d = s$ vs. $H_1 : d > s$ (testing all values of $s$ starting from $0$ until $H_0$ is not rejected or the maximum $s = r_{\max}$ is reached) $d$ would generally not be [overestimated]{}. It is likely, however, to be [underestimated]{}, if $r_x$ and $r_y$ are not chosen large enough. If $r_x$ and $r_y$ are too small, then the reduced-rank PCA descriptions do not capture all of the correlated components and thus the series of binary tests would decide for too small a $d$. This reasoning motivates the following decision rule. [Detector 1 (“max-min detector”):]{} [Choose $${\hat d} = \!\!\! \underset{ \{r_x,r_y\}=1,\ldots,r_{\max} }{\max}~ \underset{s=0,\ldots,r-1}{{\min}}\left\{s:C(r_x,r_y,s) < T(r_x,r_y,s)\right\} \label{equ:maxmin-det}$$ and choose the $r_x$ and $r_y$ that lead to $\hat{d}$ as the PCA ranks.]{} In (\[equ:maxmin-det\]) the ${\min}$-operator chooses the smallest $s$ such that the statistic $C(r_x,r_y,s)$ falls below the threshold $T(r_x,r_y,s)$, which ensures a given probability of false alarm. If there is no such $s$, then it chooses $s = r$. This step is similar to the traditional test, except that $T(r_x,r_y,s)$ depends on $r_x$ and $r_y$. The rule (\[equ:maxmin-det\]) is based on the fact that if $r_x$ and $r_y$ are not chosen optimally, the min-step might return a number smaller than $d$. Hence, the min-step is performed for all $r_x$ and $r_y$ from 1 up to $r_{\max}$, and the maximum result is chosen as $\hat{d}$. Example ------- ![Histogram of the test statistic $C(r,r,3)$ (in blue) and the probability density function of a $\chi^2$-distribution with $2(r-3)^2$ degrees of freedom (in red), for $s = d = 3$ and different PCA ranks $r = r_x = r_y$. Histograms are computed from $10^6$ independent trials. Also shown as vertical lines are the thresholds $T(r,r,3)$ for a probability of false alarm $P_{\rm FA} = 0.01$. The horizontal axis uses a logarithmic scale.[]{data-label="fig:chi2"}](s3.pdf) We will use an example to illustrate both the closeness of the $\chi^2$-approximation and the idea of the max-min detector. We consider a scenario with $m = n = 100$, $d = 3$ correlated signals, $f = f_x = f_y = 2$ stronger interfering signals, and $M = 50$ samples. The noise variance is chosen small compared to the signal variances. For $s = d = 3$ and $r = r_x = r_y$, Fig. \[fig:chi2\] compares histograms of the statistic $C(r,r,3)$ with the probability density function of a $\chi^2$-distribution with $2(r_x - d)(r_y - d) = 2(r - 3)^2$ degrees of freedom. As long as $r$ is large enough to capture all correlated components (which is the case for $r \geq d + f = 5$ since the independent signals are stronger than the correlated signals) but small compared to $M$, the statistic $C(r,r,3)$ is very well approximated by the $\chi^2$-distribution. This can be seen in subplots (b) $r = 5$ and (c) $r = 15$ (where we start to notice some divergence between the statistic and its approximation). Subplot (d) shows $r = 25$, which is not small enough with respect to $M = 50$. Here the $\chi^2$-distribution is no longer a good approximation of the test statistic. On the other hand, if $r < 5$ then the PCA step eliminates some correlated components. This can be observed in subplot (a) for $r = 4$, where the histogram of $C(4,4,3)$ does not approximate a $\chi^2$-distribution. Because the PCA steps with $r_x = r_y = 4$ keep the two stronger independent signals and only two of the three weaker correlated signals, the reduced-rank PCA descriptions $\xB_{r_x}$ and $\yB_{r_y}$ only have $\tilde{d} = 2$ correlated signals rather than $d = 3$. It can be observed in Fig. \[fig:notchi2\] (b) that $C(4,4,2)$ indeed well approximates a $\chi^2$-distribution with $2(r - \tilde{d})^2 = 2(4 - 2)^2 = 8$ degrees of freedom. ![Histogram of the test statistic $C(r,r,2)$ (in blue) and the probability density function of a $\chi^2$-distribution with $2(r-2)^2$ degrees of freedom (in red), for $d = 3$ but $s = 2$ and different PCA ranks $r = r_x = r_y$. The vertical lines are the thresholds $T(r,r,2)$ for a probability of false alarm $P_{\rm FA} = 0.01$. The horizontal axis uses a logarithmic scale.[]{data-label="fig:notchi2"}](s2.pdf) So let us look at how the max-min detector would proceed in this example. To illustrate this, we again consider Figs. \[fig:chi2\] and \[fig:notchi2\], which compare histograms of $C(r,r,s)$ with $\chi^2$-distributions with $2(r - s)^2$ degrees of freedom for $s = d = 3$ (Fig. \[fig:chi2\]) and $s = 2$ (Fig. \[fig:notchi2\]). Also shown in these figures are the thresholds $T(r,r,s)$ for a probability of false alarm $P_{\rm FA} = 0.01$. According to (\[equ:maxmin-det\]), for given $r_x$ and $r_y$, the detector needs to find the minimum $s$ (between $0$ and $r$) such that the statistic $C$ falls below the threshold $T$. Consider first $r_x = r_y = 4$, which is too small because the PCA steps eliminate one correlated component. From Fig. \[fig:notchi2\] (b), we see that it is likely that $C(4,4,2)$ falls below $T(4,4,2)$, which means that for $r_x = r_y = 4$, the min-step of the detector would likely return too small a number of correlated signals ($s = 2$).[^4] Now consider $r_x = r_y = 5$, which is large enough so that the PCA steps capture all correlated components. It can now be observed in Fig. \[fig:notchi2\] (c) that for $s = 2$, $C(5,5,2)$ will likely not fall below $T(5,5,2)$.[^5] On the other hand, Fig. \[fig:chi2\] (b) shows that it is likely that $C(5,5,3)$ falls below $T(5,5,3)$, hence returning $s = 3$ in the min-step of the detector. Finally, consider $r_x = r_y = 15$, which is larger than needed to capture all correlated components. If $r_x$ and $r_y$ are too large then it becomes increasingly difficult, as can be observed in Fig. \[fig\_cck\], to distinguish between the sample correlation coefficients that are associated with the correlated signals and those that are not. The min-step of the detector would still not generally overestimate $d$ (because the $\chi^2$-approximation remains valid under $H_0$) but it might [underestimate]{} it. This becomes clear from looking at Fig. \[fig:notchi2\] (d), which shows that there is a rather high chance that the min-step would select $s = 2$. However, an underestimating min-step is not a problem for the max-min detector because it selects the maximum of all min-step results. Order selection based on information-theoretic criterion {#sec:mdl} ======================================================== A disadvantage of the hypothesis testing approach to order selection is the requirement of selecting a probability of false alarm $P_{\rm FA}$. Setting $P_{\rm FA}$ too high will lead to a detector that tends to overfit, setting it too low will generally underfit. Achieving the best performance thus requires the right trade-off. In this section, we present two alternative approaches that do not require the manual selection of a threshold and are based on the minimum description length (MDL)-IC. The first approach will remain a hypothesis test but with automatic $P_{\rm FA}$-selection exploiting a link between the GLRT and the IC for model-order selection. The second approach will be a max-min detector based directly on the MDL-IC. Setting the threshold based on the MDL-IC ----------------------------------------- The MDL-IC for selecting the number of correlated signals in two data sets (without PCA steps) is [@StoicaSPT0496] $$\begin{aligned} I_{\rm MDL}(n,m,s)&= -\ell_{\rm max}(\XB,\YB\|\Omega_{s}) +\frac{1}{2}\ln(M)N_\Omega(n,m,s)\nonumber\\ &= M\log\prod_{i=1}^{s}\left(1-\hat{k}_i^{2}\right)+\ln(M)s(m+n-s).\label{eq:itc1}\end{aligned}$$ In this expression, the second term is the penalty term that depends on the degrees of freedom of the model[^6] and thus penalizes overly complex models. The model order chosen is the value of $s$ for which $I_{\rm MDL}(n,m,s)$ is minimized. The reduced-rank version of (\[eq:itc1\]), which accounts for the PCA steps, is $$\begin{aligned} I_{\rm MDL}&(r_x,r_y,s) \nonumber \\ % & = -\ell_{\rm max}(\XB_{r_x},\YB_{r_y}\|\Omega_{s}) + \frac{1}{2}\ln(M)N_\Omega(r_x,r_y,s) \nonumber\\ & = M\ln\prod_{i=1}^{s}\left(1-\hat{k}_i^{2}(r_x,r_y)\right) +\ln(M)s(r_x+r_y-s).\label{eq:itc2}\end{aligned}$$ As has been noted in [@StoicaSPL1110], there is the following connection between the MDL-IC and the log-likelihood ratio of the reduced-rank GLRT $H_0 : d = s$ vs. $H_1 : d > s$: $$\begin{aligned} I_{\rm MDL}&(r_x,r_y,r) - I_{\rm MDL}(r_x,r_y,s) \nonumber \\ & = \Lambda(r_x,r_y,s) + \ln(M)N_\Lambda(r_x,r_y,s) \label{equ:equivalence}\end{aligned}$$ with $N_\Lambda(r_x,r_y,s) = (r_x - s)(r_y - s)$. When choosing between model orders $s$ and $r$ based on the MDL-IC, we decide for model order $s$ if $I_{\rm MDL}(r_x,r_y,r) > I_{\rm MDL}(r_x,r_y,s)$. Because of (\[equ:equivalence\]) we can implement this decision rule also based on the GLRT. We decide for model order $s$ rather than a model order greater than $s$ if $$\Lambda(r_x,r_y,s) > \underbrace{-\ln(M)(r_x - s)(r_y - s)}_{\displaystyle T_{\rm MDL}(r_x,r_y,s)}. \label{equ:TMDL}$$ The term on the right-hand side of this inequality is thus the threshold for the GLRT, which is determined based on the MDL-IC. Note that it is unnecessary to apply the Bartlett-Lawley correction because this would amount to multiplying both sides of the inequality (\[equ:TMDL\]) with the same factor. Thus, we obtain the following max-min decision rule in terms of $\Lambda(r_x,r_y,s)$ rather than $C(r_x,r_y,s)$.\ [Detector 2 (max-min detector with threshold set by MDL-IC):]{} [Choose $$\begin{aligned} {\hat d} = \underset{ \{r_x,r_y\}=1,\ldots,r_{\max} }{\max}~ \underset{s=0,\ldots,r-1}{{\min}} \{s & :\Lambda(r_x,r_y,s) \nonumber \\ & > T_{\rm MDL}(r_x,r_y,s)\},\end{aligned}$$ where $\Lambda(r_x,r_y,s)$ is given in (\[equ:Lambda\]) and $T_{\rm MDL}(r_x,r_y,s)$ is given in (\[equ:TMDL\]), and choose the $r_x$ and $r_y$ that lead to $\hat{d}$ as the PCA ranks.]{} Min-MDL detector {#sec_minMDL} ---------------- Another approach that does not require the selection of $P_{\rm FA}$ applies the max-min idea directly to the MDL-IC. Let us first write down the decision rule and interpret it afterwards.\ [Detector 3 (“max-min MDL-IC detector”):]{} [Choose $$\hat{d} = \underset{ \{r_x,r_y\}=1,\ldots,r_{\max} }{\max}~ \underset{s=0,\ldots,r-1}{{\rm argmin}} I_{\rm MDL}(r_x,r_y,s) \label{equ:maxmin-MDL}$$ and choose the $r_x$ and $r_y$ that lead to $\hat{d}$ as the PCA ranks.]{} In order to understand this detector, we note, based on the discussion in the preceding subsection, that $$\begin{aligned} \underset{s=0,\ldots,r-1}{{\rm argmin}} & I_{\rm MDL}(r_x,r_y,s) \nonumber \\ & = \underset{s=0,\ldots,r-1}{{\rm argmax}} [- I_{\rm MDL}(r_x,r_y,s)] \nonumber \\ & = \underset{s=0,\ldots,r-1}{{\rm argmax}} [I_{\rm MDL}(r_x,r_y,r) - I_{\rm MDL}(r_x,r_y,s)] \nonumber \\ & = \underset{s=0,\ldots,r-1}{{\rm argmax}} [\Lambda(r_x,r_y,s) - T_{\rm MDL}(r_x,r_y,s)].\end{aligned}$$ The min-step in Detector 3 thus chooses the value of $s$ that [maximizes]{} the difference between the GLRT statistic and the MDL test threshold. This is different than the min-step in Detector 2, which picks the [smallest]{} $s$ for which the test statistic exceeds the threshold. Therefore, Detector 3 will never pick a $\hat{d}$ smaller, but possibly larger, than Detector 2. Performance evaluation {#sec:sim} ====================== In this section, we compare the performance of our three model-order selection schemes among each other and with competing approaches. In the absence of a competing systematic approach to the [joint]{} model-order selection in PCA-CCA, we determined the PCA ranks $r_x$ and $r_y$ separately from the number of correlated signals $d$. We used the sample eigenvalue-based (SEV) technique [@NadakuditiSPT0708] for selecting $r_x$ and $r_y$ because it is one of the few techniques that can handle the sample-poor case for a single channel. For the selection of $d$ we used the canonical correlation test (CCT) [@ChenIET-RSN1096] with $P_{\rm FA} = 0.005$, the Akaike information criterion (AIC) [@StoicaSPT0496], and the MDL criterion [@StoicaSPT0496]. Figures \[fig\_sim1\]–\[fig\_sim4\] show the probability of selecting the correct $d$ for different setups. In the first setup, shown in Figs. \[fig\_sim1\]–\[fig\_sim3\], we consider a system with $d = 2$ correlated signals (each with variance $5$ and correlation coefficients $0.8$ and $0.7$), and $f_x = 3$ and $f_y = 4$ independent signals (each with variance $1.5$). The matrices $\AB_x$ and $\AB_y$ are randomly generated unitary matrices. For each data point, we ran 1000 independent Monte Carlo trials. ![Performance of our Detectors 1, 2, 3 and competing approaches for [white]{} noise. System dimensions are $m = n = 40$.[]{data-label="fig_sim1"}](PD_vs_sample_white) ![Same setup as in Fig. \[fig\_sim1\] but with [colored]{} MA noise. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_sim2"}](PD_vs_sample_colored) We first consider a system with fixed dimension $m = n = 40$. In Fig. \[fig\_sim1\], we show the performance as a function of the number of samples $M$ when the noise is white and each noise component has unit variance. We see that the performance of Detector 1 depends on the choice of $P_{\rm FA}$: For smaller $M$, $P_{\rm FA}$ should be chosen larger, whereas for larger $M$, a smaller $P_{\rm FA}$ performs better. Detector 2 does this trade-off automatically and performs very well even for very small sample sizes. All other approaches (including Detector 3) still perform quite well but require larger sample support. The picture completely changes when we have colored rather than white noise. We now generate the noise from a spatially varying moving average (MA) process of order $3$ with coefficients $[\frac{1}{\sqrt{3}}, \frac{1}{\sqrt{3}}, \frac{1}{\sqrt{3}}]$. Before the spatial averaging, the noise components have variance $1/3$. It can be seen in Fig. \[fig\_sim2\] that methods that select $r_x$ and $r_y$ separately from $d$ completely fail. This is because a single-channel technique such as SEV cannot distinguish between signal and noise eigenvalues if the noise is colored. The performance of our detectors, on the other hand, is actually improved particularly for very small sample sizes. ![Same setup as in Fig. \[fig\_sim1\] but with varying dimensions $m = n$ and fixed sample size $M = 100$. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_sim3"}](PD_vs_dimension) In Fig. \[fig\_sim3\], we reconsider the white noise case but with varying dimensions $m = n$ and fixed sample size $M = 100$. When the noise is independent in space and time, increasing the ratio of the data dimensions $m$, $n$ to the number of samples $M$ shrinks the signal-subspace [@NadakuditiSPT0708], which worsens the detection performance. We note, however, that the decrease in performance affects the SEV + X techniques much more than our detectors. Indeed, Detector 2 again shows a very reliable performance even for large dimensions. The main reason behind this effect is that the SEV technique is designed to keep all the signal components (i.e., correlated [and]{} independent components) whereas our detectors aim to eliminate weaker independent components in the PCA step. The presence of independent components deteriorates the detection performance of the subsequent CCA step. ![Effect of the independent signals’ variance on performance. Settings: $d = 7$ correlated signals with variance 10 and correlation coefficients ($0.92$, $0.9$, $0.88$, $0.85$, $0.83$, $0.8$, $0.75$), $f_x = f_y = 2$ independent signals of varying variance, $m =n = 80$, $M = 150$, colored AR(1) noise with coefficient $0.65$. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_new_scen_1"}](Fig8.pdf) ![Same setting as in Fig. \[fig\_new\_scen\_1\], except that now there are $f_x = f_y = 4$ independent signals of varying variance. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_new_scen_2"}](Fig9.pdf) So far, we have looked at a case where the correlated signals are [stronger]{} than the independent signals. We now investigate what happens when the correlated signals are [weaker]{} than some or all of the independent signals. First consider a scenario with 2 independent signals of varying variance and 7 correlated signals of variance 10. Figure \[fig\_new\_scen\_1\] shows the probability of detection as a function of the independent signals’ variance. We see that the variance has only little effect on the performance of all techniques. Now we increase the number of independent signals to 4, leaving all other settings unchanged. The most dramatic effect that can be observed in Fig. \[fig\_new\_scen\_2\] is the failure of Detector 2 once the independent signals reach a variance close to the correlated signals’ variance. This may be explained as follows. Detector 2 sets its threshold based on MDL, which generally does not overestimate the number of correlated signals, but may [underestimate]{} it if the sample size is not sufficiently large compared to the system dimension (i.e., the PCA rank). In the case shown in Fig. \[fig\_new\_scen\_2\], there are 7 correlated signals and 4 independent signals. Once the independent signals become as strong as or stronger than the correlated signals, this leads to an optimum PCA rank of 11. As the number of samples $M = 150$ is not significantly larger than 11, MDL starts to underestimate the model order. This affects Detector 2 more severely than Detector 3 because Detector 3 will always return a model as large as, but possibly larger than, Detector 2 (see the discussion in Section \[sec\_minMDL\]). ![Same setting as in Fig. \[fig\_new\_scen\_1\], except that there are $d = 5$ correlated signals with variance 8, and $f_x = f_y = 7$ independent signals, two of which have variance 12 and 5 of which have variance 3. Performance as a function of number of samples $M$. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_new_scen_3"}](Fig10.pdf) This explanation can be validated by investigating the effect of the number of samples in a scenario where there are strong independent signals. We now consider a case with $d =5$ correlated signals with variance 8, and $f_x = f_y = 7$ independent signals, two of which have variance 12 and 5 of which have variance 3. In Fig. \[fig\_new\_scen\_3\], we look at the performance as a function of the number of samples $M$. It can be observed that, among our three detectors, Detector 2 needs the largest number of samples for satisfactory performance, followed by Detector 3. The lesson that can be learned here is that in the presence of strong independent signals, Detector 1 should be preferred if only a very small number of samples are available. ![Performance as a function of the mean correlation coefficient $\rho$. Settings: $m = n = 100$, $M = 180$, $d = 5$ correlated signals with correlation coefficients drawn from a uniform distribution between $[\rho - 0.05, \rho + 0.05]$, $f_x = f_y = 2$ stronger independent signals, AR(1) noise with coefficient $0.65$. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_new_scen_4"}](Fig11.pdf) Let us now investigate the effect that the value of the correlation coefficients among the correlated signals have. Here we consider a scenario with $d = 5$ correlated signals with variance 8, and $f_x = f_y = 2$ stronger independent signals of variance 10. In Fig. \[fig\_new\_scen\_4\], we plot the performance as function of $\rho$. The correlation coefficients for the 5 correlated signals are drawn from a uniform distribution between $[\rho - 0.05, \rho + 0.05]$. As expected, stronger correlation leads to better performance. Since the independent signals are stronger than the correlated signals, Detector 1 outperforms Detector 3, which in turn outperforms Detector 2. All of our detectors outperform the competition. ![Array processing toy example to illustrate the effect of ill-conditioned mixing matrices. For the meaning of the colored markers, please refer to the legend of Fig. \[fig\_sim1\].[]{data-label="fig_sim4"}](PD_vs_AngularSpacing) In our last example we examine an array processing toy application to see what happens if the mixing matrices $\AB_x$ and $\AB_y$ become ill-conditioned. We consider two spatially separated uniform linear arrays (ULAs) with 40 sensors (i.e., $m = n = 40$) and inter-sensor spacing of $\lambda/2$, which take $M = 60$ samples. There are $5$ fixed point-sources in the far-field emitting narrow-band Gaussian signals at wavelength $\lambda$, which impinge upon ULA 1 at angles $[\theta_{x,1},\theta_{x,2},\ldots,\theta_{x,5}] = [20^\circ, 20^\circ + \delta, ..., 20^\circ + 4\delta]$. Similarly, $6$ such signals impinge upon ULA 2 at angles $[\theta_{y,1},\theta_{y,2},\ldots,\theta_{y,6}] = [50^\circ, 50^\circ + \delta, ..., 50^\circ + 5\delta]$. Two of these signals are correlated between ULAs 1 and 2 (i.e., $d = 2$, $f_x = 3$, $f_y = 4$) with correlation coefficients $0.8$ and $0.7$. The correlated signals each have variance $5$ and the independent signals each have variance $1.5$. The noise is colored and generated as in the setup for Fig. \[fig\_sim2\]. With these assumptions, the $i$th column of $\AB_x$ is $[1, e^{j\frac{\pi}{2} \sin\theta_{x,i}}, ... , e^{j\frac{\pi}{2} (n-1) \sin\theta_{x,i}}]^T$, $i = 1, \ldots, 5$, and the $i$th column of $\AB_y$ is $[1, e^{j\frac{\pi}{2} \sin\theta_{y,i}},\ldots, e^{j\frac{\pi}{2} (m-1)\sin\theta_{y,i}}]^T, i=1,\ldots,6$. As the angular spacing $\delta$ decreases, the mixing matrices ${\bf A}_x$ and ${\bf A}_y$ become more ill-conditioned. Figure \[fig\_sim4\] shows the performance of all detectors for angular spacing $\delta$ ranging from $1^\circ$ to $10^\circ$. We can see that due to the presence of colored noise, all SEV + X methods fail irrespective of $\delta$. Our detectors, on the other hand, are able to provide very good detection rates from $\delta=4^\circ$ onward. Detectors 1 and 3 provide the best results for small $\delta$. Conclusions =========== PCA-CCA is a common approach to the analysis of correlation between two data sets when there is only small sample support. In the past, selecting the ranks of the PCA steps and identifying the number of correlated signals was often done by ad-hoc rules or based on experience. In this paper, we have presented a systematic approach to the joint order selection of PCA ranks and number of correlated signals, based on a GLRT and information-theoretic criteria. Simulation results have shown that the techniques perform very well for extremely sample-poor scenarios in particular in the presence of colored noise. Of course, it is important to remember that there is no free lunch. While we do not need many samples compared to the dimensions of the data sets, the techniques do require the number of samples to be sufficiently greater than the sum of the numbers of correlated signals and stronger independent signals (i.e., variance larger than the correlated signals). Effect of PCA on estimated canonical correlations {#app:interlace} ================================================= \[lemma\_interlace\] The estimated canonical correlation coefficients increase with increasing PCA ranks $r_x$ and $r_y$: $\hat{k}_i(\tilde{r}_1,\tilde{r}_2) \geq \hat{k}_i(r_x,r_y) ,i=1,\ldots,\min(r_x,r_y)$, for $1\leq r_x < \tilde{r}_1$ and $1\leq r_y < \tilde{r}_2$. Define the following matrices: $$\begin{aligned} {\bf G} &=& {\bf V}^H_x(:,1:\tilde{r}_1) {\bf V}_y(:,1:\tilde{r}_2) = \left[\begin{array}{c} {\bf G}_1 \\ {\bf G}_2\end{array}\right], \nonumber\\ {\bf G}_{1} &=& {\bf V}^H_x(:,1:r_x) {\bf V}_y(:,1:\tilde{r}_2) = \left[\begin{array}{cc} {\bf G}_{1,1} & {\bf G}_{1,2}\end{array}\right], \nonumber\\ {\bf G}_{2} &=& {\bf V}^H_x(:,r_x+1:\tilde{r}_1) {\bf V}_y(:,1:\tilde{r}_2), \nonumber\\ {\bf G}_{1,1} &=& {\bf V}^H_x(:,1:r_x) {\bf V}_y(:,1:r_y), \nonumber\\ {\bf G}_{1,2} &=& {\bf V}^H_x(:,1:r_x) {\bf V}_y(:,r_y+1:\tilde{r}_2). \nonumber \end{aligned}$$ According to the Cauchy interlacing theorem, we have $$\lambda_i\left({\bf GG}^H\right)= \lambda_i\left(\left[ \begin{array}{cc} {\bf G}_{1}{\bf G}_{1}^H & {\bf G}_{1}{\bf G}_{2}^H \\ {\bf G}_{2}{\bf G}_{1}^H & {\bf G}_{2}{\bf G}_{2}^H \end{array}\right]\right)\geq \lambda_i\left({\bf G}_{1}{\bf G}_{1}^H\right)$$ for $i=1,\ldots,r_x$, where $\lambda_i(\cdot)$ represents the $i$th largest eigenvalue. Furthermore, as a result of the Weyl inequality, we also have $\lambda_i\left({\bf G}_{1}{\bf G}_{1}^H\right)=\lambda_i\left({\bf G}_{1,1}{\bf G}_{1,1}^H+{\bf G}_{1,2}{\bf G}_{1,2}^H\right)\geq \lambda_i\left({\bf G}_{1,1}{\bf G}_{1,1}^H\right)$. Together with first inequality, this yields $\lambda_i\left({\bf GG}^H\right) \geq \lambda_i\left({\bf G}_{1,1}{\bf G}_{1,1}^H\right)$. As ${\bf V}_x$ and ${\bf V}_y$ represent the matrices of right-singular vectors of $\XB$ and $\YB$, respectively, it follows that the squared sample canonical correlation coefficient $\hat{k}_i^2(\tilde{r}_1,\tilde{r}_2) = \lambda_i\left({\bf GG}^H\right)$ is greater than or equal to the squared sample canonical correlation coefficient $\hat{k}_i^2(r_x,r_y) = \lambda_i\left({\bf G}_{1,1}{\bf G}_{1,1}^H\right)$. [^1]: Y. Song, P. J. Schreier, and T. Hasija are with the Signal and System Theory Group, University of Paderborn, 33098 Paderborn, Germany (e-mail: {yang.song, peter.schreier, tanuj.hasija}@sst.upb.de). D. Ramírez is with the Dpto. de Teoría de la Señal y Comunicaciones, Universidad Carlos III de Madrid, 28911 Leganés, Spain (e-mail: [email protected]). [^2]: This research was supported by the German Research Foundation (DFG) under grant SCHR 1384/3-1, and the Alfried Krupp von Bohlen und Halbach foundation under its program “Return of German scientists from abroad.” [^3]: In this difference, only the degrees of freedom associated with the cross-covariance matrix matter. [^4]: If we plotted the test statistics and thresholds also for $s = 0$ and $s = 1$ we would see that it is unlikely that a value $s < 2$ would be chosen. [^5]: As before, if we plotted the test statistics also for $s = 0$ and $s = 1$, we would see that it is even less likely that $C(5,5,s)$ falls below $T(5,5,s)$ if $s < 2$. [^6]: As before, only the degrees of freedom associated with the cross-covariance matrix are considered because the degrees of freedom associated with the auto-covariance matrices do not depend on $s$. Hence, they do not matter in the following optimization problems.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We report low-temperature magnetisation measurements on a large number of purified ropes of single wall carbon nanotubes. In spite of a large superparamagnetic contribution due to the small ferromagnetic catalytical particles still present in the sample, at low temperature ($T < 0.5K$) and low magnetic field ($H < 80 Oe$), a diamagnetic signal is detectable. This low temperature diamagnetism can be interpreted as the Meissner effect in ropes of carbon nanotubes which have previously been shown to exhibit superconductivity from transport measurements [@kociak01].' author: - 'M. Ferrier' - 'F. Ladieu' - 'M. Ocio' - 'B. Sacépé' - 'T. Vaugien' - 'V. Pichot' - 'P. Launois' - 'H. Bouchiat' title: Superconducting diamagnetic fluctuations in ropes of carbon nanotubes --- During the last decade, electronic properties of carbon nanotubes have attracted a lot of interest [@dressel]. Depending on their diameter and helicity single wall carbon nanotubes (SWNT) are either semiconducting or metallic with in the latter case only 2 conducting channels at the Fermi energy. These long molecular wires constitute a model system for the investigation of one dimensional (1D) electronic transport and its great sensitivity to electron-electron interactions. The usual Fermi liquid is expected to be unstable in 1D, with the formation of a correlated Luttinger Liquid state [@TLL]. In SWNT, interactions contain a strong repulsive Coulomb long range contribution, giving rise to a characteristic power law depression of the low energy tunneling density of states whose exponent is related to the interaction strength. Accordingly, non linear conductance measured on SWNT mounted on tunnel contacts has been interpreted as the signature of the existence of a Luttinger liquid state with repulsive interactions [@bocracth]. In this context the discovery of superconductivity below 0.5 K in transport measurements on SWNT ropes containing between 30 to 100 parallel 1.2 nm diameter tubes in good contact with normal electrodes [@kociak01; @kasu03] was a surprise. This observation, which implies the existence of attractive interactions overcoming repulsive ones, has stimulated a number of theoretical investigations [@gonzalez; @sedeki; @demartino02; @demartino03; @ferrier04] of possible, in most cases phonon mediated, attractive mechanisms. In particular the famous acoustic breathing modes of SWNT [@dressel] have been shown to provide a phonon mediated attractive electron- electron coupling in SWNT when they are assembled into ropes. Repulsive Coulomb interactions are indeed strongly screened by the nearest neighbor metallic tubes, and there is also the possibility of an important inter-tube Josephson coupling [@gonzalez; @demartino03; @ferrier04]. A good agreement between theory and experiment was achieved concerning the broadening of the transition observed on the resistance versus temperature data when decreasing the number of tubes constituting the rope. Beside transport, it is also essential to investigate thermodynamic signatures such as the Meissner effect of this unusual superconductivity. The geometry of carbon nanotubes, which are very narrow 1D cylinders, is a priori not favorable for efficient magnetic flux expulsion. Nevertheless, magnetisation measurements performed on very small (0.4 nm ) diameter SWNT, revealed a diamagnetic contribution increasing at low temperature below 6K [@tang] which was interpreted as superconducting fluctuations. However the geometry of the samples grown in zeolite matrices was not adequate for transport measurements, which did not show clear sign of superconductivity. The aim of the present work is to address this question of flux expulsion (i.e. Meissner effect) in ropes of SWNT similar to those previously studied in transport measurements[@kociak01; @kasu03]. ![Field dependence of the total magnetisation of the sample at various temperatures below 3.5K. Lower inset: field dependence of $\Delta M(T,H)= M(T,H)-M(3.5, H)$ for different temperatures showing that the hysteretic component of the magnetisation is independent of temperature below 3.5K. (This remains true up to T= 35K.) Upper inset: transmission electron microscopy image of the SWNT ropes constituting the sample.[]{data-label="hyst"}](Fig1.eps){width="8cm"} The samples consist of SWNT ropes (containing typically 10 to 100 parallel tubes whose diameter varies between 0.9 to 1.4 nm, the intertube distance being of the order of 0.3nm (see fig.\[hyst\]). They were fabricated by decomposition of Fe(CO)$_5$ according to the HiPCO process [@hipco]. Removal of the Fe catalyst was achieved through annealing under Oxygen at 300 C and subsequent treatment with hydrochloric acid. After these operations the amount of nanoparticles of Fe both in reduced or oxidized forms is estimated to be of the order of 1$\%$. These ropes were assembled into long ribbons in a PVA solution using the method described in [@poulin]. The spreading of alignment of the nanotubes within a single ribbon was determined from Xray diffraction measurements according to the method described in [@RX] and found to be $ \pm 35$ . For the purpose of this experiment these ribbons were wrapped around two 7 by 7mm high purety saphyre substrates, which were subsequently annealed 3 hours at 320 C in order to eliminate the PVA surrounding the ropes. The mass of SWNT ropes deposited in this procedure, was measured to be 5mg and 6mg on the two saphyre substrate. The spreading of alignment of the nanotubes after wiring on the saphyre is probably not better than $ \pm 45$ . We will see that this nonetheless permits a determination of the anisotropy of the magnetisation of the nanotubes. Magnetisation measurements were performed on the 2 samples in a home built extraction SQUID magnetometer [@ocio] equipped with a dilution refrigerator insert whose base temperature is 70mK. The thermalisation of the sample was done with a small bundle of high purity copper wires attached to the saphyre plates inserted into a tube made from a capton foil. The contribution of the empty sample holder, capton, saphyre and copper bundle was shown to be below 3 $10^{-8}$emu/Oe in the whole range of temperature (70mK, 300K) and magnetic field (3 to 3000 0e) investigated. The magnetic signal contains a large ferromagnetic like hysteretic contribution. One sees on Fig.\[hyst\] that this hysteresis does not change with temperature from 70mK up to 3.5K . This remains true up to 35K (not shown). At any temperature below 35K, we thus decompose the magnetic signal into a T independent contribution attributed to frozen large magnetic particles and into a temperature dependent part which is reversible (i.e non hysteretic). In the following we mainly focus on this reversible low temperature dependent part of the magnetisation of the samples. We will show that beside the contribution of unfrozen superparamagnetic particles, it is possible to identify the presence of a diamagnetic contribution growing below 0.4 K which will be interpreted as the Meissner effect of the carbon nanotubes. ![ (A)Renormalised magnetisation $M(T,H) - M(35K,H) /H_{loc}(H)$ as a function of temperature between 70mK and 35K for H=20 Oe and H=150 Oe compared to the fit with 2 assemblies of superparamagnetic particles whose adjusted volumic distributions are shown in the inset. Note the deviations at low temperature for the data taken at 20 Oe.(B)Temperature dependence of the field cooled renormalised magnetisation: $\Delta \chi(T,H)=\Delta M(T,H)/H_{loc}(H)$ for various values of magnetic field. Inset: Field dependence of the local field obtained by renormalisation of the temperature dependent data above 1K up to 3.5K (circles) and up to 30K(squares)[]{data-label="mdeT"}](Fig2.eps){width="8cm"} The magnetisation as function of temperature is plotted in fig.2-A. $M(T,H)$ increases monotonously as the temperature is lowered below 35K down to 70mK. This temperature dependent signal corresponds approximatively to 50% of the total magnetisation of the sample. The data was systematically taken in the field cooled state [@FC]. It is possible to superimpose [all]{} the curves taken at various values of magnetic field below 150 Oe and above 1K by renormalisation of the data with a field dependent factor $H_{local}(H)$ which corresponds to the effective average magnetic field inside the sample. As shown in fig.2-B $H_{local}(H)$ deviates from the applied magnetic field at low field values which suggests that the actual field inside the sample has a contribution coming from the largest particles saturating at low fields (typically lower than 30 Oe) which adds to the applied field. We have checked that the determination of $H_{local}(H)$ does [not]{} depend on the temperature range chosen between 1 and 35K to renormalise the curves. More over in this temperature range, as shown on the inset of fig.2A, the quantity $\ M(T,H_{local}(H))$ can be very well described within a simple model of a volumic distribution of superparamagnetic $Fe_2O_3$ oxyde particles as inspired from previous work in ref. [@sappey] where the anisotropy energy of $Fe_2O_3$ is taken to be K = 5.4 $10^4 J/m^3$. This distribution shown in the inset of fig.2-A, has a broad peak characteristic of a lognormal distribution around a particle volume which corresponds to typical diameters of the order of 1.5nm, most of these particles are frozen below 3-10K. To account of the very low T upturn of M(T) we have added to the previous contribution a Brillouin law with J=8. This corresponds to very small $Fe_2O_3$ aggregates containing typically 3 or 4 coupled iron spins. The magnetic moments of these particles (whose anisotropy energy is estimated to be of the order of 20mK) are unfrozen down to very low temperature. Finally, the amplitudes of these 2 superparamagnetic signals added to the ferromagnetic part (independent of temperature) corresponds to a number of typically 3 $10^{17}$ Fe atoms for a 5 mg sample i.e 1% of the sample. This is what is expected after purification, in agreement with previous magnetisation measurements made on similar samples [@mcatal] above 4K. On can see that the magnetisation signal at 150 Oe is very well described by this simple model between 60K and 70mK [@hightemp]. On the other hand important low T deviations from this model are observed for the data measured at 20 Oe, increasing below 0.5K. ![Temperature dependence of the renormalised magnetisation after substraction of the data at $H=150 Oe$: $\Delta \chi _{dia}(T,H) = \Delta \chi(T,H) - \Delta \chi(T,150). $Inset: Reconstitution, using eq.4, of the field dependence of the diamagnetic contribution for various temperatures.[]{data-label="mdia"}](Fig3.eps){width="8cm"} If we now focus on this low temperature data below 0.5 K, we can see on fig.2-B, that the temperature dependence of $\Delta \chi(T,H) = \Delta M(T,H) /H_{local}(H)$ with $\Delta M(T,H)=(M(T,H) -M(T=3.5K))$ varies substantially with the magnetic field. Note that the magnetisation of an assembly of [unfrozen]{} super-paramagnetic particles of typical moment $\mu$ is indeed expected to exhibit at low temperature a non linear field dependence starting in the field range where $\mu H \sim k_BT$. This consideration can explain the data above 150 Oe quite well, where a decrease of the T dependence of $\Delta \chi$ with magnetic field is indeed observed (see the curve at 200 0e in fig.2). On the other hand the low field data is surprising. Below 1K $\Delta \chi$ is observed (see fig.2) to [strongly decrease ]{} with decreasing magnetic field between 100 and 5 Oe. Such a behavior [cannot]{} correspond to any superparamagnetic system even in the presence of interactions between the particles leading to a spin-glass like state [@SG]. This effect, is emphasized in fig.\[mdia\] showing the quantity $\Delta \chi_{dia}(T,H)$ after performing the subtraction: $$\Delta \chi_{dia}(T,H) = \Delta \chi(T,H) - \Delta \chi(T,150)$$ This plot strongly suggests the existence of a temperature dependent diamagnetic signal increasing rather smoothly at temperatures below 0.4K, for low magnetic field, superimposed to the superparamagnetic contribution of the catalyst particles. By subtracting data at 150 Oe instead of the superparamagnetic contribution inferred from the fit of fig.2-A, we probably underestimate this diamagnetic signal, on the other hand this method is model independent. In the following we show that this diamagnetic contribution can be interpreted as screening diamagnetic supercurrents due to superconducting fluctuations in the ropes of carbon nanotubes. These ropes can be approximated by superconducting cylinders with diameter D small compared to the London penetration length $\lambda(T)$ (see for example the recent work on Pb superconducting nano-cylinders [@supnanopb]). In the linear response regime,(observed in present case up to 10 Oe) the diamagnetic susceptibility (per unit volume) of a dense assembly of superconducting cylinders can be simply expressed as a function of $D/\lambda$ as: $$M_{\\|,\bot}(H) = -H (D/\lambda_{\\|,\bot}(T))^2$$ when the field is respectively parallel, perpendicular to the axis of the nanotubes within the rope. $\lambda_{\\|}(T)$and $\lambda_{\bot}(T)$ are respectively the penetration length parallel and perpendicular to the cylinders. Considering the anisotropy of the electronic structure of a single tube, as well as the existence of an inter-tube Josephson coupling, it is reasonable to assume that $\lambda_{\\|}(T)$ and $\lambda_{\bot}(T)$ may be different from one another. In the absence of a more precise theoretical prediction, we use the simple London expression: $\lambda_{\\| , \bot}^{-2}(T)\simeq 2 \mu_0 n_S(T) e^2/m_{\bot ,\\| }$ where $ n_S (T)$ is the density of superconducting pairs in the ropes and $m_{ \bot, \\|}$ are respectively the effective masses characterizing transport parallel and perpendicular to the ropes. Since only 1/3 of the tubes within a rope are expected to be metallic, we assume that at zero temperature $n_s(0) = n_e/6$, where $n_e = \frac{4}{ a \pi d^2/4}$ is the density of electrons in the 2 conduction bands of a metallic SWNT of diameter $d$ and $a$ is the graphene hexagonal lattice size. As a result, we get a simple expression relating the London diamagnetic susceptibility to the geometry of the carbon nanotubes ropes assumed to take random orientations through the sample (anisotropy effects will be discussed later). $$M_{av}^{dia}(H) = - H\frac {8 e^2}{3 m a \pi}\frac{D^2 }{d^2} \label{chidia}$$ (m is approximated here by the free electron mass). From our measurements we can deduce the amplitude of the diamagnetic susceptibility in the linear regime to be $\chi_{meas}^{dia} = 3\pm 0.05$ $10^{-6}$ emu/Oe $(A m ^2/T)$, reasonably close to the value given by expression (\[chidia\]) (multiplied by the volume of the sample), $\chi_{cal}^{dia} = 2$ $10^{-5} emu/Oe$ for an average diameter $D= 10 nm$ per rope as estimated from transmission microscopy. This yields $\lambda(0) \simeq 0.6\mu m $ indeed much longer than the average diameter of the ropes. The temperature below which diamagnetism shows up ($~ T^* =0.4K$) is also very close to the values of transition temperature measured in transport measurements [@kociak01]. From the set of curves in fig.\[mdia\] it is possible to reconstruct the field dependence of this diamagnetic component of the magnetisation for various temperatures, as shown in the inset: $$M_{dia}(H) =\Delta M(H) - H_{local}(H)\Delta \chi(150)$$ $M_{dia}(H)$ measured at 70mK varies linearly at low field and goes through a maximum around 40 Oe, it decreases at higher field and becomes undetectable above 80 Oe. This behavior observed on both samples is typical of the magnetisation curves of a type 2 superconductor whose critical field $H_{c1}$ is of the order of $60 \pm 20$ Oe. Note that this diamagnetic contribution at the lowest investigated values of magnetic field and temperature constitute more than 2/3 of the temperature dependent magnetic signal below 3K and 1/3 of the total magnetic contribution. We have also compared the amplitude of the magnetisation when the magnetic field is parallel to the directions of wrapping of the fibers compared to the situation where it is perpendicular. In both cases the field was maintained in the plane of the saphyre plate in order to avoid any macroscopic demagnetising factor. The paramagnetic component of the sample is found to be isotropic within an accuracy of 1%. On the other hand it was possible to detect unambiguously on both samples an [ anisotropy]{} of the Meissner component of the signal of the order of 10%$\pm 1$% indicating a larger amplitude of the Meissner component along the axis of the tubes than perpendicularly by a factor 1.2$\pm 0.05$ taking an angular distribution of the tubes axis of $\pm$45 wide, (note that this number only slightly depends on the estimated spreading angle when its smaller than 60). This result is however surprising since the Josephson coupling energy $t_j$ between the tubes is expected to be much smaller (at least by a facter 100) than the hopping integral $t$ along the tubes axis [@gonzalez]. Nevertheless it is possible that the effective mass which characterize the intertube hopping of the Cooper pairs $m_{\bot}\propto 1/d^2 t_j$, is of the order and even smaller than along the tube axis $m_{\\|}\propto 1/a^2 t$ since $d/a \simeq 10$. All these findings corroborate the superconductivity observed in transport measurements on individual ropes of carbon nanotubes. Note that the diamagnetic part of the magnetisation disappears at relatively low magnetic field, more than a factor 10 lower than the critical field of the order of 1T observed in transport experiments. This can be understood considering that the weak Josephson energy coupling between the tubes within one rope is destroyed at much weaker magnetic field than superconductivity within a single tube giving rise to a diamagnetic signal per unit volume of the order of $d^2/ \lambda ^2$ (undetectable in this experiment). In all this analysis we have not taken into account the contribution of the orbital magnetism of the tubes in the normal state. Such orbital currents have been predicted and measured in multiwall nanotubes [@magMWNT]. They are related to the sensitivity of electronic eigenstates and energies to the phase of the boundary conditions modified by the magnetic field. The resulting magnetic orbital susceptibility is diamagnetic for semiconducting tubes and paramagnetic for metallic tubes [@liu] and can be estimated to be at most $ \chi_{orb}= 2 $ $10 ^{-8} emu Oe^{-1}$ for our sample, which is negligible compared to the diamagnetic contribution we have discussed. Our findings also present qualitative similarities with the results obtained by Tang et al. [@tang] on 0.4 nm diameter tubes, except that temperature and field scales are much smaller in our experiment. This could simply be related to the diameter of the tubes which are in the present case 3 times larger than in the Tang experiment. The characteristic energy of the phonon breathing mode, expected to be at the origin of superconductivity [@demartino02; @demartino03] is known to scale as the inverse tube diameter. Moreover the possibility of the existence of a Wentzel Bardeen instability [@martin] specific to electron phonon coupling in a 1D system is predicted in very small diameter tubes, [@demartino02] favoring superconducting fluctuations against the formation of a Luttinger liquid with repulsive interactions, even in the situation where they are not well screened. Finally we note that the Kondo effect, leading to a dynamical screening of the magnetic moments at low temperature, could also give rise to a low temperature decrease of the magnetic susceptibility as observed in noble metals with a low concentration of magnetic impurities. To our knowledge there is no estimate of the order of magnitude of the Kondo temperature of iron magnetic impurities in carbon nanotubes. Since the electronic density of states is very small in carbon nanotubes compared to standard metals we expect this Kondo temperature to be also extremely small. Even if such a Kondo physics scenario cannot completely be ruled out to explain the depression of the Curie law at low field and low temperature it could not provide a description of the field dependence of the data shown in fig.\[mdia\]. In conclusion, we have shown that magnetisation of ropes of SWNT, at low temperature and magnetic field, cannot be explained by the superparamagnetism of the catalytical particles present in the sample. The weaker relative increase of magnetisation at small magnetic field compared to higher one, provides a strong indication of the existence of a diamagnetic Meissner contribution in ropes of SWNT increasing below 0.4 K, in agreement with the onset of superconductivity observed in transport measurements realised on similar samples. From the alignment of the tubes in the samples, although imperfect, it was also possible to probe the anisotropy of the Meissner effect. We thank S.Gueron, R.Deblock, M.Kociak, P.Poulin, S.Senoussi, L.Lepape, E.Vincent, J.Hamman and P.Bonville for fruitful discussions and help. [99]{} M. Kociak, A. Kasumov, S. Gu[é]{}ron, B. Reulet, I. I. Khodos, Y. B. Gorbatov, V. T. Volkov, L. Vaccarini, and H. Bouchiat, Phys. Rev. Lett. [86]{}(11), 2416 (2001). R. Egger and A. O. Gogolin, Phys. Rev. Lett. [79]{}(25), 5082 (1997). M. Bockrath, D. Cobden, J. Liu, A. Rinzler, R. Smalley, L. Balents, and P. McEuen, Nature [397]{}, 598 (1999). A. Kasumov, M. Kociak, M. Ferrier, R. Deblock, S. Gu[é]{}ron, B. Reulet, I. Khodos, O. St[é]{}phan, and H. Bouchiat, Phys. Rev. B [68]{}, 214521 (2003). M. Dresselhaus, G. Dresselhaus, and P. Eklund, [Science of Fullerenes and Carbon Nanotubes]{}, Academic Press, San Diego, CA, 1996. J. Gonzalez, Phys. Rev. Lett. [88]{}(7), 76403 (2002).J. Gonzalez, Phys. Rev. B [67]{}, 014528 (2003). A. S[é]{}d[é]{}ki, L. G. Caron, and C. Bourbonnais, Phys. Rev. B [65]{}, 140515 (2002). A. D. Martino and R. Egger, Phys. Rev. B [67]{}, 235418 (2003). A. De Martino and R. Egger, Phys. Rev. B [70]{}, 014508 (2004). M. Ferrier, A. de Martino, A. Kasumov, S. Gu[é]{}ron, M. Kociak, R. Egger, and H. Bouchiat, Solid. Stat. comm.[131]{}, 615 (2004). Z. Tang, L. Zhang, N. Wang, X. Zhang, G. Wen, G. Li, J. Wang, C. Chan, and P. Sheng, Science [292]{}, 2462–2465 (2001). M. J. Bronikowski, P. A. Willis, D. T. Colbert, K. A. Smith, and R. E. Smalley, J. Vac. Sci. Technol. A [19]{},4 (2001). B. Vigolo, A. P[é]{}nicaud, C. Coulon, C. Sauder, R. Pailler, C. Journet, P. Bernier, and P. Poulin, Science [290]{}, 1331 (2000). S. Badaire, V. Pichot, C. Zakri, P. Poulin, P. Launois, J. Vavro, C. Guthy, M. Chen and J. Fischer, J. Appl. Phys. [96]{}, 7509-7513 (2004) R.Sappey, E.Vincent, M. Ocio, J.Hamman, F.Chaput, J.-P.Boilot, D. Zins, Europhysics Letters, [37]{}, 639, (1997). A small albeit systematic difference of the order of few $\%$ of $\Delta M(H,T)$ at 100mK could be noticed between the field cooled and zero field cooled data. We chose to focus on the FC data since it yields the under bound of $\vert M_{dia} \vert$. R. Sappey, E. Vincent, N. Hadacek, F. Chaput, J. P. Boilot, and D. Zins, Phys. Rev. [B 56]{}, 14551 (1997). Deviations to the fit observed at temperatures above 100K can be attributed to the largest particles whose magnetisation freeze out in this temperature range contributing to the local field and not taken into account by the 2 distributions fit we have used. F. Chen et al. Appl. Phys. Lett., [83]{}, 4602, (2003). In spin glasses at any given T, M(H)/H is a decreasing function of H. X.Y.Zhang and J.Y. Dai, Nanotechnology [15]{}, 1166(2004) X. K. Wang, R. Chang, A. Patashinski, and J. Ketterson, J. Mater. Res. [9]{}(6), 1578–1582 (1994). J. P. Lu, Phys. Rev. Lett. [74]{}(7), 1123–1126 (1995). D. Loss and T. Martin, Phys. Rev. B [50]{}, 12160 (1994).	{ "pile_set_name": "ArXiv" }
--- abstract: \| We calculate the critical value of the hopping parameter, $\kappa_c$, in Lattice QCD, up to two loops in perturbation theory. We employ the Sheikholeslami-Wohlert (clover) improved action for fermions and the Symanzik improved gluon action with 4- and 6-link loops. The quantity which we study is a typical case of a vacuum expectation value resulting in an additive renormalization; as such, it is characterized by a power (linear) divergence in the lattice spacing, and its calculation lies at the limits of applicability of perturbation theory. Our results are polynomial in $c_{\rm SW}$ (clover parameter) and cover a wide range of values for the Symanzik coefficients $c_i$. The dependence on the number of colors $N$ and the number of fermion flavors $N_f$ is shown explicitly. In order to compare our results to non perturbative evaluations of $\kappa_c$ coming from Monte Carlo simulations, we employ an improved perturbation theory method for improved actions. Lattice QCD, Lattice perturbation theory, Hopping parameter, Improved actions. 11.15.Ha, 12.38.Gc, 11.10.Gh, 12.38.Bx address: \| Department of Physics, University of Cyprus, P.O. Box 20537, Nicosia CY-1678, Cyprus\ [email: ]{}[[email protected], [email protected], [email protected]]{} author: - 'A. Skouroupathis, M. Constantinou and H. Panagopoulos' title: 'Two-loop additive mass renormalization with clover fermions and Symanzik improved gluons' --- Introduction ============ In the present work, we calculate the additive renormalization of the fermion mass in lattice QCD, using clover fermions and Symanzik improved gluons. The calculation is carried out up to two loops in perturbation theory and it is directly related to the determination of the critical value of the hopping parameter, $\kappa_c$. The clover fermion action [@SW] (SW) succesfully reduces lattice discretization effects and approaches the continuum limit faster. This justifies the extensive usage of this action in Monte Carlo simulations in recent years. The coefficient $c_{\rm SW}$ appearing in this action is a free parameter for the current work and our results will be given as a polynomial in $c_{\rm SW}$. Regarding gluon fields, we employ the Symanzik improved action [@Sym], which also aims at minimizing finite lattice spacing effects. For the coefficients parameterizing the Symanzik action, we consider several choices of values which are frequently used in the literature. The lattice discretization of fermions introduces some well known difficulties; demanding strict locality and absence of doublers leads to breaking of chiral symmetry. In order to recover this symmetry in the continuum limit one must set the renormalized fermion mass ($m_R$) equal to zero. To achieve this, the mass parameter $m_\circ$ appearing in the Lagrangian must approach a critical value $m_c$, which is nonzero due to additive renormalization. The mass parameter $m_\circ$ is directly related to the hopping parameter $\kappa$ used in simulations. Its critical value, $\kappa_c$, corresponds to chiral symmetry restoration: $$\kappa_c={1\over 2\,m_c\,a + 8\,r} \label{kappa1}$$ where $a$ is the lattice spacing and $r$ is the Wilson parameter. Using Eq.(\[kappa1\]), the non-renormalized fermion mass is given by: $$m_B\equiv m_{\rm o}-m_c=\frac{1}{2\,a}\left(\frac{1}{\kappa} -\frac{1}{\kappa_c}\right) \label{mB}$$ Thus, in order to restore chiral symmetry one must consider the limit $m_{\rm o} \to m_c$. This fact points to the necessity of an evaluation of $m_c$. The perturbative value of $m_c$ is also a necessary ingredient in higher-loop calculations of the multiplicative renormalization of operators (see, e.g., Ref. [@SP]). In mass independent schemes, such renormalizations are typically defined and calculated at zero renormalized mass, and this entails setting the value of the Lagrangian mass equal to $m_c$. Previous studies of the hopping parameter and its critical value have appeared in the literature for Wilson fermions - Wilson gluons [@FP] and for clover fermions - Wilson gluons [@PP; @CPR]. The procedure and notation in our work is the same as in the above references. Our results for $\kappa_c$ (and consequently for the critical fermion mass) depend on the number of colors ($N$) and on the number of fermion flavors ($N_f$). Besides that, there is an explicit dependence on the clover parameter $c_{\rm SW}$ which, as mentioned at the beginning, is kept as a free parameter. On the other hand, the dependence of the results on the choice of Symanzik coefficients cannot be given in closed form; instead, we present it in a list of Tables and Figures. The rest of the paper is organized as follows: In Sec. \[sec2\] we formulate the problem, define the discretized actions, and describe our calculation of the necessary Feynman diagrams. Sec. \[sec3\] is a presentation of our results. Finally, in Sec. \[sec4\] we apply to our one- and two-loop results an improvement method, proposed by us [@cactus1; @cactus2; @CPS]. This method resums a certain infinite class of subdiagrams, to all orders in perturbation theory, leading to an improved perturbative expansion. We end this section with a comparison of perturbative and non-perturbative results. Our findings are summarized in Sec. \[sec5\]. Formulation of the problem {#sec2} ========================== We begin with the Wilson formulation of the QCD action on the lattice, with $N_f$ flavors of degenerate clover (SW) [@SW] fermions. In standard notation, it reads: $$\begin{aligned} S_L =&& S_G + \sum_{f}\sum_{x} (4r+m_{\rm o})\bar{\psi}_{f}(x)\psi_f(x) \nonumber \\ &&- {1\over 2}\sum_{f}\sum_{x,\,\mu} \bigg{[}\bar{\psi}_{f}(x) \left( r - \gamma_\mu\right) U_{x,\, x+\mu}\psi_f(x+{\mu}) +\bar{\psi}_f(x+{\mu})\left( r + \gamma_\mu\right)U_{x+\mu,\,x}\psi_{f}(x)\bigg{]}\nonumber \\ &&+ {i\over 4}\,c_{\rm SW}\,\sum_{f}\sum_{x,\,\mu,\,\nu} \bar{\psi}_{f}(x) \sigma_{\mu\nu} {\hat F}_{\mu\nu}(x) \psi_f(x), \label{latact}\end{aligned}$$ $$\begin{aligned} {\rm where:}\qquad {\hat F}_{\mu\nu} &\equiv& {1\over{8}}\, (Q_{\mu\nu} - Q_{\nu\mu})\\ {\rm and:\qquad} Q_{\mu\nu} &=& U_{x,\, x+\mu}U_{x+\mu,\, x+\mu+\nu}U_{x+\mu+\nu,\, x+\nu}U_{x+\nu,\, x}\nonumber \\ &+& U_{ x,\, x+ \nu}U_{ x+ \nu,\, x+ \nu- \mu}U_{ x+ \nu- \mu,\, x- \mu}U_{ x- \mu,\, x} \nonumber \\ &+& U_{ x,\, x- \mu}U_{ x- \mu,\, x- \mu- \nu}U_{ x- \mu- \nu,\, x- \nu}U_{ x- \nu,\, x}\nonumber \\ &+& U_{ x,\, x- \nu}U_{ x- \nu,\, x- \nu+ \mu}U_{ x- \nu+ \mu,\, x+ \mu}U_{ x+ \mu,\, x} \label{latact2}\end{aligned}$$ The clover coefficient $c_{\rm SW}$ is treated here as a free parameter. Particular choices of values for $c_{\rm SW}$ have been determined both perturbatively [@SW] and non-perturbatively [@Luscher1996], so as to minimize ${\cal O}(a)$ effects. The Wilson parameter $r$ is set to $r=1$ henceforth; $f$ is a flavor index; $\sigma_{\mu\nu} =(i/2) [\gamma_\mu,\,\gamma_\nu]$. Powers of the lattice spacing $a$ have been omitted and may be directly reinserted by dimensional counting. Regarding gluons, we use the Symanzik improved gauge field action, involving Wilson loops with 4 and 6 links[^1]: $$\begin{aligned} S_G&=&\frac{2}{g^2} \Bigl[ c_0 \sum_{\rm plaquette} {\rm Re\, Tr\,}\{1-U_{\rm plaquette}\} + c_1 \sum_{\rm rectangle} {\rm Re \, Tr\,}\{1- U_{\rm rectangle}\} \nonumber \\ &&\qquad +c_2 \sum_{\rm chair} {\rm Re\, Tr\,}\{1- U_{\rm chair}\}+c_3 \sum_{\rm parallelogram} {\rm Re \,Tr\,}\{1- U_{\rm parallelogram}\}\Bigr] \label{gluonaction}\end{aligned}$$ ($g$ is the bare coupling constant). The lowest order expansion of this action, leading to the gluon propagator, is $$S_{\rm G}^{(0)} = \frac{1}{2}\int_{-\pi/a}^{\pi/a} \frac{d^4k}{(2\pi)^4} \sum_{\mu\nu} A_\mu^a(k)\left[G_{\mu\nu}(k)-\frac{\xi}{\xi-1}\hat{k}_\mu\hat{k}_\nu\right] A_\nu^a(-k)\, \label{GluonProp}$$ where $\xi$ is the gauge fixing parameter (see Eq.(\[GaugeFixing\])) and: $$\begin{aligned} G_{\mu\nu}(k) &=& \hat{k}_\mu\hat{k}_\nu + \sum_\rho \left( \hat{k}_\rho^2 \delta_{\mu\nu} - \hat{k}_\mu\hat{k}_\rho \delta_{\rho\nu} \right) \, d_{\mu\rho} , \qquad \hat{k}_\mu = \frac{2}{a}\sin\frac{ak_\mu}{2}\,, \quad \hat{k}^2 = \sum_\mu \hat{k}_\mu^2 \nonumber\\ d_{\mu\nu}&=&\left(1-\delta_{\mu\nu}\right) \left[C_0 - C_1 \, a^2 \hat{k}^2 - C_2 \, a^2( \hat{k}_\mu^2 + \hat{k}_\nu^2) \right]\end{aligned}$$ The coefficients $C_i$ are related to $c_i$ by C\_0 = c\_0 + 8 c\_1 + 16 c\_2 + 8 c\_3 , C\_1 = c\_2 + c\_3, C\_2 = c\_1 - c\_2 - c\_3 \[ContLimit\] The Symanzik coefficients must satisfy: $c_0 + 8 c_1 + 16 c_2 + 8 c_3 = 1$, in order to reach the correct classical continuum limit. Aside from this requirement, the values of $c_i$ can be chosen arbitrarily; they are normally tuned in a way as to ensure ${\cal O}(a)$ improvement. As always in perturbation theory, we must introduce an appropriate gauge-fixing term to the action; in terms of the gauge field $Q_\mu(x)$ $\left[U_{x,\,x+\mu}= \exp(i\,g\,Q_\mu(x))\right]$, it reads: $$S_{gf} = {1\over {1{-}\xi}}\,\sum_{x,\mu , \nu} \hbox{Tr} \, \bigl\{ \Delta^-_{\mu} Q_{\mu}(x) \Delta^-_{\nu} Q_{\nu}(x)\bigr\} , \qquad \Delta^-_{\mu} Q_{\nu}(x) \equiv Q_{\nu}(x - {\hat \mu}) - Q_{\nu}(x). \label{GaugeFixing}$$ Having to compute a gauge invariant quantity, we can, for convenience, choose to work either in the Feynman gauge ($\xi=0$) or in the Landau gauge ($\xi=1$). Covariant gauge fixing produces the following action for the ghost fields $\omega$ and $\overline\omega$ $$\begin{aligned} S_{gh} &=& 2 \sum_{x} \sum_{\mu} \hbox{Tr} \, \Bigl\{(\Delta^+_{\mu}\omega(x))^{\dagger} \Bigl( \Delta^+_{\mu}\omega(x) + i g \left[Q_{\mu}(x), \omega(x)\right] + \frac{1}{2} i g \left[Q_{\mu}(x), \Delta^+_{\mu}\omega(x) \right] \nonumber\\ & & \quad - \frac{1}{12} g^2 \left[Q_{\mu}(x), \left[ Q_{\mu}(x), \Delta^+_{\mu}\omega(x)\right]\right] + \cdots \Bigr)\Bigr\}, \qquad \Delta^+_{\mu}\omega(x) \equiv \omega(x + {\hat \mu}) - \omega(x). \label{GhostAction}\end{aligned}$$ Finally, the change of integration variables from links to vector fields yields a Jacobian that can be rewritten as the usual measure term $S_m$ in the action: $$S_{m} = \frac{1}{12} N g^2 \sum_{x} \sum_{\mu} \hbox{Tr} \, \{Q_{\mu}(x) Q_{\mu}(x)\} + \cdots$$ In $S_{gh}$ and $S_m$ we have written out only terms relevant to our computation. The full action is: $S = S_L + S_{gf} + S_{gh} + S_m.$ The bare fermion mass $m_B$ must be set to zero for chiral invariance in the classical continuum limit. Terms proportional to $r$ in the action, as well as the clover terms, break chiral invariance. They vanish in the classical continuum limit; at the quantum level, they induce nonvanishing, flavor-independent fermion mass corrections. Numerical simulation algorithms usually employ the hopping parameter, $$\kappa\equiv{1\over 2\,m_{\rm o}\,a + 8\,r} \label{kappaDef}$$ as an adjustable input. Its critical value, at which chiral symmetry is restored, is thus $1/8r$ classically, but gets shifted by quantum effects. The renormalized mass can be calculated in textbook fashion from the fermion self–energy. Denoting by $\Sigma^L(p,m_{\rm o},g)$ the truncated, one particle irreducible fermion two-point function, we have for the fermion propagator: $$\begin{aligned} S(p)&=& \left[ i \,\pcircslash + m(p)- \Sigma^L(p,m_{\rm o},g)\right]^{-1}\\ {\rm where:}\qquad \pcircslash &=& {1\over a}\,\sum_\mu\gamma_\mu \,\sin(ap^\mu), \quad m(p) = m_{\rm o} + {2r\over a} \sum_\mu \sin^2(ap^\mu/2)\nonumber\end{aligned}$$ To restore the explicit breaking of chiral invariance, we require that the renormalized mass vanish: $$S^{-1}(0)\Big\|_{\displaystyle m_{\rm o} \to m_c} = 0 \qquad\Longrightarrow\qquad m_c = \Sigma^L(0,m_c\,,g)$$ The above is a recursive equation for $m_c$, which can be solved order by order in perturbation theory. We denote by $dm$ the additive mass renormalization of $m_\circ$: $m_B = m_\circ - dm$. In order to obtain a zero renormalized mass, we must require $m_B \to 0$, and thus $m_\circ \to dm$. Consequently, $$m_c = dm= dm_{\rm (1-loop)}+dm_{\rm (2-loop)} \label{Totaldm}$$ At tree level, $m_c=0$. Two diagrams contribute to $dm_{\rm (1-loop)}$, shown in Fig. 1. In these diagrams, the fermion mass must be set to its tree level value, $m_{\rm o}\to 0$. 4.0cm1.0cm [ One-loop diagrams contributing to $dm_{\rm (1-loop)}$. Wavy (solid) lines represent gluons (fermions).]{} The quantity $dm_{\rm (2-loop)}$ receives contributions from a total of 26 diagrams, shown in Fig. 2. Genuine two-loop diagrams must again be evaluated at $m_{\rm o}\to 0$; in addition, one must include to this order the one-loop diagram containing an ${\cal O}(g^2)$ mass counterterm (diagram 23). Certain sets of diagrams, corresponding to one-loop renormalization of propagators, must be evaluated together in order to obtain an infrared convergent result: These are diagrams 7+8+9+10+11, 12+13, 14+15+16+17+18, 19+20, 21+22+23. 2.0cm1.0cm [ Two-loop diagrams contributing to $dm_{\rm (2-loop)}$. Wavy (solid, dotted) lines represent gluons (fermions, ghosts). Crosses denote vertices stemming from the measure part of the action; a solid circle is a fermion mass counterterm.]{} Computation and Results {#sec3} ======================= Given that the dependence of $m_c$ on the Symanzik coefficients $c_i$ cannot be expressed in closed form, we chose certain sets of values for $c_i$ , presented in Table I, which are in common use [@LWactions; @Iwasaki; @Symanzik; @LWactions2; @Alford; @Takaishi]: Plaquette, Symanzik (tree level improved), Tadpole Improved Lüscher-Weisz (TILW), Iwasaki and DBW2. Actually, since the gluon propagator contains only the combinations $C_1$ and $C_2$ (Eq.(\[ContLimit\])), all results for $m_c$ can be recast in terms of $C_1$, $C_2$ and one additional parameter, say, $c_2$; in this case the dependence on $c_2$ (at fixed $C_1$, $C_2$) is polynomial of second degree. The contribution $dm_l$ of the $l^{{\rm th}}$ one-loop diagram to $dm$, can be expressed as: $$dm_l=\frac{(N^2-1)}{N}\,g^2 \cdot \sum_{i=0}^{2}c_{{\rm SW}}^i\,\varepsilon^{(i)}_{l} \label{1loopContribution}$$ where $\varepsilon^{(i)}_{l}$ are numerical one-loop integrals whose values depend on $C_1$, $C_2$. The dependence on $c_{{\rm SW}}$ is seen to be polynomial of degree 2 ($i=0,\,1,\,2$). The contribution to $dm$ from two-loop diagrams that do not contain closed fermion loops, can be written in the form $$dm_l=\frac{(N^2-1)}{N^2}\,g^4 \cdot \sum_{i,j,k} c_{{\rm SW}}^i\,N^j\,c_2^k\,e^{(i,j,k)}_{l} \label{2loopContribution1}$$ where the index $l$ runs over all contributing diagrams, $j=0,2$ and $k=0,\,1,\,2$ (since up to two vertices from the gluon action may be present in a Feynman diagram). The dependence on $c_{{\rm SW}}$ is now polynomial of degree 4 ($i=0,\cdots,\,4$). The coefficients $e^{(i,j,k)}_{l}$ (as well as $\tilde{e}^{(i)}_{l}$ of Eq.(\[2loopContribution2\]) below) are two-loop numerical integrals; once again, they depend on $C_1$, $C_2$. Finally, the contribution to $dm$ from two-loop diagrams containing a closed fermion loop, can be expressed as $$dm_l=\frac{(N^2-1)}{N}\,N_f\,g^4 \cdot \sum_{i=0}^{4}c^{i}_{\rm SW}\,\tilde{e}^{(i)}_{l} \label{2loopContribution2}$$ where the index $l$ runs over diagrams 12-13, 19-20. Summing up the contributions of all diagrams, $dm$ assumes the form $$\begin{aligned} dm=\sum_{l}dm_l&=&\frac{(N^2-1)}{N}\,g^2 \cdot \sum_{i}c_{{\rm SW}}^i\,\varepsilon^{(i)} +\frac{(N^2-1)}{N^2}\,g^4 \cdot \sum_{i,j,k}c_{{\rm SW}}^i\,N^j\,c_2^k\,e^{(i,j,k)} \nonumber \\ &&+\frac{(N^2-1)}{N}\,N_f\,g^4 \cdot \sum_i c^{i}_{\rm SW}\,\tilde{e}^{(i)} \label{TotalContribution}\end{aligned}$$ In the above, $\varepsilon^{(i)}$, $e^{(i,j,k)}$, $\tilde{e}^{(i)}$ are the sums over all contributing diagrams of the quantities: $\varepsilon^{(i)}_{l}$, $e^{(i,j,k)}_{l}$, $\tilde{e}^{(i)}_{l}$, respectively (cf. Eqs.(\[1loopContribution\],\[2loopContribution1\],\[2loopContribution2\])). The coefficients $\varepsilon^{(i)}$ lead to the total contribution of one-loop diagrams. Their values are listed in Table \[tab2\], for the ten sets of $c_i$ values shown in Table \[tab1\]. Similarly, results for the coefficients $e^{(i,j,k)}$ and $\tilde{e}^{(i)}$ corresponding to the total contribution of two-loop diagrams, are presented in Tables \[tab3\]-\[tab7\]. In order to enable cross-checks and comparisons, numerical per-diagram values of the constants $\varepsilon^{(i)}_{l}$, $e^{(i,j,k)}_{l}$ and $\tilde{e}^{(i)}_{l}$ are presented in Tables VIII-XII, for the case of the Iwasaki action. For economy of space, several vanishing contributions to these constants have simply been omitted. A similar breakdown for other actions can be obtained from the authors upon request. The total contribution of one-loop diagrams, for $N=3$ can be written as a function of the clover parameter $c_{{\rm SW}}$. In the case of the Plaquette, Iwasaki, and DBW2 actions, we find, respectively: $$\begin{aligned} dm_{\rm (1-loop)}^{\rm Plaquette}&=&g^2\,\Big(-0.434285489(1)+0.1159547570(3)\,c_{{\rm SW}}+0.0482553833(1)\,c_{{\rm SW}}^2\Big) \label{dm1loopPlaquette}\\ dm_{\rm (1-loop)}^{\rm Iwasaki}&=&g^2\,\Big(-0.2201449497(1)+0.0761203698(3)\,c_{{\rm SW}}+0.0262264231(1)\,c_{{\rm SW}}^2\Big) \label{dm1loopIwasaki}\\ dm_{\rm (1-loop)}^{\rm DBW2}&=&g^2\,\Big(-0.0972070995(5) + 0.0421775310(1)\,c_{{\rm SW}} + 0.01141359801(1)\,c_{{\rm SW}}^2\Big) \label{dm1loopDBW2}\end{aligned}$$ A similar process can be followed for two-loop diagrams. In this case, we set $N=3$, $c_2=0$ and we use three different values for the flavor number: $N_f=0,\,2,\,3$. Thus, for the Plaquette, Iwasaki and DBW2 actions, the total contribution is, respectively: $$\begin{aligned} N_f=0:&& \quad dm_{\rm (2-loop)}^{\rm Plaquette}=g^4\,\Big(-0.1255626(2)+0.0203001(2)\,c_{{\rm SW}}+0.00108420(7)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.6cm} -\,0.00116538(2)\,c_{{\rm SW}}^3-0.0000996725(1)\,c_{{\rm SW}}^4\Big) \\ N_f=2:&& \quad dm_{\rm (2-loop)}^{\rm Plaquette}=g^4\,\Big(-0.1192361(2)+0.0173870(2)\,c_{{\rm SW}}+0.00836498(8)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.6cm} -\,0.00485727(3)\,c_{{\rm SW}}^3-0.0011561947(4)\,c_{{\rm SW}}^4\Big) \\ N_f=3:&& \quad dm_{\rm (2-loop)}^{\rm Plaquette}=g^4\,\Big(-0.1160729(2)+0.0159305(2)\,c_{{\rm SW}}+0.0120054(1)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.6cm} -\,0.00670321(3)\,c_{{\rm SW}}^3-0.0016844558(6)\,c_{{\rm SW}}^4\Big) \label{dm2loopPlaquette}\end{aligned}$$ $$\begin{aligned} N_f=0:&& \quad dm_{\rm (2-loop)}^{\rm Iwasaki}=g^4\,\Big(-0.0099523(2)-0.0024304(5)\,c_{{\rm SW}}-0.00232855(4)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.00032100(2)\,c_{{\rm SW}}^3-0.0000419365(1)\,c_{{\rm SW}}^4\Big)\\ N_f=2:&& \quad dm_{\rm (2-loop)}^{\rm Iwasaki}=g^4\,\Big(-0.0076299(2)-0.0040731(5)\,c_{{\rm SW}}+0.00102758(6)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.00242924(3)\,c_{{\rm SW}}^3-0.000457690(2)\,c_{{\rm SW}}^4\Big)\\ N_f=3:&& \quad dm_{\rm (2-loop)}^{\rm Iwasaki}=g^4\,\Big(-0.0064687(2)-0.0048944(5)\,c_{{\rm SW}}+0.00270565(7)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.00348335(3)\,c_{{\rm SW}}^3-0.000665567(2)\,c_{{\rm SW}}^4\Big) \label{dm2loopIwasaki}\end{aligned}$$ $$\begin{aligned} N_f=0:&& \quad dm_{\rm (2-loop)}^{\rm DBW2}=g^4\,\Big(+0.005099(2)-0.0053903(7)\,c_{{\rm SW}}-0.0011157(1)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.00004482(2)\,c_{{\rm SW}}^3-0.0000111470(2)\,c_{{\rm SW}}^4\Big)\\ N_f=2:&& \quad dm_{\rm (2-loop)}^{\rm DBW2}=g^4\,\Big(+0.005944(2) - 0.0061840(7)\,c_{{\rm SW}}+0.0002046(2)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.0010177(3)\,c_{{\rm SW}}^3 -0.000125065(3)\,c_{{\rm SW}}^4\Big)\\ N_f=3:&& \quad dm_{\rm (2-loop)}^{\rm DBW2}=g^4\,\Big(+0.006366(2)-0.0065809(7)\,c_{{\rm SW}}+0.0008648(2)\,c_{{\rm SW}}^2 \nonumber \\ && \hspace{3.5cm} -\,0.0015042(4)\,c_{{\rm SW}}^3 -0.000182023(5)\,c_{{\rm SW}}^4\Big) \label{dm2loopDBW2}\end{aligned}$$ In Figs. 3, 4, and 5 we present the values of $dm_{\rm (2-loop)}$ for $N_f=0,\,2,\,3$, respectively; the results are shown for all choices of Symanzik actions which we have considered, as a function of $c_{\rm SW}\,(N=3,\,c_2=0)$. In all cases, the dependence on $c_{\rm SW}$ is rather mild. One observes that $dm_{\rm (2-loop)}$ is significantly smaller for all improved actions, as compared to the plaquette action; in particular, in the case of DBW2, $dm_{\rm (2-loop)}$ is closest to zero and it vanishes exactly around $c_{\rm SW}=1$. Another feature of these results is that they change only slightly with $N_f$, especially in the range $c_{\rm SW} < 1.5$. This is due to the small contributions of diagrams with closed fermion loops (diagrams 12, 13, 19, 20). By the same token, in the case of nondegenerate flavors, $dm_{\rm (2-loop)}$ is expected to depend only weakly on the mass of the virtual fermion. [ Total contribution of two-loop diagrams, for $N=3$, $N_f=0$ and $c_2=0$. Legends appear in the same top-to-bottom order as the corresponding lines.]{} -1mm [ Total contribution of two-loop diagrams, for $N=3$, $N_f=2$ and $c_2=0$. Legends appear in the same top-to-bottom order as the corresponding lines. ]{} [ Total contribution of two-loop diagrams, for $N=3$, $N_f=3$ and $c_2=0$. Legends appear in the same top-to-bottom order as the corresponding lines.]{} Improved Perturbation Theory {#sec4} ============================ We now apply our method of improving perturbation theory [@cactus1; @cactus2; @CPS], based on resummation of an infinite subset of tadpole diagrams, termed “cactus” diagrams. In Ref. [@CPS] we show how this procedure can be applied to any action of the type we are considering here, and it provides a simple, gauge invariant way of dressing, to all orders, perturbative results at any given order (such as the one- and two-loop results of the present calculation). Some alternative ways of improving perturbation theory have been proposed in Refs. [@Parisi-81; @L-M-93]. In a nutshell, our procedure involves replacing the original values of the Symanzik and clover coefficients by improved values, which are explicitly computed in [@CPS]. Applying at first this method to one-loop diagrams, the improved (“dressed”) value $dm^{\rm dr}$ of the critical mass $(N=3,\,c_2=0)$ can be written as: $$dm^{\rm dr}_{\rm (1-loop)}=\sum_{i=0}^{2}\varepsilon^{(i)}_{dr}\,c_{{\rm SW}}^i \label{dm1loopDressed}$$ In comparing with $\varepsilon^{(i)}$ of Eq. (\[TotalContribution\]), the quantity $\varepsilon^{(i)}_{dr}$ is the result of one-loop Feynman diagrams with dressed values for the Symanzik parameters, and it has already been multiplied by $g^2\left(N^2-1\right)/N$. The dependence of $\varepsilon^{(i)}_{dr}$ on $g$ is quite complicated now, and cannot be given in closed form; instead $\varepsilon^{(i)}_{dr}$ must be computed numerically for particular choices of $g$. Listed in Table \[tab13\] are the results for $\varepsilon^{(i)}_{dr}$ along with the value of $\beta=2N/g^2$ corresponding to each one of the 16 actions used in this calculation. An attractive feature of this improvement procedure is that it can be applied also to higher loop perturbative results, with due care to avoid double counting of the cactus diagrams which were already included at one loop. Ideally, of course, one loop improvement should already be adequate enough, so as to obviate the need to consider higher loops; indeed, we find this to be the case and, consequently, we limit our discussion of two-loop improvement to only the plaquette action $(\beta=5.29,\,N=3,\,N_f=2)$, the Iwasaki action $(\beta=1.95,\,N=3,\,N_f=2)$ and the DBW2 action $(\beta=0.87\ {\rm and}\ \beta=1.04,\,N=3,\,N_f=2)$. Using these values, the contribution to $dm^{\rm dr}_{\rm (2-loop)}$ is a polynomial in $c_{\,{\rm SW}}$: $$\begin{aligned} dm^{\rm dr}_{\rm (2-loop),\,plaquette}&=& -0.77398(8)+0.16330(4)\,c_{{\rm SW}}+ 0.06224534(1)\,c_{{\rm SW}}^2 \nonumber \\ && -0.0044006(9)\,c_{{\rm SW}}^3-0.00073780(6)\,c_{{\rm SW}}^4 \label{dm2loopDressedPlaquette}\\[2ex] dm^{\rm dr}_{\rm (2-loop),\,Iwasaki}&=&-0.0813302(9)+0.043030(3)\,c_{{\rm SW}}+0.0308196(2)\,c_{{\rm SW}}^2 \nonumber \\ && -0.00767090(8)\,c_{{\rm SW}}^3-0.001160923(1)\,c_{{\rm SW}}^4 \label{dm2loopDressedIwasaki}\\[2ex] dm^{\rm dr}_{\rm (2-loop),\,DBW2(\beta=0.87)}&=&-0.044906(1)+0.029449(4)\,c_{{\rm SW}}+0.0239522(2)\,c_{{\rm SW}}^2 \nonumber \\ && -0.0082231(1)\,c_{{\rm SW}}^3 -0.001218955(4)\,c_{{\rm SW}}^4 \label{dm2loopDressedDBW2_0.87}\\[2ex] dm^{\rm dr}_{\rm (2-loop),\,DBW2(\beta=1.04)}&=&-0.031260(1)+0.021793(2)\,c_{{\rm SW}}+0.0188027(2)\,c_{{\rm SW}}^2 \nonumber \\ && -0.00705284(9)\,c_{{\rm SW}}^3 -0.001055657(1)\,c_{{\rm SW}}^4 \label{dm2loopDressedDBW2_1.04}\end{aligned}$$ The comparison between the total dressed contribution $dm^{\rm dr}=dm^{\rm dr}_{\rm (1-loop)}+dm^{\rm dr}_{\rm (2-loop)}$ and the unimproved contribution, $dm$, for the plaquette action is exhibited in Fig. 6, as a function of $c_{\rm SW}$. Similarly, $dm^{\rm dr}$ for the Iwasaki and the DBW2 actions is shown in Fig. 7 and Fig. 8, respectively. Finally, in Table \[tab14\], we present a comparison of dressed and undressed results, for some commonly used values of $\beta$, $N_f$, $c_{\rm SW}$, and we also compare with available non perturbative estimates for $\kappa_c$ [@Luscher1996; @Bowler; @UKQCD; @Sommer; @Khan]. We observe that improved perturbation theory, applied to one-loop results, already leads to a much better agreement with the non perturbative estimates. Discussion {#sec5} ========== To recapitulate, in this paper we have calculated the critical mass $m_c$, and the associated critical hopping parameter $\kappa_c$, up to two loops in perturbation theory, using the clover action for fermions and the Symanzik improved gluon action with 4- and 6-link loops. The perturbative value of $m_c$ is a necessary ingredient in the higher-loop renormalization of operators, in mass independent schemes: Such renormalizations are typically defined and calculated at vanishing renormalized mass, which amounts to setting the Lagrangian mass equal to $m_c$. In our calculations, we have chosen for the Symanzik coefficients $c_i$ a wide range of values, which are most commonly used in numerical simulations. The dependence of our results on the number of colors $N$ and the number of fermion flavors $N_f$ is shown explicitly. The dependence on the clover parameter $c_{\rm SW}$ is in the form of a fourth degree polynomial whose coefficients we compute explicitly; it is expected, of course, that the most relevant values for $c_{\rm SW}$ are those optimized for ${\cal O}(a)$ improvement, either at tree level ($c_{\rm SW}=1$), or at one loop [@SW], or non-perturbatively [@Luscher1996]. Since $m_c$ is gauge invariant, we chose to calculate it in the Feynman gauge. The propagator appearing in Feynman diagrams is the inverse of a nondiagonal matrix; while this inverse can be written down explicitly, it is more convenient, and more efficient in terms of CPU time, to perform the inversion numerically. Integrations over loop momenta were performed as momentum sums on lattices of finite size $L$, where typically $L\lsim 40$; extrapolation to $L\to\infty$ introduces a systematic error, which we estimate quite accurately. Our results for $m_c$ are significantly closer to zero in the case of Symanzik improved actions, as compared to the plaquette action. In particular, the DBW2 action stands out among the rest, in that $m_c$ vanishes exactly for a value of $c_{\rm SW}$ around 1. Thus, improved actions seem to bring us quite near the point of chiral symmetry restoration. The dependence of $m_c$ on the number of flavors is seen to be very mild. This fact would also suggest that, in the case of nondegenerate flavors, $m_c$ should depend only weakly on the mass of the virtual fermion. Finally, we have made some comparisons among perturbative and non-perturbative results for $\kappa_c$. While these are expected to differ for a power divergent additive renormalization, such as the quantity under study, we nevertheless find a reasonable agreement. This agreement is further enhanced upon using an improved perturbative scheme, which entails resumming, to all orders in the coupling constant, a dominant subclass of tadpole diagrams. The method, originally proposed for the Plaquette action (see Ref. [@cactus1]), was extended in Ref. [@CPS] to encompass all possible gluon actions made of closed Wilson loops, and can be applied at any given order in perturbation theory. As would be desirable, one-loop improvement is seen to be already adequate to give a reasonable agreement among perturbative and non-perturbative values. Indeed, our results for $\kappa_{\rm 1-loop}^{\rm dr}$ are significally closer to the non-perturbative evaluations, as shown in Table \[tab14\]; in fact, the two-loop dressing procedure introduces no further improvement to the comparison. [Acknowledgments: ]{} This work is supported in part by the Research Promotion Foundation of Cyprus (Proposal Nr: $\rm ENTA\Xi$/0504/11, $\rm ENI\Sigma X$/0505/45). ------------------------ ---- ---------- ---- --------- ---- ---------- Plaquette 1 .0 0 0 Symanzik 1 .6666667 -0 .083333 0 TILW, $\beta c_0=8.60$ 2 .3168064 -0 .151791 -0 .0128098 TILW, $\beta c_0=8.45$ 2 .3460240 -0 .154846 -0 .0134070 TILW, $\beta c_0=8.30$ 2 .3869776 -0 .159128 -0 .0142442 TILW, $\beta c_0=8.20$ 2 .4127840 -0 .161827 -0 .0147710 TILW, $\beta c_0=8.10$ 2 .4465400 -0 .165353 -0 .0154645 TILW, $\beta c_0=8.00$ 2 .4891712 -0 .169805 -0 .0163414 Iwasaki 3 .648 -0 .331 0 DBW2 12 .2688 -1 .4086 0 ------------------------ ---- ---------- ---- --------- ---- ---------- : Input parameters $c_0$, $c_1$, $c_3$ $(c_2=0)$ \[tab1\] --------------- ---- ----------------- --- ----------------- --- ------------------ Plaquette -0 .1628570582(5) 0 .0434830339(1) 0 .01809576875(4) Symanzik -0 .12805490528(8) 0 .0378314931(2) 0 .01476335801(5) TILW $(8.60)$ -0 .10821568768(4) 0 .03408560232(6) 0 .01265991972(4) TILW $(8.45)$ -0 .10749185625(3) 0 .0339409375(1) 0 .01258108895(1) TILW $(8.30)$ -0 .1064962872(3) 0 .0337409869(2) 0 .012472434543(4) TILW $(8.20)$ -0 .1058799831(2) 0 .0336166372(1) 0 .0124050416(1) TILW $(8.10)$ -0 .1050866191(1) 0 .03345591621(5) 0 .012318127134(5) TILW $(8.00)$ -0 .10410447893(3) 0 .03325593631(8) 0 .012210297749(7) Iwasaki -0 .08255435613(4) 0 .0285451387(1) 0 .00983490867(5) DBW2 -0 .0364526623(2) 0 .01581657412(5) 0 .004280099253(2) --------------- ---- ----------------- --- ----------------- --- ------------------ : Total contribution of one-loop diagrams \[tab2\] --------------- ---- -------------- ---- -------------- ---- --------------- ---- --------------- ---- ----------------- Plaquette -0 .01753602(2) 0 .00259963(2) -0 .000155894(8) -0 .000163242(2) -0 .00001721759(2) Symanzik -0 .00810366(1) 0 .00095046(2) -0 .000404510(9) -0 .000107348(2) -0 .00001275904(1) TILW $(8.60)$ -0 .00437013(7) 0 .00019403(5) -0 .00045894(1) -0 .000078117(3) -0 .00001020820(1) TILW $(8.45)$ -0 .00425575(7) 0 .00016978(6) -0 .00045962(1) -0 .000077102(3) -0 .00001011451(1) TILW $(8.30)$ -0 .00410086(7) 0 .00013682(7) -0 .00046040(1) -0 .000075713(3) -0 .00000998564(1) TILW $(8.20)$ -0 .00400636(6) 0 .00011666(8) -0 .00046080(1) -0 .000074857(3) -0 .00000990584(1) TILW $(8.10)$ -0 .00388630(6) 0 .00009097(9) -0 .00046123(1) -0 .000073760(3) -0 .00000980314(1) TILW $(8.00)$ -0 .00374009(6) 0 .00005958(9) -0 .000461601(9) -0 .000072410(3) -0 .00000967600(1) Iwasaki -0 .00112957(2) -0 .00052964(6) -0 .000436966(5) -0 .000045009(3) -0 .00000682353(1) DBW2 0 .0008481(2) -0 .00085301(8) -0 .00018540(1) -0 .000006164(3) -0 .00000173502(3) --------------- ---- -------------- ---- -------------- ---- --------------- ---- --------------- ---- ----------------- : Total contribution of two-loop diagrams containing closed fermion loops \[tab5\] --------------- ---- -------------- ---- --------------- --- --------------- --- --------------- --- ----------------- Plaquette 0 .01656633(2) -0 .00055904(1) 0 .002622771(7) 0 .000158125(2) 0 .00004282674(2) Symanzik 0 .00605656(1) 0 .000935801(6) 0 .002120980(9) 0 .000104973(2) 0 .00002971553(1) TILW $(8.60)$ 0 .00202637(3) 0 .00157890(3) 0 .001790242(9) 0 .000076167(2) 0 .00002260669(1) TILW $(8.45)$ 0 .00190729(3) 0 .00159800(3) 0 .001777415(9) 0 .000075164(3) 0 .00002235603(1) TILW $(8.30)$ 0 .00174666(3) 0 .00162375(2) 0 .001759689(9) 0 .000073791(3) 0 .00002201243(1) TILW $(8.20)$ 0 .00164901(3) 0 .00163939(2) 0 .001748661(9) 0 .000072944(3) 0 .00002180041(1) TILW $(8.10)$ 0 .00152532(3) 0 .00165917(2) 0 .001734421(9) 0 .000071859(3) 0 .00002152826(1) TILW $(8.00)$ 0 .00137535(4) 0 .00168310(3) 0 .00171671(1) 0 .000070522(3) 0 .00002119259(1) Iwasaki -0 .00103022(1) 0 .00203254(1) 0 .001313076(3) 0 .000043949(3) 0 .00001423324(1) DBW2 -0 .0018961(2) 0 .0016130(3) 0 .000413397(9) 0 .000005057(3) 0 .00000307480(3) --------------- ---- -------------- ---- --------------- --- --------------- --- --------------- --- ----------------- : Total contribution of two-loop diagrams containing closed fermion loops \[tab5\] --------------- --- -------------- ---- --------------- --- --------------- ---- --------------- ---- ----------------- Plaquette 0 .00118621(2) -0 .000546197(8) 0 .001365146(9) -0 .000692228(3) -0 .00019809791(7) Symanzik 0 .00081496(1) -0 .000448276(6) 0 .001041379(8) -0 .000574521(3) -0 .0001453370(2) TILW $(8.60)$ 0 .00063643(1) -0 .000389464(5) 0 .000857737(3) -0 .000500011(5) -0 .0001148491(1) TILW $(8.45)$ 0 .00063033(1) -0 .000387269(5) 0 .000851127(3) -0 .000497194(5) -0 .0001137544(1) TILW $(8.30)$ 0 .00062198(1) -0 .000384243(5) 0 .000842047(3) -0 .000493307(5) -0 .0001122515(1) TILW $(8.20)$ 0 .00061684(1) -0 .000382366(5) 0 .000836433(3) -0 .000490894(5) -0 .0001113227(1) TILW $(8.10)$ 0 .00061025(1) -0 .000379946(5) 0 .000829214(4) -0 .000487781(4) -0 .0001101288(1) TILW $(8.00)$ 0 .00060214(1) -0 .000376945(5) 0 .000820289(4) -0 .000483915(4) -0 .0001086536(1) Iwasaki 0 .00043546(1) -0 .00030800(1) 0 .000629274(8) -0 .000395294(3) -0 .0000779538(3) DBW2 0 .00015833(3) -0 .00014883(4) 0 .00024756(2) -0 .00018242(5) -0 .0000213595(6) --------------- --- -------------- ---- --------------- --- --------------- ---- --------------- ---- ----------------- : Total contribution of two-loop diagrams containing closed fermion loops \[tab5\] --------------- --- ------------- ---- ------------- ---- -------------- ---- ------------- ---- ------------- Plaquette 0 .077167(3) -0 .019808(3) -0 .0085415(2) -0 .047102(4) 0 .010439(3) Symanzik 0 .034929(2) -0 .010895(2) -0 .0041454(2) -0 .017940(2) 0 .004491(2) TILW $(8.60)$ 0 .020247(1) -0 .007117(2) -0 .0024559(1) -0 .008702(1) 0 .002251(1) TILW $(8.45)$ 0 .019816(1) -0 .006998(2) -0 .0024050(1) -0 .008448(1) 0 .002185(1) TILW $(8.30)$ 0 .019235(1) -0 .006835(2) -0 .0023362(1) -0 .0081078(6) 0 .0020973(9) TILW $(8.20)$ 0 .018881(1) -0 .006736(2) -0 .0022942(1) -0 .0079023(7) 0 .002044(1) TILW $(8.10)$ 0 .018433(1) -0 .006609(2) -0 .0022410(1) -0 .0076431(9) 0 .0019761(8) TILW $(8.00)$ 0 .017888(1) -0 .006454(2) -0 .0021762(1) -0 .0073300(6) 0 .0018940(6) Iwasaki 0 .0087615(7) -0 .003656(1) -0 .00107856(8) -0 .0027484(4) 0 .0006646(5) DBW2 0 .0007907(2) -0 .0004889(3) -0 .00008343(2) 0 .0001308(2) -0 .0001587(3) --------------- --- ------------- ---- ------------- ---- -------------- ---- ------------- ---- ------------- : Contribution of diagrams 3, 4, 6, for the Iwasaki action \[tab9\] -0.8cm --------------- ---- -------------- ---- ----------------- ---- --------------- --- --------------- --- ---------------- Plaquette 0 .0039245(3) -0 .0000842143(1) -0 .09448252(9) 0 .02755993(3) 0 .010521016(1) Symanzik 0 .0014622(1) -0 .0000454986(1) -0 .03417549(2) 0 .01248953(1) 0 .0041047891(2) TILW $(8.60)$ 0 .0006472(1) -0 .00002872341(6) -0 .017374635(6) 0 .007205477(3) 0 .0021218443(2) TILW $(8.45)$ 0 .0006251(1) -0 .00002818123(6) -0 .016917713(6) 0 .007049188(2) 0 .0020666192(2) TILW $(8.30)$ 0 .0005954(1) -0 .00002744385(5) -0 .016304614(5) 0 .006838088(3) 0 .0019924047(3) TILW $(8.20)$ 0 .0005775(1) -0 .00002699223(5) -0 .015933835(5) 0 .006709626(4) 0 .0019474604(2) TILW $(8.10)$ 0 .0005550(1) -0 .00002641646(5) -0 .015466270(5) 0 .006546741(4) 0 .0018907121(3) TILW $(8.00)$ 0 .0005279(1) -0 .00002571231(5) -0 .014902324(4) 0 .006348924(5) 0 .0018221643(3) Iwasaki 0 .00015719(6) -0 .00001249281(2) -0 .00596123(2) 0 .00295502(1) 0 .0007286816(4) DBW2 -0 .00002436(1) -0 .00000050404(9) -0 .00028731(2) 0 .00020317(4) 0 .0000278810(8) --------------- ---- -------------- ---- ----------------- ---- --------------- --- --------------- --- ---------------- : Contribution of diagrams 3, 4, 6, for the Iwasaki action \[tab9\] --- ---- ----------------- ---- ----------------- 0 -0 .05602636832(2) -0 .02652798781(3) 1 0 0 .0285451387(1) 2 0 0 .00983490867(5) --- ---- ----------------- ---- ----------------- : Contribution of diagrams 3, 4, 6, for the Iwasaki action \[tab9\] --- --- --- ---- ---------------- ---- ---------------- ---- ---------------- 0 0 0 -0 .0003923686(9) -0 .000743134(3) -0 .0000714882(8) 0 2 0 0 .0002615791(6) 0 .000495422(2) 0 .0000357441(4) 1 0 0 0 0 .001900337(2) 0 1 2 0 0 0 .0017774410(9) 0 2 0 0 0 -0 .0010339720(2) 0 2 2 0 0 -0 .001041123(1) 0 .0002799238(4) --- --- --- ---- ---------------- ---- ---------------- ---- ---------------- : Contribution of diagrams 3, 4, 6, for the Iwasaki action \[tab9\] --- --- --- ---- -------------- ---- ---------------- ---- ----------------- ---- ----------------- 0 0 0 0 .00042802(1) -0 .000195263(2) 0 0 0 0 1 0 .0057103(7) 0 .0030512(2) 0 0 0 2 0 -0 .00111995(2) -0 .00029748(1) 0 -0 .000298742(2) 0 2 1 -0 .0022472(3) -0 .0008718(2) 0 0 .0003705893(7) 0 2 2 -0 .00371263(2) -0 .00224859(1) 0 0 1 0 0 0 0 .00064534(1) 0 0 1 0 1 0 -0 .003656(1) 0 0 1 2 0 0 0 .00011079(6) -0 .000144897(2) 0 .000429899(1) 1 2 1 0 0 .0006450(5) 0 .000248682(4) -0 .00022905(1) 1 2 2 0 0 .00295502(1) 0 0 2 0 0 0 -0 .000000974(1) 0 0 2 0 1 0 -0 .00107856(8) 0 0 2 2 0 0 0 .000141960(3) 0 .000042314(2) 0 .0003303085(7) 2 2 1 0 0 .00039546(6) 0 .00002909398(7) -0 .000267364(2) 2 2 2 0 0 .0007286816(4) 0 0 3 2 0 0 0 0 -0 .000019835(1) 3 2 1 0 0 0 -0 .00001249281(2) --- --- --- ---- -------------- ---- ---------------- ---- ----------------- ---- ----------------- : Contribution of diagrams 21-23, 25, 27, 28, for the Iwasaki action \[tab12\] --- ---- ---------------- ---- ---------------- 0 0 .000261920(6) 0 .000173538(9) 1 -0 .0000308339(1) -0 .00027717(1) 2 0 .000370942(2) 0 .000258332(8) 3 0 -0 .000395294(3) 4 0 -0 .0000779538(3) --- ---- ---------------- ---- ---------------- : Contribution of diagrams 21-23, 25, 27, 28, for the Iwasaki action \[tab12\] --- --- --- ---- ----------------- ---- ---------------- ---- ------------------ ---- ---------------- 0 0 0 0 .000373419(3) -0 .000158621(4) -0 .000094848(3) -0 .0001759336(5) 0 2 0 -0 .000373419(3) 0 .000079311(2) 0 0 .0000879668(3) 1 0 0 -0 .000887295(1) 0 .0001396819(4) 0 .000045158(4) 0 .0001893113(5) 1 2 0 0 .000887295(1) 0 .000085189(2) 0 -0 .000120480(1) 2 0 0 0 .000194437(1) -0 .0000319392(3) 0 .000168506(2) -0 .0000509266(2) 2 2 0 -0 .000194437(1) -0 .000005787(2) 0 0 .0000098758(1) 3 0 0 0 .000059183(3) 0 -0 .000015234(1) 0 3 2 0 -0 .000059183(3) 0 .0000172022(5) 0 0 .0000168072(6) 4 0 0 0 .00000682353(1) 0 0 .000007409712(6) 0 4 2 0 -0 .00000682353(1) 0 0 0 --- --- --- ---- ----------------- ---- ---------------- ---- ------------------ ---- ---------------- : Contribution of diagrams 21-23, 25, 27, 28, for the Iwasaki action \[tab12\] --------------- -------- ---- ---------------- --- ----------------- --- ------------------ Plaquette 6.00 -0 .579221119(2) 0 .1159547570(3) 0 .03618067788(9) Symanzik 5.00 -0 .4869797578(8) 0 .1121369999(4) 0 .03538605357(4) Symanzik 5.07 -0 .478756110(2) 0 .11072412996(5) 0 .03507238306(5) Symanzik 6.00 -0 .3915226522(2) 0 .0947962001(5) 0 .03124138429(9) TILW $(8.60)$ 3.7120 -0 .5358770348(7) 0 .1265917638(3) 0 .03813963851(4) TILW $(8.45)$ 3.6018 -0 .5497415338(3) 0 .1291104644(3) 0 .0386337113(1) TILW $(8.30)$ 3.4772 -0 .5651407386(9) 0 .1319263769(1) 0 .0391695069(1) TILW $(8.20)$ 3.3985 -0 .5756111531(9) 0 .1337937558(7) 0 .03951713046(7) TILW $(8.10)$ 3.3107 -0 .5870122772(4) 0 .1358437825(6) 0 .0398899143(3) TILW $(8.00)$ 3.2139 -0 .599415804(1) 0 .138085996(2) 0 .0402877133(4) Iwasaki 1.95 -0 .757856451(1) 0 .1671007819(8) 0 .044746728234(1) Iwasaki 2.20 -0 .6555102085(5) 0 .1537748193(6) 0 .04293183656(3) Iwasaki 2.60 -0 .541348980(1) 0 .1359882440(3) 0 .03967626495(6) DBW2 0.6508 -0 .7749943512(7) 0 .1847244889(1) 0 .04731717866(3) DBW2 0.8700 -0 .574781578(1) 0 .1575688409(9) 0 .04281261980(1) DBW2 1.0400 -0 .4822863343(9) 0 .1412499230(5) 0 .039186543574(5) --------------- -------- ---- ---------------- --- ----------------- --- ------------------ : Results for $dm^{\rm dr}_{\rm (1-loop)}$ (Eq. (\[dm1loopDressed\])), with $N=3$ \[tab13\] ------------------- --- -------- --- ------- --- ------- --- ------- --- ------- --- ------- ---------------------------------------------------------------------------------------- Plaquette 0 6.00 1 .479 0 .1301 0 .1335 0 .1362 0 .1362 0.1392 [@Bowler] Plaquette 0 6.00 1 .769 0 .1275 0 .1306 0 .1337 0 .1332 0.1352 [@Luscher1996] \[1ex\] Plaquette 2 5.29 1 .9192 0 .1262 0 .1307 0 .1353 0 .1341 $\displaystyle{0.1373 \atop 0.1363}$ $\displaystyle{\cite{UKQCD} \atop \cite{Sommer}}$ \[2ex\] Iwasaki 2 1.95 1 .53 0 .1292 0 .1368 0 .1388 0 .1379 0.1421 [@Khan] TILW $(8.60)$ 0 3.7120 1 .0 0 .1339 0 .1370 0 .1378 0 .1384 TILW $(8.00)$ 0 3.2139 1 .0 0 .1348 0 .1387 0 .1397 0 .1406 DBW2 2 0.87 0 .0 0 .1502 0 .1384 0 .1460 0 .1479 DBW2 2 0.87 1 .0 0 .1352 0 .1372 0 .1379 0 .1379 DBW2 2 1.04 0 .0 0 .1454 0 .1375 0 .1421 0 .1434 DBW2 2 1.04 1 .0 0 .1334 0 .1348 0 .1352 0 .1352 ------------------- --- -------- --- ------- --- ------- --- ------- --- ------- --- ------- ---------------------------------------------------------------------------------------- : One- and two-loop results, and non-perturbative estimates for $\kappa_c$ \[tab14\] B. Sheikholeslami and R. Wohlert, Nucl. Phys. [B259]{}, 572 (1985). K. Symanzik, Nucl. Phys. [B226]{}, 187 (1983); K. Symanzik, Nucl. Phys. [B226]{}, 205 (1983). A. Skouroupathis and H. Panagopoulos, Phys. Rev. [D76]{}, 094514 (2007) \[arXiv:0707.2906\]. E. Follana and H. Panagopoulos, Phys. Rev. [D63]{}, 017501 (2000) \[hep-lat/0006001\]. H. Panagopoulos and Y. Proestos, Phys. Rev. [D65]{}, 014511 (2001) \[hep-lat/0108021\]. S. Caracciolo, A. Pelissetto and A. Rago, Phys. Rev. [D64]{}, 094506 (2001) . H. Panagopoulos and E. Vicari, Phys. Rev. [D58]{}, 114501 (1998) \[hep-lat/9806009\]. H. Panagopoulos and E. Vicari, Phys. Rev. [D59]{}, 057503 (1999) \[hep-lat/9809007\]. M. Constantinou, H. Panagopoulos and A. Skouroupathis, Phys. Rev. [D74]{}, 074503 (2006) \[hep-lat/0606001\]. M. Lüscher et al., Nucl. Phys. [B491]{}, 323 (1997) \[hep-lat/9609035\]. M. Lüscher and P. Weisz, Phys. Lett. [B158]{}, 250 (1985). Y. Iwasaki, Univ. of Tsukuba Report UTHEP-118 (1983). K. Symanzik, Nucl. Phys. [B226]{}, 187 (1983). M. Lüscher and P. Weisz, Commun. Math. Phys. [97]{}, 59 (1985); Erratum-ibid. [98]{}, 433 (1985). M.G. Alford et al., Phys. Lett. [B361]{}, 87 (1995) \[hep-lat/9507010\]. T. Takaishi, Phys. Rev. [D54]{}, 1050 (1996). G. Parisi, in: [Proceedings of the 20th International Conference on High Energy Physics]{}, Madison, 1980, ed. L. Durand and L. G. Pondrom (AIP, New York, 1981). G. P. Lepage and P. B. Mackenzie, Phys. Rev. [D48]{}, 2250 (1993) \[hep-lat/9209022\]. UKQCD Collaboration (K.C. Bowler et al.), Phys. Rev. [D62]{}, 054506 (2000). UKQCD Collaboration (C.R. Allton et al.), Phys. Rev. [D65]{}, 054502 (2002) . ALPHA Collaboration (M. Della Morte et al.), JHEP [0507]{}, 007 (2005) . CP-PACS Collaboration (A. Ali Khan et al.), Phys. Rev. [D65]{}, 054505 (2002); Erratum-ibid. [D67]{}, 059901 \[hep-lat/0105015\]. [^1]: $1\times 1$ [plaquette]{}, $1\times 2$ [rectangle]{}, $1\times 2$ [chair]{} (bent rectangle), and $1\times 1\times 1$ [parallelogram]{} wrapped around an elementary 3-d cube.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Sloan Digital Sky Survey (SDSS) automated spectroscopic reduction pipeline provides $>$1.5 million intermediate resolution, $R \simeq 2000$, moderate signal-to-noise ratio (SNR), SNR$\sim 15$, astronomical spectra of unprecedented homogeneity that cover the wavelength range $3800-9200\,$Å. However, there remain significant systematic residuals in many spectra due to the sub-optimal subtraction of the strong OH sky emission lines longward of $6700\,$Å. The OH sky lines extend over almost half the wavelength range of the SDSS spectra, and the SNR over substantial wavelength regions in many spectra is reduced by more than a factor two over that expected from photon counting statistics. Following the OH line subtraction procedure presented in @2005MNRAS.358.1083W [WH05], we make available to the community sky-residual subtracted spectra for the Sloan Digital Sky Survey Data Relase 7. Here we summarise briefly the method, including minor changes in the implementation of the procedure with respect to WH05. The spectra are suitable for many science applications but we highlight some limitations for certain investigations. Details of the data model for the sky-residual subtracted spectra and instructions on how to access the spectra are provided.' author: - \| Vivienne Wild$^{1,2}$[^1], Paul C. Hewett$^3$\ \ $^1$Institut d’astrophysique de Paris, 98bis Boulevard Arago, 75014 Paris, France\ $^2$Institute for Astronomy, Blackford Hill, Edinburgh, EH9 3HJ, UK\ $^3$Institute of Astronomy, Madingley Road, Cambridge, CB3 0HA, UK bibliography: - 'refs\_all.bib' title: 'Peering through the OH Forest: public release of sky-residual subtracted spectra for SDSS DR7' --- Introduction ============ The Sloan Digital Sky Survey [SDSS; @2000AJ....120.1579Y] represents the most impressive application yet of wide-field multi-object spectrographs to provide spectroscopic samples of quasars, galaxies and stars of unprecedented size. The final data release [DR7; @Abazajian:2009p5430] provides $>$1.5 million science-target spectra, nearly 1.3 million of which form part of the SDSS Legacy Survey. Despite the high quality of the SDSS spectra, visual inspection reveals significant systematic sky-subtraction residuals longward of $6700\,$Å in many spectra. For fainter objects in particular, the sky-subtraction residuals, which exhibit strong non-Gaussian properties, are the dominant source of uncertainty over a wavelength interval of some $2000\,$Å. The wavelength range affected includes features of significant astrophysical interest; examples include the Calcium triplet (8500, 8544, $8665\,$Å), a powerful diagnostic of stellar populations in low-redshift galaxies, and the H$\beta$ + \[OIII\] 4960, $5008\,$Å emission region in active galactic nuclei (AGN) in the redshift interval $0.4 < z < 0.8$. The fundamental problem is the difficulty associated with the removal of OH emission features from spectra in which the line profiles are barely resolved. Combined with sub-pixel changes in the pixel-to-wavelength calibration between spectra, sharp (positive/negative) residuals often remain following sky subtraction. The sky-subtraction residuals arise from the subtraction of two essentially identical tooth-comb signatures that have been very slightly misaligned, leading to well-defined residual patterns. In WH05 we presented a method to remove the dominant OH sky-subtraction residuals in the SDSS spectra based on Principal Component Analysis (PCA). PCA takes advantage of the correlations present as a function of wavelength, offering significant advantages over simply masking the affected pixels. In this document we describe the release to the astronomical community of FITS-files containing the sky-residual subtracted fluxes. We briefly summarise the main steps of the procedure in Section \[sec:method\], including details of the object-feature masks used and minor alterations to the original method. In Section \[sec:model\] we describe the data model and how to access the data. The core code, written in the IDL language, is also released for reference. Method {#sec:method} ====== Here we outline the main points of the procedure, and refer the reader to WH05 for full details. For each observation of a SDSS spectroscopic plate, 32 fibres, 16 for each of the 2 spectrographs, are normally assigned to blank sky regions, selected from areas containing no detected objects in the SDSS imaging survey. During the SDSS pipeline processing, the sky spectra are combined to create a “master-sky” spectrum for each spectroscopic plate, which is then scaled and subtracted from each of the 640 spectra, producing 608 sky-subtracted object spectra and 32 sky-subtracted sky spectra, all of which are part of the standard SDSS data releases. Throughout the remainder of the paper we will refer to the sky-subtracted sky spectra as “sky spectra”, and the sky-subtracted object spectra as “object spectra”. The purpose of our technique is to remove the tooth-comb residuals which remain after the SDSS pipeline master sky subtraction. The analysis that follows is applied only to the red part of the SDSS spectra ($6700-9180\,$Å), as the strong emission features blueward of $6700\,$Åare small in number and easily masked. Additionally, only pixels which may have been contaminated by night sky lines, termed “sky-pixels”, are cleaned. Those pixels which lie between sky lines, termed “non-sky pixels”, are not altered by the procedure and are used simply as reference points during the reconstruction of the OH features in the sky-pixels. The eigensystem --------------- The SDSS sky spectra are used to calculate the eigensystem. The first stage is to form an estimate of the Poisson noise spectrum for each plate from the median of the error arrays of the 32 sky spectra on that plate. The reason for such an approach, rather than using the individual noise arrays of each sky spectrum, is the difficulty of separating the sky noise component from the object noise component in the object error arrays. Thus, a single Poisson noise spectrum is used for all objects on the same plate. This Poisson noise spectrum is corrected for the rescaling applied by the SDSS reduction pipeline at the position of the OH lines (see WH05). With a robust estimate of the Poisson noise spectrum available for each plate, each sky spectrum is then normalised by the expected Poisson noise. Failure to do so would result in over-subtraction of flux in sky pixels, with “corrected” pixels apparently exhibiting fluctuations below the Poisson limit (see WH05). The sky-subtraction method then proceeds by performing a PCA of the Poisson error normalised sky spectra to produce a set of orthogonal components that provides a compact representation of the systematic residuals caused by the OH emission lines. In our updated method, a robust and iterative PCA procedure is used, which is presented in @Budavari:2009p4530. The algorithm essentially replaces the least-squares minimisation solved by traditional PCA with the minimisation of a new robust function of the data, which employs a robust Cauchy-type function to limit the impact of outliers. The iterative nature of the algorithm allows us to greatly reduce the computational time for the eigensystem; only $\sim3000$ sky spectra are required for the eigensystem to converge. A single set of eigenspectra are created for the entire SDSS DR7. During development of this release, attempts were made to create sets of eigenspectra for different plates, and for groups of fibres which exhibit similar fluctuations. However, no substantial improvement in the final result was found and therefore the simplest (and original) method was adopted. Cleaning of object spectra {#sec:cleaning} -------------------------- Once the eigensystem has been created from the sky spectra, the second step of the procedure involves adding the components in linear combinations to remove the systematic sky residuals from the spectra of target objects, such as galaxies, quasars and stars. The number of components to use in the reconstruction is not well defined. The use of too many components results in the artificial suppression of noise below the Poisson limit, with the PCA acting as an (undesirable for most science applications) high-frequency filter. The use of too few components means that the removal of sky residuals is sub-optimal and, in some cases, the overall quality of the spectra can decrease. The reduction of the noise in the sky pixels below the noise in the non-sky pixels is clearly unphysical, and we therefore estimate the number of components to employ in the reconstruction of each spectrum by adopting the non-sky pixels of that spectrum as a reference. The number of components is determined using the root-mean-square (RMS) of the (flux-continuum)/error array for both the sky, and non-sky pixels. All RMS-values are calculated after subtraction of the object continuum, using a 71-pixel running median filter, masking of potential line features as described below, and removal of bad (error=0) pixels. Despite these precautions, a robust estimator of the RMS is still required, therefore the 90th percentile of the absolute (flux-continuum)/error array is used as our “RMS” estimator. This simplifies slightly the procedure of WH05. The error array used in this section has been rescaled to empirically remove the rescaling applied during the SDSS pipeline reduction at the position of the OH lines (see WH05 for more details). All three RMS values are included in the FITS-headers described in Section \[sec:model\]. The reconstruction of a spectrum proceeds one component at a time, with the RMS-ratio calculated between the sky-residual subtracted spectrum and the non-sky pixels as reference. The reconstruction is stopped when the RMS-ratio reaches unity, i.e. when the noise weighted flux RMS is the same for the sky and non-sky pixels. The scheme is largely self-calibrating. For example, in spectra with high SNR the systematic sky residuals typically contribute only marginally to the sky-pixel noise, the RMS-ratio thus starts close to unity and only a small number of components are necessary to achieve equality in the RMS-ratio. During projection of the eigenspectra onto the object spectra, bad pixels (error=0) and pixels which may have been affected by a narrow emission or absorption feature, not removed by the 71-pixel median filter, are masked (Section \[sec:masks\]). However, OH emission line residuals are still removed from these pixels via subtraction of the contiguous PCA components, the amplitudes of which are determined using the rest of the wavelength range. The effective correction of pixels which include astrophysical features is a major benefit of the PCA subtraction scheme compared to the application of a high-pass filter or masking out pixels affected by OH residuals completely. As in WH05 we have imposed an upper limit on the number of components used of 300. Any spectrum which reaches this upper limit (see header keyword [NRECON]{} in the FITS files, see Table \[tab:header\]) should be treated with suspicion, as it is possible that the stopping criteria has failed due to some real unmasked feature in the spectrum, such as strong intervening absorption lines in quasars. In a very small number of spectra, the final reconstruction after 300 components has a sky pixel RMS that is worse than prior to subtraction. No correction is applied to such spectra, which may be identified by inspection of the FITS header keywords [SKYVAR1]{} and [SKYVAR2]{} (i.e. [ SKYVAR]{}1=[SKYVAR]{}2, see Table \[tab:header\]). The reliability of the PCA reconstructions deteriorates significantly when a substantial fraction of all the pixels are unavailable, therefore no correction is applied to spectra where the combined fraction of bad and masked pixels in the range $6700-9180\,$Å exceeds 20 per cent. Samples ------- In the inital public release (Version 1.0) we have made available spectra for two science samples: galaxies and quasars. Galaxies have been selected as all objects in DR7 possessing spSpec FITS-files, with header keyword [SPEC\_CLN]{}$=2$. Quasars are selected from the FITS header keyword [SPEC\_CLN]{}$=3$ or $4$. Additionally, spectra of additional objects not selected by the latter criterion, but included in the DR7 quasar catalogue [@Schneider:2010p5407] have also been treated as quasars. Sky spectra ([SPEC\_CLN]{}$=5$) and spectra without classification ([SPEC\_CLN]{}$=0$) have also been processed and are included in the release. Masks {#sec:masks} ----- Narrow features, intrinsic to object spectra, that are not subtracted by the median-filter derived “continua” (Section \[sec:cleaning\]) can effect the determination of the sky- and non-sky-pixel RMS values and bias the reconstruction of the sky-residuals. To mitigate both effects, narrow features, notably emission and absorption lines, are masked during the projection of the eigenspectra and calculation of the RMS values. The masked pixels are still subject to the sky-residual correction but their presence is not allowed to influence the projection of the eigenspectra. Details of the mask applied for each spectrum are included in the third row of the FITS data files (see Table \[tab:data\]). Individually tailored masks can be generated for each object spectrum, depending on the specific science application. However, for the majority of science applications, the use of a liberally-defined mask results in only a minor degredation in the quality of the sky-residual subtraction while minimising the probability of undesirable bias, due to the influence of unmasked features on the projections. Thus, for galaxies, a single mask has been used which includes all the common emission and absorption features found in galaxy spectra that are detectable at the resolution and SNR of the SDSS spectra. Details of the features, their vacuum wavelengths and the rest-frame extent of the masked region are given in Table \[tab:galwav\]. Redshifts, necessary to calculate the observed-frame wavelengths of the masked regions, are taken from the SDSS FITS file primary headers. The same philosophy has been applied for the quasar sample although redshifts are taken from the DR7 quasar catalogue of @Schneider:2010p5407 where available. Vacuum wavelengths and the rest-frame extent of the masked regions for quasars are given in Table \[tab:qsowav\][^2]. The advantage of using the sky-residual subtracted spectra can be appreciated by inspection of Figs. \[fig:eggal\], \[fig:egqso\] and \[fig:qsocomp\]. It is important to recognise that the spectra resulting from the use of the generic masking procedure described above are not suitable for science investigations where additional narrow features are present at a different redshift to the science object. Examples include quasar absorption line catalog generation and the identification of spectra consisting of multiple object components, such as galaxies with gravitationally lensed background sources. The potential problems are demonstrated via inspection strong MgII absorption systems at $z_{abs}$=2.191 and $z_{abs}$=1.485 in the original and sky-residual-subtracted spectra of spSpec-52282-0518-495 and spSpec-52282-0328-472 respectively. In the former case, the absorption system is virtually eliminated, while in the latter case, the absorber doublet ratio alters dramatically. -------------------- ------ ------------------ Central Wavelength Size Species (Å) (Å) 3728.30 20.0 O [ii]{} 3798.4 15.0 H$\theta$ 3836.2 15.0 H$\eta$ 3869.9 15.0 \[Ne [iii]{}\] 3889.00 15.0 H$\zeta$ 3934.8 33.0 Ca [K]{} 3969.6 33.0 ca [H]{} 3971.19 15.0 H$\epsilon$ 4102.9 20.0 H$\delta$ 4305.6 30.0 G–band 4341.7 20.0 H$\gamma$ 4364.44 15.0 \[O [iii]{}\] 4472.5 15.0 He [i]{} 4862.68 30.0 H$\beta$ 4960.30 15.0 \[O [iii]{}\] 5008.24 20.0 \[O [iii]{}\] 5176.7 30.0 Mg [b]{} 5200.0 10.0 \[N [i]{}\] 5272.0 20.0 5877.4 15.0 He [i]{} 5891.58 20.0 Na [i]{} 5897.56 20.0 Na [i]{} 6302.3 15.0 \[N [ii]{}\] 6498.0 20.0 6549.86 15.0 \[N [ii]{}\] 6564.6 30.0 H$\alpha$ 6585.27 15.0 \[N [ii]{}\] 6680.4 15.0 6718.3 20.0 \[S [ii]{}\] 6732.7 20.0 \[S [ii]{}\] 7066.4 15.0 He [i]{} 7138.4 15.0 \[Ar [iii]{}\] 7321.5 15.0 \[O [ii]{}\] 7331.6 15.0 \[Ar [sc iv]{}\] 7753.5 10.0 \[Ar [iii]{}\] 8500.36 20.0 Ca [ii]{} 8544.38 25.0 Ca [ii]{} 8664.52 25.0 Ca [ii]{} 9018.0 15.0 9072.0 15.0 \[S [iii]{}\] -------------------- ------ ------------------ : \[tab:galwav\] Galaxy Mask -------------------- ------ -------------------- Central Wavelength Size Species (Å) (Å) 1216.0 40.0 Ly$\alpha$ 1241.0 40.0 N [v]{} 1305.0 40.0 O [i]{} 1400.0 40.0 Si [iv]{}+O [iv]{} 1549.0 40.0 C [iv]{} 1909.0 40.0 C [iii]{}\] 2798.0 40.0 Mg [ii]{} 3427.0 20.0 \[Ne [v]{}\] 3729.0 20.0 O [ii]{} 3870.0 20.0 \[Ne [iii]{}\] 4103.0 40.0 H$\gamma$ 4342.0 40.0 H$\delta$ 4863.0 40.0 H$\beta$ 4960.0 40.0 \[O [iii]{}\] 4984.0 40.0 — 5008.0 40.0 \[O [iii]{}\] 5878.0 20.0 He [i]{} 6302.0 20.0 \[O [i]{}\] 6550.0 20.0 \[N [ii]{}\] 6565.0 40.0 H$\alpha$ 6585.0 20.0 \[N [ii]{}\] 6725.0 40.0 \[S [ii]{}\] 7139.0 20.0 \[Ar [iii]{}\] -------------------- ------ -------------------- : \[tab:qsowav\] Quasar Mask Data model {#sec:model} ========== Conscious of the significant data volume of the SDSS Legacy Survey spectra we have deliberately opted for a data model that minimises the duplication of information contained in the original Legacy Survey FITS spSpec files. Our new files include a direct copy of the primary FITS header from the SDSS spSpec file, with the addition of the FITS header keywords listed in Table \[tab:header\]. Three [NAXIS1]{}-pixel floating-point data arrays, of the same length as the original SDSS spectra, are included: the sky-residual subtracted spectrum flux array as the first row; the re-scaled error array as the second row; the mask information as the third row (see Table \[tab:data\]). Only those pixels where the sky-residual correction has been applied have had their flux and error values altered, all other pixels remain unchanged from the original SDSS spSpec file. The sky-residual subtracted spectra have the same naming convention as the SDSS spSpec files, i.e. one file per spectrum, labeled by the mjd, plate and fiberid of the spectrum. The file names include the tag “\_skysub” to prevent any confusion with the original SDSS spSpec files. Individual files are stored in directories according to their plate number. The reduction of the number of data arrays per spectrum from five to three, combined with the elimination of the extensive FITS-header units (containing information about emission and absorption features detected in the spectrum) results in a file size just under half that of the original SDSS Legacy Survey spSpec files. Keyword Description ------------- -------------------------------------------- [SKYVAR0]{} Non-sky pixel variance [SKYVAR1]{} Sky pixel variance before correction error [SKYVAR2]{} Sky pixel variance after correction [NRECON]{} Number of components in reconstruction : \[tab:header\] FITS file primary header keywords Data row Content Equivalent SDSS data row ---------- --------------------- -------------------------- 1 Object flux array 1 2 Object error array 3 3 Sky-correction mask – : \[tab:data\] FITS file data content Mask value Meaning ------------ ----------------------------------------------------------------------------- 0 outside wavelength range $6700-9180\,$Å for sky-correction 1 non-sky pixels: inside wavelength range but not subject to sky-correction 2 sky pixels: inside wavelength range and subject to sky-correction 3 masked non-sky pixels: pixel masked due to object feature or SDSS bad pixel 4 masked sky pixels: pixel masked due to object feature or SDSS bad pixel : \[tab:mask\] FITS sky-correction mask values and their meanings Data Access ----------- The data files may be accessed through the Johns Hopkins University SDSS server at the address http://www.sdss.jhu.edu/skypca/spSpec Subsequent versions will be stored in the subdirectories v1/, v2/ etc. with the main directory given above pointing to the latest version at all times. The files are easily accessible via the “wget” command. Several file lists have been prepared for the convenience of the user and can be found in the main directory given above: wget\_all.lis : To download all reduced spectra. wget\_dr7qsocat.lis : To download the spectra of the DR7 quasar catalog. wget\_specobj\_specclass2.lis : All spectra spectroscopically classified as galaxies ([SPECCLASS=2]{}) in the CAS SPECOBJ view, except for 20 objects classified as galaxies in the CAS catalog, but not in their FITS file headers. wget\_specobjall\_specclass2.lis : All spectra spectroscopically classified as galaxies ([SPECCLASS=2]{}) in the CAS SPECOBJALL table (i.e. includes duplicates), except for objects on plate 1631 for which the spectra of the same mjd as in the catalog do not exist, and the 20 objects noted above. After downloading the individual file list to your computer or creating your own, a command such as: wget -x -nH --cut-dir=2 -a logfile.txt -i wget_all.lis will download all the files in that file list to your machine, creating the file logfile.txt to keep track of the progress. The DR7 QSOs may additionally be accessed from the DR7 Value Added Catalogue web page:\ http://www.sdss.org/dr7/products/value\_added/index.html\#quasars Conclusion {#sec:conclude} ========== We have released to the community a set of SDSS DR7 spectra with significantly improved subtraction of OH sky lines. The PCA-based technique employed takes advantage of the correlation of OH lines as a function of wavelength and sky subtraction close to the Poisson noise limit is achieved for many spectra. Our procedure is generally more effective than simply employing a high-pass filter or simple feature-exclusion mask. Release Version 1.0 includes galaxies, quasars, sky and “unknown” spectra. Future releases will incorporate improved redshifts for the quasars and the masking of intervening absorption features in quasar spectra. Acknowledgements ================ We would like to thank Alex Szalay and Tamas Budavari at Johns Hopkins University for kindly organising the hosting of the data files for public access, and Jan Vandenberg and Rich Ercolani for their time setting up the web access. Gordon Richards enabled the spectra for the SDSS DR7 quasar catalogue to be made available as part of the SDSS DR7 Legacy Release website ‘Added Value’ data page. Figures ======= Here we include a repeat of some of the figures presented in WH05, for easy reference, and to allow direct comparison between the results before and after the re-reduction of the entire data set by the SDSS team. ![[Left:]{} Histogram of the number of components used to reconstruct the sky signal in the sky, unknown, galaxy and quasar samples. The y-axis is truncated in order to show the majority of the objects. Twenty-three per cent of the unknown spectra require zero components, demonstrating the importance of employing appropriate object-specific masks (Section \[sec:masks\]). [Right:]{} Histogram of the ratio between the root mean square noise of the pixels affected by OH lines before and after application of the sky-residual subtraction procedure. []{data-label="fig:nrecon"}](nrecon.ps "fig:") ![[Left:]{} Histogram of the number of components used to reconstruct the sky signal in the sky, unknown, galaxy and quasar samples. The y-axis is truncated in order to show the majority of the objects. Twenty-three per cent of the unknown spectra require zero components, demonstrating the importance of employing appropriate object-specific masks (Section \[sec:masks\]). [Right:]{} Histogram of the ratio between the root mean square noise of the pixels affected by OH lines before and after application of the sky-residual subtraction procedure. []{data-label="fig:nrecon"}](rms.ps "fig:") ![Examples of sky residual subtraction applied to galaxy spectra. In each panel the lower spectrum is the raw SDSS data and the upper spectrum is after application of the sky-residual subtraction procedure. The upper sky-residual subtracted spectrum is offset for clarity. []{data-label="fig:eggal"}](eg_galaxy.ps) ![Same as Figure \[fig:eggal\], but for quasars[]{data-label="fig:egqso"}](eg_quasar.ps) ![Composite spectra of $z\sim4$ quasars. The observed frame spectra fall at wavelengths $>6700\AA$. For each pair of spectra, the lower composite is created directly from the SDSS spectra, and the upper composite from the same spectra after subtracting the sky residuals.[]{data-label="fig:qsocomp"}](highzcomp.ps) [^1]: [email protected] [^2]: A region centred on 4984Å is masked to exclude artificial “absorption” resulting from the continuum-subtracted spectra of objects possessing strong \[O [iii]{}\]$\lambda\lambda$4960,5008 emission.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We generalize results by Wakabayashi and Orevkov about rational cuspidal curves on the projective plane to that on ${\mbox{$\mathbb Q$}}$-homology projective planes. It turns out that the result is exactly the same as the projective plane case under suitable assumptions. We also provide examples which demonstrate sharpness of the results. The ambient surface is singular in these examples.' author: - 'R.V. Gurjar, DongSeon Hwang and Sagar Kolte' bibliography: - 'biblio.bib' title: 'A note on rational cuspidal curves on ${\mbox{$\mathbb Q$}}$-homology projective planes [^1]' --- Introduction ================ The study of rational cuspidal plane curves has had a long and interesting history. One approach is to apply the theory of open surfaces to the complement ${\mbox{$\mathbb P$}}^2-C$. In particular, Wakabayashi proved some sharp bounds on the number of cusps of $C$ in terms of the log Kodaira dimension of ${\mbox{$\mathbb P$}}^2-C$. $($[@waka78]$)$ Let $C$ be a rational cuspidal plane curve. Then, we have the following: 1. If ${{\overline}\kappa}({\mbox{$\mathbb P$}}^2-C) = -\infty$, then $C$ has at most one cusp. 2. If ${{\overline}\kappa}({\mbox{$\mathbb P$}}^2-C) \leq 1$, then $C$ has at most two cusps. Orevkov showed that the log Kodaira dimension of ${\mbox{$\mathbb P$}}^2-C$ is always non-zero. $($[@orevkov02 Theorem B(c)]$)$ Let $C$ be a rational cuspidal plane curve. Then, ${{\overline}\kappa}({\mbox{$\mathbb P$}}^2-C) \neq 0$. The aim of this note is to generalize these results to the case where the ambient surface is a ${\mbox{$\mathbb Q$}}$-homology projective plane instead of the usual projective plane. Here, a -homology projective plane is defined as a normal projective surface $S$ with quotient singularities having the same rational homology as ${\mbox{$\mathbb P$}}^2$. It turns out that the generalized results give us the same conclusions as the results of Wakabayashi and Orevkov under the assumption that the cuspidal curve $C$ does not pass through the singularities of $S$.\ \[main\] Let $C$ be a rational cuspidal curve on a ${\mbox{$\mathbb Q$}}$-homology projective plane $S$. Assume that $C$ is contained in the smooth locus of $S$. Then, we have the following: 1. If ${{\overline}\kappa}(S-C)= -\infty$, then $C$ has at most one cusp. 2. If ${{\overline}\kappa}(S-C) \leq 1$, then $C$ has at most two cusps. 3. ${{\overline}\kappa}(S^0-C) \neq 0 $. The above inequalities are sharp. See Examples in Section \[example\]. Example \[ExB\] also shows that we can have ${{\overline}\kappa}(S-C) = 0 $. Thus, the assumption that $C$ is contained in the smooth locus of $S$ is necessary to have a right generalization of the results on plane cuspidal curves.\ The study of rational cuspidal curves on ${\mbox{$\mathbb Q$}}$-homology projective planes are motivated by a problem posed by Kollár asking the classification of pairs $(S, C)$ where $C$ is a rational cuspidal curve on a ${\mbox{$\mathbb Q$}}$-homology projective plane $S$([@kollar08 Problem 33]).\ Note that $S$ in Theorem \[main\] is necessarily rational. Indeed, since $S$ is of Picard number one, the complement $X:=S-C$ is ${\mbox{$\mathbb Q$}}$-acyclic, i.e., a ${\mbox{$\mathbb Q$}}$-homology plane. ${\mbox{$\mathbb Q$}}$-homology planes have been studied extensively. In particular, R.V. Gurjar, C.R. Pradeep and A.R. Shastri have proved that every ${\mbox{$\mathbb Q$}}$-homology plane with quotient singularities is rational ([@pradeep97] and [@gurjar97]). We also note that, by a result of T. Fujita, -homology planes are affine (cf. [@palka13 Theorem 1.1]).\ Notation and Preliminaries ============================== Let $S$ be a smooth projective surface over the field $\mathbb{C}$ of complex numbers. A divisor on $S$ is a $\mathbb{Z}$-linear combination of irreducible curves on $S$. Similarly, a ${\mbox{$\mathbb Q$}}$-divisor on $S$ is defined to be a ${\mbox{$\mathbb Q$}}$-linear combination of irreducible curves on $S$. A ${\mbox{$\mathbb Q$}}$-divisor $D$ is said to be nef if $D.C \geq 0$ for every irreducible curve $C$ on $S$. A ${\mbox{$\mathbb Q$}}$-divisor $D$ is said to be pseudo-effective if $D.H \geq 0$ for any nef ${\mbox{$\mathbb Q$}}$-divisor $H$. Clearly, every effective divisor is pseudo-effective. For any (possibly reducible) curve $\Delta$ on $S$, by a component of $\Delta$, we mean an irreducible component of $\Delta$.\ We recall some terminology from [@fujita82 Section 6]. For an irreducible component $Y$ of $D$, we denote $Y.(D-Y)$ by $\beta_D(Y)$. This is called the branching number of $Y$ in $D$. The component $Y$ is called a tip of $D$ if $\beta_D(Y)=1$. A sequence $C_1,...,C_r$ of irreducible components of $D$ is called a rational twig $T$ of $D$ if each $C_i$ is a non-singular rational curve, $\beta_D(C_1)=1$, $\beta_D(C_j)=2$ and $C_{j-i} \cdot C_j=1$ for $2 \leq j \leq r$. $C_1$ is called the tip of $T$. Since $\beta_D(C_r)=2$, there is a component $Y$ of $D$, not in $T$, such that $C_{r} \cdot Y=1$. If $Y$ is a rational tip of $D$, then $T'= T + Y$ is a connected component of $D$ and will be called a rational club of $D$. A component $Y$ of $D$ such that $\beta_D(Y)=0$ is also called a rational club of $D$. When the above $Y$ is rational and $\beta_D(Y)=2$, $T'$ is again a rational twig of $D$. If $\beta_D(Y) \geq 3$ or $Y$ is non-rational then $T$ is called a maximal rational twig of $D$ and $Y$ is called the branching component of $T$. For an effective ${\mbox{$\mathbb Q$}}$-divisor $D$, we denote by ${\mbox{$\mathbb Q$}}(D)$ the ${\mbox{$\mathbb Q$}}$-vector space generated by the irreducible components of $D$.\ If $T$ is a rational twig of $D$ which is a contractible rational twig of $D$, the element $N \in {\mbox{$\mathbb Q$}}(T)$ such that $N \cdot C_1=-1$ and $N \cdot C_j=0$ for $j \geq 2$ is called the bark of $T$. If $T'=C_1+...+C_r+Y$ is a contractible rational club of $D$, the bark of $T'$ is define to be the ${\mbox{$\mathbb Q$}}$-divisor $N'$ in ${\mbox{$\mathbb Q$}}(T')$ such that $N' \cdot C_1=N' \cdot Y=-1$ and $N' \cdot C_j=0$ for $2 \leq j \leq r$. For an isolated rational normal curve $Y$ with $Y^2<0$, its bark is defined to be $2(-Y^2)^{-1}Y$. If all the rational clubs and maximal twigs of $D$ are contractible, then the sum of their barks is denoted by $Bk(D)$ and is called the bark of $D$. For the definition of bark of a non-linear tree of smooth rational curves, see [@fujita82].\ The next results originally due to Fujita [@fujita82] are useful in this work. $($[@fujita82 Corollary 6.14]$)$\[contractible\] Let ${\widetilde{S}}$ be a smooth projective surface and $D$ be an effective reduced divisor on it. If $D$ has a rational twig or a club which is not contractible, then $\kappa({\widetilde{S}}, K_{{\widetilde{S}}}+D)= {\overline}\kappa(S)= -\infty$. \[Fujita1\]$($[@fujita82 Lemma 6.20], [@gurjar95 Lemma 1]$)$ Let ${\widetilde{S}}$ be a smooth projective surface and $D$ be an effective reduced SNC divisor on it. Assume that ${\overline}\kappa({\widetilde{S}}-D) \geq 0$ and $N \neq Bk(D)$. Then, there exists a component $L$ of $N$ which is a $(-1)$-curve, not contained in $D$, satisfying one of the following conditions: 1. $D \cdot L=0$, i.e. $D \cap L= \emptyset$ 2. $D \cdot L=1$ and $L$ meets a twig of $D$. 3. $D \cdot L=2$ and $L$ meets exactly two components of $D$, one of which is an irreducible component of a twig of $D$ and the other one is a tip of a rational club of $D$. Furthermore, ${{\overline}\kappa}(S-D-L)={\overline}\kappa(S-D)$. \[Fujita2\] $($[@fujita82 Corollary 8.8], [@gurjar12 Lemma 4]$)$ Let ${\widetilde{S}}$ be a smooth projective surface and $D$ be an effective reduced SNC divisor on it. Assume that ${\overline}\kappa({\widetilde{S}}-D)=0$. Assume also that every $(-1)$-curve in $D$ meets at least three other components of $D$. If $Bk(D)$ = $N$, then any connected component of $D$ is one of the following. 1. A minimal resolution of a quotient singular point. 2. A tree of $\mathbb{P}^1$’s with exactly two branching components such that the branching components are connected by a (possibly empty) linear chain of $\mathbb{P}^1$’s and each branching component meets exactly two other (-2)-curves which are tips of $D$. 3. A simple loop of $\mathbb{P}^1$’s. 4. A tree of $\mathbb{P}^1$’s with a unique branching component which meets three linear trees defining cyclic quotient singular points at one of their end points. Furthermore, the absolute values $d_1,d_2,d_3$ of the discriminants of the three trees satisfy $\Sigma 1/d_i=1$. 5. A tree of five $\mathbb{P}^1$’s with a unique branching component which intersects the other four curves transversally in one point each, and such that the four curves are all (-2)-curves. 6. A smooth elliptic curve. We will repeatedly use the well-known properties of singular fibers of a ${\mbox{$\mathbb P$}}^1$-fibration on a smooth surface [@miya81 Chapter I, Lemma 4.4.1]. \[1sing\] Let $C$ be a rational cuspidal curve on a smooth projective surface $S'$. If ${\overline}\kappa(S'-C) = -\infty$, then $C$ has at most one singular point. Let $\widetilde{S} \rightarrow S'$ be a resolution of singularities on $C$ and let $D$ be the total transform of $C$. Since $D$ is connected, by a result of Miyanishi-Sugie and Russell(see [@miya81 Theorem I.3.13]), there is an ${\mbox{$\mathbb A$}}^1$- fibration on $\widetilde{S}-D$. Let $h: \widetilde{S}' \rightarrow \tilde{S}$ be a birational morphism such that the ${\mbox{$\mathbb A$}}^1$-fibration extends to a ${\mbox{$\mathbb P$}}^1$-fibration on $\widetilde{S}'$, and $D'$ be the proper transform of $D$ under $h$.\ Assume that $C$ has at least two cusps. Then, $D'$ has at least two $(-1)$-curves with branching number $3$. Indeed, those $(-1)$-curves arise from the final blow-ups in order to resolve the cusps of $C$ and make $D$ an SNC divisor. Note that at least one of the singular fiber of the ${\mbox{$\mathbb P$}}^1$-fibration on $\widetilde{S}'$ has a $(-1)$-curve with branching number $3$ in the fiber. This is not possible by well-known properties of singular fibers of a ${\mbox{$\mathbb P$}}^1$-fibration on a smooth projective surface.\ From now on, let $C$ be a rational cuspidal curve on a ${\mbox{$\mathbb Q$}}$-homology projective plane $S$. Let ${\widetilde{S}}{\rightarrow}S$ be a resolution of singularities of both $S$ and $C$ such that the total transform $D$ of $C$ is a simple normal crossing divisor. Let $C'$ be the proper transform of $C$. We always assume that the map is minimal, i.e., the Picard number of ${\widetilde{S}}$ is the least possible. The map ${\widetilde{S}}{\rightarrow}S$ factors through the minimal resolution $S' \rightarrow S$ of singularities of $S$. Let $E$ be the reduced exceptional divisor for the resolution of singularities of $S$. Since $S$ has only quotient singularities, $E$ is a simple normal crossing divisor. The smooth locus of $S$ is denoted by $S^0$.\ The following lemma is useful in proving Theorem \[main\] (3). \[inf\] If ${{\overline}\kappa}({\widetilde{S}}-D)=-\infty$, then ${{\overline}\kappa}({\widetilde{S}}-D-E)=-\infty$. Before we begin the proof, we describe the structure of $D$ when $C$ is uni-cuspidal. There is a unique $(-1)$-curve in $D$, which we denote by $H$, such that its branching number is three. The proper transform $C'$ of $C$ is one of the branches of $H$. Let $T_2$ and $T_3$ denote the other two branches of $H$. It is easy to see that one of these two (say $T_2$) is a linear chain of rational curves and the other branch, say $T_3$, is a tree of rational curves such that every branch of $T_3$ has branching number three([@orevkov02], Proposition 3.2). We will use this description implicitly in the arguments that follow.\ By a result of Koras and Palka([@koras13]), ${{\overline}\kappa}({\widetilde{S}}-D-E) < 2$. Suppose that ${{\overline}\kappa}({\widetilde{S}}-D-E) \geq 0$. By [@palka12 Theorem 5.1], $S-C$ has an untwisted ${\mbox{$\mathbb C$}}^$-fibration $f$ over ${\mbox{$\mathbb A$}}^1$ such that the fiber at infinity can be assumed to be smooth and $f$ has two singular fibers. One of the singular fibers is a punctured affine line ${\mbox{$\mathbb C$}}^$ with suitable multiplicity and the other consists of two ${\mbox{$\mathbb C$}}$’s meeting in a cyclic singular point. Furthermore, both ${\mbox{$\mathbb C$}}$’s have multiplicity at least two. The ${\mbox{$\mathbb C$}}^$-fibration is naturally extended to ${\widetilde{S}}-D$ and the extended action is also denote by $f$. We will derive a contradiction by showing that such a ${\mbox{$\mathbb C$}}^$-fibration cannot exist on ${\widetilde{S}}-D$.\ We claim that $f$ is again extended to a ${\mbox{$\mathbb P$}}^1$-fibration on ${\widetilde{S}}$. If not, then $f$ has a base point on $D$. Let $D'$ be the total transform of $D$ under the resolution of base locus. It is clear that the $(-1)$-curve resulting from the last blow up will be one of the horizontal components of $D'$ to $f$ which has branching number at most two. This is a contradiction to the description of the ${\mbox{$\mathbb C$}}^$-fibration by Palka (Figure 3, pp.450, [@palka12]) where each horizontal component has branching number three. Thus $f$ has no base points on $D$.\ The extended ${\mbox{$\mathbb P$}}^1$-fibration on ${\widetilde{S}}$ is also denoted by $f$. A fiber of $f$ is contained in $D$. We denote this fiber by $R$. The fiber $R$ is reducible because $D$ does not contain an irreducible component with self intersection 0 and branching number two. Clearly $R$ contains a $(-1)$-curve. Furthermore, the only components of $D$ that can possibly have self intersection $-1$ are $C'$ and $H$.\ We claim that $R= C' \cup H$. First note that $H$ is contained in $R$. Indeed, if $H$ is not contained in $R$, it forces $(C')^2=-1$ and $C' \subset R$. But $C'$ is a tip of $D$ and $R$ is a rivet of $f$. Thus, the unique component of $D$ which is adjacent to $C'$ (that is $H$) has to be in $R$, a contradiction. Thus, $H \subset R$. Now at least one of the irreducible components of $D$ adjacent to $H$ must be horizontal to the fibration, since otherwise we will have a $(-1)$-curve with branching number three in a singular fiber of a ${\mbox{$\mathbb P$}}^1$ fibration. Next, if one of the irreducible components of $D$ adjacent to $H$ is a cross-section then $H$ is reduced in $R$. Therefore $H$ cannot have branching number two in $R$. This shows that $H$ meets both sections of $f$ in $D$. This forces $C'^2=-1$ and $R=H \cup C'$. This proves our claim.\ Using the claim and the fact that $T_2$ is a linear chain of rational curves, we see that the branching number of one of the horizontal components of $f$ is two. By [@palka12 Theorem 5.1], ${{\overline}\kappa}({\widetilde{S}}-D) \geq 0$, a contradiction. Proof of Theorem \[main\] (1),(2) ================================= Theorem \[main\] (1) directly follows from Lemma \[1sing\] by taking $S'$ as a minimal resolution of $S$. Now we prove Theorem \[main\] (2). The case ${\overline}\kappa(\tilde{S}-D) =0$ -------------------------------------------- If $C$ has at least three cusps, then $C'$ has more than 2 branches in $D$. Furthermore, each branching component is attached to $C'$ by means of a $(-1)$-curve each having branching number 3 in $D$. By Lemma \[Fujita2\], $Bk(D) \neq N$. Then, there is a $(-1)$-curve not contained in $D$ satisfying one of the three conditions in Lemma \[Fujita1\]. By blowing down this curve and any other $(-1)$-curves that arise as a result, $D$ maps to one of the divisors listed in Lemma \[Fujita2\]. In our case, only the cases (2) and (4) in Lemma \[Fujita2\] is possible. The image of $C'$ under the blowing downs have branching number at least $3$ as the branches of $C'$ are attached to it by $(-1)$-curves. Hence, by Lemma $\ref{Fujita2}$ (2) and (4), at least two of the branches of the image of $C'$ will have to be contractible linear chains of rational curves whose tips meet the image of $C'$. This is not possible as the branches of $C'$ are attached to $C'$ by a $(-1)$-curve which is not a tip of a rational twig.\ Assume that $C$ has more than two cusps. Then $D$ has more than two $(-1)$-curves with branching number 3. By Lemma \[Fujita2\], $Bk(D) \neq N$. Then, by Lemma \[Fujita1\], there is a $(-1)$-curve not contained in $D$ satisfying one of the three conditions. By blowing down this curve and any other $(-1)$-curves that arise as a result, $D$ maps to one of the divisors listed in Lemma \[Fujita2\]. But all the possibilities listed in Lemma \[Fujita2\] have at most two branching component points with branching number 3. From this we deduce that at least one of the $(-1)$-curves in $D$ will be mapped to a curve with non-negative self-intersection, We denote this curve by $J$. Thus $J$ cannot be in a twig. Thus $J$ cannot have branching number 1, hence it has branching number 2. The only possibility for a curve of non-negative self-intersection to have branching number 2, by \[Fujita2\], is if $J$ is adjacent to a $(-1)$-curve $J'$ with branching number 3. But the pre-image of $J$ in $D$ was not adjacent to such a curve, hence the inverse image of $J'$ in $D$ was not a $(-1)$-curve. Thus are at least two $(-1)$ cruves in $D$ which are mapped to curves with non-negative self intersection. Both are not in a twig.\ The case ${\overline}\kappa(\tilde{S}-D) =1$ -------------------------------------------- In this case, because $X$ is affine, by [@miyanishi01 Chapter 3, Theorem 1.7.1] originally due to Y. Kawamata, there is a ${\mbox{$\mathbb C$}}^$-fibration on ${\widetilde{S}}-D$.\ Assume that $C$ has at least three cusps. Then, $D$ has at least three $(-1)$-curves each of which is a branching component with branching number three. Furthermore, none of them are adjacent to each other.\ The ${\mbox{$\mathbb C$}}^$-fibration has no base points on $D$. Indeed, if otherwise, there are at most two base points thus there is at least one $(-1)$-curve in $D$ on which there is no base point. Thus, after resolution of base locus, this $(-1)$-curve is in a fiber of a ${\mbox{$\mathbb P$}}^1$-fibration and has branching number $3$ in the fiber. This is not possible.\ Now that the ${\mbox{$\mathbb C$}}^$-fibration extends to a ${\mbox{$\mathbb P$}}^1$-fibration on ${\widetilde{S}}$, we claim that none of the $(-1)$-curves in $D$ can be contained in a fiber: If a $(-1)$-curve $l$ from $D$ is contained in a fiber $F$ of the ${\mbox{$\mathbb P$}}^1$-fibration $f$ then $l$ cannot have branching number three in $F$. Thus at least one of the curves in $D$ adjacent to $l$ is horizontal to $f$. But this makes $l$ a reduced $(-1)$-curve in $F$. Hence $l$ cannot have branching number two in $F$. Thus, two curves adjacent to $l$ are horizontal to $f$. None of them are $(-1)$-curves, hence the other $(-1)$-curves in $D$ are in the fibers of $f$ as there are only two components of $D$ which are horizontal to $f$. Let $l'$ be the other $(-1)$-curve in $D$ which is contained in a fiber $F'$ of $f$. By a similar reasoning, two of the branching components of $l'$ are horizontal to $f$. Thus, $D$ contains a loop. This is not possible as $D$ is a tree of rational curves. Hence none of the $(-1)$-curves are in fibers of $f$ thus all the $(-1)$-curves in $D$ are horizontal to $f$. But there are exactly two components of $D$ which are horizontal to $f$. Hence $D$ cannot have more than two $(-1)$-curves which is a contradiction. Proof of Theorem 1.3 (3) ============================ We prove Theorem \[main\] (3). Assume that ${{\overline}\kappa}({\widetilde{S}}-D-E) = 0$. By Palka (Theorem 6.1 [@palka12]), if ${{\overline}\kappa}({\widetilde{S}}-D-E)=0$, then there is a twisted ${\mbox{$\mathbb C$}}^$-fibration on ${\widetilde{S}}-D-E$ or $S-C$ is an exceptional ${\mbox{$\mathbb Q$}}$-homology plane as described in [@palka11]. It is easy to check that exceptional ${\mbox{$\mathbb Q$}}$-homology planes do not occur since $D$ is the SNC resolution of a cuspidal rational curve([@palka11 Proposition 4.4]). Now the following lemma gives a contradiction. \[untwisted\] If $ {{\overline}\kappa}({\widetilde{S}}-D-E) =0$ or $1$, then every ${\mbox{$\mathbb C$}}^$-fibration on ${\widetilde{S}}-D-E$ is untwisted. The ${\mbox{$\mathbb C$}}^$-fibration $f$ on ${\widetilde{S}}-D-E$ can be extended to a ${\mbox{$\mathbb C$}}^$-fibration on ${\widetilde{S}}-D$. That is to say, $E$ is contained in fibers of $f$. Indeed, if otherwise, an irreducible component of $E$ is horizontal for $f$. Thus, $\tilde{S} - D$ contains a cylinder-like open set. Hence, by ([@miya81], Chapter 1, Theorem 3.13) that ${{\overline}\kappa}({\widetilde{S}}-D)=-\infty$. By Lemma \[inf\], this implies ${{\overline}\kappa}({\widetilde{S}}-D-E)=-\infty$, a contradiction.\ Since $ {{\overline}\kappa}({\widetilde{S}}-D-E) < 2$, we have ${{\overline}\kappa}({\widetilde{S}}-D) <2$, thus $C$ has at most two cusps by Theorem 1.3 (1) and (2).\ Case 1. C is uni-cuspidal.\ We use the notation in Lemma \[inf\] in which a description of $D$, when $C$ is uni-cuspidal, is given.\ We claim that $C'$ is not a $(-1)$-curve. Indeed, if $C'$ is a $(-1)$-curve, then we can blow it down and map $D$ to a divisor with a non-contractible twig, which would imply by Lemma \[contractible\] that ${{\overline}\kappa}({\widetilde{S}}-D)=-\infty$. But, then, Lemma \[inf\] implies that ${{\overline}\kappa}({\widetilde{S}}-D-E)=-\infty$, a contradiction.\ Assume that $f$ is a twisted ${\mbox{$\mathbb C$}}^$-fibration on ${\widetilde{S}}-D-E$. By [@miya91 Lemma 2.10], the base of the ${\mbox{$\mathbb C$}}^$-fibration $f$ is ${\mbox{$\mathbb A$}}^1$. Furthermore, all singular fibers of $f$ are irreducible and exactly one of them is an affine line (under the reduced structure). By [@palka12 Lemma 4.1], the fiber at infinity is of the type $[2,1,2]$.\ We claim that $f$ has no base points on $D$. If $f$ has a base point on $D$, it has to be unique. Indeed, if $f$ has two base points on $D$, then $f$ is an untwisted fibration. Furthermore, the base point of $f$ will have to lie on $H$ as $H$ is the unique $(-1)$-curve in $D$ with branching number three. Let $D'$ be the total transform of $D$ under the resolution of base locus. As the base point is on $H$, $D'$ has a unique $(-1)$-curve which is horizontal to $f$. But the fiber at infinity must also contain a $(-1)$-curve, a contradiction.\ Since $f$ has no base points on $D$, $H$ is the unique $(-1)$-curve in $D$ and is contained in the fiber at infinity which is of the type $[2,1,2]$. Furthermore, $C'$ is a $(-2)$-curve contained in the fiber at infinity and the linear chain $T_2$ is an irreducible $(-2)$-curve. Let $\theta$ denote the 2-section of $f$ which is an irreducible component of $D$ adjacent to $H$. As $D$ can be blown down to $C$, it is easy to deduce that $\theta$ is a $(-3)$-curve.\ Recall that $\theta$ has branching number at most three. First, we show that $\theta$ has at most two branches. Assume that $\theta$ has branching number three. By $M_1$ and $M_2$ we denote the two branches of $\theta$ in $D$ other than $H$. Note that $M_1$ and $M_2$ contain no $(-1)$-curves. If $M_1$ and $M_2$ are in separate fibers, then both fibers will contain an affine line ${\mbox{$\mathbb A$}}^1$ for the fibration on ${\widetilde{S}}-D-E$. This contradicts [@miya91 Lemma 2.10]. Hence $M_1$ and $M_2$ lie in the same singular fiber of $f$. Let $F_0$ be the fiber of $f$ containing both $M_1$ and $M_2$. Then, by [@miya91 Proof of Lemma 1.4], $F_0 \cap ({\widetilde{S}}-D)$ contains an affine line ${\mbox{$\mathbb A$}}^1$. Thus, $F_0$ contains at least one more irreducible component which meets both $M_1$ and $M_2$ as the fiber is connected. Thus, $F_0$ is reducible which again contradicts to [@miya91 Lemma 2.10].\ Now we show that $\theta$ has only one branch, that is, $H$. In other words, $T_3$ is irreducible. Let $M_1$ be the branch of $\theta$ other than $H$. Let $F_{M_1}$ be the fiber containing $M_1$. This fiber meets $\theta$ in only one distinct point, hence the component of $F_{M_1}$ which meets $\theta$ has multiplicity 2 and hence has branching number 2 in $F_{M_1}$. There is a $(-1)$-curve in $F_{M_1}$ which meets $M_1$. We blow this curve down and the subsequent $(-1)$-curves until the image of $D$ is minimally SNC. By Lemma \[Fujita2\], $F_{M_1}$ maps to a fiber of the type $[2,1,2]$ and $\theta$ maps to a $(-3)$-curve as it is not affected during the process of blowing down. If we blow down the singular fibers of the resulting ${\mbox{$\mathbb P$}}^1$-fibration to irreducible curves, the image of $\theta$ is a smooth 2-section with self-intersection 1. By elementary properties of ruled surfaces this is not possible. Thus, $\theta$ has only one branch.\ This shows that the number of singular fibers of the ${\mbox{$\mathbb P$}}^1$-fibration $f$ is exactly one. Hence we have an open subset of ${\widetilde{S}}-D-E$ which is a twisted ${\mbox{$\mathbb C$}}^$-bundle over ${\mbox{$\mathbb C$}}^$. Either this is a trivial ${\mbox{$\mathbb C$}}^$-bundle or a quotient of a trivial ${\mbox{$\mathbb C$}}^$-bundle modulo a fixed-point free action of a group of order $2$. Thus, ${{\overline}\kappa}({\widetilde{S}}-D-E)=0$.\ Let $M_1$ be the branch of $\theta$ other than the one containing $H$. There is a $(-1)$-curve meeting $M_1$. We can blow down this $(-1)$-curve and subsequent $(-1)$-curves to obtain an irreducible fiber of $f$. Hence $D$ has a unique branching component, i.e., the image of $H$. Moreover, the image of $H$ has three branches such that $\Sigma 1/d_i > 1$. This is impossible by Lemma \[Fujita2\].\ Case 2. C is bi-cuspidal.\ We begin with a description of $D$ when $C$ is bi-cuspidal. The branching number of $C'$ is two. It is adjacent to two $(-1)$-curves $H_1$ and $H_2$ both having branching number three. By $B_{11}$ and $B_{12}$ we denote the branches of $H_1$ other than $C'$. By $B_{22}$ and $B_{21}$ we denote the branches of $H_2$ other than $C'$. We can assume $B_{11}$ and $B_{22}$ to be linear. Both branches of $C'$ can be blown down to smooth points.\ Recall that we have a twisted fibration $f$ on $\tilde{S}-D-E$. We claim that the fibration $f$ has no base point on $D$. Indeed, if otherwise, as we argued in Case 1, the base point of $f$ on $D$ have to be unique because $f$ is twisted. Also, the base point of $f$ has to lie on one of the $(-1)$-curves with branching number three. This forces the other $(-1)$-curve with branching number three to be in a singular fiber and have branching number three in the fiber. This is not possible.\ Now that $f$ has no base points on $D$, by [@palka12 Lemma 4.1], there is a fiber at infinity of type $[2,1,2]$. This is impossible by the description of $D$ above. Examples {#example} ============ In this section we give three examples of pairs $(S,C)$ such that ${{\overline}\kappa}(S-C)<2$. Example A {#ExA} --------- We start with the Hirzebruch surface $\Sigma_0$. We fix a fiber $C'$ of one of the ${\mbox{$\mathbb P$}}^1$-fibrations, say $f$, and a cross-section $H_0$ on $\Sigma_0$. By choosing two other fibers of $f$ we blow up suitably to produce linear chains of the type $[2,2,1,3]$ and $[4,2,1,3,2,2]$. We contract the subchains $[3,2,2]$ and $[3]$ in these subchains. The proper transform $H$ of $H_0$ is a $(-1)$-curve. Let $S$ be the surface obtained by blowing down $H$ and the two linear chains $[4,2]$ and $[2,2]$ to a smooth point on a normal surface $S$. The image of $C'$ in $S$ is a unicuspidal rational curve $C$ contained in $S^0$. Both singular points of $S$ are not rational double points. One can see that ${{\overline}\kappa}({\widetilde{S}}-D)=-\infty$ and ${{\overline}\kappa}({\widetilde{S}}-D-E)=-\infty$. Example B {#ExB} --------- We start with the Hirzebruch Surface $\Sigma_1$. Let $E$ be the minimal section with negative self-intersection. There exists a smooth rational curve $C$ which meets every fiber twice and is disjoint from $E$. Let $F_1$ and $F_2$ be the fibers that meet $C$ tangentially. Let the points of intersection of $F_1$ and $F_2$ with $E$ be $p_1$ and $p_2$ respectively. We perform elementary transforms by blowing up at $p_1$ and $p_2$ and then blowing down the proper transforms of $F_1$ and $F_2$. In this process $C$ acquires two cusps and $E^2=-3$. Now blow down $E$ to a singular point. One can see that ${{\overline}\kappa}({\widetilde{S}}-D)=0$ and ${{\overline}\kappa}({\widetilde{S}}-D-E)=1$. Example C {#ExC} --------- Consider a cubic $C$ on ${\mbox{$\mathbb C$}}{\mbox{$\mathbb P$}}^2$ with a single cusp. Let $T$ be the tangent line to $C$ at the cusp. Let $L$ be the tangent to $C$ at a smooth point of inflection. Let $p= T \cap L$, clearly $p$ is not on $C$. We blow up at $p$. We have a fibration with the proper transforms of $T$ and $L$ are fibers. Let $q$ be the point where the proper transform $L'$ of $L$ and $E$ meet. Here $E$ is the exceptional curve. We perform elementary transforms by blowing up at the point $q$ and blowing down the proper transform of $L$. Thus, $C$ has two cusps. One can see that ${{\overline}\kappa}({\widetilde{S}}-D)=1$ and ${{\overline}\kappa}({\widetilde{S}}-D-E)=1$. Acknowledgement {#acknowledgement .unnumbered} =============== D. Hwang was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(NRF-2015R1D1A1A01060179). R.V. Gurjar\ Tata Institute of Fundamental Research, Mumbai.\ [email protected]\ DongSeon Hwang\ Ajou University, Republic of Korea.\ [email protected]\ Sagar Kolte\ Tata Institute of Fundamental Research, Mumbai.\ [email protected]\ [^1]: Mathematics Subject Classification: 14J26, 14R25, 14H45	{ "pile_set_name": "ArXiv" }
--- abstract: 'A modular, maintainable and extensible particle beam simulation architecture is presented. Design considerations for single particle, multi particle, and rms envelope simulations (in two and three dimensions) are outlined. Envelope simulation results have been validated against Trace3D. Hybridization with a physics-centric contol-system abstraction provides a convenient environment for rapid deployment of applications employing model-reference control strategies.' author: - \| Nicholas D. Pattengale Christopher K. Allen\ `nickp,[email protected]` date: 'June 10, 2002' title: \| LA-UR-02-3761\ The Software Anatomy of a Flexible Accelerator Simulation Engine --- Background ========== Discovering a Simulation Architecture ------------------------------------- Our group has designed and implemented a unified accelerator Application Programming Interface (API) called XAL[@XAL]. XAL is designed to aid in the development of science control applications for beam physics. Accordingly, the XAL API is a physics-centric software programming interface. The physics applications interact with a model of an accelerator that resides in computer memory. XAL also contains the software infrastructure that creates the accelerator model. XAL loads a text-based (XML) description of an accelerator and assembles software objects such that an accurate model of the accelerator exists in computer memory. XAL is based on UAL [@UAL], the Unified Accelerator Library. The original motivation for XAL was to provide an accelerator independent interface for applications to interact with I/O from a live accelerator. This allows physicists to write beam physics control applications (Orbit Correctors, Beam Profile Monitors, RF Tuners, etc.) to the XAL API so that they can run on any accelerator. Some pseudo-code illustrating the principles of an XAL-based Orbit Correction Application may illustrate the essence of the concept. Accelerator theAccel = XALFactory.newAccel(``sns.xml'') BPM[] theBPMs = theAccel.getNodesOfType(BPM) HorzDipole[] theCorrectors = theAccel.getNodesOfType(DCH) for each BPM in theBPMs read BPM.avgPos() and set a corrector magnet accordingly To aid in writing applications that take into account design values, the accelerator description file contains all design information for the accelerator. This condition allows, for example, a physics application to compare the design field of a quadrupole with its read-back (runtime) field. With all design information incorporated into a software model of an accelerator, we have discovered an excellent simulation engine. As long as the software accelerator has a convenient means for traversing beam-line devices in a spatially sequential manner, we can use design values along the way to simulate beam-dynamics. This scenario allows for a drastic departure from traditional accelerator simulation codes. Traditionally simulators have been isolated software products. They load some type of lattice description of an accelerator and apply predefined beam-dynamics to an initial beam. Ultimately this design has led to huge codes (to account for various beam-line element types). Further, these codes typically operate with only one type of simulation (multi-particle or rms envelope, but not both). The architecture presented here contains a novel approach to the simulation domain. It is our conjecture that the method presented here better captures reality in that there is some sort of software beam actually traversing a software model of a real accelerator. The Architecture ---------------- Our approach is based upon the Element-Algorithm-Probe Design Pattern [@EAP]. The core concept of this design pattern is the separation of beam-dynamics code from the actual beam-line elements. It is desirable to keep the code that corresponds to beam-line elements as simple as possible so that the application writer has a clean interface to a beam-line element. The Element-Algorithm-Probe pattern enforces this concept by requiring beam-dynamics code to exist in a separate entity, called an IAlgorithm . Deferred until runtime is the binding of beam-dynamics to actual beam-line elements. This deployment strategy allows for conceptually correct simulations. First it is truly modular. The three concepts, beam-line elements, beam-dynamics, and the beam are compartmentalized into separate code. Second it is truly maintainable. To support a new beam type or new beam-line element type does not cause code bloat. Finally it is truly extensible. Via the mechanism of a Java interface, various beam-dynamics algorithms can be written for the same type of beam-line element and switched at will at runtime. Modularity, maintainability, and extensibility provide true power and flexibility to our architecture. IProbe, IAlgorithm, and IElement ================================ Technology Introduction ----------------------- It may help to understand the facets of Java that we exploit in order to implement the Element-Algorithm-Probe Design Pattern. At the center of the Element-Algorithm-Probe pattern is the concept of a Java interface. Essentially, an interface is a contract between a user and an implementor. The contract says that the implementor of an interface is required to provide an implementation of the methods defined in the interface. For example, consider the interface public interface Thermometer { public double getTemperature(); } Using this interface, a programmer can assume being able to perform operations on a thermometer no matter how the thermometer actually obtains the temperature. This is desirable because a thermometer implementor can change how the temperature is actually obtained (if, say, a new sensor system was installed) without requiring all thermometer users to recompile their code. We use the same idea with beam-dynamics code. Beam-dynamics reside in files that implement (the computer science term for acknowledging involvement in the contract from the implementors point of view) the IAlgorithm interface. Since the simulation engine knows how to do beam-dynamics calculations solely by interacting with IAlgorithms, it is trivial to swap beam-dynamics algorithms at will. The IAlgorithm interface looks like this. public interface IAlgorithm { public void propagate(IElement, IProbe); public Class legalElementType(); public Class legalProbeType(); } Conceptually, an IAlgorithm implementor is required provide an implementation of the method propagate() to modify the the beam (IProbe) according to the beam dynamics of the the beam-line element (IElement). In essence, all that the simulation engine knows about are the three data types (all defined in interfaces) IAlgorithm, IProbe, and IElement. The beauty of this design is that there are separate code locations for beam-line elements, beam-dynamics, and the beam itself. Probes ------ The IProbe interface should ideally contain the bare minimum information to fully represent a beam. Such beam information consists of beam current, beam charge, particle charge, particle rest energy, particle kinetic energy, etc. Further, since a probe represents the state of the beam at a position in the beam-line, a probe also contains a beam-line-position attribute. The current IProbe specification serves the purpose of representing a beam for a single particle , particle ensemble, and envelope simulations (in both two and three dimensions). Figure 1 represents a suitable inheritance hierarchy of probe types to handle the aforementioned simulation types. Propagation ----------- It is important to note that there are various approaches toward simulating beam-dynamics. For example, accurate approaches may involve slicing nodes up into small pieces. An aggregate of approximations done on sufficiently small elements is typically more accurate than one overall approximation. However, normally this is only practical in elements that have special behavior. So the question arises: How are probes propagated through elements? It is the responsibility of the IAlgorithm implementor to handle all beam-dynamics, including the propagation mechanism. Sample propagation mechanisms will be presented later in this paper. However, keep in mind the most appropriate propagation mechanism when implementing algorithms for the particular problem at hand. Algorithms ---------- We have already introduced the concept of the IAlgorithm interface. Now let us pursue a few more details regarding its implementation. The IAlgorithm interface provides a generic way of assembling algorithms in a simulation engine. In practice any particular IAlgorithm implementation only makes sense in the context of a particular beam-line element type and probe type. For example, a hypothetical IAlgorithm called QuadParticleMapper would expect a Quadrupole as its IElement and a Particle as its IProbe . Providing such specificity is the job of the legalElementType() and legalProbeType() methods. An implementation of the QuadParticleMapper could look like this. public class QuadParticleMapper(IElement p_elem, IProbe p_probe){ public Class legalElementType(){return Quadrupole.class;} public Class legalProbeType(){return Particle.class;} public void propagate(){Quadrupole/Particle beam dynamics} } By providing these methods, the simulation engine can do type checking upon algorithm binding. It would not make sense to bind this algorithm to a WireScanner. Providing these methods helps to avoid that condition. Design of a Single Particle Simulation ====================================== Designing an actual simulation merely involves putting together Elements, Algorithms, and Probes in a semantically meaningful way. It turns out that, to the first order, the beam dynamics through a particular node type can be captured by a transfer matrix. This property allows for a straight-forward means of simulating a particle traveling down a beam-line. An object-oriented approach would be to create a ParticleMapper class that transforms the Particle probe by the simple vector-matrix multiplication $$\vec{\bf{z}}_{n + 1} = \Phi_n \cdot \vec{\bf{z}}_n$$ where $\vec{\bf{z}}_n$ is the coordinate vector of the particle $\left(\begin{array}{cccccc} x & x' & y & y' & z & z' \end{array}\right)^T$ at the start of the node, and $\Phi_n$ is the transfer matrix of the node. Further, $\Phi_n$ can be obtained by the ParticleMapper via the use of an abstract method that is implemented by beam-dynamics algorithms for individual nodes (QuadrupoleParticleMapper, RFCavityParticleMapper, etc.). A suitable class design can be seen in Figure 2. To further illustrate some of these concepts, the basic layout of the ParticleMapper class looks like this. abstract public Matrix computeTransferMatrix(); public void propagate(IElement pElem, IProbe pProbe) { //type-cast the probe and element to what we expect Particle theProbe = ((Particle)pProbe); AcceleratorNode theNode = ((AcceleratorNode)pElem); //do the vector-matrix multiplication theProbe.setCoords(computeTransferMatrix .times(theProbe.getCoords())); //advance the probe the length of the node theProbe.advancePosition(theNode.getLength()); return; } Once the computeTransferMatrix() operations are implemented for the node-specific dynamics, all that remains is writing a driver program. A driver program binds algorithms to nodes and injects the probe. Here is a pseudo-code driver. //instantiate the XAL accelerator model Accelerator theAccel = XALFactory.newAccel(``sns.xml'') //bind the algorithms Quadrupole[] theQuads = theAccel.getNodesOfType(QUAD) RFCavity[] theCavities = theAccel.getNodesOfType(RFC) for each QUAD in theQuads bind a QuadParticleMapper instance to QUAD for each RFCav in theCavities bind a RFCavityParticleMapper instance to RFCav //instantiate a probe Particle p1 = new Particle(initial conditions...) //run the probe down the beam-line AcceleratorNode[] theNodes = theAccel.getAllNodes(); for each NODE in theNodes NODE.propagate(p1) And that is it! The particle probe will be transformed by each beam-line element according the the bound algorithm. Note that the pseudo-code is a basic proof of concept and does not contain the code necessary to broadcast probe increment intermediate data to produce, for example, a plot. The single particle simulation can be applied to a two-dimensional case by only considering the first four elements of $\vec{\bf{z}}$. Further, the single particle simulation can be extended to a multi-particle simulation (in two and three dimensions) by constructing a container of particle probes and writing beam-dynamics algorithms that properly transform the collection. The only matter that complicates (and complicate it does!) a multi-particle simulation is the concept of space-charge. Before biting off this task, however, a presentation of another type of simulation that accounts for space-charge is warranted. Design of an RMS Envelope Simulation ==================================== The Concept ----------- The RMS qualities of a beam can be represented by the 6x6 symmetric matrix $\sigma$ that statistically expresses the boundaries of a beam in transverse, longitudinal, and phase space by using moments of the beam distribution. RMS Envelopes are convenient because applying beam-dynamics involves a simple matrix operation. Namely, the same transfer matrix $\Phi$ used in single particle simulations can propagate rms envelopes according to the conjugation $$\sigma_{n+1} = \Phi \cdot \sigma_n \cdot \Phi^T$$ The other important concept in this simulation is space charge. A $\sigma$ matrix is a statistical representation of a beam, which is a multi-particle entity. Therefore, each particle in the beam is aware (electromagnetically) of all other particles in the beam. It turns out that to the first order the effects of space charge can be captured in a $\Phi$ matrix. While it may not be mathematically trivial to calculate the matrix, having the calculation in such a form makes the integration into our simulation engine simple. However it should not be overlooked that this quantity is very important to the correctness of simulation. Propagation ----------- The envelope simulation is more complex than a single-particle simulation in that we will propagate envelopes through elements using more than one propagation mechanism. Specifically, we may be able to compute a better approximation of behavior through quadrupoles than RF Cavities. This condition is due to the fact that the $\Phi$ matrix for a quadrupole adheres to the semi-group property. $$\Phi(\Delta s_1 + \Delta s_2) = \Phi(\Delta s_1) \cdot \Phi(\Delta s_2)$$ or $$\Phi(n \cdot \Delta s) = \Phi^n(\Delta s)$$ where $\Delta s$ is the length of the quadrupole being considered. To more accurately consider space charge, we take advantage of the semi-group property of the transfer matrices. In the propagate() method of the SemiGroupEnvelopeMapper (see Figure 3) we subsection the node (e.g., a quadrupole) into $n$ slices of length $\Delta s = {l \over n}$ where $l$ is the length of the quadrupole. Then we run the probe through these $n$ slices, applying space charge kicks after every subsection (See Figure 4). Since RF Cavity transfer ($\Phi$) matrices do not in general adhere to a semi-group property, we are forced to take a more simplistic approach toward transforming the envelope. We will slice the node in two, treating each half as a drift-space (to account for space charge) and hit the envelope in the middle of the node with the numerically approximated $\Phi$ matrix (see Figure 5). Design of a Particle Ensemble Simulation ======================================== As a final exercise it will be useful to consider the design of a multi-particle simulation. The true complication of designing a multiple-particle (ensemble) simulation is the computation of space-charge effects. Unfortunately, to model multiple particles, space-charge effects cannot be accurately captured by a transfer matrix. On the other hand, the architecture outlined in this paper keeps the details of the space-charge calculations from interfering with code cleanliness. The two core concepts of a multi-particle (ensemble) simulation are - Calculation of the electric self-fields of the ensemble - Using the calculated fields to update the particle coordinates. There are various approaches that can be taken for both tasks. All that we attempt to show here is that by correctly isolating these concepts, a clean software architecture can be maintained. Namely, an Ensemble probe should encapsulate the logic necessary to obtain the electric self fields of the ensemble. That being the case, various Ensemble probe implementations could be swapped at will to employ different field calculation techniques. For example, many electric field calculation techniques involve solving Poisson’s equation to obtain the electric potential of the ensemble. By hiding this code from the simulation engine (it is contained within the Ensemble probe implementation), the implementor could exploit parallel processing facilities (See Figure 6). By moving the calculation of electric fields out of the beam-dynamics code, the beam-dynamics algorithm developer is free to choose space-charge consideration techniques with minimal impact to code clarity. One may decide to take the “thin lens kick” approach that has been used previously in this paper. One may alternatively decide to apply a “trajectory integration” based approach. The key point here is that by separating codes into their logical components allows for a high degree of flexibility in simulation technique. Conclusions and Future Directions ================================= We are enthusiastic to report that the results obtained in the RMS Envelope Simulation have been validated against Trace3D [@Trace3D] Figure 7 shows agreement between simulation results of the SNS Medium Energy Beam Transport (MEBT) using both Trace3D and the XAL simulation engine. It is encouraging that a problem domain with so many interdependencies(particle physics) can be simulated with a clean architecture. As we move toward the future, we are anticipating the ability to implement model reference control techniques. That is, within the XAL model there is access to a live accelerator and a simulated accelerator. Having both at hand allows the comparison of live behavior with simulated behavior to develop control strategies. The key to effectively implementing an environment conducive to model reference control is architectural discipline when designing both the I/O and simulation aspects of XAL. As long as the interface to the two are respectively clean, hybridization of the two will be a straightforward extension. [2]{} J. Galambos, C.M. Chu, T.A. Pelaia, A. Shishlo, C.K. Allen, N. Pattengale. “SNS Application Programming Environment”, EPAC 2002 N. Malitsky and R. Talman. “Unified Accelerator Libraries”, AIP 391(1996) N. Malitsky and R. Talman. “The Framework of Unified Accelerator Libraries”, ICAP 1998 C. K. Allen and N. D. Pattengale. “Simulation of Bunched Beams with Ellipsoidal Symmetry and Linear Space Charge Effects”, LANL Technical Report 2002 K. Crandall, D. P. Rusthoi, “TRACE 3-D Documentation,” Los Alamos National Laboratory report LA-UR-97-886, May 1997.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Indoor localization for autonomous micro aerial vehicles (MAVs) requires specific localization techniques, since the Global Positioning System (GPS) is usually not available. We present an efficient onboard computer vision approach that estimates 2D positions of an MAV in real-time. This global localization system does not suffer from error accumulation over time and uses a $k$-Nearest Neighbors ($k$-NN) algorithm to predict positions based on textons—small characteristic image patches that capture the texture of an environment. A particle filter aggregates the estimates and resolves positional ambiguities. To predict the performance of the approach in a given setting, we developed an evaluation technique that compares environments and identifies critical areas within them. We conducted flight tests to demonstrate the applicability of our approach. The algorithm has a localization accuracy of approximately 0.6m on a 5m$\times$5m area at a runtime of 32 ms on board of an MAV. Based on random sampling, its computational effort is scalable to different platforms, trading off speed and accuracy.' author: - 'V. Strobel' - 'R. Meertens' - 'G.C.H.E. de Croon' bibliography: - 'library.bib' title: - 'Preparation of Papers for IEEE Sponsored Conferences & Symposia\* ' - Efficient Global Indoor Localization for Micro Aerial Vehicles --- INTRODUCTION {#sec:introduction} ============ Accurate onboard localization is a key challenge for micro aerial vehicles (MAV). In confined spaces, specific localization algorithms are essential, since the Global Positioning System (GPS) is usually not available. While light-weight MAVs could be employed in various indoor tasks, they cannot fall back on standard localization approaches due to their limited payload and processing power. To address this issue, this paper presents an efficient indoor localization technique. ![[\[fig:highleveloverview\] The figure illustrates the presented system from a high-level perspective. A feature vector—the texton histogram—is extracted from the current camera image. The vector is forwarded to a machine learning model that uses a $k$-Nearest Neighbors algorithm to output $k$ $x,y$-position estimates. These estimates are passed to a particle filter, which filters position estimates over time and outputs a final position estimate (red point). The expected loss shows regions in the map where a lower localization accuracy is expected. The average expected loss can be used as “fitness value” of a given map. ]{}](nutshell){width="\columnwidth"} Our contribution is a machine learning-based indoor localization system that runs onboard of an MAV paving the way to an autonomous system. In the presented approach, computational power is shifted to an offline training phase to achieve high-speed during live operation. In contrast to visual SLAM frameworks, this project considers scenarios in which the environment is known beforehand or can even be actively modified. The approach is based on the occurrence of textons, which are small characteristic image patches. With textons as image features and a $k$-Nearest Neighbors ($k$-NN) algorithm, we obtain 2D positions in real-time within a known indoor environment. A particle filter was developed that handles the estimates of the $k$-NN algorithm and resolves positional ambiguities. We consider settings in which the MAV moves at an approximately constant height, such that the estimation of height is not necessary. In contrast to existing approaches that use active sensors, the developed approach only uses a passive monocular downward-looking camera. While carrying active sensors, such as laser range finders, is too demanding for a light-weight MAV, onboard cameras can typically be attached. Additionally, we developed a technique for evaluating the suitability of a given environment for the presented algorithm. It identifies critical areas and assigns a global loss value to an environment. This allows for comparing different potential maps and identifying regions with low expected localization accuracy. The developed global localization system does not suffer from error accumulation over time. Since it is intended to further reduce the size of MAVs, lightweight and scalable position estimation algorithms are needed. Onboard processing helps to reduce errors and delays introduced by wireless communication, and ensures a high versatility on the way to an autonomous system. The validity of the approach is evaluated in flight experiments. An overview of the presented approach can be seen in Figure \[fig:highleveloverview\]. The remainder of this paper is structured as follows. Section \[sec:related\] surveys existing indoor localization approaches. In Section \[sec:methods\], the developed texton-based approach is presented and its components, the $k$-NN algorithm and the particle filter, are introduced. Section \[sec:experiments\] describes the setup and results of the flight experiments. The results are discussed in Section \[sec:discussion\] and we draw conclusions in Section \[sec:conclusion\]. RELATED WORK {#sec:related} ============= While a wide range of methods for indoor localization exists, we only consider methods in this section that use the same technical and conceptual setup—localization with a monocular camera. One distinguishes two types of robot localization: local techniques and global techniques [@fox1999monte]. Local techniques need an initial reference point and estimate coordinates based on the change in position over time. Once they lost track, the position can typically not be recovered. The approaches also suffer from “drift” since errors are accumulating over time. Global techniques are more powerful and do not need an initial reference point. They can recover when temporarily losing track and address the kidnapped robot problem, in which a robot is carried to an arbitrary location [@engelson1992error]. Optical Flow {#sec:opticalflow} ------------ Optical flow algorithms estimate the apparent motion between successive images. The most popular optical flow methods are gradient based approaches and keypoint-based methods [@chao2013survey]. Optical flow methods belong to the class of local localization techniques and can only estimate the position relative to an initial reference point. The approaches suffer from accumulating errors over time and typically do not provide a means for correcting these errors. Most approaches are computationally rather complex [@chao2013survey]. Fiducial Markers {#sec:fiducialmarkers} ---------------- Fiducial markers have been used for UAV localization and landing [@eberli2011vision; @bebop2015]. The markers encode information by the spatial arrangement of black and white or colored image patches. Their corners can be used for estimating the camera pose at a high frequency. An advantage of fiducial markers is their widespread use, leading to technically mature and open-source libraries. A drawback of the approach is that motion blur, which frequently occurs during flight, can hinder the detection of markers [@albasiouny2015mean]. Furthermore, partial occlusion of the markers through objects or shadows break the detection. Another downside is that markers might be considered as visually unpleasant and may not fit into a product or environmental design [@chu2013halftone]. Homography Determination & Keypoint Matching {#sec:keypointmatching} -------------------------------------------- A standard approach for estimating camera pose is detecting and describing keypoints of the current view and a reference image [@se2002global], using algorithms such as Scale-invariant feature transform (<span style="font-variant:small-caps;">Sift</span>) [@lowe1999object], followed by finding a homography—a perspective transformation—between both keypoint sets. A keypoint is a salient image location described by a feature vector. Depending on the algorithm, it is invariant to different viewing angles and scaling. This homography-based approach is employed in frameworks for visual Simultaneous Localization and Mapping (SLAM) but the pipeline of feature detection, description, matching, and pose estimation is computationally complex [@kendall2015posenet]. While the approach has been employed for global localization for UAVs, the required processing power is still too high for small MAVs [@de2009design]. Convolutional Neural Networks ----------------------------- Convolutional neural networks (CNNs) are a specialized machine learning method for image processing [@lecun1998gradient]. The supervised method has outperformed other approaches in many computer vision challenges [@dosovitskiy2014discriminative]. While their training is usually time-consuming, predictions with CNNs often takes only few milliseconds, shifting computational effort from the test phase to the training phase. CNNs have been used as robust alternative for keypoint detection and description if images were perturbed [@dosovitskiy2014discriminative] but needed more computation time than <span style="font-variant:small-caps;">Sift</span>. In recent work, @kendall2015posenet present a framework for regressing camera positions based on CNNs [@kendall2015posenet]. The approach is rather robust to different lighting settings, motion blur, and varying camera intrinsics. The approach predicts positions on a modern desktop computer in short time. Texton-based Methods {#sec:textonbasedapproaches} -------------------- Textons [@varma2005statistical] are small characteristic image patches; their frequency in an image can be used as image feature vector. A texton histogram is obtained by extracting patches from an image and comparing them to all textons in a “texton dictionary”. The frequency of the most similar texton is then incremented in the histogram. Texton histograms are flexible image features and their extraction requires little processing time, which makes them suitable for MAV on-board algorithms. The approach allows for adjusting the computational effort by modifying the amount of extracted image patches, resulting in a trade-off between accuracy and execution frequency [@de2012sub]. @de2009design [@de2009design] use textons as image features to distinguish between three height classes of the MAV during flight. Using a nearest neighbor classifier, their approach achieves a height classification accuracy of approximately 78% on a hold-out test set. This enables a flapping-wing MAV to roughly hold its height during an experiment. In another work, @de2012appearance [@de2012appearance] introduce the appearance variation cue, which is based on textons, for estimating the proximity to objects [@de2012appearance]. Using this method, the MAV achieves a high accuracy for collision detection and can avoid obstacles in a $5\,m \times 5\,m$ office space. METHODS {#sec:methods} ======= The pseudo code in Algorithm \[alg:trexton\_run\] shows a high-level overview of the parts of the framework. Details are given in the following sections. $t \gets 0$ $\mathcal{X}_0 \gets$ $t \gets t+1$ $I_t \gets$ $\mathcal{H}_t \gets$ $\mathbf{z}_t \gets$ $\mathcal{X}_t \gets$ $x_t, y_t \gets$ end Hardware and Software --------------------- We decided to use the quadcopter Parrot Bebop Drone as a prototype for all our tests. The developed approach makes use of the bottom camera only, which has a resolution of 640 $\times$ 480 pixels with a frequency of 30 frames per second. Dataset Generation ------------------ A main idea of the presented method is to shift computational effort to a pre-flight phase. Since the MAV will be used in a fixed environment, the results of these pre-calculations can be employed during the actual flight phase. Supervised machine learning methods need a training set to find a mapping from features to target values. In this first step, the goal is to label images with the physical $x,y$-position of the UAV at the time of taking the image. One possible way to create the data set is to align the images with high-precision position estimates from a motion tracking system, which yields high-quality training sets. Major disadvantages of the approach are that motion tracking systems are usually expensive and time-consuming to move to different environments. As an alternative, we sought a low-budget and more flexible solution. Out of the presented approaches in Section \[sec:related\], the homography-based approach (Section \[sec:keypointmatching\]) promises the highest flexibility with a good accuracy but also requires the most processing time. Since fast processing time is not relevant during the pre-flight phase, the approach is well-suited for the problem. The required image dataset can be obtained by using images gathered during manual flight or by recording images with a hand-held camera. To get a hyperspatial image of the scene for creating a map, the images from the dataset have to be stitched together. With certain software packages the images can be orthrectified by estimating the most probable viewing angle based on the set of all images. However, since a downward-looking camera is attached to the UAV, most images will be roughly aligned with the z-axis, given slow flight [@blosch2010vision]. For the stitching process, we used the freeware software Microsoft Image Composite Editor (ICE) [@ice]. Keypoints of the current image and the map image are detected and described using the <span style="font-variant:small-caps;">Sift</span> algorithm. This is followed by a matching process, that identifies corresponding keypoints between both images. These matches allow for finding a homography between both images. For determining the $x, y$-position of the current image, its center is projected on the reference image using the homography matrix. Texton Dictionary Generation ---------------------------- For learning a suitable texton dictionary for an environment, image patches were clustered. The resulting cluster centers—the prototypes of the clustering result—are the textons [@varma2003texture]. The clustering was performed with a Kohonen network [@kohonen1990self]. The first 100 images of each dataset were used to generate the dictionary. From each image, 1000 randomly selected image patches of size $w \times h = 6 \times 6$ pixels were extracted, yielding $N = 100\,000$ image patches in total that were clustered. For our approach, we also used the color channels U and V from the camera to obtain color textons. Histogram Extraction -------------------- The images from the preliminary dataset are converted to the final training set that consists of texton histograms and $x,y$-values. To extract histograms in the full sampling setting, a small window—or kernel—is convolved across the width and height of an image and patches are extracted from all positions. Each patch is compared with all textons in the dictionary and is labeled with the nearest match based on Euclidean distance. The histogram is normalized by dividing the number of cases in each bin by the total number of extracted patches, to yield the relative frequency of each texton. The convolution is a time-consuming step, since all possible combinations of width and height are considered: $(640 - w + 1) \cdot (480 - h + 1) = 301\,625$ samples are extracted. To speed up the time requirements of the histogram extraction step, the kernel can be applied only to randomly sampled image position instead [@de2012sub]. This sampling step speeds up the creation of the histograms and permits a trade-off between speed and accuracy. The random sampling step introduces random effects into the approach. Therefore, for generating the training dataset, no random sampling was used to obtain high-quality feature vectors. $k$-Nearest Neighbors ($k$-NN) algorithm {#sec:knn} ---------------------------------------- The $k$-Nearest Neighbors ($k$-NN) algorithm is the “machine learning-core” of the developed algorithm. Taking a texton histogram as input, the algorithm measures the Euclidean distance of this histogram to all histograms in the training dataset and outputs the $k$ most similar training histograms and the corresponding $x,y$-positions. While the $k$-NN algorithm is one of the simplest machine learning algorithms, it offers several advantages [@kordos2010we]: it is non-parametric, allowing for the modeling of arbitrary distributions. Its capability to output multiple predictions enables neat integration with the developed particle filter. Additionally, $k$-NN regression often outperforms more sophisticated algorithms [@knn]. A frequent point of criticism is its increasing computational complexity with an increasing size of the training dataset. While the used training datasets consisted of fewer than 1000 images, resulting in short prediction times (see also Figure \[fig:freqhist\]), time complexity can be reduced by storing and searching the training examples in an efficient manner, for example, with tree structures [@bhatia2010survey]. Filtering {#sec:filtering} --------- Computer vision-based estimations are often noisy or ambiguous. Texton histograms obtained during flight will not perfectly match the ones in the training dataset: blur, lighting settings, viewing angles, and, other variables change the shape of the histograms. A popular filter choice is the Kalman filter. However, the Kalman filter is not able to represent multimodal probability distributions. This makes it rather unsuitable for the presented global localization approach. The “naive” $k$-NN regression calculates the mean of the $k$ outputs and forwards this value to the Kalman Filter. We decided to use a more sophisticated method to capture multimodal distributions. Given an adequate measurement model, a general Bayesian filter can simultaneously maintain multiple possible locations and resolve the ambiguity as soon as one location can be favored. In this case, the predictions of the $k$ neighbors can be directly fed into the filter without averaging them first. However, a general Bayesian filter is computationally intractable. Therefore, a variant based on random sampling was used: the particle filter. While its computational complexity is still high compared to a Kalman filter, one can modify the amount of particles to trade off speed and accuracy and adapt the computational payload to the used processor. The weighted particles are a discrete approximation of the probability density function ($pdf$) of the state vector ($x,y$-position of the MAV). Estimating the filtered position of the MAV can be described as $p(X_t \mid Z_t)$, where $X_t$ is the state vector at time $t$ and $Z_t = \mathbf{z}_1, ..., \mathbf{z}_t$ are all outputs of the $k$-NN algorithm up to time $t$, with each $\mathbf{z}_i$ representing the $k$ $x,y$-outputs of the algorithm at time $i$. The used particle filter is initialized with particles at random $x, y$-positions. To incorporate the measurement noise for each of the $k$ estimates from the $k$-NN algorithm, we developed a two-dimensional Gaussian Mixture Model (GMM) as measurement model. The GMM is parameterized by the variances $\Sigma^{[j]}, j \in \{1, \ldots, k\}$ that are dependent on the rank $j$ of the prediction of the $k$-NN algorithm (for example, $j = 2$ is the second nearest neighbor). The variance matrix $\Sigma^{[j]}$ specifies the variances of the deviations in $x$-direction and $y$-direction and the correlation $\rho$ between the deviations. The values for $\Sigma^{[j]}$ were determined by calculating the variance-covariance matrix for the difference between the ground truth $T$ from the motion tracking system and the predictions $P_j$ of the $k$-NN algorithm: $\Sigma^{[j]} := \text{Var}(T-P_j)$. In contrast to the measurement model, the used motion model is simple. It is solely based on Gaussian process noise and does not consider velocity estimates, headings, or control inputs. Its mean and variance are dependent on the expected velocity of the MAV. We used the forward difference $T_t - T_{t-1}$ to estimate the average movement and its variance-covariance matrix $\Sigma_{\text{process}}$ between timesteps $t$ and $t-1$. The algorithm of the developed particle filter is presented in the pseudo code in Algorithm \[alg:particle\_filter\]. In the pseudo code, $\mathcal{X}$ is the list of particles, $f$ the two-dimensional Gaussian probability density function, $z_t^{[i]}$ the $i$th neighbor from the $k$NN prediction, $x_t^{[m]}$ the $m$th particle at time $t$, and $w_t^{[m]}$ its corresponding weight. Initialize particle list $\mathcal{X}_{\text{temp}} := \varnothing$ $x_t^{[m]} \gets x_t^{[m]} + \mathcal{N}(0, \Sigma_{\text{process}})$ $w \gets 0$ $w \gets w + f(z_t^{[i]} ; x_t^{[m]}, \Sigma_{\text{measurement}}^{[i]})$ $\mathcal{X}_{\text{temp}} := \mathcal{X}_{\text{temp}} \cup (x_t^{[m]}, w)$ $\mathcal{X}_t \gets$ $\mathcal{X}_t$ The “resampling wheel” [@thrun] performs the importance resampling step. With the GMM, the information of all $k$ neighbors can be used, yielding a possibly multimodal distribution. While a multimodal distribution allows for keeping track of several possible positions, certain subsystems—for example a control loop—often need one point estimate. Using a weighted average of the particles would again introduce the problem that it could fall into a low density region (an unlikely position). Instead, we used a maximum a posteriori (MAP) estimate, as described by @driessen2008map [@driessen2008map]. The estimation of uncertainty was modeled using the spread of the particles—as expressed by their variance in $x$-direction and $y$-direction. Map evaluation -------------- The performance of the developed method depends on the environment: a texture-rich environment without repeating patterns will be better suited than a texture-poor environment. Ideally, one would like to know if the algorithm will work in a given environment. Therefore, we propose an evaluation scheme that can compare different environments and areas within an environment. This scheme assigns a global fitness value or global loss value to a “map”—expressed as dataset $\mathcal{D}$ consisting of $N$ texton histograms $h_i$ and corresponding $x,y$-coordinates $\text{pos}_i = (x_i, y_i)$. The fitness value is intended to be proportional to the accuracy that can be expected when using this dataset as training set for the developed localization algorithm. The scheme allows for inspecting the dataset and detecting regions within the map that are responsible for the overall fitness value. The idea behind the global loss function $L$ is that histograms $h_i$ and $h_j$ in closeby areas should be similar and the similarity should decrease with increasing distance of the corresponding $x,y$-coordinates $\text{pos}_i$ and $\text{pos}_j$. Therefore, the approach is based on the difference between actual and ideal texton histogram similarities in a dataset. The ideal texton similarity distribution is modeled as a two-dimensional Gaussian distribution around each $x,y$-position in the dataset. Using this idea, a histogram is compared to all others by comparing expected similarities to actual similarities. This results in a loss value per sample of the dataset (local loss). Applying the algorithm to each sample in the dataset yields the global loss of a dataset. The method uses the cosine similarity ($CS$) to compare histograms: $$\begin{aligned} CS(h_i, h_j) = \frac{h_i^Th_j}{\|\|h_i\|\|\,\|\|h_j\|\|}\end{aligned}$$ The cosine similarity has the convenient property that its values are bounded between $-1$ and $1$. In the present case, since the elements of the histograms are non-negative, it is even bounded between $0$ and $1$. Let the function $f$ describe the non-normalized one-dimensional Gaussian probability density function: $$\begin{aligned} f(x; \mu, \sigma) = e^{- \frac{(x - \mu)^2}{2 \sigma ^ 2}} \end{aligned}$$ Since we assume that the ideal similarity in $x$-position is independent of the $y$-position, the ideal two-dimensional similarity function $d_e(\text{pos}_i, \text{pos}_j; \Sigma)$ can be modeled as the product of the respective one-dimensional function $f$: $$\begin{aligned} d_e(\text{pos}_i, \text{pos}_j; \Sigma) = f(x_i; x_j, \sigma_x) \cdot f(y_i; y_j, \sigma_y)\end{aligned}$$ This function is also bounded between $0$ and $1$, which makes the functions $d_e$ and $CS$—ideal similarity and actual similarity—easily comparable. In summary, we propose the following global loss function ($L$) for evaluating a given dataset ($\mathcal{D}$): $$\begin{aligned} L(\mathcal{D}) &= \frac{1}{N^2}\sum_{i = 1}^{N} \sum_{j = 1}^{N} CS(h_i, h_j) - f(x_i; x_j, \sigma_x) \cdot f(y_i; y_j, \sigma_y) \end{aligned}$$ The simple difference—in contrast to least absolute deviations or least square errors—ensures that similarities that are less similar than the ideal similarity reduce the loss. Therefore, a high variation in texture is always seen as “positive”. The variances $\sigma_x$ and $\sigma_y$ specify the dimension of the region, where similar histograms are desired. The lower their value, the more focused the ideal similarity will be, requiring a high texture variety for getting a low loss value. A high value might overestimate the suitability of a dataset. While the approach is relatively robust to the choice of the parameter values, we still need to find a heuristic for suitable values. ![The figure shows the loss of a map: the regions that did not follow the ideal similarity pattern are displayed in red. For the visualization, the loss values per sample in the dataset were smoothed with a Gaussian filter. This assigns a loss value to each $x,y$-position of the map.[]{data-label="fig:globalloss"}](ggg){width="50.00000%"} ANALYSIS {#sec:experiments} ======== In the experiments, the MAV was guided along flight plans using the motion tracking systen. If not otherwise stated, we used the following default values for the parameters in our framework. - number of samples in the histogram extraction step: 400 - number of textons in the dictionary: 20 - number of particles of the particle filter: 50 - number of histograms / images in the training set: 800 - number of histograms / images in the test set: 415 - number of neighbors in the $k$-NN algorithm: 5 Map-dependent texton dictionaries were used and created by conducting an initial flight over the respective maps. Baseline: Homography-based Approach {#sec:siftvsoptitrack} ----------------------------------- To find a baseline for our approach and to provide a homography-based training set, we used the homography-based approach to estimate $x,y$-coordinates in the same environment and based on the same images as the texton-based framework. The required hyperspatial image (Figure \[fig:mapexp\]) of the environment was stitched together using 800 images and the software Microsoft ICE. ![[\[fig:mapexp\]The created map (size: approximately $5 \times 5$ meters) that was stitched together using 800 images. The “non-mapped” area in the center of the image is a result of the set flight path. An image distortion can be seen at the right-hand side, where the landing spot appears twice, while in reality, only one circle was visible. ]{}](map_rotated){width="0.7\columnwidth"} We estimated the $x,y$-coordinates of the 415 test images using the homography-based approach and compared the predictions to the ground truth. The predictions were not filtered. The results can be found in the following table. x-position y-position ------------- ------------ ------------ -- Error in cm 31 59 STD in cm 68 77 \[tab:homoerror\] Training Set based on Motion Tracking System {#sec:experiment-real} -------------------------------------------- In this experiment, the position estimates were calculated on board of the MAV using the texton-based approach with the particle filter. The Euclidean distances between the estimates of the motion tracking system and the texton-based approach were measured in $x$-direction and $y$-direction. The training dataset was composed of 800 texton histograms with corresponding $x,y$-coordinates that were obtained from the motion tracking system. The images were recorded in a $5 \times 5$ meter area at a height of approximately one meter in a time span of one hour before the experiment to keep environmental factors roughly the same. The results can be found in the following table. They are based on 415 images, which corresponds to a flight time of approximately 35 seconds. x-position y-position ------------- ------------ ------------ -- Error in cm 61 59 STD in cm 39 39 Training Set based on Homography-finding Method {#sec:traininghomo} ----------------------------------------------- In this experiment, the training dataset was created by estimating the $x,y$-positions of the 800 training images using the homography-finding method from the previous section and the same hyperspatial image. Apart from that, the settings are the same as in the previous experiment. x-position y-position ------------- ------------ ------------ -- Error in cm 54 97 STD in cm 41 61 Triggered Landing {#sec:triggered} ----------------- For the triggered landing experiment, the MAV was guided along random flight paths, which covered a $5 \times 5$ meter area; during navigation, the MAV was programmed to land as soon as its position estimates were in a “landing zone”: an $x,y$-position with a specified radius $r$. A safety criterion was introduced such that the landing is only performed if the standard deviations of the particles in $x$-direction and $y$-direction are below thresholds $\theta_x$ and $\theta_y$. We set the parameters to $\theta_x = \theta_y = 60$cm. The $x,y$-coordinate of the circle was specified in the flight plan; the radius was set to $r = 60$cm. We performed six triggered landings; after each landing, the $x,y$-center of the zone was randomly set to another position in the map. For the texton framework, the same training set as in Experiment \[sec:experiment-real\] was used. Four out of six landings were correctly performed in the landing area. The distances of the two outliers were 14cm and 18cm, measured as distance to the circumference of the landing area. Speed versus Accuracy Trade-Off {#sec:speedvsacc} ------------------------------- Adapting the frequency of the main loop of the developed approach to make it suitable for different platforms with varying processing power is one of its core parts. Figures \[fig:tradepart\] and \[fig:tradesamples\] show the speed versus accuracy trade-off as a function of the used particles and of the used samples in the histogram extraction step, respectively. As a reference, the frequency using full sampling in the histogram extraction step was 0.1Hz. The above stated default values were used for the ceteris paribus assumption, when varying the parameters. While the bottom camera of the Parrot Bebop Drone has a frequency of 30Hz, the Paparazzi software currently only receives the images with a frequency of 12.5Hz. Therefore, the maximum achievable frequency without further image processing is 12.5Hz, which is the baseline for the conducted experiments. Figure \[fig:freqhist\] illustrates the frequency as a function of the used histograms in the $k$-NN algorithm. We did not compare the frequency to the distance between ground truth and the predictions, since our training dataset did not contain more than 800 histograms. After having received the image, the processing time of the presented algorithm using the default parameter values is 32ms, which includes the histogram extraction (16ms) as well as the $k$-NN predictions, the filtering and the output of the best $x,y$-coordinate (16ms). [0.33]{} ![image](Error_x_particles-crop){width="100.00000%"} [0.33]{} ![image](SD_x_particles-crop){width="100.00000%"} [0.33]{} ![image](particles_vs_freq-crop){width="100.00000%"} [0.33]{} ![image](Error_x_samples-crop){width="100.00000%"} [0.33]{} ![image](SD_x_samples-crop){width="100.00000%"} [0.33]{} ![image](samples_vs_freq-crop){width="100.00000%"} ![image](histograms_vs_freq-crop){width="33.00000%"} DISCUSSION {#sec:discussion} ========== The flight tests show initial evidence for the real-world suitability of the method. It yields slightly less accurate results than the unfiltered homography-finding method. While we did not test the frequency of the homography-based approach on board of an MAV, on a desktop computer, it took 200 ms per image. Therefore, the developed algorithm runs at a much higher frequency. The training set generation based on the homography method yielded higher errors in the flight test. Filtering the estimates of the homography-method first could improve the accuracy. The triggered landing (Experiment 4.4.4) showed good accuracy: while most landings were triggered inside the landing zone, two out of the six landings were outliers. However, their distance to the landing area were rather small, with an average distance of 16 cm. The experiments addressing the “Speed versus accuracy trade-of” show that with an increasing accuracy of the approach, the frequency of the algorithm decreases. However, the errors reach a plateau after which no large improvements can be expected at the lower end of parameter ranges. By optimizing the parameters, one can obtain localization errors below 50cm with the developed approach. While we compared the settings of different parameters, there are no generally optimal parameters for the presented framework: setting the number of textons, the number of images patches, or the number of neighbors is dependent on the environment and the size of the training dataset. The parameters have to be adapted to the particular environment. Despite the overall promising results of our localization algorithm, we noticed drawbacks during the flight tests and identified several directions for future research that are described in what follows. The accuracy could be further improved by combining the presented global localization technique with a local technique. To this end, odometry estimates using optical flow or the inclusion of data from the inertial measurement unit (IMU) could be suitable. Our current implementation assumes constant height up to few centimeters and only small rotations of the MAV. While a quadroter can move in every direction without performing yaw movements, other MAVs or the use of the front camera for obstacle avoidance could require them. The inclusions of images of arbitrary yaw movements into the dataset would inflate its size to a great extent. This could lead to a deterioration of the accuracy and increase the time-complexity of the k-NN algorithm. Instead, a “derotation” of the camera image based on IMU data could be performed to align it with the underlying images of the dataset. CONCLUSION {#sec:conclusion} ========== This paper presented an approach for lightweight indoor localization of MAVs. We pursued an onboard design to foster real-world use. The conducted experiments underline the applicability of the system. Promising results were obtained for position estimates and accurate landing in the indoor environment. An important step in the approach is to shift computational effort to a pre-flight phase. This provides the advantages of sophisticated algorithms, without affecting performance during flight. The approach can trade off speed with accuracy to use it on a wide range of models. The map evaluation technique allows for predicting and improving the quality of the approach.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We summarize unusual bound or localized states in quantum mechanics. Our guide through these intriguing phenomena is the classical physics of the upside–down pendulum, taking advantage of the analogy between the corresponding Newton’s equation of motion and the time independent Schrödinger equation. We discuss the zero–energy ground state in a three–dimensional, spatially oscillating, potential. Moreover, we focus on the effect of the attractive quantum anti–centrifugal potential that only occurs in a two–dimensional situation.' address: \| $^{a}$ Abteilung für Quantenphysik, Universität Ulm, D-89069 Ulm, Germany\ $^{b}$ Department of Physics, North Texas State University, Denton, TX\ author: - 'M. A. Cirone$^{a},\; $G. Metikas$^{a}$ and W. P. Schleich$^{a,b}$' title: Unusual bound or localized states --- Quantum Mechanics; Bound States; Parametric Oscillator; Periodic Potential. 1. The pendulum: a guide to quantum physics {#the-pendulum-a-guide-to-quantum-physics .unnumbered} =========================================== Determining the energy levels of a quantum system was a desperate enterprise before the advent of the Bohr–Sommerfeld quantization conditions. Max Planck had emphasized that the energy of the harmonic oscillator is quantized in units of a fundamental energy given by the product of what we now call Dirac’s constant and the frequency of the oscillator. However, this rule failed miserably when used in other quantum systems. Paul Ehrenfest’s adiabatic principle applied to the pendulum [@ehr] whose length slowly changes as a function of time, made clear that it is not the energy, but the action that is quantized, but why? When we change the length of the pendulum adiabatically, the amplitude of oscillation does not stay constant, neither does the frequency nor the energy. What stays constant is the action, that is, the area in phase space. For the founding fathers of quantum mechanics it must have been a miracle, an amazing fact symbolizing in post–Schrödinger language that the number of nodes in an energy wave function stays constant under adiabatic changes. The classical dynamics of the pendulum serves as an excellent guide for many quantum phenomena. For example, the upside–down pendulum yields insight into the energy wave function of a periodic potential [@whe]. Here we do not focus on the familiar Bloch states which appear when the potential enjoys a strict periodicity over the whole space. Our states occur when the modulation of the potential extends only over a finite domain of space. Due to the shape of this potential we refer to it as the [accordion potential]{}. Extensions of these one–dimensional considerations to two and three dimensions lead us to the effect of the quantum anti–centrifugal force [@cir]. In the present paper we take seriously the joke ‘physics takes mathematics and makes it understandable’, therefore we shall suppress all the mathematics and highlight the essential ideas. This approach is justified by the fact that we are not inventing or applying new mathematics, but attempt to draw together phenomena of different fields exposing a common thread. We shall focus on a general point of view but emphasize that the modern tools of cold atoms in a standing electromagnetic wave can demonstrate these unusual bound or localized states predicted in this paper. We follow one ‘Leitmotif’, that is, a theme common to all phenomena discussed in this article: The classical physics of the upside–down pendulum. Moreover, we develop a little side theme summarized by the phrase “wave mechanics in quantum potentials”. Our paper is organized as follows: In Sec.2 we briefly summarize the classical physics of the upside–down pendulum and then turn in Sec.3 to related phenomena, such as the Paul trap, the helium atom and field induced multipoles. We devote Sec.4 to the transition to quantum mechanics and in particular, discuss the shape of the ground state wave function in the accordion potential. In Sec.5 we show that in three dimensions a localized ground state can exist – even at zero energy. The quantum anti–centrifugal potential emerges when in Sec.6 we focus on the Schrödinger equation in two dimensions. We conclude in Sec.7 with an outlook and present extensions of the concepts discussed in this paper. An Appendix summarizes the mathematical aspects of the upside–down pendulum and the Schrödinger equation of the particle in a periodic potential. 2. The upside–down pendulum {#the-upsidedown-pendulum .unnumbered} =========================== In the mid-forties it was recognized [@bet1; @bet2] that the cyclotron would never be able to accelerate particles to energies above a critical value. The problem lies in the fact that in free space static electric or magnetic or both fields together cannot focus a beam of charged particles in all planes through the axis of the beam. Fortunately, this difficulty, which applies to linear accelerators, can be circumvented to some degree in circular systems. However, the degree of focusing permitted by the effect of the centrifugal force is extremely weak. These considerations suggested that energies higher than 10 MeV are not attainable without unreasonable expenditures of money and materials. However, in 1949 Nick Christofilos from Athens, who at that time was an elevator operator, dreamt up the principle of strong focusing. Although he got a patent, his work was unpublished and the principle was re–discovered [@cou] in 1952. The central idea of the principle of strong focusing is to use a sequence of strong focusing and defocusing fields to achieve a net focusing effect. This idea is best illustrated using the principle of the upside–down pendulum [@lei]. Consider the motion of a point mass on one end of a massless rod. The opposite end of the rod is connected via a hinge to another rod that is fixed to a support. This pendulum experiences a constant gravitational force pulling downwards. When the point of suspension is such that the pendulum hangs down and we concentrate on the limit of small displacements from the stable hanging position we can approximate the motion of the mass by that of a harmonic oscillator. However, when the pendulum is upside–down and stands straight against the gravitational force, the motion is unstable and the mass tends to fall down. In this case we face an inverted harmonic oscillator of negative steepness. In order to stabilize it, Pyotr L. Kapitza [@lan] in 1951 suggested a rapid vertical modulation of the foundation of the pendulum. When the modulation frequency is above a critical value determined by the length of the pendulum, the gravitational acceleration and the modulation amplitude, the motion of the pendulum becomes stable. Why is this so? Many mathematical arguments offer themselves. They range from the method of averaging [@bog] via the secular growth theory [@kev] to the Floquet theorem [@whi]. However, none of these mathematical techniques provide deeper insight into the physics of the stabilizing mechanism. For the sake of argument and to illustrate this lack of insight, we follow the Floquet reasoning. The equation of motion for the phase angle $ \varphi = \varphi (t)$ of the pendulum, measured from its upside–down unstable position, is the Mathieu equation [@abr], as discussed in the Appendix. The Floquet theorem [@whi] determines the domains of stability and instability of this equation. Hence, it is the stability chart [@abr] of the Mathieu equation that governs the parameter regime for which the upside–down pendulum is stable. What is the mechanism of this stabilization? When the foundation accelerates the pendulum upwards, the gravitational force which pulls the pendulum down and makes its motion unstable, is reduced. A push up therefore corresponds to a stabilizing force. On the other hand, a motion of the foundation downwards increases the effective gravitational force and thereby enhances the instability of the pendulum: A pull down corresponds to a de–stabilizing force. On the first sight one might think that during a complete cycle of modulation, the two effects average out, however, during a majority of time there is a pushing up effect. Why is this? In the stable mode the pendulum performs two types of motion, the secular and the micro–motion. The secular motion corresponds to a slow oscillation of the mass between two extremes of phase angles. In addition, the mass undergoes a rapid motion with a frequency identical to the modulation frequency. It is this micro–motion, superimposed on the secular motion, which makes the pendulum stable and which breaks the symmetry between the upward and the downward push. Indeed, due to the geometry of the pendulum, the angular acceleration is not only determined by the acceleration due to the modulation, but by the product of this acceleration and the instantaneous phase angle. The corresponding force is therefore a tidal force: No displacement of the oscillator — no force; large displacement — large force. The modulation of the foundation translates into a modulation of the phase angle. Since the effective acceleration is the product of the acceleration due to the modulation and the phase angle we find the product of two oscillatory functions with the same phase. The time average of such a quantity is a positive constant providing a harmonic oscillator potential of positive steepness for the slow motion. 3. Paul trap, helium atom and negative ions {#paul-trap-helium-atom-and-negative-ions .unnumbered} =========================================== Many other applications of the upside–down pendulum come to mind, for example the Paul trap [@pau] that allows us to store and manipulate single ions in a controlled way. It is worth mentioning that Paul traps play an important role in the recent proposals for a quantum computer. In the Paul trap the need for a time dependent force is dictated by the Poisson equation of electrostatics making it impossible to create binding forces in all three dimensions of space. Indeed, the Poisson equation enforces the feature that even when two spatial directions enjoy a binding potential the third one is anti–binding. To overcome the instability of the trap in the third direction two possibilities offer themselves: apply a time independent homogeneous magnetic field that confines the ion in the third direction or apply an alternating voltage to the trap. The first approach corresponds to the Penning trap, the second to the Paul trap. In the most elementary case the Paul trap consists of a quadrupole field, giving rise to harmonic oscillator potentials. The alternating voltage applied to the electrodes creates a dynamical binding in three dimensions. In each dimension the physics of this binding phenomenon is identical to the upside–down pendulum. The helium atom represents another interesting application of this concept of dynamical binding. For more than one hundred years physicists have tried to understand the motion of two electrons around the nucleus. A one–dimensional model in which the electrons move along a line with the nucleus at the origin brings out the essential physics. On first sight, a situation in which the electrons are on different sides of the nucleus seems to be preferable since in this way the electrons can avoid each other in a most effective way. However, an extremely interesting situation [@ric] occurs when both electrons are on the same side. Indeed, here one electron is close to the nucleus and oscillates rapidly between two equilibrium points. The second electron is far away from the nucleus and moves slowly feeling the attractive force of the nucleus and the repulsion from the inner electron. Indeed, the fast motion of the inner electron creates a time averaged repulsive potential, which superposes with the attractive Coulomb potential of the nucleus, giving rise to a local potential minimum for the outer electron. In this case, the electron–electron repulsion, together with the rapid motion of the inner electron, forms a time dependent barrier to stop the unstable motion of the outer electron caused by the attraction towards the nucleus. This phenomenon is the upside–down pendulum in disguise [@sai]. Our last illustration of effective potentials, arising from time averaging, is the effect of an induced dipole or multipole. It helps us to make the transition to unusual bound states in quantum mechanics. Two examples offer themselves, ([i]{}) the binding of an electron in a negative ion and ([ii]{}) the atom in a standing electromagnetic wave. The field of an electron can polarize a neutral atom even when the electron is at a distance large compared with the atomic dimensions. This interaction between the electron and the atom leads to a force of attraction. This attraction is the reason why some atoms, such as hydrogen or halogen, are able to form negative ions by an attachment of an electron. We now consider the motion of an atom in a standing electromagnetic wave. Indeed, in quantum optics many such experiments have been and are being performed. The light induces a dipole moment in the atom and this dipole again interacts with the light field. Since the interaction energy is the product of the induced dipole and the field, it enters quadratically. Consequently, the position dependence of the interaction energy follows from the position dependence of the square of the field that is, from the square of the mode function thereby creating the effective potential for the motion of the atom. 4. Energy ground state in the accordion potential {#energy-ground-state-in-the-accordion-potential .unnumbered} ================================================= All examples in the previous sections illustrate how dynamical binding can occur in classical physics. We now turn our attention from Newton’s equation of motion $$\frac{d^2 \varphi(t)}{dt^2}+\left[\xi_{0}+\xi(t) \right]\varphi(t)=0$$ for a harmonic oscillator with time dependent steepness $\xi(t)$ to the time independent Schrödinger equation $$\frac{d^2u(x)}{dx^2} + \frac{2M}{\hbar^2} \ [E - V(x)] \ u(x)= 0 \label{Schrodinger}$$ for a point particle of mass $M$ in a position dependent potential $V(x)$; both equations are identical in form. The role of the angle coordinate of the classical upside-down pendulum is now played by the energy wave function $u$. Moreover, the time variable $t$ is replaced by the coordinate variable $x$. Indeed, Newton’s equation of motion is an equation of second order in time for the position and the Schrödinger equation is an equation of second order in position for the wave function. We emphasize that this analogy between the two equations only holds true for the classical harmonic oscillator. Whereas the Schrödinger equation is always linear, a feature independent of the form of the potential, the form of Newton’s equation strongly depends on the shape of the potential and in general is nonlinear. We focus on a periodic potential in space. The particular shape of this potential is of no importance as long as it is averages out to zero over space; it is as often positive as it is negative. For our analysis it is crucial that the oscillations of the potential start at a given point in space and end thereby connecting a flat space through a domain of wiggles back to a flat space. We measure energy relative to flat space. Fig.1 shows an example of such a potential. It is not surprising that under these circumstances we find bound states of negative energy, however, under appropriate conditions, this potential can also display a localized state of zero energy. Thus, it is again the physics of the upside–down pendulum that allows us to understand this phenomenon. We first concentrate on the one–dimensional case. Guided by the analogy between the Newton and the Schrödinger equation we expect the corresponding ground state energy wave function to exhibit a slowly varying envelope that is modulated with the period of the periodic potential; but where does the binding come from? The origin of this bound state lies in the modulation of the envelope—without the modulation there is no binding. The modulation can be easily understood: the particle is more likely to be in a stable minimum of the potential rather than in an unstable maximum. The curvature of the wave function at a given position, expressed by its second derivative, is determined, not only by the potential, but also by the wave function at that point. Indeed, according to the Schrödinger equation (\[Schrodinger\]), it is the product of the potential and the wave function that governs the curvature. This quantity contains, apart from other terms, the product of the oscillatory potential and the modulations of the wave function. The simplest case of a cosine modulation gives rise to the square of the cosine, thus creating a constant potential and a part that varies in space with twice the modulation frequency. We focus on the constant part and note that this constant is always negative. In order to explain this feature we recall that the modulation of the envelope of the wave function is such that at the minima of the potential we have local maxima of the wave function and vice versa. Hence, the modulation and the potential defer in their sign and their product creating the potential well is negative. The fact that the potential only wiggles over a finite domain makes this constant negative potential into an attractive potential well whose width is determined by the range of the oscillations. The depth of the well involves, apart from other parameters, the square of the height of the periodic potential. For more mathematical treatment and explicit expressions for the well, we refer to the Appendix. There is a close analogy between the ground state in the periodic potential and the problem of the upside–down pendulum. In both cases the binding effect is a consequence of the average of the product of two oscillatory functions: for the pendulum it is the time dependent modulation of the foundation and the time dependent angle of oscillation, which creates a harmonic oscillator of positive steepness. When considering the particle in the periodic potential, it is the product of the position dependent potential and the wave function modulated in space that gives rise to a well. However, there is a dramatic difference: whereas the modulation of the pendulum is on for all times, the spatial modulation is confined to a certain domain of space. It is this confinement in space, together with the creation of an effective potential, which defines the bound state. 5. Zero-energy ground state {#zero-energy-ground-state .unnumbered} =========================== The spatial domain, over which the wiggles are present, and the depth of the potential well, determine the ground state energy. Is it possible to choose these parameters so as to achieve a ground state of zero energy? The answer is: it depends! In one dimension the answer is a flat no, but in three dimensions it is possible, as we shall now show. We start by considering the situation in one dimension. As we decrease the depth of the well and the energy of the ground state approaches zero, the tails of the wave function reach more and more into the forbidden region outside of the well. This feature is a consequence of the fact that the decay is governed by the square root of the absolute value of the energy. In the limit of zero energy the wave function is no longer localized; no mechanism prevents the wave from flooding all space. Consequently, there is no ground state of zero energy in one dimension. Is it possible to construct such a zero-energy bound state in higher dimensions? In order to answer this question, we first concentrate on the three-dimensional case and consider a potential $V=V(r)$ that wiggles along the radial direction. Beyond a final radius $r_{0}$, the potential vanishes. The time independent Schrödinger equation $$\frac{d^2 u_{0}(r)}{dr^2} + \frac{2M}{\hbar^2} \ [E- V(r)] \ u_{0}(r)=0 \label{radial}$$ for the radial energy wave function $u_{0}=u_{0}(r)$ corresponding to vanishing angular momentum is then identical to the one-dimensional case. Due to the product of the oscillatory potential and the wave function, and the finite domain of oscillations, we find an attractive potential well of radius $r_{0}$. There is however, a subtle difference to the one-dimensional case: it is the origin where the radial wave function has to vanish [@rau] corresponding to an infinitely high potential wall at the origin. This feature is a consequence of the spherical symmetry. Therefore, the infinite wall at the origin provides one side of the effective well creating even a node of the wave function. The other wall of finite height is due to the effective potential together with the finite range of the oscillatory domain. For appropriate parameters such as the final radius and the depth of the modulation of the potential we can fit a ground state of negative energy into this well, as shown in Fig.2. The corresponding wave function displays a node at the origin and an exponential decay in the classically forbidden domain beyond the critical radius, with a single maximum lying in between. Can we now take the limit of zero energy and thus realize a zero-energy ground state? The answer is yes! Indeed, we can arrange the depth of the well such that exactly one quarter of a period of a sine oscillation fits into the potential well. Why a quarter of a period? Since the decay of the wave function outside of the potential is governed by the square root of the energy we find that for zero energy the wave function is constant. Due to the condition of continuity on the wave function it must take on the value at the wall. The radial wave function displays a node at the origin and increases like a sine function to a constant value. However, the sine function must merge smoothly into the constant, which is only possible at a point where the sine function has an extremum. Since we want to have the ground state without a node, except the one at the origin, the only possibility for such a merger appears at one quarter of a period [@bri] of the sine function. Obviously, this zero–energy ground state radial wave function does not display any localization. The localization becomes apparent when we recall that we have to divide the radial wave function by the radial variable in order to find the total wave function, that is $$\psi_{0}(r) =\frac{u_{0}(r)}{r}.$$ Indeed, now the total wave function enjoys a maximum at the origin and decays with the inverse power of the radius. We emphasize however, that the total wave function is not square integrable since the volume element $4 \pi r^2 dr$ brings in the square of the radius. This contribution cancels exactly the factor creating the localization in the probability density, that is, in the absolute value squared of the wave function. For this reason we refer to this state as localized state rather than bound state. We conclude this discussion by briefly addressing the case of positive energies. Here, the exponential decay into the classically forbidden region of a wave function corresponding to negative energy turns into right and left running plane waves of positive energy. Obviously, plane waves are not localized wave functions, nevertheless, the total wave function is concentrated at the origin due to the inverse power of the radius; in this sense we have localized wave functions in the continuum. 6. Quantum anti-centrifugal force {#quantum-anti-centrifugal-force .unnumbered} ================================= The physics of a bound state in two dimensions is very different from the one in other dimensions; and this for many reasons. In [@cir] we have studied this case in more detail. In the present paper we only highlight and motivate the results of [@cir]. For detailed derivations we refer to that article. To bring out the peculiarities of the two–dimensional case we first consider the case of no external potential. In two dimensions a vanishing angular momentum does not imply [@flu] a vanishing centrifugal potential as in three dimensions. Indeed here the radial Schrödinger equation reads $$\frac{d^2 u_{0}(r)}{dr^2} + \frac{2M}{\hbar^2} \ [E-V_{Q}(r)] \ u_{0}(r)=0. \label{polar}$$ We can trace the origin of the new potential $$V_{Q}(r) \equiv - \frac{\hbar^2}{2M} \ \frac{1}{4 r^2}$$ back to the non-vanishing commutation relation between the operators of momentum and the radial unit vector. This potential is, therefore, a true quantum potential. Since, in addition, it is attractive rather than repulsive we have named it the quantum anti-centrifugal potential [@cir]. Moreover, the potential is attractive only in a two–dimensional world. For one or three dimensions it vanishes, but for higher than three dimensions it becomes repulsive. The quantum anti-centrifugal potential has brought us into a rather unusual situation. Usually, we perform wave mechanics in a classical external potential. In the present context however, we have a quantum potential. This feature stems from the reduction of space from three to two dimensions, reminiscent of the Born-Oppenheimer approximation in molecular physics. A molecule is a quantum system consisting of several degrees of freedom. In the framework of the Born–Oppenheimer approximation we solve the Schrödinger equation for the electrons for a fixed position of the nuclei and then use the resulting potential curves, that is, the electronic energies as a function of the separation of the nuclei to determine their relative motion. In this sense we indeed perform wave mechanics on potentials that are not due to external forces but due to quantum mechanics. In the example of the nuclear motion in a molecule, the appearance of wave mechanics in a quantum potential is a consequence of the reduction of the degrees of freedom [@bri2]. The attractive quantum anti-centrifugal potential manifests itself even in the energy wave function of a free particle, that is in the absence of any other external potential. The ordinary Bessel functions $J_{0}$ and $N_{0}$ are two independent solutions of the time independent Schrödinger equation for positive energy. Both show a bunching of nodes towards the origin, as suggested by an attractive potential leading to an acceleration of the particle towards the origin and demonstrated for the example of $J_{0}$ in Fig.3. We have to compare and contrast this situation with the case of a particle with one unit of angular momentum, where the centrifugal potential is indeed repulsive. It is worth noting that the quantum anti-centrifugal potential reduces the strength of repulsion of the centrifugal potential and the effective potential is not as repulsive as in the classical case. In the case of one unit of angular momentum, the solutions of the time independent Schrödinger equation are the ordinary Bessel functions $J_{1}$ and $N_{1}$. As expected by the repulsive potential enforcing a deceleration of the particle as it approaches the origin, the nodes of the Bessel functions show anti-bunching towards the origin. One might wonder if the quantum anti-centrifugal potential is strong enough to support a bound state with negative energy. In the transition from positive to negative energies real wave numbers transform into purely imaginary ones as a consequence of the quadratic dispersion relation between energy and wave number. Therefore, the ordinary Bessel functions $J_{0}$ and $N_{0}$ turn into modified Bessel functions $I_{0}$ and $K_{0}$. Since $I_{0}$ explodes for large arguments [@abr], that is, large radial distances, this solution does not satisfy the requirement of exponential decay enforced by the attractive quantum anti-centrifugal potential. Only the function $K_{0}$ achieves this goal, but it has another disease: it explodes at the origin like a logarithm. This singularity at the origin indicates that the solution $$u_{0}(r)=\sqrt{\frac{k}{\pi}}\sqrt{kr}K_{0}(kr)$$ of the Schrödinger equation, shown in Fig.4, does not describe a truly free particle but that there is a delta function potential present. The strength of this external potential determines [@wod] the energy eigenvalue of this bound state. Moreover, due to the quantum anti-centrifugal potential the probability of finding the particle within a given area in space is concentrated along a band around the location of the delta function; a remarkable feature unique to two dimensions. In a similar arrangement in one or three dimensions the maximum of the probability is at the location of the potential minimum. This property might have interesting applications for guiding atoms along wires [@den] or electro-magnetic waves in wave-guides [@col]. We now return briefly to the case of an energy eigenstate in a two–dimensional oscillatory potential of finite range. As emphasized in the previous paragraph, the case of vanishing angular momentum contains the most quantum effects. Here, the ground state wave function of negative energy feels the combination of the quantum anti-centrifugal potential and the potential well created by the product of the wave function and the potential. 7. Light induced potentials, Trojan wave packets and the lone electron {#light-induced-potentials-trojan-wave-packets-and-the-lone-electron .unnumbered} ====================================================================== The examples discussed in the preceding sections represent but only a small subsection of the class of unusual bound states. They have been selected because they follow or are motivated by the physics of the upside-down pendulum. However, there are many more intriguing bound states in atomic and molecular physics [@gw]. We conclude our paper by briefly summarizing three examples starting with one that is still closely related to our main theme of the upside-down pendulum and gradually moving away towards new frontiers. Consider a diatomic molecule with various electronic states. Within the Born–Oppenheimer approximation the electronic states provide the potentials for the relative motion of the two nuclei. A strong external time dependent laser field interacts with the electrons and couples the individual electronic potential surfaces. The wave functions for the vibratory motion of the nuclei are therefore coupled through a time dependent periodic interaction Hamiltonian. The product of the interaction Hamiltonian and the nuclear wave functions govern the time evolution of the wave functions; a feature that is again reminiscent of the upside-down pendulum. In the latter the product consists of the time dependent steepness of the oscillator and the phase angle. The only difference lies in the fact that in the classical case the Newton equation is of second order, whereas the Schrödinger equation is a first order. However, this distinction is superficial since we are dealing with the nuclear wave functions corresponding to different electronic states and hence, with a system of differential equations of first order. Due to this product of two periodic functions with the same period (interaction Hamiltonian and wave function) we find again a constant which lifts up the electronic potential by integer multiples of the energy of the absorbed photon. In mathematical terms this energy elevation is a consequence of Floquet states. However, it can be viewed as just one more illustration of the upside-down pendulum with counter-intuitive consequences. Indeed, when we consider a binding and a repulsive potential this constant shift in energy can result in a crossing of the potential curves. Since in a diatomic molecule there exists a non-crossing rule we obtain an avoided crossing forming new potentials; one is binding and one is repulsive. However, these potentials are rather peculiar because each of them consists of parts of the two original potentials. The formation of new bound states [@bes] in this light induced molecule has been observed experimentally [@san]. Trojan asteroids are celestial bodies that are held in their positions by the gravitational forces of the sun and Jupiter. A similar phenomenon can occur [@bia] in an atom where the electron plays the role of a Trojan asteroid, and the nucleus substitutes for one of the two planets. A circularly polarized electromagnetic field simulates the effect of the second planet. Indeed, the electromagnetic interaction of the electron with the field and the Coulomb attraction between the electron and the nucleus replace the gravitational attraction of the celestial bodies. In a frame rotating with the electromagnetic field, the motion of the electron is in a plane and governed by three potentials which depend on the two-dimensional radial direction: the Coulomb attraction of the nucleus, the centrifugal potential of the circular motion and the linear potential arising from the interaction of the electron with the electromagnetic field. For short distances the Coulomb potential dominates, whereas for large distances the repulsive potential of the circular motion prevails. Consequently, for intermediate distances, an unstable potential maximum occurs. This feature is in complete analogy with the upside–down pendulum. But where does the stabilizing drive come from? The answer to this question lies in the fact that the radial motion is coupled to the angular motion. The latter is locked to the rotating electromagnetic field and the electron performs angular vibrations around this stable point of equilibrium. These vibrations couple into the radial motion creating a dynamically binding potential. This classical picture also holds true for the quantum case as verified by extensive studies of wave packets propagating in this minimum. In our last example of unusual bound states it is the balance of static forces that is responsible for the formation of a bound state of a lone electron in a Rydberg atom in crossed electric and magnetic fields [@fau]. The crossed-field situation is of particular appeal in the field of quantum chaology since in this system energy is the only conserved quantity. The classical system displays chaos and is therefore of interest in the search for fingerprints of chaos in the corresponding quantum system. Apart from these questions of quantum chaos there is another quite interesting aspect of this system. The electron in the atom experiences, not only the attractive Coulomb potential of the nucleus, but also a linear potential due to a constant electric field and a binding harmonic oscillator potential due to the magnetic field. The superposition of all three potentials creates a local potential minimum far away from the nucleus. An electron bound in this minimum displays a large dipole moment that can be observed by sending the atom through an inhomogeneous electric field. Indeed, the experiments [@den] have confirmed the existence of this far outside lying minimum. We have come a long way on our journey into unusual bound states. Starting from the classical physics of the upside-down pendulum we have been led to the phenomenon of a zero-energy ground state in a periodic potential in three dimensions. The two-dimensional world has even more surprises in store: the attractive quantum anti-centrifugal force that manifests itself in a free particle through the bunching of the nodes or even in a bound state. Here the probability of finding the particle is concentrated in the domain where no force is acting. This feature resulting from the action of the quantum anti-centrifugal potential could be useful in guiding atoms along wires or electromagnetic waves along fibers. Experiments on trapping cold atoms by single photons [@hoo] or along a whispering gallery mode of a glass sphere resonator [@ver] are yet more realizations of the physics of the upside-down pendulum and give us confidence that our predictions of unusual bound and localized states can be verified experimentally in the near future. Acknowledgments {#acknowledgments .unnumbered} =============== This work started when I (WPS) had the great privilege to be a postdoc with John A. Wheeler more than a decade ago. Since that time, John and I have frequently returned to the topics addressed in the present paper. These discussions took place in trains, planes and automobiles and at various locations such as Hightstown, High Island, and Ulm. I thank John for this wonderful time, his outstanding hospitality at Hightstown and High Island and especially for the unique experience to work with him. The present paper is partially based on notes and sketches of figures prepared jointly. I am grateful for many fruitful discussions and John’s deep insights and, in particular, for allowing us to use the material that had originally been obtained in close collaboration with him. The proceedings of the Lake Garda Conference are a most welcome opportunity to finally summarize this project started many years ago. Moreover, we are grateful to I. Białynicki-Birula for a critical reading of the manuscript. Two of us (MAC and WPS) thank R. Bonifacio for his hospitality and for organizing a most interesting conference in the splendid surroundings of Lake Garda. The work of WPS is partially supported by the Deutsche Forschungsgemeinschaft, moreover, he gratefully acknowledges a travel grant from the Universität Ulm which made part of this research possible. Appendix {#appendix .unnumbered} ======== In contrast to the main body of the paper, we pursue in the present Appendix a more mathematical approach. In particular, we emphasize the similarities between the classical equation of motion for the upside–down pendulum and the Schrödinger equation for a nonrelativistic particle in a periodic potential. For this purpose, we first cast the corresponding equations into a dimensionless form. This approach allows us to simultaneously derive equations for the dynamics contained in the macro– and micro–motion of the pendulum or the envelope function and the modulation of the ground state wave function. We start by summarizing the classical equation of motion for the upside–down pendulum. For the sake of simplicity we consider the limit of small angles $\varphi$ measured relative to the vertical position. A vertical acceleration $\ddot{f}$ of the foundation translates into a vertical acceleration of the mass $M$ of the pendulum and adds to the gravitational acceleration $g$ pointing downwards. Here, the sign of $\ddot{f}$ is crucial: when $\ddot{f}$ is positive, that is, when the pendulum is accelerated upwards, the gravitational acceleration pointing downwards is reduced. Likewise, when $\ddot{f}$ is negative, that is, when the pendulum is accelerated downwards, the gravitational acceleration is increased. Consequently, the total acceleration is $g-\ddot{f}$. Hence for small angles $\varphi$ the force tangentially to the rod of the pendulum of length $L$ reads $$M L \frac{d^2\varphi}{dt^2} = M \left( g-\ddot{f}\: \right) \varphi,$$ giving rise to $$\frac{d^2\varphi(t)}{dt^2}+\left[-\Omega^2+\frac{1}{L}\ddot{f} \right] \varphi(t)=0,$$ where $\Omega^2\equiv g/L$. It is reasonable to assume that the modulation $f$ is a periodic function of period $T\equiv 2\pi / \nu$. This time scale allows us to introduce a dimensionless variable $\theta \equiv \nu t$ and the differential equation of the upside–down pendulum takes the form $$\frac{d^2\varphi}{d\theta^2} + \left[ \kappa +a(\theta)\right]\varphi=0. \label{eight}$$ Here we have introduced the abbreviations $\kappa\equiv -(\Omega/\nu)^2$ and $a(\theta)\equiv f''(\theta)/L$ for the steepness of the inverted harmonic oscillator and the acceleration due to the modulation, respectively. Moreover, prime denotes differentiation with respect to $\theta$. We now make the connection to the time independent Schrödinger equation $$\frac{d^2 u(x)}{dx^2}+\frac{2M}{\hbar^2}\left[E-V(x)\right] u(x)=0 \label{sch1}$$ for a particle of mass $M$ in the potential $V(x)$, which is periodic in space with period $\lambda\equiv 2\pi / k$. When we introduce the recoil energy $E_{r}\equiv (\hbar k)^2/(2M)$ and the dimensionless coordinate $\theta\equiv kx$, the Schrödinger equation takes the form $$\frac{d^2 u}{d\theta^2}+\left[ \eta-v(\theta) \right] u=0 \label{ten}$$ where the dimensionless energy eigenvalue $\eta\equiv E/E_{r}$ and the potential $v(\theta)\equiv V(\theta/k)/E_{r}$ are scaled in units of the recoil energy. The two dimensionless equations of the driven pendulum (\[eight\]) and the quantum particle in the periodic potential (\[ten\]) have identical structure. In Secs.2 and 4 we have presented qualitative arguments to explain the motion of the pendulum as a superposition of a slow motion (macromotion) and a rapid motion (micromotion) or, in the language of wave functions, to decompose the energy wave function into an envelope and a modulation. We now take advantage of the analogy between the two systems to support these qualitative arguments by rigorous mathematics. For this purpose we start from the equation $$u''(\theta)+\left[\eta-v(\theta)\right] u(\theta)=0. \label{schacc}$$ Again, we emphasize that $u$ can either be the phase angle of the pendulum or the wave function. Likewise, $\theta$ represents time or position and the drive $v(\theta)$ is either the acceleration of the foundation or the potential. In order to find the solution of (\[schacc\]), we make the [Ansatz]{} $$u(\theta)={\cal A}(\theta)\left[ 1-\epsilon v(\theta) \right], \label{ansatz}$$ consisting of the product of the slowly varying envelope $\cal{A}(\theta)$ and the modulation caused by the drive $v(\theta)$. We substitute the [Ansatz]{} (\[ansatz\]) into (\[schacc\]), which yields $$\begin{aligned} & & {\cal A}''+\left[ \eta+\epsilon v^2\right]{\cal A} \nonumber \\ & & -\left[ {\cal A}''+\left( \eta +\frac{1}{\epsilon}+ \frac{v''}{v}\right) {\cal A}\right] \epsilon v \nonumber \\ & & -2\epsilon {\cal A}' v'=0. \label{star}\end{aligned}$$ So far, our analysis is exact. We now solve the equation (\[star\]) in an approximate way. For this purpose, we first neglect the last contribution in (\[star\]) involving the product of the first derivatives of the envelope function and the potential. This approximation is justified since $v'$ is out of phase with $v$ and cannot therefore lead to a significant contribution. Moreover, we choose $\epsilon$ such that $$\frac{1}{\epsilon}+\frac{v''}{v}=0.$$ We emphasize that this is not possible in a strict sense, since $v''/v$ is not necessarily a constant. However, in the oscillatory domain of the potential the most elementary model $v(\theta)\propto \cos(\theta)$ suggests the estimate $v''/v\simeq -1$ and hence we find $$\epsilon=+1.$$ Provided the envelope ${\cal A}$ satisfies the equation $${\cal A}''+\left[ \eta+\epsilon v^2\right]{\cal A}=0$$ also the second contribution in Eq.(\[star\]) vanishes to this order in $\epsilon$. This is due to the fact that the second contribution is already multiplied by $\epsilon$. For the case of the upside–down pendulum, the parameter $\eta$ is negative. In the absence of the drive, the negative value of $\eta$ leads to an exponential growth of ${\cal A}$. However, the square of the drive is always positive and can create an overall positive coefficient in front of ${\cal A}$, providing an oscillation rather that an exponential explosion. In case of the quantum particle, it is advantageous to write the equation for ${\cal A}$ in the form of a Schrödinger equation $${\cal A}''(\theta)+\left[ \eta-v_{eff}(\theta)\right] {\cal A}(\theta)=0.$$ Here we have averaged the square of the potential over a period of the oscillations giving rise to a smooth effective potential $$v_{eff}(\theta)\equiv -\frac{1}{2\pi}\int_{\theta}^{\theta+2\pi} d\theta' v^2(\theta').$$ We note that this potential is always negative. When the modulation of the potential is confined in space, for example by a window function $\mu(\theta)$, and the potential $v(\theta)$ is of the form $$v(\theta)\approx \mu(\theta)C(\theta)$$ where $C$ is a periodic function, we can approximate the effective potential by a potential well $$v_{eff}(\theta)=-\frac{1}{2} \mu^2(\theta). \label{well}$$ In this case, the envelope ${\cal A}$ is an energy wave function of this potential well and the parameter $\eta$ is the dimensionless energy eigenvalue. Hence, $\eta$ is not determined from the outside, but rather by the well (\[well\]) itself. If the well is deep enough, it can support, apart from the ground state, also excited states with negative energy. P. Ehrenfest, Naturwiss. [11]{}, 543 (1923); for an excellent introduction into and comprehensive summary of the early quantum mechanics we refer to D. ter Haar, The Old Quantum Theory, Pergamon Press, Oxford, 1967, p. 44. A classic in this field is M. Born, Atommechanik, Springer, Heidelberg, 1925; see also J. Duck and E. C. G. Sudarshan, 100 Years of Planck’s Quantum, World Scientific, Singapore, 2000. J. A. Wheeler, in C. and B. S. De Witt (eds), Relativity, Groups and Topology, Gordon and Breach, New York, 1964. M. A. Cirone, K. Rzazewski, W. P. Schleich, F. Straub and J. A. Wheeler, to be published. For a summary of the problem of acceleration of charged particles to high energies see the article by J. P. Blewett, in E. U. Condon and H. Odishaw (eds), Handbook of Physics, McGraw-Hill, New York, 1958, chapter 9, p. 153. J. A. Wheeler in many discussions with one of us (WPS) repeatedly credited H. Bethe with the discovery of the problems leading to the principle of strong focusing. Unfortunately, we have not been able to locate the appropriate reference. We have talked to Prof. Bethe, but also he was not able to point out the article. E. D. Courant, M. S. Livingston and H. S. Snyder, Phys. Rev. [88]{}, 1190 (1952) . For a treatment of the upside-down pendulum and the double pendulum in terms of scalar and matrix continued fractions see Th. Leiber and H. Risken, Phys. Lett. A [129]{}, 214 (1988). L. D. Landau and E. M . Lifshitz, Mechanics, Pergamon Press, Oxford, 1969, p. 93-95. N. N. Bogoljubov and Yu. A. Mitropolsky, Asymptotic Methods in the Theory of Non-linear Oscillations, Hindustan Publishing, Delhi, 1961. Y. Kevorkian, and J. D. Cole, Perturbation in Applied Mathematics, Springer, Heidelberg, 1981. E. T. Whittaker and G. N. Watson, A Course of Modern Analysis, Cambridge University Press, Cambridge, 1973. M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, Dover Publications, New York, 1965; for a particularly illuminating method to motivate the stability chart of the Mathieu equation using WKB wave functions, see M. J. Richardson, Am. J. Phys. [39]{}, 560 (1971). W. Paul, Rev. Mod. Phys. [62]{}, 531 (1990). A pedagogical introduction into the physics of Paul traps emphasizing the concept of the effective potential is given by P. E. Toschek, in G. Grynberg and R. Stora (eds), New trends in atomic physics, North-Holland, Amsterdam, 1984, p. 390. For a summary of current activities in Paul traps see H. Walther, in B. Bederson and H. Walther (eds), Advances in Atomic, Molecular and Optical Physics, Academic Press, Boston, 1995. K. Richter and D. Wintgen, Phys. Rev. Lett. [65]{}, 1965 (1990); J. Phys. B [24]{}, L565 (1991): for an overview of the full classical and quantum dynamics of two-electron systems see K. Richter, G. Tanner, and D. Wintgen, Phys. Rev. A [48]{}, 4182 (1993). We can also interpret the system of the two electrons in the helium atom as one realization of the Fermi accelerator. In the most elementary version of this device an oscillating wall confines the unbounded motion of a particle in a linear potential. For a review of the Fermi accelerator see F. Saif, I. Białynicki-Birula, M. Fortunato and W. P. Schleich, Physics Reports, to be published. L. D. Landau and E. M. Lifshitz, Quantum Mechanics, Non-Relativistic Theory, Pergamon Press, Oxford, 1965, p. 267; There exists a large amount of literature on the physics of negative ions. See for example the classic book H. Massey, Negative Ions, Cambridge University Press, Cambridge, 1976. For the most recent activities in this field see the Springer series, Production and Neutralization of Negative Ions and Beams. A similar effect appears also for an electron in the field of a super-heavy nuclei, see for example F. G. Werner and J. A. Wheeler, Phys. Rev. [109]{}, 126 (1958). For summary see C. N. Cohen-Tannoudji, in J. Dalibard et al. (eds), Fundamental Systems in Quantum Optics, Elsevier Science, Amsterdam, 1992; see also the Nobel Lectures by S. Chu, C.N. Cohen-Tannoudji, and W. D. Phillips, Rev. Mod. Phys. [70]{}, 685 (1998). For a most striking and counter-intuitive bound state resulting from the peculiar behavior of the radial wave function at the origin of three-dimensional space we refer to J. Rauch and M. Reed, Comm. Math. Phys. [29]{}, 105 (1973); and M. Reed, and B. Simon, Methods of Modern Mathematical Analysis II, Academic Press, New York, 1975. In this example the potential is a sequence of appropriately constructed steps that lead continuously downwards as the radial variable increases. Classically a particle of given energy has to fall down the steps. However, the reflections of the quantum wave at the individual steps interfere in a way as to localize the particle. Likewise, the same authors discuss a potential consisting of an infinite sequence of potential spikes that classically would keep a particle trapped, however, due to the tunneling effect the quantum particle escapes. A similar reasoning appears [@whe] in general relativity, that is in geometrodynamics, when we determine the metric coefficients of a system. They follow from Brill’s equation [@whe], which is similar to the time independent Schrödinger equation for zero energy. In contrast to quantum mechanics where the Schrödinger equation determines the energy eigenvalues we now have to solve the equation under the constraints that the wave is not allowed to have nodes and corresponds to zero energy. S. Flügge, Practical Quantum Mechanics, Springer, Heidelberg, 1971. This observation is in complete accordance with a recent paper arguing that the time dependent Schrödinger equation is an approximation of the time independent Schrödinger equation resulting from the elimination of degrees of freedom. See for example J. S. Briggs and J. M. Rost, EPJD [10]{}, 311 (2000). K. Wódkiewicz, Phys. Rev. A. [43]{}, 68 (1991). J. Denschlag, D. Cassettari and J. Schmiedmayer, Phys. Rev. Lett. [82]{}, 2014 (1999). R. E. Collin, Field Theory of Guided Waves, McGraw Hill, New York, 1960 p.477. Two examples illustrate the unusual bound states that originate from the mutual interaction of the electrons in a heavy atom: The Thomas-Fermi potential together with the centrifugal potential can form a second potential minimum, which is very deep and located close to the nucleus. This effect occurs provided the atomic number is larger than 57 and we are dealing with an energy eigenstate corresponding to the angular momentum quantum number $l=3$, see for example M. Goeppert-Mayer, Phys. Rev. [60]{}, 184 (1941). Since A. Sommerfeld we associate the motion of an electron in an atom with an ellipse or a circle. However, the electron at the top of the sea of filled atomic states moves in an effective screened potential giving rise to a necklace orbit as pointed out by J. A. Wheeler, in E. H. Lieb et al. (eds), Studies in Mathematical Physics, Princeton University Press, Princeton, 1976, p. 383. J. Bestle, Stabilisierung zweiatomiger Moleküle in intensiven Laserfeldern, Shaker Verlag, Aachen, 1997. K. Sändig, H. Figger and T. W. Hänsch, Phys. Rev. Lett. [85]{}, 4876 (2000). I. Białynicki-Birula, M. Kaliński and J. H. Eberly, Phys. Rev. Lett. [73]{}, 1777 (1994). M. Fauth, H. Walther and E. Werner, Z. Phys. D [7]{}, 293 (1987); G. Raithel, M. Fauth, and H. Walther, Phys. Rev. A. [47]{}, 419 (1993). C. J. Hood, T. W. Lynn, A. C. Doherty, A. S. Parkins and H. J. Kimble, Science [287]{}, 1447 (2000); P. W. H. Pinkse, T. Fischer, P. Maunz and G. Rempe, Nature [404]{}, 365 (2000). D. W. Vernooy and H. J. Kimble, Phys. Rev. A [55]{}, 1239 (1997): ibid [56]{}, 4287 (1997).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We aim at demonstrating the influence of diversity in the ensemble of CNNs on the detection of black-box adversarial instances and hardening the generation of white-box adversarial attacks. To this end, we propose an ensemble of diverse specialized CNNs along with a simple voting mechanism. The diversity in this ensemble creates a gap between the predictive confidences of adversaries and those of clean samples, making adversaries detectable. We then analyze how diversity in such an ensemble of specialists may mitigate the risk of the black-box and white-box adversarial examples. Using MNIST and CIFAR-10, we empirically verify the ability of our ensemble to detect a large portion of well-known black-box adversarial examples, which leads to a significant reduction in the risk rate of adversaries, at the expense of a small increase in the risk rate of clean samples. Moreover, we show that the success rate of generating white-box attacks by our ensemble is remarkably decreased compared to a vanilla CNN and an ensemble of vanilla CNNs, highlighting the beneficial role of diversity in the ensemble for developing more robust models.' author: - Mahdieh Abbasi - Arezoo Rajabi - Christian Gagné - 'Rakesh B. Bobba' title: Toward Adversarial Robustness by Diversity in an Ensemble of Specialized Deep Neural Networks ---	{ "pile_set_name": "ArXiv" }
Introduction ============ Thirty five years ago Wilson and Fisher [@01] emphasized the relevance of the $\varphi^4$-theory to understanding the critical phenomena. Since then, the theory has become one of the most appealing theoretical tools for studying the critical phenomena in a wide variety of systems in statistical physics. In the strong coupling limit, the $\varphi^4$-theory develops domain walls, a phenomenon which is of great interest in the classical and quantum field theories \[2-10\]. The dynamical $\varphi^4-$theory - what is usually referred to as the time-dependent Ginzburg-Landau (GL) theory - provides a phenomenological approach to, and plays an important role in, understanding dynamical phase transitions and calculating the associated dynamical exponent \[11-16\]. The time-dependent GL theory for superconductors was presented phenomenologically only in 1968 by Schmid [@Schmid] (and derived from microscopic theory shortly thereafter [@Gorkov]), when the first modulational theory was derived in the context of Rayleigh-Benard convection [@Newell; @Segel]. Moreover, the GL equation with an additional noise term has been studied intensively as a model of phase transitions in equilibrium systems; see, for example, [@Hohenberg]. In the present paper we consider the stochastic $\varphi^4$-theory in the strong coupling limit. This limit is singular in the sense that, the equation that describes the dynamics of the system develops singularities. Therefore, starting with a smooth initial condition, the domain-wall singularities are dynamically developed after a finite time. At the singular points the field $\varphi(x,t)$ is not continuous. We derive master equations for the joint probability density functions (PDF) of $\varphi$ and its increments in $d$ dimensions. It is shown that in the stationary state, where the singularities are fully developed, the relaxation term in the strong coupling limit leads to an unclosed term in the PDF equations. Using the boundary layer method [@BO; @16], we show that the unclosed term makes no finite contribution (anomaly) in the strong coupling limit, and derive the PDF of $\varphi$ and its moments, $\langle\varphi^n\rangle$, in the same limit. We also investigate the scaling behavior of the moments of the field’s increments defined by, $\delta\varphi=\varphi(x_2)-\varphi(x_1)$, and show that when $\|x_2-x_1\|$ is small, fluctuations of the $\varphi$ field have a bi-fractal structure and are intermittent. The intermittency implies that the structure function defined by, $C_b=\langle\|\varphi(x_2)-\varphi(x_1)\|^b\rangle$, scales as $\|x_2-x_1\|^{\zeta_b}$, where $\zeta_b$ is a nonlinear function of $b$. It is also shown numerically that the moments of the field’s increments, $\langle\|\varphi(x_2)-\varphi(x_1)\|^b\rangle$, behave as, $\|x_2-x_1\|^{\xi_b}$, where $x_b=b$ for $b\leq 1$, and $\xi_b=1$ for $b\geq1$. The rest of this paper is organized as follows. In the next section we present the model that we wish to study, and describe some of its properties by solving it numerically. In III and IV we derive master equations for the order parameter of the model, and for the field’s increments and its PDF tail. The numerical simulations for extracting the scaling exponents are described in V. The paper is summarized in VI, while the Appendices provide some technical details of the work that we present in the main part of the paper. The Model and the Coupling Constant =================================== The standard GL $\varphi^4$-theory describes a second-order phase transition in any system with a one-component order parameter $\varphi(x)$ and the $\varphi\to -\varphi$ symmetry in a zero external field. The theory is described by the following action, $$S=\int\left[-\frac{k}{2}(\nabla\varphi)^2+\frac{\tau}{2!}\varphi^2 -\frac{g}{4!}\varphi^4\right]d^dx\;,$$ where $\tau=T-T_c$, with $T_c$ being the critical temperature, and $k$ is the diffusion coefficient. We consider the case in which $\tau > 0$. For $d\geq 2$, the critical temperature $T_c$ is finite, while in one dimension (1D), $T_c=0$. The parameter $g$ characterizes the strength of the fluctuation interaction, or the coupling constant. The equation of motion is given by, $$k\nabla^2\varphi+\tau\varphi-\frac{g}{6}\varphi^3=0.$$ The critical dynamics of the system is described by a stochastic equation of a particular form - the Langevin equation (for a comprehensive review on Langevin type equations see [@Risken] )- given by $$\frac{\partial\varphi}{\partial t}=k\nabla^2\varphi+\tau\varphi-\frac{g}{6} \varphi^3+\eta({\bf x},t)\;,$$ where $\eta({\bf x},t)$ is a Gaussian-distributed noise with zero average and the correlation function, $$\langle\eta({\bf x},t)\eta({\bf x}',t')\rangle=D_0D({\bf x}-{\bf x}') \delta(t-t')\;,$$ with $D({\bf x}-{\bf x}')$ being an arbitrary smooth function. Typically, the spatial correlation of the forcing term is considered to be a delta function in order to mimic short-range correlations. Here, though, the spatial correlation is defined by $$\label{} D({\bf x}-{\bf x}')=\frac{1}{({\pi\sigma^2})^{d/2}}\exp\left[-\frac{({\bf x}- {\bf x}')^2}{\sigma^2}\right]\;,$$ where $\sigma\ll L$ endows a short-range character to the random forcing. It is useful to rescale Eq. (3) by writing, $\varphi'=\varphi/\varphi_0$, $x'=x/x_0$, and $t'=t/t_0$. If we let $t_0=1/\tau$, $\varphi_0=(6\tau/g)^{1/2}$, and, $x_0= [(d_0g)/(6\tau^2)]^{1/d}$, all the parameters are eliminated in Eq. (3) except for, $k'=[6/(D_0g)]^{2/d}\tau^{4/d-1}k$, and one finds that $$\label{eqphi4} \frac{\partial\varphi}{\partial t}=k'\nabla^2\varphi+\varphi-\varphi^3 +\eta({\bf x},t)\;,$$ where $k'$ is now the effective coupling constant of the theory. The weak and strong coupling limits of the theory are then defined, respectively, by, $k'\to \infty$ and $k'\to 0$. In the weak coupling limit one can use numerical simulations and the Feynman diagrams to calculate the critical exponents. On the other hand, to solve the problem in the strong coupling limit we need other techniques to derive the stochastic properties of the fluctuation field \[16\]. The nonlinearity of Eq. (6) in the strong coupling limit gives rise to the possibility of formation of singularity in a finite time. This means that there is a competition between the smoothing effect of diffusion (the Laplacian term) and the $\varphi^3$ term. Let us now describe the main properties of the GL theory in the limit, $k'\to 0$. i\) The unforced GL model \[$\eta({\bf x},t)=0$\], with given initial conditions, develops singularities in any spatial dimension. In one spatial dimension (1D) the singularities are developed in a finite time $t_c$ as $k'\to 0$. At such singular points the field $\varphi$, representing an order parameter, is not continuous. In 2D the unforced GL model develops domain walls, characterized by singular lines with finite lengths (that depend on the initial condition). Under these conditions, the field $\varphi$ is discontinuous when crossing the singular lines. In three and higher dimensions the structure of the singularities can be more complex. For example, in 3D the singularities are domain walls where the field $\varphi$ is discontinuous. In Figs. 1 we show the time evolution of the order parameter $\varphi$ of the unforced GL model in 2D, in the limit $k'\to 0$ \[Eq. (6) with $\eta=0$\]. We have used the finite-element method to numerically solve the Langevin equation with $k'\to 0$ and the initial condition, $\varphi(x,y,0)=\sin x\sin y$. Such initial conditions are typical, and were used only for simplicity. The time scale for reaching the singularity is of the order of $k'^{-1/2}$. As Figs. 1a and 1b indicate, it is evident that at times $t<t_c$ (in the limit, $k'\to 0$) the $\varphi$ field is continuous. At $t=t_c$ the $\varphi$ field becomes singular; see Fig. 1c. \(ii) Similarly, for a forcing term which is white noise in time and smooth in space, singularities are developed in any spatial dimension in the strong coupling limit and in a finite time, $t_\eta =t_{c,\eta}$, as $k'\to 0$. For example, in 2D the boundaries of the domain walls are smooth curves. In Fig. 2 we demonstrate the time evolution of the order parameter $\varphi$ of the forced GL model in 2D in the limit $k'\to 0$. Starting from a smooth initial condition, as shown in Figs. 2a and 2b, it is evident that for times $t<t_{c,\eta}$ the $\varphi$ field is continuous. At $t=t_{c,\eta}$ the field becomes singular; see Fig. 2c. \[fig01\] =6truecm =6truecm =6truecm \[fig02\] =6truecm =6truecm =6truecm Master Equation of the Order Parameter ====================================== In this section we derive a master equation to describe the time evolution of the PDF $P(\varphi,t)$ of the order parameter $\varphi$. Defining a one-point generating function by, $Z(\lambda)=\langle\Theta\rangle$, where $\Theta$ is defined by, $\Theta=\exp[-i\lambda\varphi(x,t)]$. Using Eq. (6), the time evolution of $Z$ is governed by $$\begin{aligned} Z_t=-i\lambda k'\langle\nabla^2\varphi\exp[-i\lambda\varphi(x,t)]\rangle -i\lambda\langle\varphi\Theta\rangle\nonumber\\+i\lambda\langle\varphi^3\Theta \rangle-i\lambda\langle\eta\Theta\rangle-\lambda^2k(0)Z\;,\end{aligned}$$ where, $k(x)=D_0D({\bf x})$, and we have invoked Novikov’s theorem (see Appendix I), which is expressed via the relation, $$\label{nov} \langle\eta\exp[-i\lambda\varphi(x,t)]\rangle=-i\lambda k(0) Z\;.$$ Now, using the identities, $-i\lambda\langle\varphi\exp[-i\lambda\varphi(x,t)] \rangle=\lambda Z_\lambda$, and $-i\lambda\langle\varphi^3\exp[-i\lambda \varphi({\bf x},t)]\rangle=\lambda Z_\lambda\lambda\lambda$, the generating function $Z$ satisfies the following unclosed master equation $$\begin{aligned} \label{ZFG} Z_t=-i\lambda k'\langle\nabla^2\varphi\exp[-i\lambda\varphi(x,t)]\rangle +\lambda Z_\lambda+\lambda Z_{\lambda\lambda\lambda}\nonumber\\ -\lambda^2 k(0)Z\;.\end{aligned}$$ The $-i\lambda k'\langle\nabla^2\varphi\exp[-i\lambda\varphi(x,t)]\rangle$ term of Eq. (\[ZFG\]) is the only one which is not closed with respect to $Z$. The PDF of order parameter $P(\varphi)$ is constructed by Fourier transforming the generating function $Z$: $$P(\varphi,t)=\int\frac{d\lambda}{2\pi}\exp(i\lambda\varphi)Z(\lambda,t)\;.$$ Thus, $$\begin{aligned} P_{t}&=&-[(\varphi-\varphi^3)P]_{\varphi}+k(0)P_{\varphi\varphi}\nonumber\\ &-&ik'\int\frac{d\lambda}{2\pi}\lambda\exp(i\lambda\varphi) \langle\nabla^2\varphi\exp[-i\lambda\varphi(x,t)]\rangle\;.\end{aligned}$$ It is evident that the governing equation for $P(\varphi,t)$ is also not closed. Let us now use the boundary layer technique to prove that the unclosed term \[the last term of Eq. (11)\] makes, in the strong coupling limit, no contribution to the governing equation for the PDF[@BO; @16]. We consider two different time scales in the limit, $k'\to 0$. (i) Early stages before developing the singularities ($t<t_{c,\eta}$), and (ii) in the regime of established stationary state with fully-developed sharp singularities ($t \geq t_{c,\eta}$). In regime (i), ignoring the relaxation term in the governing equation for the PDF, one finds, in the limit $k'\to 0$, the exact equation for the time evolution of the PDF for the order parameter (see below for more details). In contrast, the limit $k'\to 0$ is singular in regime (ii), leading to an unclosed term (the relaxation term) in the equation for the PDF. However, we show that the unclosed term scales as $k'^{1/2}$, implying that this term, in the strong coupling limit, makes no finite contribution or anomaly to the solution of Eq. (11). It is known for such time scales (the stationary state) that the $\varphi$-field, which satisfies the Langevin equation, gives rise to discontinuous solutions in the limit, $k'\to 0$. Consequently, for finite $\sigma$ the singular solutions form a set of points where the domain walls are located, and are continuously connected. We should note that $k' \varphi_{xx}$, in the limit $k'\to 0$, is zero at those points at which there is no singularity. Therefore, in the limit $k'\to 0$ only small intervals around the walls contribute to the integral in Eq. (11). Within these intervals, a boundary layer analysis may be used for obtaining accurate approximation of $\varphi({\bf x},t)$. Generally speaking, the boundary layer analysis deals with problems in which the perturbations are operative over very narrow regions, across which the dependent variables undergo very rapid changes. The narrow regions, usually referred to as the domain walls, frequently adjoin the boundaries of the domain of interest, due to the fact that a small parameter ($k'$ in the present problem) multiplies the highest derivative. A powerful method for treating the boundary layer problems is the method of matched asymptotic expansions. The basic idea underlying this method is that, an approximate solution to a given problem is sought, not as a single expansion in terms of a single scale, but as two or more separate expansions in terms of two or more scales, each of which is valid in some part of the domain. The scales are selected such that the expansion as a whole covers the entire domain of interest, and the domains of validity of neighboring expansions overlap. In order to handle the rapid variations in the domain walls’ layers, we define a suitable magnified or stretched scale and expand the functions in terms of it in the domain walls’ regions. For this purpose, we split $\varphi$ into a sum of inner solution near the domain walls and an outer solution away from the singularity lines, and use systematic matched asymptotics to construct a uniform approximation of $\varphi$. For the outer solution, we look for an approximation in the form of a series in $k'$, $$\varphi=\varphi^{\rm out}=\varphi_0+k'\varphi_1+O(k'^2)\;,$$ where $\varphi_0$ satisfies the following equation $$\varphi_{0t}=\varphi_0-\varphi_0^3+\eta({\bf x},t)\;.$$ Indeed, $\varphi_0$ satisfies Eq. (6) with $k'=0$. Far from the singular points or lines, the PDF of $\varphi_0$ satisfies the Fokker-Planck equation, with the drift and diffusion coefficients being, $D^{(1)}(\varphi_0,t)=\varphi_0- \varphi_0^3 $, and $D^{(2)}=k(0)$, respectively. Reference \[16\] gives the solution of the time-dependent Fokker-Planck equation with such drift and diffusion coefficients. At long times and in the area far from the singular points or lines, the PDF of $\varphi_0$ will have two maxima at $\pm 1$. This means that we are dealing with the smooth areas in Fig. 2c in the stationary state. In order to deal with the inner solution around the domain walls, we consider the $x$ component [normal]{} to the domain wall or singularity line, and decompose the operator $\nabla^2$ as $\partial_{xx}+\nabla_{d-1}^2$. In the strong coupling limit, $k'\to 0$, the term $\nabla_{d-1}^2\varphi$ makes no contribution to the PDF equation, whereas the term $\partial_{xx}\varphi$ is singular. To derive the long-time solution of Eq. (6), we rescale $x$ to $z\equiv\frac{x}{\sqrt{2k'}}$ and suppose that complete solution of Eq. (6) has the form, $\varphi(z,t)=f(z,t)+\tanh(z)$. All the effects of the initial condition and time-dependence of $\varphi(x,t)$ will then be contained in $f(z,t)$. We now rewrite Eq. (6) with the new variables to obtain $$\begin{aligned} \label{eqnnphi4} \partial_t f(z,t)=&&\frac{1}{2}\partial_{zz}f(z,t)+f(z,t)-f^3(z,t)\nonumber\\ && -3f(z,t)\tanh(z)[f(z,t)+\tanh(z)]\nonumber\\&&+\sqrt[4]{2k'}\eta(z,t)\;.\end{aligned}$$ The last term of Eq. (14) is zero in the limit $k'\to 0$. Multiplying Eq. (14) by $f(z,t)$ and integrating over $z$, one finds that, $$\begin{aligned} && \partial_t\int dz f^2(z,t)=- \int dz (\partial_{z}f(z,t))^2 + 2 \int dz f^2(z,t)\nonumber\\&&- 2 \int dz f^4(z,t) -6\int dz f^3(z,t)a(z) \nonumber \\ &&- 6\int dz a^2(z)f(z,t)^2\;,\end{aligned}$$ where $a(z)=\tanh(z)$. We show in Fig. (3) the time variations of $\int dz f^2(z,t) $ verses $t$ with different types of initial conditions. The results show that $\int dz f^2(z,t)$ vanishes at long times. Therefore, $\varphi(z,t)\to\tanh(z)$, in the limit of a stationary state. Let us now compute the contribution of the unclosed term in Eq. (11) in the stationary state, i.e., $$\begin{aligned} && -k'\int \frac{d\lambda}{2\pi}i\lambda e^{i\lambda\varphi}\langle\nabla^2\varphi e^{-i\lambda\varphi(x,t)}\rangle\cr \nonumber\\&& = -k'\left(\int \frac{d\lambda}{2\pi} e^{i\lambda\varphi}\langle\nabla^2\varphi e^{-i\lambda\varphi(x,t)}\rangle\right)_\varphi\nonumber\\&& =-k'\langle\nabla^2\varphi\delta[\varphi-\varphi(x,t)]\rangle_\varphi\end{aligned}$$ In the second line of Eq. (16), we have replaced $i\lambda$ with differentiation with respect to $\varphi$ and in the third line the integration of $\lambda$ has been carried through. Now, assuming ergodicity, the term $k'\langle\nabla^2\varphi\delta(\varphi-\varphi(x,t))\rangle$ is converted to, $$=-k'\lim_{V\to\infty}\frac{1}{V}\int_V dxdv_{d-1}\nabla^2\varphi \delta[\varphi-\varphi(x,t)]\;.$$ In the limit $k'\to 0$, only at points where we have singularity the above term is not zero. Therefore, we approach the domain walls’ regions as $$\begin{aligned} = -k'\lim_{V\to\infty}\frac{1}{V}\sum_j\int_{\Omega_j} dxdv_{d-1}(\varphi_{xx}+\nabla^{2}_{d-1}\varphi)\nonumber\\ \times\delta[\varphi-\varphi(x,t)]\;,\end{aligned}$$ where $\Omega_j$ is the space close to the domain walls. Therefore, Eq. (16) is written as $$= -k'\lim_{V\to\infty}\frac{1}{V}\sum_j\int_{\Omega_j} dxdv_{d-1} \varphi_{xx}\delta[\varphi-\varphi(x,t)]\;.$$ Changing the variables from $x$ to $z$ and integrating over $dv_{d-1}$, we have $$\begin{aligned} = -k'\lim_{V\to\infty}\frac{V_{d-1}}{V}\sum_j\int_{-\infty}^{+\infty} \epsilon dz \frac{1}{\epsilon^2}\varphi_{zz}\delta[\varphi-\varphi(z,t)]\\ = -\frac{k'}{\epsilon}\lim_{V\to\infty}\frac{V_{d-1}}{V}\sum_j \int_{-\infty}^{+\infty}dz\varphi_{zz}\delta[\varphi-\varphi(z,t)]\\\end{aligned}$$ where $\epsilon=(2k')^{1/2}$. Assuming statistical homogeneity, one finds, $$= -\frac{k'}{\epsilon}\lim_{V\to\infty}\frac{N\times V_{d-1}}{V} \int_{-\infty}^{+\infty}dz \varphi_{zz}\delta[\varphi-\varphi(z,t)]\\$$ where $N$ is number of singular lines. The quantity $NV_{d-1}/V$ is the density of the singular lines, and $k'/\epsilon=(k'/2)^{1/2}$. In the limit, $V\to \infty$, we denote the density of the singularities by $\rho$. Therefore, $$=-(k'/2)^{1/2}\rho\int_{-\infty}^{+\infty}dz \varphi_{zz} \delta[\varphi-\varphi(z,t)]\;.$$ Now, by changing the integration variable form $z$ to $\varphi$, we can determine the integral exactly. We find that, $$\int_{-\infty}^{+\infty}dz \varphi_{zz}\delta[\varphi-\varphi(z,t)] =\int_{-1}^{+1} d\varphi \frac{\varphi_{zz}}{\varphi_{z}}\delta[\varphi-\varphi(z,t)]\;.$$ Using Eq. (6) in the limit, $t\to\infty$, we determine $\varphi_{zz}$ and $\varphi_z$ in terms of $\varphi$. Multiplying Eq. (6) by $\varphi_z$ and integrating over $z$, we obtain, $$\varphi_z^2=\frac{1}{2}\varphi^4-\varphi^2+C\;,$$ where $C$ is an integration constant. In the limit, $z\to\pm\infty$, $\varphi_z=\varphi=\pm 1$. Therefore, $C=1/2$, and $\varphi_{zz}/\varphi_z$ is written as, $$\begin{aligned} \frac{\varphi_{zz}}{\varphi_z}=\frac{\varphi^3-\varphi}{\sqrt{\|\frac{1}{2} \varphi^4-\varphi^2+\frac{1}{2}\|}}\\ =\frac{\sqrt{2}\varphi(\varphi^2-1)}{\|\varphi^2-1\|}\\ =\sqrt{2}\varphi\times sign(\varphi^2-1)\;.\end{aligned}$$ The integral in Eq. (22) is now given by, $$\begin{aligned} &&\int_{-1}^{+1}d\varphi\frac{\varphi_{zz}}{\varphi_z} \delta[\varphi-\varphi(z,t)]\\ &=&\int_{-1}^{+1} d\varphi(z) \sqrt{2}\varphi(z)\times sign(\varphi(z)^2-1)\delta(\varphi-\varphi(z,t))\\ &=&\sqrt{2}\varphi \hskip 0.1cm sign(\varphi^2 -1) \theta(1-\varphi^2).\end{aligned}$$ Now, the unclosed term in Eq. (16) is written as $$\begin{aligned} &&-k'\langle\nabla^2\varphi\delta[\varphi-\varphi(x,t)]\rangle_{\varphi}=\\ &=& -(k'/2)^{1/2}\rho\{A+2\sqrt{2}\varphi^2\delta(\varphi^2-1)\},\end{aligned}$$ where, $A=2\sqrt{2}\theta(1-\varphi^2)\theta(\varphi^2-1)-\sqrt{2} \theta(1-\varphi^2)$. Therefore, in the limit $k'\to 0$, the master equation takes on the following form $$\begin{aligned} 0&=&-[(\varphi-\varphi^3)P]_\varphi+k(0)P_{\varphi\varphi}- \cr \nonumber \\ &&(k'/2)^{1/2}\rho[A+2\sqrt{2}\varphi^2\delta(\varphi^2-1)]\;,\end{aligned}$$ where we set $P_t=0$ in the stationary state. The PDF $P$ is continuous at $\varphi=\pm 1$, but its derivative is not. By integrating Eq. (24) in the interval $[1-\epsilon,1+\epsilon]$ (or $[-1-\epsilon, -1+\epsilon]$), one finds $$\Delta P_\varphi\|_{\varphi=\pm 1}=\frac{ (k')^{1/2}\rho}{k(0)}\;.$$ In the limit, $k'\to 0$ the derivative of the PDF will also be continuous. Considering the factor $(k')^{1/2}$ in Eq. (24), we conclude that, in the strong coupling limit, the unclosed term is identically zero. This means that there is no anomaly or finite term in the strong coupling limit in the master equation for the PDF of the order parameter $\varphi$. The stationary solution of Eq. (24), in the limit, $k'\to 0$, takes on the following expression, $$P_{st}=N\exp\left[\frac{-\varphi^4+2\varphi^2}{4k(0)}\right]\;,$$ where the normalization constant is given by $$\frac{1}{N}=\frac{1}{\sqrt{2}}\exp\left[\frac{1}{8k(0)}\right] K_{\frac{1}{4}} \left(\frac{1}{8k(0)}\right)\;.$$ where the $K_ a (b)$ is the modified Bessel functions. To derive the moments of $\langle\varphi^n\rangle$ in the stationary state, we multiply Eq. (25) by $\varphi^n$ and integrate the result over $\varphi$ to obtain $$\begin{aligned} n\langle\varphi^n\rangle-n\langle\varphi^{n+2}\rangle\nonumber\\ +n(n-1)k(0)n\langle\varphi^{n-2}\rangle=0\;.\end{aligned}$$ Equation (28) is a recursive equation for computing all the moments in terms of the second-order one. Direct calculation then shows that, $$\langle\varphi^2\rangle=-\frac{K_{\frac{1}{4}} \left(\frac{1}{8k(0)}\right) - K_{\frac{3}{4}} \left(\frac{1}{8k(0)}\right)}{2 K_{\frac{1}{4}} \left(\frac{1}{8k(0)}\right)}\;,$$ and all the odd moments vanish, $\langle\varphi^{2k+1}\rangle=0$. Therefore, using Eqs. (28) and (29), we are able to derive all the moments of the order parameter in the $d-$dimensional Ginzburg - Landau theory in the strong coupling limit. =9truecm Master equation for the increments and their PDF tail ===================================================== In this section we derive the PDF and the scaling properties of the moments of the increments, $P(\varphi(x_2)-\varphi(x_1))$, for the $\varphi^4$-theory in the strong-coupling limit. Defining the two-point generating function by, $Z(\lambda)=\langle\Theta\rangle$, where $\Theta$ is defined as $$\Theta= e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)}$$ the time evolution of $Z$ is related to that of $\varphi$ by $$\begin{aligned} &&Z_t=-i\lambda_1 \langle\varphi_t(x_1,t)e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)} \rangle\nonumber\\ &&-i\lambda_2 \langle\varphi_t(x_2,t)e^{-i\lambda_2\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)} \rangle\;.\end{aligned}$$ Substituting $\varphi_t (x_1,t)$ and $\varphi_t (x_2,t)$ from Eq. (6), the governing equation for the generating function satisfies, $$\begin{aligned} &&Z_t=\lambda_1 Z_{\lambda_1}+\lambda_1 Z_{\lambda_1\lambda_1\lambda_1}+\lambda_2 z_{\lambda_2}+\lambda_2 Z_{\lambda_2\lambda_2\lambda_2}\nonumber\\ &&-(\lambda_1^2+\lambda_2^2)k(0)Z-2\lambda_1\lambda_2k(x)Z\;,\end{aligned}$$ Where, $x=x_1-x_2$, and we have used the fact that in the strong-coupling limit, the Laplacian term makes no contribution to the PDF equation (see Appendix II for more details). Moreover, we have invoked the generalized Novikov’s theorem for the two-point generating function, according to which [@16] (see also Appendix I), $$\begin{aligned} &&-i\lambda_1\langle\eta(x_1) e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)}\rangle\nonumber\\ &&-i\lambda_2\langle\eta(x_2) e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)}\rangle\nonumber\\ &&=-(\lambda_1^2+\lambda_2^2)k(0)Z-2\lambda_1\lambda_2k(x)Z\;.\end{aligned}$$ Fourier transforming Eq. (32), the governing equation of the joint PDF will be given by $$\begin{aligned} &&P_t(\varphi_1,\varphi_2)= -[(\varphi_1-\varphi_1^3)P]_{\varphi_1}\nonumber\\ &&-[(\varphi_2-\varphi_2^3)P]_{\varphi_2} +k(0)(P_{\varphi_1,\varphi_1}+P_{\varphi_2,\varphi_2})\nonumber\\ &&+2k(x)P_{\varphi_1,\varphi_2}\;.\end{aligned}$$ It is useful to change the variables as, $ \varphi_1=(w-u)/2$, and, $\varphi_2= (w+u)/2$, and, therefore, $d/d\varphi_1=d/dw-d/du$, and, $d/d\varphi_2= d/dw +d/du$. Now, Eq. (34) can be written as $$\begin{aligned} && P_t(w,u)=-(wP)_w-(uP)_u +\nonumber\\ && [\frac{1}{4}(w^3+3wu^2)P]_w+[\frac{1}{4}(u^3+3uw^2)P]_u\nonumber\\ && 2k(0)(P_{ww}+P_{uu})+2k(x)(P_{ww}-P_{uu})\;.\end{aligned}$$ To derive the governing equation for the PDF of the increments, $u=\varphi_2-\varphi_1$, we integrate over $w$ to find that, $$\begin{aligned} && P_t(u)=-(uP)_u+\frac{1}{4}(u^3P)_u\nonumber\\ && \frac{3}{4}(u\langle w^2\mid u\rangle P)_u+2[k(0)-k(x)]P_{uu}\;,\end{aligned}$$ where we have used the fact that the joint PDF $P(w,u)$ can be written as, $P(w\mid u)P(u)$. It is evident that we cannot derive a closed equation for the PDF of $u$. Indeed, to determine $P(u)$ we need to know the conditional averaging $\langle w^2\mid u\rangle$. However, one can derive the tail (both the left and right ones) of the PDF in the limit, $u\to\infty$. To determine the tail we note that only near the singularities, in small separation in space, one finds a large difference in the field $\varphi$ and, hence, large $u$. On the other hand, near such points or lines, the field $w$ will be very small. Therefore, in the limit $u\to\infty$, we can ignore the conditional averaging to find that, $$\lim_{u\to\infty}\langle w^2\mid u\rangle\simeq 0\;.$$ Therefore, in the limit, $u\to\infty$, we obtain the following behavior for the tails of the $P(u)$ in the stationary state, $$P_{st}(\phi_2-\phi_1\to\infty)\sim\exp\left\{-\frac{(\phi_2-\phi_1)^4} {32[k(0)-k(x)]}\right\}\;.$$ To derive the scaling behavior of the moments $\langle u^n\rangle$ one needs to know the entire range of the behavior of the increments’ PDF. Here, we are able to only derive the equation for the shape of the PDF tails. In the next section, we investigate by numerical simulation the scaling behavior of the moments $\langle u^n\rangle$ vs. the separation $x$. scaling exponents of the moments: numerical simulation ====================================================== To calculate numerically the scaling behavior of the moments with $x$, when $x << 1$, we shall use here the initial-value problem for the two-dimensional Langevin equation, Eq. (6), in the limit, $k'\to 0$, when the force is concentrated at discrete times \[17-20\]: $$f(x,y,t)=\sum_j f_j(x,y)\;\delta(t-t_j), \label{kickforce}$$ where both the “impulses” $f_j(x,y)$ and the “kicking times” $t_j$ are prescribed (deterministic or random). The kicking times are ordered and form a finite or an infinite sequence. The impulses are always taken to be smooth and acting only at large scales. The precise meaning that we ascribe to the dynamical Langevin equation with such a forcing is that, at time $t_j$, the solution $\varphi(x,y,t)$ changes discontinuously by the amount $f_j(x,y)$, $$\varphi(x,y,t_{j+})=\varphi(x,y,t_{j-})+f_j(x,y)\;, \label{discon}$$ whereas between $t_{j+}$ and $t_{(j+1)-}$ the solution evolves according to the unforced $\varphi^4$ equation, $$\partial_t\varphi= \varphi - \varphi^3 + k'\nabla^2 \varphi. \label{equnf}$$ Without loss of generality, we may assume that the earliest kicking time is, $t_{j_0}=t_0$, provided that we set, $f_{j_0}=f_0$, and, $\varphi(x,y,t_{j_0-})=\varphi(x,y,t_{0-})$ for $t<t_0$. Therefore, starting from $t_0$, according to Eq. (\[discon\]) we obtain $$\varphi(x,,y,t_{0+})= \varphi(x,y,t_{0-})+f_0(x,y)\;, \label{eqdiscon}$$ and beyond that up to $t_{1-}$, according to Eq. (\[equnf\]), $$\begin{aligned} && \varphi(x,y,t_{1-})=(1+h)\varphi(x,y,t_{0+})-\varphi^3(x,y,t_{0+}), \nonumber\\ && h=t_1-t_0 \label{eqalgorithm}\end{aligned}$$ where, $h=t_1-t_0$. =9truecm It is clear that any force $f(x,y,t)$ which is continuously acting in time can be approximated in this way by selecting the kicking times sufficiently close. Hereafter, we shall consider exclusively the case where the kicking is periodic in both space and time. Specifically, we assume that the force in the $\varphi^4$ equation is given by $$\begin{aligned} f(x,y,t) &=& g(x,y)\sum_{j=-\infty}^{+\infty} \delta(t-jT),\label{forceperiodic}\\ g(x,y)&\equiv& -\nabla G(x,y), \label{defg}\end{aligned}$$ where $G(x,y)$, the kicking potential, is a deterministic function of ($x,y$) which is periodic and sufficiently smooth (e.g., analytic), and $T$ is the kicking period. The numerical experiments reported hereafter were made with the kicking potential $G(x,y)=G_1(x) G_1(y)$, where $G_1(q)$ is given by $$G_1(q)= \frac{1}{3}\sin 3q +\cos q, \label{defG}$$ and the kicking period, $T=10^{-6}$. The number of collocation points chosen for our simulations is generally $N_x=10^3$. In Fig. 4 we plot the PDF of $\varphi$ according to Eq. (26), and compare it with the numerical results. In Fig. 5, the moments of $\varphi$ increments, $\langle\|\varphi(x_2+\delta x)-\varphi(x_1)\|^b\rangle$, are calculated numerically as a function of $x=\|x_2-x_1\|$ for several values of $b$ (with $0< b < 1$, and $1\leq b $) and its scaling exponents $\xi_b$ for $ x << 1$ are checked. The results indicate that with good precision $\langle \|\delta\varphi\|^b\rangle$ scales with $x$ with an exponent $1$ for $b> 1$; otherwise, it scales with $x$ with exponents $\xi_b=b$. Values of $\xi_b$ are given in Fig. 6. The bi-fractal behavior of the exponents is a consequence of the presence of the domain walls. Indeed, the structure function, $$C_b =\langle\|\varphi(x_2)-\varphi(x_1)\|^b\rangle\;,$$ for $b > 0$ behaves, for small $\Delta x=\|x_2-x_1\|$ as, $$C_b \sim A_b \|\Delta x\|^b + A'_b \|\Delta x\|$$ where the first term is due to the regular (smooth) parts of the order parameter $\varphi$, while the second one is contributed by the $O(\|\Delta x\|)$ probability to have a domain wall somewhere in an interval of length $\|\Delta x\|$. For $0 < b <1$ the first term dominates as $\|\Delta x\|\to 0$, while, for $b>1$ it is the second term that does so. Summary ======= We studied the domain wall-type solutions in the $\varphi^4$-theory in the strong-coupling limit, $k'\to 0$, in which the equation develops singularities. The scaling behavior of the moments of differences of $\varphi$, $\delta\varphi=\varphi(x_2)-\varphi(x_1)$, and the PDF of $\varphi$, i.e., $P(\varphi)$, were all determined. It was shown that in the stationary state, where the singularities are fully developed, the relaxation term in the strong-coupling limit leads to an unclosed term in the equation for the PDF. However, we showed that the unclosed term can be omitted in the strong-coupling limit. We proved that to leading order, when $\|x_2-x_1\|$ is small, fluctuation of the $\varphi$ field is intermittent for $b\geq 1$. The intermittency implies that, $C_b=\langle\|\varphi(x_1)-\varphi(x_2)\|^b>$ scales as $\|x_1-x_2\|^{\xi_b}$, where $\xi_b$ is a constant. It was shown, numerically, that for the space scale $\|x_2 - x_1\|$ and $b\geq 1$, the exponents $\xi_b$ are equal to $1$. APPENDIX I ========== In this appendix we provide a proof of Novikov’s theorem. Consider the general stochastic differential equation with the following form, $$\label{eq1} \frac{\partial}{\partial t}\varphi=-\frac{1}{2}L[\varphi(x,t)]+\eta(x,t)\;.$$ where $L$ is an operator acting on $\varphi$, and $\eta$ is a Gaussian noise with the correlation, $$\langle\eta(x,t)\eta(x',t')\rangle=k(x-x')\delta(t-t')\;.$$ The PDF of the random noise has the following form, $$\begin{aligned} &&[d\rho(\eta)]=[d\eta]\times\nonumber\\ &&exp[-\frac{1}{2}\int d^dx d^dx'dt dt' \eta(x,t)B(x-x')\delta(t-t')\eta(x',t')]\;,\nonumber\end{aligned}$$ where $B(x-x')$ is the inverse of $k(x-x')$, so that, $$\label{eqortogonal} \int k(x-x')B(x'-x'')d^d x=\delta(x-x'')\;.$$ we write the average of $\eta(x,t)F(\eta)$ over the noise realization as: $$\langle\eta(x,t)F(\eta)\rangle=\int \eta(x,t)F(\eta)[d\rho(\eta)]\;.$$ By integrating by parts and using Eq. (\[eqortogonal\]), one finds that, $$\begin{aligned} && \langle\eta(x,t)F(\eta)\rangle=\nonumber\\ &&\int d^dx'' dt'' \langle\eta(x,t)\eta(x'',t'')\rangle\langle\frac{\partial F}{\partial\eta(x'',t'')}\rangle\;.\end{aligned}$$ Now, let us assume the function $F$ to have the following form, $$F[\eta]=exp(-i\lambda\varphi(x',t))\;.$$ so that one finds, $$\frac{\partial F}{\partial\eta(x'',t'')}=-i\lambda\frac{\partial \varphi(x',t)}{\partial\eta(x'',t'')}F[\eta]\;,$$ Integrating Eq. (\[eq1\]) with respect to $t$, we find that, $$\begin{aligned} &&\varphi(x',t)=\nonumber\\ &&\varphi(x',t_0)-\frac{1}{2}\int_{t_0}^{t} dt'' L[\varphi(x',t'')]+\int_{t_0}^{t} dt''\eta(x',t'')\;.\end{aligned}$$ This allows us to show that, $$\begin{aligned} \label{eq2} &&\frac{\partial\varphi(x',t)}{\partial\eta(x'',t'')}=\nonumber\\ &&-\frac{1}{2}\int_{t_0}^{t} dt'''\frac{\partial L[\varphi(x,t''')]}{\partial\eta(x'',t'')}+\delta(x'-x'')\theta(t-t'')\;.\end{aligned}$$ where, in the limit $t''\rightarrow t$, the first term of the right-hand side of Eq. (\[eq2\]) will vanish, and we can write $$\begin{aligned} &&\langle\eta(x,t)\exp[-i\lambda\varphi(x',t)]\rangle\nonumber\\ &&=(-i\lambda)k(x-x')\langle\exp(-i\lambda\varphi(x',t))\rangle\;.\end{aligned}$$ where we used, $\theta(0)=1$ appendix II =========== In this Appendix we prove that, for example in Eq. (24), the relaxation term $k'\nabla^2\varphi$ makes no contribution or anomaly to the PDF of the increments, in the limit $k'\to 0$. =9truecm =9truecm The joint probability distribution $P(\varphi_1,\varphi_2)$ satisfies the following equation, $$\begin{aligned} && P_t(\varphi_1,\varphi_2)=-[(\varphi_1-\varphi_1^3)P]_{\varphi_1}\nonumber\\ &&-[(\varphi_2-\varphi_2^3)P]_{\varphi_2}+k(0)(P_{\varphi_1,\varphi_1} +P_{\varphi_2,\varphi_2})\nonumber\\ &&+2k(x)P_{\varphi_1,\varphi_2}\nonumber\\ &&-i k'\int \frac{d\lambda_1}{2\pi}\frac{d\lambda_2}{2\pi} \lambda_1\exp(i\lambda_1\varphi_1+i\lambda_2\varphi_2)\nonumber\\ &&\times\langle\nabla^2\varphi(x_1) \exp[-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)]\rangle\nonumber\\ &&-i k'\int \frac{d\lambda_1}{2\pi}\frac{d\lambda_2}{2\pi} \lambda_2\exp(i\lambda_1\varphi_1+i\lambda_2\varphi_2)\nonumber\\ &&\times\langle\nabla^2\varphi(x_2) \exp[-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)]\rangle\;.\end{aligned}$$ The last two terms in Eq. (47) are not closed with respect to the PDF. Let us then compute the contribution of the unclosed terms. They can be written as, $$\begin{aligned} &&-i k^{'}\int \frac{d\lambda_1}{2\pi}\frac{d\lambda_2}{2\pi} \lambda_i exp(i\lambda_1\varphi_1+i\lambda_2\varphi_2)\\\nonumber &&\times<\nabla^2\varphi(x_i) e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)}>\cr \nonumber\\&& =-i k^{'}(\int \frac{d\lambda_1}{2\pi}\frac{d\lambda_2}{2\pi} exp(i\lambda_1\varphi_1+i\lambda_2\varphi_2)\\\nonumber &&\times<\nabla^2\varphi(x_i) e^{-i\lambda_1\varphi(x_1,t)-i\lambda_2\varphi(x_2,t)}>)_{\varphi_i} \nonumber\\&& =-k'\langle\nabla^2\varphi(x_i)\delta(\varphi_1-\varphi(x_1,t))\\\nonumber &&\times\delta(\varphi_2-\varphi(x_2,t))\rangle_{\varphi_i}\end{aligned}$$ Consider one of the terms in the above equation, for example, $-k'\langle\nabla^2\varphi(x_i)\delta[\varphi_1-\varphi(x_1,t)] \times\delta[\varphi_2-\varphi(x_2,t)]\rangle_{\varphi_i}$. Assuming ergodicity, it is written as, $$= k'\lim_{V\to\infty}\frac{1}{V}\int_V dx_i dv^i_{d-1}\nabla^2\varphi(x_i)\delta[\varphi_i-\varphi(x_i,t)]$$ in the limit, $k'\to 0$ limit, only at the points where we have singularity this term is not zero. Therefore, we restrict ourselves to the space near the domain walls, $$\begin{aligned} = -k'\lim_{V\to\infty}\frac{1}{V} \sum_{j}\int_{\Omega_{j}} dx_idv^i_{d-1}(\varphi_{x_ix_i}+\nabla^{2}_{d-1}\varphi)\nonumber\\ \times\delta[\varphi_i-\varphi(x_i,t)]\end{aligned}$$ where $\Omega_j$ is the space close to the domain walls. Therefore, Eq. (52), in the limit, $k'\to 0$, is written as, $$=-k'\lim_{V\to\infty}\frac{1}{V}\sum_j\int_{\Omega_{j}} dx_idv^i_{d-1}\varphi_{x_ix_i}\delta[\varphi_i-\varphi(x_i,t)]\;.$$ Changing the variables from $x_i$ to $z_i$ and integrating over $dv^i_{d-1}$, one finds, $$\begin{aligned} = -k'\lim_{V\to\infty} \frac{V_{d-1}}{V}\sum_j\int_{-\infty}^{+\infty} \epsilon dz_i\frac{1}{\epsilon^2}\varphi_{z_iz_i} \delta[\varphi_i-\varphi(z_i,t)]\nonumber\\ =-\frac{k'}{\epsilon}\lim_{V\to\infty}\frac{V_{d-1}}{V}\sum_j \int_{-\infty}^{+\infty}dz_i \varphi_{z_iz_i}\delta[\varphi_i-\varphi(z_i,t)]\\\end{aligned}$$ where $\epsilon=(2k')^{1/2}$. Assuming statistical homogeneity, we have $$=-\frac{k'}{\epsilon}\lim_{V\to\infty}\frac{NV_{d-1}}{V} \int_{-\infty}^{+\infty}dz_i \varphi_{z_iz_i}\delta[\varphi_i-\varphi(z_i,t)]\\$$ where $N$ is number of singular lines. Moreover, $\frac{k'}{\epsilon}= (k')^{1/2}$, and, $\frac{NV_{d-1}}{V}$ is the density of the singular lines which, in the limit, $V\to\infty$, is simply the singularity density $\rho$. Therefore, $$= -(k')^{1/2}\rho\int_{-\infty}^{+\infty}dz_i \varphi_{z_iz_i} \delta[\varphi_i-\varphi(z_i,t)]\;.$$ In the same way in, for example Eq. (21), by changing the integration variable from $z_i$ to $\varphi_i$, we calculate the integral exactly, $$\begin{aligned} =\sqrt{k'/2}\varphi\{A+2\sqrt{2}\phi_i ^2\delta(\phi_i -1)\}\end{aligned}$$ where $A=2\sqrt{2}\theta(1-\phi_i^2)\theta(\phi_i^2-1)-\sqrt{2}\theta(1-\phi_i^2)$. Therefore, in the limit, $k'\to 0$ the master equation will give Eq. (35). [99]{} K. G. Wilson, Phys. Rev. B [4]{}, 3174, 3184 (1971); K. G. Wilson and M. E. Fisher, Phys. Rev. Lett. [28]{}, 240 (1972); K. G. Wilson, [ibid.]{} [28]{}, 548 (1972); A. A. Migdal, Sov. Phys. JETP [32]{}, 552 (1971). H. Kleinert and V. Schulte-Frohlinde, [Critical Properties of $\phi^4$-Theories]{} (World Scientific, Singapore, 2001). J. Berges and J. Cox, Phys. Lett. B [517]{}, 369 (2001). J. Berges, Nucl. Phys. A [699]{}, 847 (2002). K. Blagoev, F. Cooper, J. Dawson, and B. Mihaila, Phys. Rev. D [64]{}, 125003 (2001). G. Aarts, D. Ahrensmeier, R. Baier, J. Berges, and J. Serreau, Phys. Rev. D [66]{}, 045008 (2002). F. Cooper, J. F. Dawson, and B. Mihaila, Phys. Rev. D [67]{}, 051901 (2003). J. Berges and J. Serreau, hep-ph/0208070. F. Cooper, J. F. Dawson, and B. Mihaila, Phys. Rev. D [67]{}, 056003 (2003). J. Baacke and A. Heinen, hep-ph/0212312. J. Dreger, A. Pelster, and B. Hamperecht, Eur. Phys. J. B [45]{}, 355 (2005). A.-L. Barabasi and H. E. Stanley, [Fractal Concepts in Surface Growth]{} (Cambridge University Press, New York, 1995). NATO Advanced Study Institute on Formation and Interactions of Topological Defects, edited by A.-C. Davis and R. Brandenberger (Plenum, New York, 1995). L. P. Gorkov, Sov. Phys. JETP [9]{}, 1364 (1959). J. Bardeen, L. N. Cooper, J. R. Schrieffer, Phys. Rev. [108]{}, 1175 (1957). V. L. Ginzburg and L. D. Landau, Zh. Eksp. Teor. Fiz. [20]{} 1064 (1950); reprinted in L. D. Landau, [Collected Papers]{} (Pergamon, London, 1965), p. 546. A. Schmid, Z. Phys. [215]{}, 210 (1968). L. P. Gorkov and G. M. Eliashberg, Sov. Phys. JETP [27]{}, 338 (1968). A. C. Newell and J. A. Whitehead, J. Fluid Mech. [38]{}, 279 (1969). L. A. Segel, J. Fluid Mech. [38]{}, 203 (1969). P. C. Hohenberg and B. I. Halperin, Rev. Mod. Phys. [49]{}, 435 (1977). C. M. Bender and S. A. Orzag,[Advanced Mathematical Methods for Scientists and Engineers]{}, 3rd International Edition (McGraw-Hill, New York, 1987). A. A. Masoudi, F. Shahbazi, J. Davoudi and M. R. Rahimi Tabar, Phys. Rev. E [65]{} 026132 (2002). H. Risken, The Fokker-Planck equation, Springer-Verlag Berlin, 1984. U. Frisch and J. Bec, in [Les Houches 2000: New Trends in Turbulence]{}, edited by M. Lesieur, A. Yaglom, and F. David (Springer EDP-Sciences, Berlin, 2001), p. 341. R. Peyret, [Computational Fluid Mechanics]{} (Academic Press, San Diego, 2000). R. W. Hockney and J. W. Eastwood, [Computer Simulation Using Particles]{} (Institute of Physics Publishing, London, 1992). J. Bec, U. Frisch, and K. Khanin, J. Fluid Mech. [416]{}, 239 (2000).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Collapse and reverse to collapse explosion transition in self-gravitating systems are studied by molecular dynamics simulations. A microcanonical ensemble of point particles confined to a spherical box is considered; the particles interact via an attractive soft Coulomb potential. It is observed that the collapse in the particle system indeed takes place when the energy of the uniform state is put near or below the metastability-instability threshold (collapse energy), predicted by the mean-field theory. Similarly, the explosion in the particle system occurs when the energy of the core-halo state is increased above the explosion energy, where according to the mean field predictions the core-halo state becomes unstable. For a system consisting of 125 – 500 particles, the collapse takes about $10^5$ single particle crossing times to complete, while a typical explosion is by an order of magnitude faster. A finite lifetime of metastable states is observed. It is also found that the mean-field description of the uniform and the core-halo states is exact within the statistical uncertainty of the molecular dynamics data.' author: - 'I. Ispolatov' - 'M. Karttunen' title: 'Collapses and explosions in self-gravitating systems' --- ł Introduction {#sec_intro} ============ Systems of particles interacting via a potential with attractive nonintegrable large $r$ asymptotics, $U(r)\sim r^{-\a}$, $0<\a<3$, and sufficiently short-range small $r$ regularization exhibit gravitational phase transition between a relatively uniform high energy state and a low-energy state with a core-halo structure [@pr; @ki2; @ch1; @usg; @dv; @ch; @chs; @chi]. Extensive mean-field (MF) studies of the equilibrium properties of such systems [@pr; @ki2; @ch1; @usg; @dv; @ch; @chs; @chi] revealed that in a microcanonical ensemble during such a transition entropy has to undergo a discontinuous jump from a state that just ceases to be a local entropy maximum to a state with the same energy but different temperature, which is the global entropy maximum. Due to the long-range nature of gravitational interaction, the MF studies are believed to provide asymptotically (in the infinite system limit) exact information about the density and the velocity distributions and other thermodynamical parameters of the uniform state. The applicability of the MF theory to the description of the core-halo state is less obvious as the properties of a core are controlled by the short-range asymptotics of the potential. Relatively little is known about how such a transition actually occurs, however. Youngkins and Miller [@bm2] performed a Molecular Dynamics (MD) study of a one-dimensional system consisting of concentric spherical shells. Their main emphasis was to check the MF description of the stable and metastable states rather than to study the dynamics of the phase transition itself. Cerruti-Sola, Cipriani, and Pettini [@pet] studied the phase diagram of a more realistic 3-dimensional particle system by using Monte Carlo and MD methods. Their studies again focused on the equilibrium properties of the system rather than on the dynamics of the transitions. In addition, their general conclusions about the second order of the gravitational phase transition apparently contradict the MF results [@ki2; @usg; @dv; @chi]. Here, we attempt to resolve this contradiction. A MF description of the dynamics of collapse in ensembles of self-gravitating Brownian particles with a bare $1/r$ interaction based on a Smoluchowski equation was developed by Chavanis et al [@chs]. It predicts a self-similar evolution of the central part of density distribution to a finite-time singularity. However, the precise nature of the random force and friction terms in the corresponding Fokker-Plank equation as well as the applicability of the overdamped limit used to reduce the Fokker-Plank equation to the Smoluchowski equation are not entirely understood. A more rigorous approach based on the Fokker-Plank equation with Landau collision integral was used by Lancellotti and Kiessling [@ki3] to prove a scaling property of the central part of the density profile. The model considered there allows the particles to escape to infinity and therefore does not have an equilibrium or even a metastable state. There exists a vast amount of literature on cosmologically- and astrophysically- motivated studies of the temporal evolution of naturally occurring self-gravitating systems (see, e.g., Ref. [@cos] and references therein). The selection of systems and their initial and final conditions made in such studies are typically astrophysically-motivated; the considered systems are often too complex to make a general conclusions about the phase diagrams and phase transitions in such systems. In this paper we present MD studies of gravitational collapse, and reverse to collapse, i.e., explosion, transitions in a microcanonical ensemble of self-attracting particles. Besides their pure statistical mechanical implications, these studies represent our attempt to bridge a gap between the usually complicated MD and hydrodynamic simulations of the realistic astrophysical systems and the MF analysis of the phase diagram of simple self-gravitating models. A system with soft Coulomb potential $-(r^2+r_0^2)^{-1/2}$ is considered. Such systems are have been studied using both MF theory (see, e.g., Refs. [@usg; @chi]) and simulations [@pet]. We chose the microcanonical ensemble as the most fundamental one for the long-range interacting systems. It has to be noted that the considered system is strongly ensemble-dependent: While the nature of the uniform state is the same in both microcanonical and canonical ensembles (apart form the difference in their stability range), the core-halo states and the collapse itself in these ensembles have very little in common with each other [@dv; @chs]. A MF phase diagram of the considered self-attracting microcanonical system is presented in Fig. (\[fig\_mf\]) [@usg; @chi]. ![\[fig\_mf\] Plots of entropy $s(\e)$ (solid line) and inverse temperature $\b(\e)=ds/d\e$ (dashed line) vs. energy $\e$ for a system with a gravitational phase transition and a short-range cutoff.](fig1.eps){width=".45\textwidth"} High- and low-energy branches terminating at the energies $\e_{coll}$ and $\e_{expl}$ correspond to the uniform and core-halo states. The energy $\e^$ where the entropies of the core-halo and uniform states are equal is the energy of the true phase transition; the uniform and the core-halo states are metastable in the energy intervals $(\e_{coll},\e^)$ and $(\e^,\e_{expl})$, respectively. However, for the phase transition to occur at or near $\e^$, a macroscopic-scale fluctuation with prohibitively low entropy is required. Consequently, the metastable branches are stable everywhere except at the vicinity of $\Delta \e \sim N^{-2/3}$ of their end-points $\e_{coll}$ and $\e_{expl}$ [@ka2; @chi], where $N$ is the number of particles in the system. Hence it is natural to assume that once the energy of the system in the uniform state is set sufficiently near above or below $\e_{coll}$, the system will undergo a collapse to a core-halo state with the same energy and higher entropy. Similarly, if the energy of the core-halo system is set sufficiently near below or above $\e_{expl}$, the system will undergo an explosion bringing it to a uniform state with the same energy and higher entropy. Our goal here is to study if and how such collapse and explosion proceed in a realistic three-dimensional N-particle dynamical system. The paper is organized as follows. In the next section we formally introduce the system, outline the MF analysis and describe the MD setup. Then, we present the simulation results for the equilibrium uniform and core-halo states and compare them to the MF predictions. After that we describe and interpret the observed dynamics of the collapse and the explosion transitions. A discussion of the obtained results concludes the paper. simulation ========== We consider a system consisting of $N$ identical particles of unit mass confined to a spherical container of radius $R$ with reflecting walls. The particles interact via the attractive soft Coulomb pair potential $-(r^2+r_0^2)^{-1/2}$. Using a traditional convention for self-gravitating systems in which the equilibrium properties of such systems become universal, we define rescaled energy $\e$, inverse temperature $\b$, distance $x$, velocity $u$, and time $\t$ as $$\begin{aligned} \label{def} \nonumber \e\equiv E{R\over N^2}\\ \nonumber x\equiv{r\over R}\\ \b\equiv{N\over RT}\\ \nonumber u \equiv v \sqrt{R \over N}\\ \nonumber \t\equiv t {N^{1/2}\over R^{3/2}}.\end{aligned}$$ The unit of time, often referred to as crossing time, $[t]=\frac {R^{3/2}} {N^{1/2}}$, is obtained by dividing the unit of length $R$ by the unit of velocity $\sqrt{N/R}$. This unit of time is also proportional to the period of plasma oscillations in a medium with the charge concentration $N/R^3$. As this time unit has purely kinematic origin, we do not expect the evolution of systems having different $N$ and $R$ to be universal in time $\t$. The evolution, assuming that it is collisional, is expected to be universal in the relaxation time $\t_{r}=\t{\ln N\over N}$ [@bt], where the factor $N/ \ln N$ is proportional to the number of crossings a typical particle needs to change its velocity by a factor of 2 through weak Coulomb scattering events. The soft core radius $x_0=r_0/R=5\times 10^{-3}$ is chosen to be well below the critical value $x_{gr}\approx 0.021$, above which the collapse-explosion transition is replaced by a normal first-order phase transition [@chi]. The MF theory of the system is described in detail in Ref.[@usg]. The equilibrium velocity distribution is Maxwellian and isothermal, while the equilibrium (saddle point) density profile $\r({ x})$ corresponding a stable or a metastable state, is a spherically symmetric solution of the integral equation (\[extr\]). This equation replaces the Poisson-Boltzmann differential equation for the self-consistent potential (see, for example, Ref.[@pr]) since the interparticle interaction considered here is not pure Coulombic. ![\[fig\_rho\] MF density profiles $\r_u({x})$ of a uniform state (dashed line) and $\r_{c-h}({x})$ of a core-halo state (solid line) for $\e=\e_{coll}$ ](fig2.eps){width=".45\textwidth"} $$\begin{aligned} \label{extr} \nonumber \r({\bf x})=\r_0 F[\r(.), {\bf x}]\\ \nonumber F[\r(.),{\bf x}]=\exp\left [\b \int {\r({\bf x}')\over \sqrt{({\bf x}-{\bf x}')^2+x_0^2}}d^3{\bf x}'\right ]\\ %\nonumber \b={3\over 2} \left[\e + {1\over 2}\int \int{\r({\bf x}_1) \r({\bf x}_2) \over \sqrt{({\bf x}_1-{\bf x}_2)^2+x_0^2} }d^3{\bf x}_1 d^3{\bf x}_2 \right ]^{-1}\\ \nonumber \r_0=\left\{\int F[\r_s(.), {\bf x}]d^3{\bf x}\right\}^{-1}\end{aligned}$$ The equilibrium density profile $\r({x})$ obtained from this equation is then used to calculate the entropy and the pressure. The MF phase diagram of the system is presented in Fig. \[fig\_mf\]. The collapse and explosion energies are $\e_{coll}\approx -0.339$ and $\e_{expl}\approx 0.267$. Examples of the uniform and the core-halo density profiles for $\e=\e_{coll}$ are shown in Fig. \[fig\_rho\]. In the MD simulations we consider a system consisting of $N=125$ – 500 particles in a spherical container of the radius $R=1$. All interparticle forces are calculated directly at each time step $dt$. This is done to avoid any mean-field-like effects inevitably present in any truncated multipole or Fourier potential expansion. The particle velocities and coordinates are updated according to the velocity-Verlet algorithm which is symplectic and reversible. The system is initialized by randomly distributing particles according to a spherically symmetric density profile; typically the appropriate MF density profile $\r(x)$ was used. The potential energy ($U$) of the initial configuration is calculated, and the target kinetic energy $E_k=E-U$ is determined. The particle velocities are randomly generated from some (usually Maxwell) distribution with the appropriate square average. Finally, the deviation of the total energy from its target value, caused by a stochasticity in velocity assignment, is determined, and the velocities are rescaled to fine-tune the total energy. Due to the isotropicity of the random velocity assignment, we have always obtained the states with sufficiently low total angular momentum which collapsed to single-core states rather than to binaries [@grv]. To implement the reflective boundary condition, at each time step the normal components $v_{\perp}$ of the velocities of all particles which had escaped from the container were reversed. Values of the normal components were stored to evaluate the pressure on the wall $P$. $$\label{P} P(t)=\frac{\sum_{t'=t-t''/2}^{t'=t+t''/2}v_{\perp}(t')}{2\pi R^2 t''}$$ During each simulation run we measured such characteristics as the kinetic energy $\e_{k in}=\frac{3}{2\b}$, the virial variable $\s$ (dimension of energy) quantifying deviations from the virial theorem, $\s \equiv \e + \e_{kin} -\frac{3 PV R}{N^2}$ (where $P$ is the pressure on the wall, $V=4\pi R^3/3$ is volume of the container, and the factor $N^2/R$ rescales the volume-pressure term to the unit of energy introduced in (\[def\])), ratio of the velocity moments, $ \frac{19 \langle v^2 \rangle ^2}{ 5 \langle v^4 \rangle }$ (which should be 1 for the Maxwell distribution), and the number of particles in the core $N_c$ of the prescribed radius $x_c$. For the last measurement we count the number of particles $N_i$ that are within $x_c$ from the $i$th particle and find the particle which has the largest $N_i$. In addition, we measured the histograms of the velocity distribution and radial distribution functions, $W(u)$ and $C(x)$, respectively. The latter was defined as the number of particles in the spherical layer of the radius $x$ around each particle, normalized by the volume of such layer, disregarding the nonuniformity of the system and the boundary effect. The measurements of the ”scalar” quantities such as energy, kinetic energy, pressure, and velocity distribution moments were taken in time intervals $\t_{meas}$, which were selected sufficiently long to avoid measuring the unchanged configuration repeatedly and sufficiently short not to miss the important details of the system evolution. We usually pick $\t_{meas}$ of the order of the uniform density sphere crossing time $\t_{cross}^u=\pi$, which is a half period of the oscillation of a particle released with zero velocity at the container wall. The histogram data, such as the velocity distribution and the radial distribution functions, was incremented at each $\t_{meas}$ and accumulated over a longer time period $\t_{hist}$, $\t_{hist}\sim 10$ – $10^3 \times \t_{meas}$. Our attempts to resolve the high-density part of the radial density profile of the system turned out to be fruitless due to the strong fluctuations in the positions of this part. This fluctuations result in smearing the central peak in both core-halo and low energy uniform states. Considering the center of mass system of reference does not resolve this difficulty, as, despite being dense, the core typically contains only 10 – 20% of the total system mass (see below) and the positions of the core and the center of mass of the system do not usually coincide. To control the quality of the simulation, we monitored the total energy $\e$ and the total angular momentum $L$. We selected timestep $dt$ small enough to keep the total energy variation within 0.01% of its initial value, usually we used $dt =10^{-5}$, or in rescaled units, $d\t \sim 10^{-4}$. For such time steps, the relative deviation of the angular momentum was within $10^{-14}$. All the measurements below are presented in the rescaled dimensionless units as defined in Eq. (\[def\]). uniform and core-halo equilibrium states: comparison to the MF ============================================================== To check our simulation procedure and possibly resolve the apparent contradiction between the MF and the particle simulation results [@pet], we first considered the system in what we expected to be a stable or a metastable states far away from a transition point. Since we were interested in the equilibrium properties, we were initiating the MD systems according the corresponding MF predictions. It meant that the density profiles were seeded according to the MF profiles and the velocities were assigned according to the Maxwell distribution. We observed that the MF density initiation virtually eliminates the transitory period, while the method of velocity assignment was practically unimportant, provided that it gave the correct value for the total kinetic energy. For example, it takes a system initialized with a flat $W({\bf u})=const$ about $\t\sim \t_r$ to evolve to the Maxwell distribution. A typical plot of the steady state time dependence of the kinetic energy, virial variable, and the total energy is presented in Fig. \[fig\_td\]. ![\[fig\_td\] Plots of time dependence of kinetic energy $e_{kin}$ (solid line), virial variable $\s$ (dashed line), and total energy $\e$ (dotted line) for a uniform system of $N=250$ particles at $e=-0.3$. ](fig3.eps){width=".45\textwidth"} The comparison between the MD measurements and the MF results for the uniform and core-halo states is presented in Tables \[tab\_un\] and \[tab\_ch\] and reveals a perfect agreement between these two sets of data. -------------------------------------------------------------------- \[tab\_un\] MD MF ------------------------------- ---------------------------- ------- $\e$ $-0.3 \pm 5\times 10^{-7}$ -0.3 $\e_{kin}$ $0.66 \pm 0.05$ 0.644 $\s$ $0\pm 0.03$ 0.012 $ 19 \langle v^2 \rangle ^2 / $1.01\pm0.04$ 1 5 \langle v^4 \rangle $ -------------------------------------------------------------------- : Equilibrium MD and MF results for a uniform state for $\e=-0.3$, $N=250$, and $0\leq \t \leq 5000 $ ------------------------------------------------------------------------ \[tab\_ch\] MD MF ------------------------------- ------------------------------- -------- $\e$ $-0.3392 \pm 2\times 10^{-4}$ -0.339 $\e_{kin}$ $2.9 \pm 0.1$ 2.94 $\s$ $-1.5 \pm 0.1$ -1.46 $ 19 \langle v^2 \rangle ^2 / $0.99 \pm 0.03$ 1 5 \langle v^4 \rangle $ $N_{core}$ $48\pm 2$ 47.6 ------------------------------------------------------------------------ : Equilibrium MD and MF results for a core-halo state at $\e=-0.339 $, $N=250$, and $0\leq \t \leq 1500 $ To obtain the expression for the MF virial variable $$\label{vir} \s_{MF}=\e+\e_{kin}(1-8\pi\r(1)/3)$$ we write for the pressure at the container wall $P=2\rho(x=1)\e_{kin}/3$ implying that the system is isothermal. Since the interparticle potential is not pure Coulombic, the virial variable is non-zero. The difference is especially prominent for the core-halo states where more particles ”probe” the short-range part of the potential. To evaluate the core radius and number of core particles of the core-halo system, we considered an integrated MF density profile, $f(x)=\int_0^x 4 \pi y^2 \r(y) dy$ (see Fig. \[fig\_irho\]). ![\[fig\_irho\] Integrated MF density profiles $f_u({x})$ of a uniform state (dashed line) and $f_{c-h}({x})$ of a core-halo state (solid line) for $\e=\e_{coll}$ ](fig4.eps){width=".45\textwidth"} As it follows from the Figure, the MF core-halo state indeed contains a distinct core with a sharp boundary of the radius $x_c\approx10^{-2}$ relatively insensitive to the energy in the range we considered, $\|\e\|<0.5$. Using this MF core radius, we located cores in the MD core-halo systems which contained very similar to the MF cores number of particles (see Table \[tab\_ch\]). Using smaller core radius resulted in significant reduction in the number of observed core particles. A reasonably small over-estimation of the core radius did not affect the results of the MD measurements: we observed that even in the sphere twice the core radius the number of particles is only marginally (at most by 8%) larger than in the core. To check if the system has more than one core, we performed search for the second-largest core of the same radius $x_c$. We looked for a largest group of particles which are within $x_c$ from a single particle with none of these particles belonging to the first, largest core. We never observed the second-largest core containing more than 2 particles; most of the time it contained only a single one. In Fig. \[fig\_vel\] we present the MD velocity distribution functions $W(u)$ for a core-halo and uniform states; shown $W(u)$ confirm the MF prediction for the Maxwellian form of these distributions. ![\[fig\_vel\] The MD velocity distribution functions $W(u)$ of a core-halo state with $\e=-0.339$ (solid line) and a uniform state with $\e=-0.3$ (dashed line). In both cases $N=250$.](fig5.eps){width=".45\textwidth"} As we mentioned in the previous section, we were unable to resolve the high-density part of the radial density profile due to the core motion. However, an indirect comparison between the radial distribution of particles in the MF and MD was made using the radial distribution function. The MF radial distribution function $C_{MF}(x)$ was computed as $$\label{rdf} C_{MF}(x)= \frac{1}{4\pi x^2}\int\r({\bf x}')\r({\bf x + x}')d {\bf x}'.$$ The good agreement between the MF and the MD radial distribution functions is illustrated in Fig. \[fig\_rdf\]. This indicates that the mutual distribution of particles is correctly predicted by the MF theory. ![\[fig\_rdf\] MF (dashed line) and MD (solid line) radial distribution functions $C(x)$ of a core-halo state with $\e=0.25$. The step at $x=1$ in the MF $C(x)$ is caused by the localization of the core exactly at $x=0$ and a sharp boundary of the container.](fig6.eps){width=".45\textwidth"} To summarize, for all the quantities considered, we observed no systematic deviations between the MF theory and the MD data. Collapse ======== According to the MF theory, if the energy of the uniform state becomes lower than $\e_{coll}\approx-0.339$, the system should undergo a collapse to a core-halo state. To study the collapse, we considered several uniform systems with the energies ranging between $\e=-0.5$ and $\e=-0.3$. The systems were initialized according to the MF density distributions. For systems with $\e<\e_{coll}$ the particles were distributed according to the MF density profile for $\e_{coll}$. In a perfect agreement with the MF theory, a uniform state with $\e<\e_{coll}$ undergoes a gradual transition to a core-halo state with a typical timescale of $\t_{coll}\sim 10^4$ for $N=125$ – 250 particles. An example of the time dependence of the kinetic energy and the virial variable for a collapsing system is shown in Fig. \[fig\_coll\]. ![\[fig\_coll\] Time dependence of the kinetic energy $\e_{kin}$ (top) and the virial variable $\s$ (bottom) of the collapsing uniform state with $\e=-0.5$ and $N=125$. The dashed horizontal lines indicate the equilibrium values of $\e_{kin}$ and $\s$ of the corresponding core-halo state. The data is averaged over $\d \t=100$ time intervals.](fig7.eps){width=".45\textwidth"} We observe that if the number of particles is increased but the rescaled energy $\e$ is kept fixed, it takes generally longer time for the collapse to be complete. Our results (see Fig. \[fig\_N\]) qualitatively confirm that the characteristic time for the full collapse scales as $\t_r$ [@bt]. A quantitative study of the dependence of the collapse dynamics on the number of particles requires much faster simulation code, however. ![\[fig\_N\] Collapse in systems with $\e=-0.5$ and different numbers of particles, $N=125$ (dashed line) and $N=250$ (solid line), shown in relaxation time units, $\t\ln N/N$ [@bt]. Horizontal line shows the kinetic energy of the target core-halo state. The data is averaged over $\d \t=100$ time intervals.](fig8.eps){width=".45\textwidth"} In the above examples, the energy was set to $\e=-0.5$ which is well below $\e_{coll}\approx-0.339$, and as a consequence the collapse started immediately at $\t=0$ in all simulation runs. If the system energy is $\e_{coll}$, the noticeable increase in kinetic energy and decrease of the virial variable, characteristic for collapse, start not exactly at $\t=0$ but with a small delay (Fig. \[fig\_ec1\]) which varies from run to run from almost zero to about $\t\approx 1500$. This indicates that the MD system is able to overcome the metastability at or near $\e_{coll}$. The observed uncertainty is likely due to the relatively small number of particles. ![\[fig\_ec1\] Plots of the kinetic energy $\e_{kin}$ (top) and the virial variable $\s$ (bottom) vs time $\t$ for a system with $\e=\e_{coll}\approx-0.339$ and $N=250$.The data is averaged over $\d \t=100$ time intervals.](fig9.eps){width=".45\textwidth"} As we increase the energy above $\e_{coll}$, the stability of the uniform state increases which results in a longer lifetime of such state with respect to collapse. In Fig. \[fig\_met\], an evolution of a system with $\e=-0.3$ is shown. The system stays in the uniform state for about $\d \t \approx 5000$ before the collapse starts, after which the evolution proceeds qualitatively similar to the collapses in systems with lower energies. ![\[fig\_met\] Plots of the kinetic energy $\e_{kin}$ (top) and virial variable $\s$ (bottom) vs time $\t$ for a system with $\e=-0.3$ and $N=250$. The data is averaged over $\d \t=100$ time intervals.](fig10.eps){width=".45\textwidth"} To compare the temporal evolution of the kinetic energy, virial variable, and the number of core particles, the relative variables $\e'_{kin}(\t)$, $\s'_{kin}(\t)$, and $N'_{core}(\t)$, all defined as $\e'_{kin}(t)=[\e_{kin}(t)-\e_{kin}(u)]/[\e_{kin}(c-h)-\e_{kin}(u)]$, are plotted in Fig.\[fig\_ec\]. The values $\e_{kin}(u)$ and $\e_{kin}(c-h)$ correspond to the uniform and core-halo states in equilibrium. ![\[fig\_ec\] Plots of the relative values of (from top to the bottom) number of core particles $N'_{core}(\t)$, virial variable $\s'_{kin}(\t)$, and kinetic energy $\e'_{kin}(\t)$ for the system with $\e=-0.339$ and $N=250$.](fig11.eps){width=".45\textwidth"} Figure \[fig\_ec\] indicates that during the initial stages of collapse the core grows faster than the kinetic energy and the virial variable. In addition, one can notice large reversible fluctuations in the number of core particles (the core grows up to 12% of its final value and then disappears) that are not matched by comparable scale fluctuations in the kinetic energy or virial variable. All these observations suggest that the density evolution causing the core formation plays the leading role in the process of collapse while the relaxation of kinetic energy follows. Once the collapse has started, the core grows to about a half of its final size in only $\d\t_{core}\sim 10^3$ for systems with $N=125$ – 500 particles, while the changes in kinetic energy during this interval of time are small. After this rapid initial stage the system relaxes more slowly, and finally after $\t_{coll}\sim 10^5$ reaches the equilibrium core-halo state. Our observations strongly suggest that the growth of the core takes place through a sequential absorption of single particles rather than through hierarchical merging of smaller cores: We never detected other cores containing more than two particles. Although the kinetic energy relaxation trails behind the the core formation, the velocity distribution function remains Maxwellian throughout the whole evolution with the temperature corresponding to the corresponding value of the kinetic energy. This is caused by the fast velocity relaxation ($\t_{vel} \leq 1$) as discussed in the previous section. Explosion ========= It is natural to assume that if a system exhibits a collapse, it should also exhibit an explosion which is the reverse to the collapse transition. According to the MF theory, such explosion should take place when the core-halo state becomes unstable, i.e., when $\e \geq \e_{expl}\approx 0.267$. To check this prediction, we initialized the MD system according to the MF equilibrium core-halo state and followed its evolution. As in the study of the collapse, for initial states with $\e > \e_{expl}$ we used the MF density profiles of the highest energy locally stable state, i.e. of the state with $\e = \e_{expl}$. We observe that a system with sufficiently high energy, such as $\e =0.5$ in Fig. \[fig\_expl\] or $\e =0.4$ in Fig. \[fig\_expla\], indeed undergoes an explosion which brings it to the uniform equilibrium state. During such an explosion, the state variables such as kinetic energy and virial variable continuously change from their equilibrium core-halo state values to the uniform state ones, and the core gradually sheds particles until only one particle is left. ![\[fig\_expl\] Plots of the kinetic energy $\e'_{kin}(\t)$ (top) and relative number of core particles $N'_{core}(\t)$ (bottom) (defined as in Fig. \[fig\_ec\]) vs time $\t$ for a system with $\e=0.5$ and $N=250$. The kinetic energy is averaged over $\d \t=100$ time intervals. ](fig12.eps){width=".45\textwidth"} ![\[fig\_expla\] Same as in Fig. \[fig\_expla\] but for $\e=0.4$. ](fig13.eps){width=".45\textwidth"} The main features of an explosion (Figs. \[fig\_expl\] and \[fig\_expla\]) resemble those of a time-reversed collapse. The kinetic energy evolves relatively uniformly, while the number of core particles changes only slightly during the first stages of evolution and rapidly decreases at the final stages. In the example presented in Fig. \[fig\_expl\], the explosion is complete after the time $t_{expl}\approx 15000$, which is noticeably less than the time for a collapse $t_{coll}\approx 10^5$ (see Fig. \[fig\_N\]) for a system having the same number of particles ($N=250$). However, the latter is rather vaguely defined due to larger fluctuations in a core-halo than in a uniform state. Similarly to a collapse, the system remains thermalized in the velocity space during an explosion. The velocity distribution remains Maxwellian throughout the evolution with the temperature corresponding to the current value of kinetic energy. As an illustration, Fig. \[fig\_velm\] shows the ratio of the moments of velocity distribution, $19 \langle v^2 \rangle ^2 / 5 \langle v^4 \rangle$, which should be 1 for a Gaussian distribution. ![\[fig\_velm\] Plot of he ratio of the moments of velocity distribution, $19 \langle v^2 \rangle ^2 / 5 \langle v^4 \rangle$, vs.time $\t$ for a system with $\e=0.5$ and $N=250$.](fig14.eps){width=".45\textwidth"} However, as it is evident from a comparison between the Figs. \[fig\_expl\] and \[fig\_expla\], as $\e$ gets closer to $\e_{expl}$ the explosion takes longer to initiate. We have observed, that even for $\e=0.3$, which is noticeably larger than $\e_{expl} \approx 0.267$, the explosion does not happen during the first $\t=30000$ of evolution. This suggests that either the MF value for $\e_{expl}$ is incorrect, or during the incitation of the system we somehow prepare the system not exactly in the equilibrium (metastable) core-halo state. If the latter is the case, a deviation from the equilibrium most probably takes place in the core, as because of its compactness, its equilibration with the rest of the system may take a rather long time. Using the current MD setup, we were unable to determine a reason for this apparent discrepancy. CONCLUSION ========== In the previous sections we have presented the following molecular dynamics results for the self-attracting systems with soft Coulomb potential: [.]{}[ ]{} A collapse from a uniform to a core-halo state was observed. The timescale for the collapse in systems consisting of 125 – 500 particles is of order of $10^5$ crossing times and is by the same factor longer than the timescale of the velocity relaxation. The collapse starts with a fast growth of a core via absorption of single particles and continues with more gradual relaxation towards an equilibrium core-halo state. Metastable states exhibit a finite lifetime before collapsing. A reverse to collapse, i.e., an explosion transition from a core-halo to a uniform state was observed. The explosion time is considerably shorter than the collapse time, being of the order of $10^4$ crossing times (125 – 500 particles). An explosion resembles a time-reversed collapse; the core decrease, which happens by shedding individual particles, is trailing the kinetic energy evolution till the last stages, when the core rapidly disappears. Such molecular dynamics characteristics of the equilibrium or the metastable uniform and core-halo states as kinetic energy, wall pressure, number of core particles, particle-particle radial distribution function and velocity distribution function, are found to be equal within the statistical uncertainty of the molecular dynamics measurements to the corresponding mean field predictions. The long collapse time observed in our simulations appears to be an explanation for the apparent discrepancy between the phase diagram presented in [@pet] and the mean field phase diagram (see, for example, [@chi]). The relaxation time allowed in [@pet] before the measurements of what was considered to be a steady state, $t_{rel}=3N/\|EN\|^{3/2}$, which is apparently equivalent to $\t_{rel}<1$, is by far insufficient for a system to collapse. Therefore, the discontinuities in caloric curves $\beta$ vs $\e$, typical for collapse and explosion gravitational transitions, were not observed in [@pet]. Although we considered systems only with the soft Coulomb potential, we speculate that a likewise similarity between the mean field and molecular dynamics equilibrium properties of the core-halo state exists for all ”soft” long-range (like a Fourier-truncated Coulomb) potentials. This is so because all soft potentials are effectively longer-ranged than the bare Coulomb one. However, the core-halo state in the system with a ”harder” short-range cutoff may have completely different properties from the one considered above, and its mean field theory may be inadequate. As for the uniform states, their properties are virtually independent on the nature of the cutoff (see, for example, [@chi]) and their mean-field description is universally correct. The main goal in the paper was to check the existence of collapses and explosions and the validity of the mean field data for the self-gravitating systems with short-range cutoff. For this goal one or few molecular dynamics runs for each considered system were sufficient. However, to be able to study the dynamical features of collapses and explosion in more detail and to compare the simulations results to various theoretical models, one needs to study the relaxation averaged over many initial configuration. For example, an interesting question is whether a collapse (or an explosion) indeed consists of two stages; the first fast stage of collisionless ”violent relaxation” with particle number-independent rate, and the slower second stage of soft collisional relaxation with characteristic time $\t_r$ (see, for example, [@bt] and references therein). Another important question is to resolve the apparent contradiction between the mean field prediction for $\e_{expl}$ and the molecular dynamics observations, outlined at the end of the previous section. Such studies require a more efficient computation code. The main improvement possibly coming from a better force calculator that may include various mean-field-like potential expansions, which are qualitatively justified by this study. We leave this for the future research. acknowledgments =============== The authors are thankful to P.-H. Chavanis and E. G. D. Cohen for helpful and inspiring discussions and gratefully acknowledge the support of Chilean FONDECYT under grants 1020052 and 7020052. M.K. would like to thank the Department of Physics at Universidad de Santiago for warm hospitality. T. Padmanabhan, Phys. Rep. [188]{}, 285 (1990). B. Stahl, M. K.-H. Kiessling, and K. Schindler, Planet. Space Sci. [43]{}, 271 (1995). P.-H. Chavanis and J. Sommeria, Mon. Not. R. Astr. Soc. [296]{}, 569 (1998). V. P. Youngkins and B. N. Miller, Phys. Rev. E. [62]{}, 4583 (2000). I. Ispolatov and E. G. D. Cohen, Phys. Rev. E. [64]{}, 056103 (2001), Phys. Rev. Lett. [87]{}, 210601 (2001). H. J. de Vega and N. Sánchez, Nucl. Phys. B [625]{}, 409 (2002) P.H. Chavanis, Phys. Rev. E. [65]{}, 056123 (2002) P.H. Chavanis, C. Rosier, and C. Sire, Phys. Rev. E. [66]{}, 036105 (2002). S. Chandrasekhar, [An Introduction to the Study of Stellar Structure]{} (Dover Publications, 1958), Ch. 11. P.H. Chavanis, I. Ispolatov, Phys. Rev. E. [64]{}, 056103 (2001) J. Binney and S. Tremaine [Galactic Dynamics]{} (Princeton Series in Astrophysics, 1987). J. Katz and I. Okamoto, Mon. Not. R. Astr. Soc. [317]{}, 163 (2000). M. Cerruti-Sola, P. Cipriani, and M. Pettini, Mon. Not. R. Astr. Soc. [328]{}, 339 (2001). P. Hut, M. M. Shara, S. J. Aarseth, R. S. Klessen, J. C. Lombardi Jr., J. Makino, S. McMillan, O. R. Pols, P. J. Teuben, R. F. Webbink, to appear in New Astronomy, astro-ph/0207318. C. Lancellotti and M. Kiessling, Astrophys. J. [549]{}, L93 (2001). E. V. Votyakov, H. I. Hidmi, A. De Martino, D. H. E. Gross, Phys. Rev. Lett., [89]{}, 031101 (2002).	{ "pile_set_name": "ArXiv" }
--- abstract: \| For a closed topological manifold $M$ with $\dim (M) \geq 5$ the topological structure set ${\mathcal{S}}(M)$ admits an abelian group structure which may be identified with the algebraic structure group of $M$ as defined by Ranicki. If $\dim (M) = 2d-1$, $M$ is oriented and $M$ is equipped with a map to the classifying space of a finite group $G$, then the reduced $\rho$-invariant defines a function, $${\widetilde{\rho}} : {\mathcal{S}}(M) \to {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d},$$ to a certain sub-quotient of the complex representation ring of $G$. We show that the function ${\widetilde{\rho}}$ is a homomorphism when $2d-1 \geq 5$. Along the way we give a detailed proof that a geometrically defined map due to Cappell and Weinberger realises the 8-fold Siebenmann periodicity map in topological surgery. address: - \| Hausdorff Research Institute for Mathematics\ Poppelsdorfer Allee 82\ D-53115 Bonn\ Germany - \| Mathematisches Institut\ Universität Münster\ Einsteinstra[ß]{}e 62\ Münster, D-48149\ Germany\ and Matematický Ústav SAV\ Štefánikova 49\ Bratislava, SK-81473\ Slovakia author: - 'Diarmuid Crowley, Tibor Macko' bibliography: - 'rho-add.bib' title: \| The additivity of the $\rho$-invariant and periodicity\ in topological surgery --- Introduction {#sec:intro} ============ Let $M$ be a closed oriented $(2d-1)$-dimensional topological manifold and let $\lambda (M) {\colon\!}M {\rightarrow}BG$ be a map to the classifying space of a finite group $G$. The $\rho$-invariant of $(M, \lambda (M))$, $$\rho (M,\lambda (M)) \in {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d},$$ lies in a certain sub-quotient of the rationalised complex representation ring of $G$ (see Section \[subsec:rho-inv\] for details). It is a powerful invariant of odd-dimensional manifolds with torsion elements in their fundamental group. To mention just two examples: it was used by Atiyah and Bott to show that two smooth lens spaces which are $h$-cobordant are diffeomorphic [@Atiyah-Bott(1967)]. It also plays a key role in Wall’s classification results for fake lens spaces in the piecewise linear and topological categories [@Wall(1999) Chapter 14]. Assume now that $2d-1\geq 5$ and consider ${\mathcal{S}}(M)$, the topological structure set of $M$. The elements of ${\mathcal{S}}(M)$ are homotopy equivalences $h {\colon\!}N {\rightarrow}M$ of closed manifolds modulo the $h$-cobordism relation in the source.[^1] We define the reduced $\rho$-invariant by $$\label{formula:reduced-rho} {\widetilde{\rho}}{\colon\!}{\mathcal{S}}(M) {\longrightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d}, ~~~~~[h: N {\rightarrow}M] ~ \longmapsto ~ \rho (N, \lambda \circ h) - \rho(M,\lambda).$$ A feature of topological surgery is that ${\mathcal{S}}(M)$ admits the structure of an abelian group which is natural in some sense [@Siebenmann(1977); @Ranicki(1978)]. Since this group structure on ${\mathcal{S}}(M)$ is mysterious from the geometric point of view it is not clear whether ${\widetilde{\rho}}$ is a homomorphism of abelian groups. It is clear, however, that ${\widetilde{\rho}}$ is additive with respect to the action of the $L$-group on ${\mathcal{S}}(M)$. Let $\pi = \pi_1(M)$ and recall that the surgery group $L_{2d}(\pi)$ acts on ${\mathcal{S}}(M)$ via Wall-realisation and also that the induced homomorphism $\lambda (M)_* : \pi \to G$ together with the $G$-signature define a homomorphism $\sigma_{\lambda (M)} {\colon\!}L_{2d}(\pi) \to {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d} $. It is well known that this action is additive with respect to ${\widetilde{\rho}}$ ([@Petrie(1970)]): if $x \in L_{2d}(\pi)$ and $[h] \in {\mathcal{S}}(M)$ then $$\label{eqn:additivity-w-r-t-action-of-L} {\widetilde{\rho}}([h] + x) = {\widetilde{\rho}}([h]) + \sigma_\lambda(x).$$ Moreover, calculations in [@Wall(1999) Chapter 14E] and [@Macko-Wegner(2008)] show that ${\widetilde{\rho}}$ is a homomorphism when $M$ is a lens space. Wolfgang L[ü]{}ck asked whether this is true in general and a positive answer to this question is our main theorem. \[thm:main\] Let $M$ be a closed oriented topological manifold of dimension $2d-1 \geq 5$ with a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ where $G$ is a finite group. Then the map $${\widetilde{\rho}}{\colon\!}{\mathcal{S}}(M) {\longrightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d}$$ is a homomorphism of abelian groups. We see that Theorem \[thm:main\] is a generalisation of the long standing identity (\[eqn:additivity-w-r-t-action-of-L\]). One may also take the point of view that it sheds light on the group structure on ${\mathcal{S}}(M)$. Clearly it has the potential to aid in computations of ${\mathcal{S}}(M)$ and this is shown to be the case in a forthcoming paper of Davis and L[ü]{}ck [@Davis-Lueck(2010)] about torus bundles over lens spaces. Clearly we also have The map ${\widetilde{\rho}}$ factors through ${\mathcal{S}}(M) {\longrightarrow} {\mathcal{S}}(M) \otimes {\mathbb{Q}}$. The outline of the proof of Theorem \[thm:main\] {#sec:outline} ------------------------------------------------ To describe the essential ideas of the proof, we first sketch the topological definition of the $\rho$-invariant which we use throughout the paper. Let $(M, \lambda (M))$ be as above. If $Z$ is a compact oriented $2d$-dimensional manifold with a map $\lambda(Z) {\colon\!}Z \to BG$, we call it an $r$-coboundary for $(M, \lambda (M))$ if ${\partial}(Z, \lambda(Z)) = \sqcup_r (M, \lambda (M))$ is the disjoint union of $r$ copies of $(M, \lambda (M))$ for some $r \geq 1$. From bordism theory we know that $r$-coboundaries always exist for some $r$. The $G$-signature of the induced $G$-covering $\widetilde Z$ is an element in the complex representation ring $R(G)$. It follows from the Atiyah-Singer $G$-index theorem [@Atiyah-Singer-III(1968)], [@Wall(1999) Chapter 14B] that the expression $$\rho(M,\lambda(M)) : = (1/r) \cdot {\textup{G-sign}}(\widetilde Z)$$ becomes independent of the choice of $Z$ and $r \geq 1$ after passing to a certain subquotient of the rationalisation of $R(G)$ (see Definition \[defn:rho-1\] for a precise statement). Suppose now that we have structures $h_0 {\colon\!}N_0 \to M$ and $h_1 {\colon\!}N_1 \to M$ representing two elements in ${\mathcal{S}}(M)$. Unless $[h_1] = [{\rm id}] + x$ for some $x \in L_{2d}(\pi)$, a geometric description of the structure $[h_0] + [h_1]$ in terms of $[h_0]$ and $[h_1]$ is not known at present. Thus it is not a-priori clear how to relate $r$-coboundaries for $N_0$ and $N_1$ and one sees that the additivity of the function ${\widetilde{\rho}}$ from Theorem \[formula:reduced-rho\] is not obvious. On the other hand, the situation becomes much simpler if we replace the closed manifold $M$ by $M \times D^l$ for some $l \geq 1$ as we now describe. The rel boundary structure set of $M \times D^l$, ${\mathcal{S}}_{\partial}(M \times D^{l})$, consists of equivalence classes of homotopy equivalences of manifolds with boundary $h {\colon\!}(N,{\partial}) \to (M \times D^l,{\partial})$, such that the restriction to the boundary is a homeomorphism ${\partial}h {\colon\!}{\partial}N \cong M \times S^{l-1}$. The equivalence relation is given by $h$-cobordism of pairs in the source. For $l \geq 1$ there is a geometrically defined group structure using “stacking” which is easy to understand: see Definition \[defn:stacking\]. Suppose that $n+l = \dim (M) + l = 2d-1$. Then following [@Madsen-Rothenberg-II(1989)], we define the rel boundary reduced $\rho$-invariant $$\label{formula:relative--reduced-rho} {\widetilde{\rho}}_{\partial}: \mathcal{S}_{\partial}^{}(M \times D^l) {\longrightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d}, ~~~~~[h {\colon\!}N {\rightarrow}M \times D^l] ~ \longmapsto ~ \rho(N \cup_{{\partial}h} (M \times D^l)).$$ The reference maps are left out of the notation. Notice that $N \cup_{{\partial}h} (M \times D^l)$ is a closed oriented $(2d-1)$-dimensional manifold and so the formula makes sense. Using a certain generalised connected sum operation we prove \[prop-C\] Let $M$ be a closed oriented topological manifold of dimension $n$ with a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ for a finite group $G$, and let $n+l = 2d-1 \geq 5$. Then the map $${\widetilde{\rho}}_{\partial}{\colon\!}{\mathcal{S}}_\partial (M \times D^l) {\longrightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d}$$ is a homomorphism of abelian groups. For reasons that will become apparent later we choose $l = 4j$, and contemplate the following diagram. $$\xymatrix{ {\mathcal{S}}(M) \ar[dr]_{{\widetilde{\rho}}} & ??? & {\mathcal{S}}_\partial (M \times D^{4j}) \ar[dl]^{{\widetilde{\rho}}_\partial} \\ & {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d} }$$ If we can find a homomorphism ${\mathcal{S}}(M) \to {\mathcal{S}}_{\partial}(M \times D^{4j})$ making the above diagram commute then ${\widetilde{\rho}}$ is a homomorphism and we are done. This brings us to periodicity in topological surgery which we discuss in more detail in Section \[subsec:periodhist\] below. For now we simply note that there is an injective near periodicity map $P^j : {\mathcal{S}}(M) \to \mathcal{S}_{\partial}(M \times D^{4j})$ defined in [@Siebenmann(1977)] and in a different way in [@Ranicki(1978)] and [@Ranicki(1992)]. However both definitions are complicated and require one to travel a long journey away from the geometry of a structure $[h : N \to M] \in {\mathcal{S}}(M)$. The distance is large enough that we lose sight of $r$-coboundaries and so of the $\rho$-invariant. A geometric passage from ${\mathcal{S}}(M)$ to ${\mathcal{S}}_{\partial}(M \times D^{4j})$ remained unclear until [@Cappell-Weinberger(1985)] where Cappell and Weinberger sketched maps $CW^j {\colon\!}{\mathcal{S}}(M) \to {\mathcal{S}}_{\partial}(M \times D^{4j})$ for $j = 1, 2$ or $4$. However, their construction was given using piecewise linear techniques and so strictly applies only when all manifolds involved are triangulable, although the authors hinted at the generalisations needed for the topological case. They claimed that $CW^j = P^j$ but their proof uses Sullivan’s Characteristic Variety Theorem which was never published in sufficient generality. Later, Hutt tried to address these issues [@Hutt(1998)]. He gave a construction of the map $CW^1$ for topological manifolds. However Hutt’s proof of near $4$-periodicity uses his own theory of Poincar[é]{} sheaves which was never published. Much of the work in this paper goes into giving a detailed proof that the Hutt construction adapted to the map $CW^2$ indeed realises the near periodicity map $P^2$. In particular we replace Hutt’s use of Poincaré sheaves with algebraic surgery from [@Ranicki(1992)] and thereby prove \[thm:A\] Let $M$ be a closed topological manifold of dimension $n \geq 5$. The Hutt description of the Cappell-Weinberger map gives an exact sequence of homomorphisms of abelian groups: $$0 {\longrightarrow}{\mathcal{S}}(M) \stackrel{CW^2}{{\longrightarrow}} {\mathcal{S}}_\partial (M \times D^{8}) {\longrightarrow}H_0(M; {\mathbb{Z}}).$$ The details of the Hutt construction of the map $CW^j$ allow us to do the following: given $(Z, \lambda(Z))$, an $r$-coboundary for $N$, the domain of a structure $[h: N \to M] \in {\mathcal{S}}(M)$, we can construct an $r$-coboundary for the domain of $CW^j([h])$. This then allows us to prove \[thm:B\] Let $M$ be a closed topological manifold of dimension $(2d-1) \geq 5$ with a reference map $\lambda {\colon\!}M {\rightarrow}BG$ for a finite group $G$. Then the following diagram commutes. $$\xymatrix{ {\mathcal{S}}(M) \ar[rr]^{CW^2} \ar[dr]_{{\widetilde{\rho}}} & & {\mathcal{S}}_\partial (M \times D^8) \ar[dl]^{{\widetilde{\rho}}_\partial} \\ & {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d} }$$ Theorem \[thm:main\] now follows immediately from Proposition \[prop-C\] and Theorems \[thm:A\] and \[thm:B\] since together they show that ${\widetilde{\rho}}= {\widetilde{\rho}}_{\partial}\circ CW^2$ is a composition of homomorphisms. The idea of understanding the group structure on ${\mathcal{S}}(M)$ via the stacking group structure on ${\mathcal{S}}_{\partial}(M \times D^{4j})$ and periodicity is very natural. For example in [@Jahren-Kwasik(2008)] Jahren and Kwasik used this method to solve an extension problem for ${\mathcal{S}}(S^1 \times {\mathbb{R}}P^n)$ related to the Browder-Livesay invariant, a close cousin of the $\rho$-invariant. Periodicity in topological surgery {#subsec:periodhist} ---------------------------------- In this subsection we briefly recall the history of periodicity in topological surgery as well as describing how this paper adds to the detailed proof of near periodicity. Let $M$ be a closed topological manifold of dimension $n \geq 5$. The source of periodicity in topological surgery is the 4-fold periodicity of the homotopy groups $\pi_i(G/{\textup{TOP}}) \cong \pi_{i+4}(G/{\textup{TOP}})$ for $i \geq 1$. However, it took Quinn’s theory of surgery spaces [@Quinn(1970)] to see how this periodicity could be extended to the structure set. Once the surgery exact sequence was identified as the long exact homotopy sequence of a fibration, Siebenmann [@Siebenmann(1977)] could define injective maps $P^j : {\mathcal{S}}(M) \to {\mathcal{S}}_{\partial}(M \times D^{4j}) $[^2]. He used these maps to define an abelian group structure on ${\mathcal{S}}(M)$. In [@Ranicki(1978)] and [@Ranicki(1992)] Ranicki produced algebraic versions of surgery theory which translate Quinn’s theory into a category of chain complexes. In particular bijections $$s : {\mathcal{S}}_{\partial}(M \times D^l) \to {\mathbb{S}}_{n+l+1}(M)$$ are defined where ${\mathbb{S}}_{n+l+1}(M)$ is an abelian group. Moreover, with respect to the stacking group structure on ${\mathcal{S}}_{\partial}(M \times D^l)$ this map is an isomorphism if $l \geq 1$. Since the algebraic groups are nearly $4$-periodic almost by definition, Ranicki was able to give an algebraic proof of Siebenmann’s periodicity theorem. In particular the algebraic theory of surgery so closely mirrors surgery spaces that the two group structures defined on ${\mathcal{S}}(M)$ agree. As we have seen, for certain purposes the abstract descriptions of the maps $P^j$ do not suffice and the papers of [@Cappell-Weinberger(1985)] and [@Hutt(1998)] were written to fill this gap. For reasons mentioned above, however, neither of these papers gives a water tight proof that the maps $CW^j : {\mathcal{S}}(M) \to {\mathcal{S}}_{\partial}(M \times D^{4j})$ realise the periodicity maps $P^j$. In the end, to give a detailed proof that $CW^2 = P^2$ we have had to combine important ideas from both papers and add some of our own. For the outline of the proof of periodicity we were able to follow [@Hutt(1998)]. However to Hutt’s arguments one must add foundational results of [@Hughes-Taylor-Williams(1990)] and a folk theorem proved in [@Hughes(1999)] about mapping cylinder neighbourhoods, MCNs, and manifold approximate fibrations, MAFs. We summarise these results in Theorem \[thm:mafs-vs-mcns\] and Corollary \[cor:relative-mafs-vs-mcns\] and use them to show that Hutt’s map is defined. Then one has to take more care than Hutt to show that the map is well-defined. To show that the now well-defined map $CW^j$ is indeed $P^j$ we use algebraic surgery which requires an inductive dissection of a topological manifold similar to, but not in general the same as, a simplicial decomposition. In particular algebraic surgery requires that we apply Hutt’s construction inductively to each space in such a dissection. Then one discovers that Theorem \[thm:mafs-vs-mcns\] concerning MCNs and MAFs has dimension restrictions which can only be satisfied for 8-periodicity. Thus we show that $CW^2 = P^2$ and this is sufficient to prove the additivity of the reduced $\rho$-invariant. We hope that the work in this paper might serve as a foundation to at last give a detailed proof that $CW^1 = P^1$. \[rem:decorations\] All the results of the present paper work equally well for structure sets and simple structure sets. Thus in the familiar notation, the reader may substitute ${\mathcal{S}}^s(M)$ or ${\mathcal{S}}^h(M)$ for ${\mathcal{S}}(M)$ and its variants throughout the paper. To justify this we note that the $\rho$-invariant is an $h$-cobordism invariant and hence defines a function of both versions of the structure set. Moreover the forgetful map $${\mathcal{S}}^s(M) \to {\mathcal{S}}^h(M)$$ is a homomorphism. In particular the results of [@Quinn(1970)], [@Siebenmann(1977)], [@Ranicki(1978)] and [@Ranicki(1992)] work equally well for each torsion decoration. With regard to the periodicity maps $CW^j$ our treatment is also simultaneous for both decorations: we use $h$-cobordisms throughout, but the arguments are verbally the same with $s$-cobordisms. In this direction our work generalises [@Cappell-Weinberger(1985)] and [@Hutt(1998)] who only deal with the $s$-decoration. The rest of the paper is organised as follows. In Section \[sec:rho-inv\] we define the $\rho$-invariant and its reduced variations. We also recall the group structure on ${\mathcal{S}}_{\partial}(M \times D^l)$ for $l \geq 1$ and we prove Proposition \[prop-C\]. The proof of Theorem \[thm:A\] occupies Sections \[sec:mafs-and-mcns\]-\[sec:sieb-per\]. In the preparatory Section \[sec:mafs-and-mcns\] we recall and reformulate essential facts about MCNs and MAFs. In Section \[sec:cw-map\] we review Hutt’s account of the construction of the Cappell-Weinberger map. In Section \[sec:alg-sur\] we review the framework of the algebraic theory of surgery from [@Ranicki(1992)] which is the key tool in the proof of Theorem \[thm:A\] in Section \[sec:sieb-per\]. The proof of Theorem \[thm:B\] occupies the last two sections. Section \[sec:bordism-groups\] is again preparatory and the proof is completed in Section \[sec:completion\]. [Acknowledgements.]{} We would like to thank Jim Davis and Andrew Ranicki for helpful discussions, Bruce Hughes for enlightening correspondence at a critical juncture and Wolfgang L[ü]{}ck for raising the motivating question of the paper as well as helpful discussions. The $\rho$-invariant {#sec:rho-inv} ==================== Let $M$ be a closed oriented topological manifold of dimension $n = 2d-1 \geq 5$ with a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ where $G$ is a finite group. In this section we recall the definition of the reduced $\rho$-invariant function, denoted ${\widetilde{\rho}}$, defined on the structure set ${\mathcal{S}}(M)$ as well as a relative $\rho$-invariant, denoted ${\widetilde{\rho}}_{\partial}$, which is defined on the rel boundary structure set $ {\mathcal{S}}_{\partial}(M \times D^{2j})$. The main outcome of the section is the proof of Proposition \[prop-C\] which states that ${\widetilde{\rho}}_{\partial}$ is a homomorphism. The $\rho$-invariant. {#subsec:rho-inv} --------------------- The $\rho$-invariant is an invariant of odd-dimensional manifolds associated to the $G$-signature of cobounding even-dimensional manifolds. We first briefly recall the $G$-signature. G-signature. Let $G$ be a finite group acting smoothly on a smooth manifold $Z^{2d}$. The rational intersection form of $Z$ is then a non-degenerate $(-1)^d$-symmetric bilinear form on which $G$ acts. One can complexify the form and consider the positive and negative definite ${\mathbb{C}}$-vector subspaces. These are $G$-invariant and hence define $G$-representations which can be subtracted in the representation ring $R_{\mathbb{C}}(G)$. The virtual representation thus obtained is denoted by ${\textup{G-sign}}(Z)$. Complex conjugation induces an involution on $R_{{\mathbb{C}}} (G)$ with $(\pm 1)$-eigenspaces. In terms of characters the $(+1)$-eigenspace corresponds to real characters and the $(-1)$-eigenspace corresponds to purely imaginary characters. We will denote $$R_{{\mathbb{C}}}^{\pm} (G) : = \{ \chi \pm \chi^{-1} \; \| \; \chi \in R_{{\mathbb{C}}} (G) \}.$$ One can also show that ${\textup{G-sign}}(Z) \in R^{(-1)^d} (G)$ which in terms of characters means that we obtain a real (purely imaginary) character, which will be denoted as ${\textup{G-sign}}(-,Z) {\colon\!}g \in G \mapsto {\textup{G-sign}}(g,Z) \in {\mathbb{C}}$. The cohomological version of the Atiyah-Singer $G$-index theorem, [@Atiyah-Singer-III(1968) Theorem 6.12], tells us that if $Z$ is closed then for all $g \in G$ $$\label{ASGIT} {\textup{G-sign}}(g,Z) = L(g,Z) \in {\mathbb{C}},$$ where $L(g,Z)$ is an expression obtained by evaluating certain cohomological classes on the fundamental classes of the $g$-fixed point submanifolds $Z^g$ of $Z$. In particular if the action is free then ${\textup{G-sign}}(g,Z) = 0$ if $g \neq 1$. This means that ${\textup{G-sign}}(Z)$ is a multiple of the regular representation. This theorem was generalised by Wall to topological semifree actions on topological manifolds, which is the case we will need in this paper [@Wall(1999) Chapter 14B]. The assumption that $Z$ is closed is essential here, and motivates the definition of the $\rho$-invariant. Bordism groups. To define the $\rho$-invariant one also needs the following result which starts with the work of Conner and Floyd [@Conner-Floyd(1964)] on smooth bordism, proceeds through [@Williamson(1966)] for piecewise linear bordism and finishes with [@Madsen-Milgram(1979)] for topological bordism. \[thm:bord\] Let $G$ be a finite group with classifying space $BG$ and let $\Omega_{n}^{{\textup{STOP}}} (BG)$ denote bordism group of $n$-dimensional closed oriented topological manifolds with a reference map to $BG$. Then for $2d-1 \geq 1$, $$\Omega_{2d-1}^{{\textup{STOP}}} (BG) \otimes {\mathbb{Q}}= 0.$$ Let $N$ be a closed $(2d-1)$-dimensional manifold with a reference map $\lambda (N) {\colon\!}N {\rightarrow}BG$ inducing a homomorphism $\lambda (N)_\ast {\colon\!}\pi_1 (N) {\rightarrow}G$. The above result means that there exists a $2d$-dimensional manifold with boundary $Z$ with a reference map $\lambda (Z) {\colon\!}Z {\rightarrow}BG$ inducing a homomorphism $\lambda (Z)_\ast {\colon\!}\pi_1 (Z) {\rightarrow}G$ such that ${\partial}Z = r \cdot N$ for some $r \geq 1$ and such that the restriction $\lambda (Z)\|_{{\partial}Z} = r \cdot \lambda (N)$. Then we also have the induced $G$-covering $\widetilde Z$ on which the group $G$ acts freely via deck transformations. It is a manifold with boundary $r \cdot \widetilde Y$, $r$ copies of the induced $G$-covering of $Y$. The above considerations make it possible to make the following definition. [[@Atiyah-Singer-III(1968) Section 7]]{} \[defn:rho-1\] Let $N$ be a closed topological $(2d-1)$-dimensional manifold with a reference map $\lambda (N) {\colon\!}N {\rightarrow}BG$ where $G$ is a finite group. Define $$\rho (N,\lambda (N)) : = \frac{1}{r} \cdot {\textup{G-sign}}(\widetilde Z) \in {\mathbb{Q}}R^{(-1)^d} (G)/ \langle \textup{reg} \rangle =: {\mathbb{Q}}{R_{\widehat G}}^{(-1)^d}$$ for some $r \in {\mathbb{N}}$ and $(Z,\partial Z)$ such that $\partial Z = r \cdot N$ and there is $\lambda (Z) {\colon\!}Z {\rightarrow}BG$ restricting to $r \cdot \lambda (N)$ on ${\partial}Z$. The symbol $\langle \textup{reg} \rangle$ denotes the ideal generated by the regular representation, the symbol ${\mathbb{Q}}R^\pm (G)$ means ${\mathbb{Q}}\otimes R^\pm (G)$. The $\rho$-invariant is well defined by the Atiyah-Singer $G$-index theorem [@Atiyah-Singer-III(1968) Theorem (6.12)] and its topological generalisation [@Wall(1999) Chapter 14B]. When the reference map is clear we will often leave out the map $\lambda (N)$ from the notation and simply write $\rho (N)$. Structure sets -------------- The structure set of a compact topological manifold is the basic object of study in surgery theory. When calculated it gives us understanding of the manifolds in the homotopy type of that given manifold, as shown for example in [@Wall(1999) Part 3]. \[defn:structure-set\] Let $M$ be a compact $n$-dimensional manifold with boundary $\partial M$ (which may be empty). A (simple) manifold structure on $M$ relative to ${\partial}M$ consists of a (simple) homotopy equivalence of $n$-dimensional compact manifold with boundary $$(h,\partial h) {\colon\!}(N,\partial N) {\rightarrow}(M,\partial M)$$ such that $\partial h$ is a homeomorphism. Two such structures $(h_1,\partial h_1)$ and $(h_2, \partial h_2)$ are equivalent if there exists a (simple) homotopy equivalence of $(n+1)$-dimensional manifold $4$-ads $H {\colon\!}(W,N_1,N_2,W_{\partial}) {\rightarrow}(M \times I,M \times \{0\},M \times\{1\},{\partial}M \times I)$, such that $H\|_{N_1} = h_1$ and $H\|_{N_2} = h_2$ and $H\|_{W_{\partial}}$ are homeomorphisms. The (simple) structure set ${\mathcal{S}}_\partial (M)$ is defined as the set of equivalence classes of (simple) manifold structures on $M$ relative to ${\partial}M$. In the case where ${\partial}M$ is empty we write ${\mathcal{S}}(M)$. More generally, assume that $(M, {\partial}_1 M, {\partial}_2 M)$ is a manifold $3$-ad (see [@Wall(1999) Chapter 0], note that one or both of ${\partial}_i M$ may be empty). A (simple) manifold structure on $M$ relative to ${\partial}_1 M$ consists of a (simple) homotopy equivalence of $n$-dimensional compact manifold $3$-ads $$(h, \partial_1 h, \partial_2 h) {\colon\!}(N, {\partial}_1 N, {\partial}_2 N) {\rightarrow}(M, {\partial}_1 M, {\partial}_2 M)$$ such that ${\partial}_1 h$ is a homeomorphism. So one allows more flexibility on the part of the boundary ${\partial}_2 M$. There is a corresponding equivalence relation which allows to one define the (simple) structure set in this setting, which is denoted ${\mathcal{S}}_{{\partial}_1 M} (M)$. Hence ${\partial}_2 M$ does not appear in the notation, which usually does not cause a confusion. Also note that the $s$-cobordism theorem entails that, if $\dim (M) = n \geq 5$, then two simple manifold structures $h_1$ and $h_2$ are equivalent if and only if there exists a homeomorphism $f {\colon\!}N_1 {\rightarrow}N_2$ such that $h_2 \circ f \simeq h_1$ rel ${\partial}$. All the results of the present paper work equally well for structure sets and simple structure sets. Thus in the familiar notation, the reader may substitute ${\mathcal{S}}^s(M)$ or ${\mathcal{S}}^h(M)$ for ${\mathcal{S}}(M)$ and its variants throughout the paper. To keep the language simple we will work with structure sets and manifold structures. The main tool for determining ${\mathcal{S}}_{\partial}(M)$ for a specific manifold $M$, with $\dim (M) = n \geq 5$, is the surgery exact sequence (see [@Wall(1999) Chapter 10], [@Kirby-Siebenmann(1977)] for definitions and details): $$\label{eqn:ses} \cdots {\rightarrow}L_{n+1} ({\mathbb{Z}}[\pi_1(M)]) {\rightarrow}{\mathcal{S}}_{\partial}(M) {\rightarrow}[M/{\partial}M; {\textup{G}}/{\textup{TOP}}] {\rightarrow}L_n ({\mathbb{Z}}[\pi_1(M)]).$$ We remark that the expression “exact sequence” makes sense, as explained in [@Wall(1999) Chapter 10], despite the fact that the structure set, as defined, is only a pointed set with the base point the identity ${\textup{id}}{\colon\!}(M,\partial M) {\rightarrow}(M,\partial M)$. On the other hand, as pointed out in the introduction, one can endow ${\mathcal{S}}_{\partial}(M)$ with the structure of an abelian group, which is natural in the sense that the above sequence indeed becomes an exact sequence of abelian groups. This follows from the identification of the surgery exact sequence (\[eqn:ses\]) with the algebraic surgery exact sequence (\[alg-sur-seq-concrete\]) which will be discussed in detail in Section \[sec:alg-sur\]. In this section we only want to discuss the case when the compact manifold in question is of the form $M \times D^k$, for $M$ closed with $k \geq 1$. Then there is an easy geometric way of defining the structure of a group on ${\mathcal{S}}_\partial (M \times D^k)$ which is abelian if $k \geq 2$. Abstractly, this follows from the observation that ${\mathcal{S}}_\partial (M \times D^k)$ is the $k$-th homotopy group of a certain space, as explained for example in [@Wall(1999) Chapter 17A]. But we also need an explicit description of the addition. For this denote $$S^{k-1}_\pm : = \{ x = (x_1, \ldots , x_{k}) \in S^{k-1} \; \| \; \pm x_1 \geq 0 \} \\$$ and note that each element in ${\mathcal{S}}_\partial (M \times D^k)$ can be represented as a homotopy equivalence of manifold triads $$(h, \partial_+ h, \partial_- h) {\colon\!}(N, {\partial}_+ N, {\partial}_- N) {\rightarrow}(M \times D^k, M \times S_+^{k-1}, M \times S_-^{k-1})$$ where ${\partial}_\pm h$ are homeomorphisms. Further denote $$D^{k}_\pm : = \{ x = (x_1,\ldots ,x_{k}) \in D^{k} \; \| \; \pm x_1 \geq 0 \} \\$$ and choose suitable homeomorphisms $(D^k, S^{k-1}_+, S^{k-1}_-) \cong (D^k_+, S^{k-1}_+, D^{k-1})$ and $(D^k, S^{k-1}_+, S^{k-1}_-) \cong (D^k_-, D^{k-1}, S^{k-1}_-)$. Also note $D^k = D^k_+ \cup_{D^{k-1}} D^k_-$. \[defn:stacking\] Let $h_i {\colon\!}N_i {\rightarrow}M \times D^k$ with $i = 1,2$ be maps which represent elements in ${\mathcal{S}}_\partial (M \times D^k)$. Define $h_1+h_2 = h$ by $$\label{defn:addition} h = h_1 \cup h_2 {\colon\!}N = N_1 \cup_g N_2 {\rightarrow}M \times D^k = M \times D^k_+ \cup M \times D^k_-$$ where $g {\colon\!}{\partial}_+ N_1 {\rightarrow}{\partial}_- N_2$ is given by $g = ({\partial}_- h_2)^{-1} \circ {\partial}_+ h_1$. Structure sets and the $\rho$-invariant --------------------------------------- Next we define the reduced $\rho$-invariant functions. \[defn:reduced-rho\] Let $M$ be a closed oriented manifold of dimension $n = (2d-1) \geq 5$ with a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ where $G$ is a finite group. Define the function $${\widetilde{\rho}}{\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d} \quad \textup{by} \quad {\widetilde{\rho}}([h]) = \rho (N,\lambda (M) \circ h) - \rho (M,\lambda (M)),$$ where the orientation on $N$ is chosen so that the homotopy equivalence $h {\colon\!}N {\rightarrow}M$ is a map of degree $1$. Since the $\rho$-invariant is an $h$-cobordism invariant [@Atiyah-Singer-III(1968) Corollary 7.5], the function ${\widetilde{\rho}}$ is well-defined. The definition in the relative setting comes from [@Madsen-Rothenberg-II(1989) Section 3]. We need a little preparation. Consider an element $[h]$ in ${\mathcal{S}}_\partial (M \times D^k)$. Let $M(h)$ be a closed manifold given by $$\label{defn:M(h)} M (h) := N \cup_{\partial h} (M \times D^k).$$ If $h$ is the identity we obtain $M ({\textup{id}}) \cong M \times S^k$, in general the map $h$ induces $M(h) \simeq M \times S^k$, and if $k \geq 2$ then $\pi_1(M(h)) \cong \pi_1 (M)$. When $M$ is oriented we equip $N$ with an orientation so that $h$ is a map of degree $1$. The orientation on the closed manifold $M(h)$ can then be chosen so that it agrees with the given orientation on $N$ and it reverses the orientation on $M \times D^k$. If $M$ possess a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ then we obtain a reference map $\lambda (M(h)) {\colon\!}M(h) \simeq M \times S^k {\rightarrow}M {\rightarrow}BG$. \[defn:reduced-rho-del\] Let $M$ be a closed oriented manifold of dimension $n$ with a reference map $\lambda (M) {\colon\!}M {\rightarrow}BG$ where $G$ is a finite group and let $k \geq 1$ be such that $n+k = 2d-1 \geq 5$. Define the function $${\widetilde{\rho}}_\partial {\colon\!}{\mathcal{S}}_\partial (M \times D^k) {\rightarrow}{\mathbb{Q}}{R_{\widehat G}}^{(-1)^d} \quad \textup{by} \quad {\widetilde{\rho}}_\partial ([h]) := \rho (M(h),\lambda(M(h))).$$ Again this well-defined. Also notice that if $k \geq 1$ then ${\widetilde{\rho}}_\partial ([{\textup{id}}]) = 0$. Now we would like to understand the behaviour of ${\widetilde{\rho}}_\partial$ with respect to $+$ defined in \[defn:addition\]. First a definition and then an observation. \[defn:ctd-sum-along-M\] Let $h_i {\colon\!}N_i {\rightarrow}M \times D^k$ with $i = 1,2$ be maps which represent elements in ${\mathcal{S}}_\partial (M \times D^k)$. Consider $\bar M (h_i) : = M(h_i) \smallsetminus \textup{int} (M \times D^{k-1} \times [-\epsilon,\epsilon])$, for small $\epsilon > 0$, where $D^{k-1} = D^k_+ \cap D^k_-$. Identify ${\partial}\bar M(h_i)$ with $M \times S^{k-1}$. Define the closed oriented manifold $M(h_1) \#_M M(h_2)$ by: $$M(h_1) \#_M M(h_2) := \bar M (h_1) \cup_{{\textup{id}}\times r} \bar M (h_2)$$ where $r$ is an orientation reversing homeomorphism of $S^{k-1}$. \[plus-vs-par-ctd-sum\] There is a homeomorphism of oriented manifolds $$M(h_1) \#_M M(h_2) \cong M(h_1+h_2).$$ Both sides can be identified with the union: $$N_1 \cup_{g_1} M \times S^{k-1} \times I \cup_{g_2} N_2$$ where $g_1 = {\partial}h_1 {\colon\!}{\partial}N_1 {\rightarrow}M \times S^{k-1}$ and $g_2 = r \circ {\partial}h_2 {\colon\!}{\partial}N_2 {\rightarrow}M \times S^{k-1}$. From the definition ${\widetilde{\rho}}_{\partial}([h]) = \rho(M(h))$ we see that Proposition \[prop-C\] is equivalent to the following \[rho-add-for-bdry\] There is an equality $$\rho (M(h_1+h_2)) = \rho (M(h_1)) + \rho (M(h_2)).$$ Let $Z(h_1)$ be such that $\partial Z(h_1) = k \cdot M(h_1)$ and let $Z(h_2)$ be such that $\partial Z(h_2) = l \cdot M(h_2)$. Then $\partial l \cdot Z(h_1) = kl \cdot M(h_1)$ and $\partial k Z(h_2) = kl \cdot M(h_2)$, so we can assume $k=l$. In fact we will assume $k=l=1$, which makes the notation simple, the easy generalisation is left for the reader. Using Lemma \[plus-vs-par-ctd-sum\] we build a coboundary for $M(h_1+h_2)$ from the coboundaries $Z(h_1)$ and $Z(h_2)$ by the following construction. Note that the manifold $M(h_1)$, as a boundary component of $Z(h_1)$, has a collar. Denote by $U(h_1) \subset Z(h_1)$ the portion of that collar along $M \times D^{k-1} \times [-\epsilon,\epsilon] \subset M(h_1)$. Construct the manifold $\bar Z(h_1)$ by removing $U(h_1)$ (the interior and a suitable part of the boundary). Then the boundary of $\bar Z(h_1)$ is decomposed as $$\bar M(h_1) \cup \partial' \bar Z(h_1) = \bar M(h_1) \cup (M \times S^{k-1}) \times I \; \cup \; (M \times D^{k-1} \times [-\epsilon,\epsilon]) \times \{1\}.$$ Similarly for $h_2$ instead of $h_1$. Recall also the orientation reversing homeomorphism $r {\colon\!}S^{k-1} {\rightarrow}S^{k-1}$ which we can extend to $r : D^k {\rightarrow}D^k$ and identify $D^{k-1} \times [-\epsilon,\epsilon] \cong D^k$. We define $$Z := \bar Z(h_1) \cup_{\bar r} \bar Z(h_2)$$ where $\bar r {\colon\!}{\partial}' Z(h_1) {\rightarrow}{\partial}' Z(h_2)$ is the homeomorphism given by $$\bar r = ({\textup{id}}\times r \times {\textup{id}}) \cup ({\textup{id}}\times r \times \{1\}).$$ The following picture depicts the situation. $$\begin{tikzpicture}[scale=0.5] \draw (3cm,1cm) arc (0:90:2cm); \draw (-1cm,3cm) arc (90:180:2cm); \draw (-3cm,-1cm) arc (180:270:2cm); \draw (1cm,-3cm) arc (270:360:2cm); \draw (2.9,1) -- (3.1,1); \draw (1,2.9) -- (1,3.1); \draw (-1,2.9) -- (-1,3.1); \draw (-2.9,1) -- (-3.1,1); \draw (-2.9,-1) -- (-3.1,-1); \draw (-1,-2.9) -- (-1,-3.1); \draw (1,-2.9) -- (1,-3.1); \draw (2.9,-1) -- (3.1,-1); \draw[dashed][->] (-3,-0.9) -- (-3,0.9); \draw (-3,0) node[anchor=east] {$\partial h_1$}; \draw[dashed][->] (3,0.9) -- (3,-0.9); \draw (3,0) node[anchor=west] {$\partial h_2$}; \draw[dashed][->] (-0.9,3) -- (0.9,3); \draw (-0.4,2.2) node[anchor=south] {$r$}; \draw[dashed][->] (-0.9,-3) -- (0.9,-3); \draw (-0.4,-3.1) node[anchor=north] {$r$}; \draw (-1,-3) -- (-2,-2) -- (-2,4) -- (-1,3); \draw (1,-3) -- (0,-2) -- (0,4) -- (1,3); \draw[dashed][->] (-1.9,0.5) -- (-0.1,0.5); \draw (-1,0.5) node[anchor=north] {$\bar r$}; \draw (0,4) .. controls (1,5) and (2,5) .. (3,5); \draw (0,-2) .. controls (1,-2) and (2,-2) .. (3,-3); \draw (-2,4) .. controls (-2.5,5) and (-3.25,5) .. (-4,5); \draw (-2,-2) .. controls (-2.5,-2) and (-3.25,-2) .. (-4,-3); \end{tikzpicture}$$ By construction the boundary of $Z$ is $M(h_1+h_2) = M(h_1) \#_M M(h_2)$. Then Novikov additivity gives $$\rho(M(h_1+h_2)) = {\textup{G-sign}}(Z) = {\textup{G-sign}}(Z(h_1)) + {\textup{G-sign}}(Z(h_2)) = \rho (M(h_1)) + \rho (M(h_2)).$$ MAFs and MCNs {#sec:mafs-and-mcns} ============= This section contains preparatory material about mapping cylinder neighbourhoods which will be used in the construction of the map $CW^j$ in the following section. Let $X^n \subset Y^{n+q}$, $q \geq 1$, be a locally flat submanifold. A [mapping cylinder neighbourhood]{}, MCN, of $X$ is a codimension-$0$ submanifold with boundary of $Y$, $(N,{\partial}N)$, such that $X \subset {\textup{int}}(N)$ and there is a deformation retraction $p {\colon\!}N {\rightarrow}X$ such that ${\partial}p := p\|_{{\partial}N} {\colon\!}{\partial}N {\rightarrow}X$ satisfies $({\textup{cyl}}({\partial}p),{\partial}N) \cong (N,{\partial}N)$, where ${\textup{cyl}}({\partial}p))$ denotes the mapping cylinder of ${\partial}p$. See [@Quinn(1979) Section 3] for more information about the existence of MCNs. Theorem \[thm:mafs-vs-mcns\] below, which is taken from [@Hughes(1999)], recalls a characterisation of MCNs using manifold approximate fibrations (MAFs). An [approximate fibration]{} $p {\colon\!}P {\rightarrow}N$ is a map which has an approximate homotopy lifting property. It is called a MAF if both $P$ and $N$ are manifolds. For more information we refer the reader to [@Hughes-Taylor-Williams(1990) Section 1]. We will need the following two properties. Firstly, it follows easily from the definitions that a composition of MAFs is a MAF. Secondly, being a MAF is a local property: this means that in order to determine whether a map $p {\colon\!}P {\rightarrow}N$ between closed manifold is a MAF it is enough to check this property in a neighbourhood of each point of $N$, see [@Hughes-Taylor-Williams(1990) Corollary 12.14] and [@Chapman(1980) Proposition 2.2]. [[@Hughes(1999) Theorem 6.1]]{} \[thm:mafs-vs-mcns\] Let $p {\colon\!}P {\rightarrow}N$ be a map between closed manifolds with $\dim (P) = m$ and $\dim (N) = n$. If ${\textup{cyl}}(p)$ is a manifold with $N$ a locally flat submanifold then $p$ is a MAF with ${\textup{hofib}}(p) \simeq S^{m-n}$. The converse is also true if $m \geq 5$. The above theorem is stated in [@Hughes(1999)] with the dimension restriction for both implications. However, it is clear from the proof that it is used only in one direction. We need an generalisation of the above statement for compact manifolds with boundary. Let $(p,{\partial}p) {\colon\!}(P,{\partial}P) {\rightarrow}(N,{\partial}N)$ be a map of compact manifolds with boundary. If ${\partial}N$ has more than one connected component, we index these by $\alpha \in {\mathcal{I}}$ and denote ${\partial}_\alpha p {\colon\!}{\partial}_\alpha P {\rightarrow}{\partial}_\alpha N$ the corresponding restrictions of ${\partial}p$. The mapping cylinder ${\textup{cyl}}({\partial}p)$ is a subspace of the mapping cylinder ${\textup{cyl}}(p)$ and we have $${\partial}{\textup{cyl}}(p) = {\textup{cyl}}({\partial}p) \cup_{{\partial}P \times \{0\}} P \times \{0\}.$$ The triple $({\textup{cyl}}(p); {\textup{cyl}}({\partial}p), P \times \{0\})$ defines a triad of spaces. \[cor:relative-mafs-vs-mcns\] Let $(p,{\partial}p) {\colon\!}(P,{\partial}P) {\rightarrow}(N,{\partial}N)$ be a map of compact manifolds with boundary, $\dim (P) = m$, $\dim (N) = n$. Assume in addition that on a collar of ${\partial}P$ the map $p$ is the product map of ${\partial}p$ with the identity in the collar direction. If the triad $({\textup{cyl}}(p); {\textup{cyl}}({\partial}p), P \times \{0\})$ is a manifold triad with $(N,{\partial}N)$ a locally flat submanifold (with boundary) of $({\textup{cyl}}(p),{\textup{cyl}}({\partial}p))$ then $(p,{\partial}p)$ is a pair of MAFs and we have for each connected component $\alpha \in {\mathcal{I}}$ $${\textup{hofib}}({\partial}_\alpha p) \simeq {\textup{hofib}}(p) \simeq S^{m-n}.$$ The converse is also true if $m \geq 6$. By a pair of MAFs we just mean that both ${\partial}p$ and $p$ are MAFs. For a manifold triad see [@Wall(1999) Chapter 0]. The following two observations are used. Firstly, the fact recalled above that being a MAF is a local property. Secondly, for a map $p {\colon\!}P^m {\rightarrow}N^n$ between closed manifold, we have that ${\textup{hofib}}(p) \simeq S^{m-n}$ if and only if $({\textup{cyl}}(p),P)$ is a Poincaré pair. Also, for $(p,{\partial}p) {\colon\!}(P,{\partial}P) {\rightarrow}(N,{\partial}N)$, a map between compact manifolds with boundary, we have that ${\textup{hofib}}({\partial}_\alpha p) \simeq {\textup{hofib}}(p) \simeq S^{m-n}$ for each $\alpha \in {\mathcal{I}}$ if and only if $({\textup{cyl}}(p);P,{\textup{cyl}}({\partial}p))$ is a Poincaré triad. The general idea is to reduce the proof to considerations about the map $$\bar p := p \cup_{{\partial}p} p {\colon\!}(P \cup_{{\partial}P} P) \longrightarrow (N \cup_{{\partial}N} N).$$ To prove the if part note that $\bar p$ fulfils the assumptions of the if part of Theorem \[thm:mafs-vs-mcns\]. It is a MAF because being a MAF is a local property and both parts are MAFs. We have ${\textup{hofib}}(\bar p) \simeq S^{m-n}$, since $({\textup{cyl}}(\bar p),P)$ is a Poincaré pair, because it is obtained by gluing two Poincaré triads. Hence by Theorem \[thm:mafs-vs-mcns\], ${\textup{cyl}}(\bar p)$ is a closed manifold with $N \cup N$ locally flatly embedded. But this implies that $({\textup{cyl}}(p),{\textup{cyl}}({\partial}p) \subset {\textup{cyl}}(\bar p)$ is a codimension-$0$ submanifold with boundary with $(N,{\partial}N)$ locally flatly embedded. To prove the only if part observe immediately that ${\textup{cyl}}(\bar p)$ fulfils the assumptions of Theorem \[thm:mafs-vs-mcns\]. Hence $\bar p$ is a MAF with ${\textup{hofib}}(\bar p) \simeq S^{m-n}$. We obtain that $(p,{\partial}p)$ is a MAF pair by the locality of the MAF condition. The statement about the homotopy fibre follows immediately from the assumptions even without invoking $\bar p$. Recall from surgery theory that the normal invariants of a manifold $N$, that means the bordism set of degree one normal maps with target $N$, are in one-to-one correspondence with $[N,{\textup{G}}/{\textup{TOP}}]$, homotopy classes of maps from $N$ to the space ${\textup{G}}/{\textup{TOP}}$. If instead of $N$ we consider the manifold with boundary $N \times D^{r+1}$ and if $r \geq 2$, then we have a $(\pi\!-\!\pi)-$situation and hence $${\mathcal{S}}(N \times D^{r+1},{\partial}) \cong [N \times D^{r+1} ; {\textup{G}}/{\textup{TOP}}] \cong [N;{\textup{G}}/{\textup{TOP}}].$$ Therefore we can “realise” elements of $[N;{\textup{G}}/{\textup{TOP}}]$ as homotopy equivalences of manifolds with boundary $f {\colon\!}(W,{\partial}W) {\rightarrow}(N \times D^{r+1},{\partial})$. As such $(W,{\partial}W)$ is just some manifold with boundary homotopy equivalent to $N \times D^{r+1}$. The following proposition says that there is more structure in this situation: the manifold $M$ can be identified as a MCN of $N$ and the map $f$ as the cylinder of the restriction of $f$ to the boundary. \[prop:constructing-neighborhoods\] Let $r \geq 2$, let $n + r \geq 5$ and let $N$ be a closed manifold of dimension $n$. For any element $\chi \in [N,{\textup{G}}/{\textup{TOP}}]$ there exists a commutative diagram $$\xymatrix{ (W,{\partial}) \ar[d]_{p} \ar[r]^(0.4){\omega} & (N \times D^{r+1},{\partial}) \ar[d]^{{\textup{pr}}_1} \\ N \ar[r]_{{\textup{id}}} & N }$$ where $p {\colon\!}W {\rightarrow}N$ is a MCN with $N$ a locally flat submanifold and the map $\omega {\colon\!}W {\rightarrow}N \times D^{r+1}$ is induced from a fibre homotopy equivalence of $r$-dimensional spherical fibrations associated to $p$ and ${\textup{pr}}_1$. In particular $\omega$ is a homotopy equivalence of pairs such that $[\omega] = \chi \in {\mathcal{S}}(N \times D^{r+1},{\partial}) \cong [N;{\textup{G}}/{\textup{TOP}}]$. This is proved by Hutt in [@Hutt(1998) Section 1]. He closely follows Pedersen in [@Pedersen(1975) Proof of Lemma 9] with some additional ingredients from [@Quinn(1979)]. Pedersen’s arguments are in turn based on [@Rourke-Sanderson(1970)]. In [@Rourke-Sanderson(1970) Section 0] a simplicial group ${\bf T}op_{r+1}$ is defined with the following properties: (1) the set of equivalence classes of germs of codimension $(r+1)$ topological neighbourhoods of a topological manifold $N$ is in bijective correspondence with the set of homotopy classes of maps $[N, B{\bf T}op_{r+1}]$, (2) there is a map ${\bf T}op_{r+1} \to G_{r+1}$, the group of self-equivalences of $S^r$, such that the inclusions $G_{r+1} \to {\textup{G}}$ and ${\bf T}op_{r+1} \to {\textup{TOP}}$ give rise to a homotopy equivalence $G_{r+1}/{\bf T}op_{r+1} \simeq {\textup{G}}/{\textup{TOP}}$. Let $x_{r+1} : N \to G_{r+1}/{\bf T}op_{r+1}$ be a map representing $\chi_{r+1} \in [N, G_{r+1}/{\bf T}op_{r+1}]$ where $\chi_{r+1}$ corresponds to $\chi$ under the above equivalence. If $i: G_{r+1}/{\bf T}op_{r+1} \to B{\bf T}op_{r+1}$ is the canonical map then $i \circ x_{r+1} : N \to B{\bf T}op_{r+1}$ defines the germ of a codimension-$(r+1)$ topological neighbourhood of $N$. Let $V \supset N$ be such a neighbourhood. By [@Quinn(1979) Theorem 3.1.1] there is a MCN of $N$ in $V$, $W \supset N$ and we let $p : {\partial}W \to N$ be the MAF associated to $W \supset N$. The homotopy fibre of $p$ is $S^r$ and we convert $p$ into a spherical fibration $S(p) : S({\partial}W) \to N$. So far we have only used $i \circ x_{r+1}$: by definition the map $x_{r+1}$ defines a fibre homotopy trivialisation of $S(p)$ and so we obtain a homotopy equivalence $\tau : {\partial}W \simeq N \times S^r$. We set $\omega := {\textup{cyl}}(\tau): (W, {\partial}W) \simeq (N \times D^{r+1}, {\partial})$. It remains to show that the normal invariant of $\omega$ is $\chi$. As discussed above, we have a canonical identification of the normal invariant set of $N \times D^{r+1}$ with $[N, {\textup{G}}/{\textup{TOP}}]$. The normal invariant of any degree one normal map $f: (M, {\partial}M) \to (N \times D^{r+1}, {\partial})$ can be computed from the normal invariant of the splitting obstruction along $N \times \{ 0 \} \subset N \times D^{r+1}$. But by definition, the splitting obstruction to $\omega$ along $N \times 0$ has normal invariant $\chi$: see [@Madsen-Milgram(1979) Theorem 2.23] for this identification which was stated there for the $PL$-category but remains true in the topological category given the later proof of topological transversality [@Kirby-Siebenmann(1977)], [@Freedman-Quinn(1990) Ch. 9]. We will also need a relative version of Proposition \[prop:constructing-neighborhoods\]. In that case we are given a manifold with boundary $(N,{\partial}N)$ of dimension $n$. The product $N \times D^{r+1}$ becomes a manifold triad and we define ${\partial}_0 (N \times D^{r+1}) := N \times S^r$ and ${\partial}_1 (N \times D^{r+1}) := {\partial}N \times D^{r+1}$. Then ${\partial}{\partial}_1 = {\partial}{\partial}_0 = {\partial}N \times S^r$. If $r \geq 2$ we are again in a $(\pi\!-\!\pi)-$situation and hence ${\mathcal{S}}(N \times D^{r+1},{\partial}_0,{\partial}_1) \cong [N \times D^{r+1} ; {\textup{G}}/{\textup{TOP}}] \cong [N;{\textup{G}}/{\textup{TOP}}]$. \[prop:constructing-neighborhoods-relative\] Let $r \geq 2$, $n + r \geq 5$ and let $(N,{\partial}N)$ be a compact manifold of dimension $n$ with boundary. Suppose given a MCN $({\partial}_1 W, {\partial}{\partial}_1 W) {\rightarrow}{\partial}N$, with ${\partial}N$ a locally flat submanifold, whose associated $r$-spherical fibration is fibre homotopically trivialised by a map $({\partial}_1 W, {\partial}{\partial}_1 W) {\rightarrow}({\partial}N \times D^{r+1},{\partial}N \times S^r)$ which is represented by a map $\xi {\colon\!}{\partial}N {\rightarrow}{\textup{G}}/{\textup{TOP}}$. For any element $\chi \in [N,{\textup{G}}/{\textup{TOP}}]$, such that $\chi\|_{{\partial}N} = \xi$ there exists a diagram $$\xymatrix{ (W,{\partial}) \ar[d]_{p} \ar[r]^(0.4){\omega} & (N \times D^{r+1},{\partial}) \ar[d]^{{\textup{pr}}_1} \\ N \ar[r]_{{\textup{id}}} & N }$$ where $p {\colon\!}W {\rightarrow}N$ is a MCN with $N$ a locally flat submanifold and the map $\omega {\colon\!}W {\rightarrow}N \times D^{r+1}$ is induced from a fibre homotopy equivalence of $r$-dimensional spherical fibrations associated to $p$ and ${\textup{pr}}_1$. In particular $\omega$ is a homotopy equivalence of pairs, and restrictions of everything to the appropriate parts of boundary agree with the given structures, such that $[\omega] = \chi$. The theorems of [@Rourke-Sanderson(1970)] and [@Quinn(1979)] and [@Madsen-Milgram(1979)] used in the proof of Proposition \[prop:constructing-neighborhoods\] have suitable relative versions. The proof is then a routine modification of the arguments in Proposition \[prop:constructing-neighborhoods\]. In Proposition \[prop:constructing-neighborhoods-relative\] we had another possibility: to start with a map $N {\rightarrow}{\textup{G}}/{\textup{TOP}}$ without specifying the MCN over the boundary. We could then use the absolute version to produce such a MCN and then use the relative version. In that case, however, the dimension restrictions would have to be relaxed by $1$ which would be inconvenient later. The Cappell-Weinberger map {#sec:cw-map} ========================== In this section we recall Hutt’s description of the Cappell-Weinberger map and prove some basic facts about this map. We present the map both for the usual $4$-periodicity and also for $8$-periodicity: a possibility pointed out in [@Cappell-Weinberger(1985) p. 48]. Thus let ${\mathbb{F}}= {\mathbb{C}}$ or ${\mathbb{H}}$ and let $k = 2$ or $4$ be the dimension of ${\mathbb{F}}$ over ${\mathbb{R}}$. Let $h : N \to M$ be a homotopy equivalence of closed topological manifolds of dimension $n \geq 5$ representing $[h] \in \mathcal{S}(M)$. From $h$ the Hutt construction produces a homotopy equivalence $h' {\colon\!}N' {\rightarrow}M \times D^{2k}$ of manifolds with boundary defined by (\[def:N-prime\]) and (\[def:h-prime\]) below. The restriction of $h'$ to the boundary of $N'$ is a homeomorphism and so $h'$ represents an element $[h'] \in {\mathcal{S}}_{\partial}(M \times D^{2k})$. The mapping of structures, $[h] \mapsto [h']$, is the $CW$-map of Definition \[def:CW-map\]. The rest of the first subsection is devoted to proving that this map is well defined. In the second subsection we will review the construction of a homotopy equivalence of closed manifolds $\widehat h {\colon\!}\widehat N {\rightarrow}M \times {\mathbb{F}}P^2$ given by “extension by a homeomorphism” from $h'$. In Lemma \[lem:phi\] we will show that this structure is equivalent in ${\mathcal{S}}(M \times {\mathbb{F}}P^2)$ to another structure, $\bar h {\colon\!}\bar N {\rightarrow}M \times {\mathbb{F}}P^2$ which has a certain factorisation, whereas $h'$ does not possess an analogous factorisation. It is not immediately clear that the Hutt construction produces a periodicity map. We prove this later in Section \[sec:sieb-per\] for ${\mathbb{F}}= {\mathbb{H}}$. We note that an essential component of the construction is the use of certain $S^{k-1}$-branched coverings. This permits the extension by a homeomorphism mentioned above, which is a key ingredient in the proof of the fact that the $CW$-map is a periodicity map. It is useful to observe that the constructions of $h'$, $\widehat h$ and $\bar h$ have two components. It is one thing to construct the sources $N'$, $\widehat N$ and $\bar N$ of the above maps and it is another issue to construct the maps to $M \times D^{2k}$ and $M \times {\mathbb{F}}P^2$. In particular it is easier to construct the sources. We use this point in the last Section \[sec:completion\], where we apply a version of the Hutt construction for manifolds with boundary to construct a coboundary for the closed manifold $\bar N$. Definition ---------- Let $h : N \to M$ be a homotopy equivalence of closed topological manifolds of dimension $n \geq 5$ representing $[h] \in \mathcal{S}(M)$. Our starting point is the following [[@Hutt(1998) Section 1]]{} \[lem:hutt-1\] There is a commutative diagram of maps of pairs $$\begin{diagram} \node{(\bar W, {\partial}\bar W)} \arrow{e,t}{\omega} \arrow{s,r}{} \node{(N \times D^{k+1}, N \times S^k)} \arrow{e,t}{h \times {{\textup{id}}}} \arrow{s,r}{} \node{(M \times D^{k+1}, M \times S^k)} \arrow{s,r}{} \\ \node{N} \arrow{e,t}{{{\textup{id}}}} \node{N} \arrow{e,t}{h} \node{M} \end{diagram}$$ where $\bar W$ is a mapping cylinder neighbourhood of $N$ in which $N$ is locally flatly embedded and $\omega : \bar W \to N \times D^{k+1}$ is a homotopy equivalence of pairs such that the composite $\bar \psi : = (h \times {{\textup{id}}}) \circ \omega$ is $h$-cobordant, as a map of pairs, to a homeomorphism. The homotopy equivalence $h \times {{\textup{id}}}$ defines an element of $\mathcal{S}(M \times D^{k+1})$ which, by the $(\pi\!-\!\pi)-$theorem [@Wall(1999) Chapter 4], is isomorphic to the normal invariant set $\mathcal{N}(M \times D^{k+1}) \cong [M,{\textup{G}}/{\textup{TOP}}]$. Hence $h \times {{\textup{id}}}$ defines an element $[h \times {{\textup{id}}}] \in [M,{\textup{G}}/{\textup{TOP}}]$. We choose $\chi \in [N,{\textup{G}}/{\textup{TOP}}]$ so that $(h^{-1} \times {{\textup{id}}})^\ast (\chi) = - [h \times {{\textup{id}}}]$: here we take the negative with the Whitney sum group structure on ${\textup{G}}/{\textup{TOP}}$ and note that since $(h^{-1})^ \ast$ induces an isomorphism of $[N, {\textup{G}}/{\textup{TOP}}] \cong [M, {\textup{G}}/{\textup{TOP}}]$, such a $\chi$ exists. By Proposition \[prop:constructing-neighborhoods\] the element $\chi \in [N,{\textup{G}}/{\textup{TOP}}]$ gives rise to a homotopy equivalence of pairs $$\omega : (\bar W, {\partial}\bar W) \to (N \times D^{k+1}, N \times S^k)$$ with the claimed properties: $\bar W$ is a mapping cylinder neighbourhood of $N$ in which $N$ is locally flatly embedded and $\omega(N) = N \times \{0\} \subset N \times D^{k+1}$. Finally, by the composition formula of [@Madsen-Taylor-Williams(1980)]\[Lemma 2.5\] $(h \times {{\textup{id}}}) \circ \omega$ has trivial normal invariant and hence $\bar \psi : = (h \times {{\textup{id}}}) \circ \omega$ is h-cobordant to a homeomorphism as required. Lemma \[lem:hutt-1\] states in particular that there is a homotopy equivalence $$\label{map:H} H: (U;\bar W,W,U_{\partial}) \to (M \times D^{k+1} \times [0,1];\{0\},\{1\}, M \times S^k \times [0,1])$$ where $U$ is an $h$-cobordism of manifolds with boundary; we denote ${\partial}U = \bar W \cup W \cup U_{\partial}$, where $U_{\partial}$ is an $h$-cobordism between ${\partial}\bar W$ and ${\partial}W$. Also $H\|_{\bar W} = \bar \psi$ and $H\|_{W}$ is a homeomorphism $H\|_{W} =: \psi : (W, {\partial}W) \cong (M \times D^{k+1}, M \times S^{k})$. For later use, we write $H_{\partial}: U_{\partial}\to M \times S^{k} \times [0,1]$ for the restriction of $H$ to the boundary part $U_{\partial}$. Note that this differs from Hutt who only uses $H_{\partial}$ and calls it $H$. As Hutt observes, $\bar \psi$ and $\psi$ are maps with contrasting properties: $\bar \psi$ is a map over $h$ which is not in general a homeomorphism whereas $\psi$ is a homeomorphism but not in general a map over $h$.[^3] We shall need to exploit these two properties in the following construction and therefore slide between them via the map $H$. The key part of the construction is the following “$S^{k-1}$-branched cover plus $h$-cobordism construction” of [@Cappell-Weinberger(1985) Section 1]. Let $\gamma$ denote the Hopf bundle $S^{2k-1} \to S^{k}$ and for a manifold $X$ let $\gamma_X := {\textup{id}}_X \times \gamma {\colon\!}X \times S^{2k-1} {\rightarrow}X \times S^k$. Further let ${\partial}\psi = \psi\|_{{\partial}W}$, ${\partial}\bar \psi = \bar \psi\|_{{\partial}\bar W}$ and ${\partial}\omega = \omega\|_{{\partial}\bar W}$. Define $$\label{defn:del-W-prime} {\partial}\bar W' := ({\partial}\bar \psi)^(\gamma_M) \cong ({\partial}\omega)^ (\gamma_N), \, \, {\partial}W' := ({\partial}\psi)^(\gamma_M), \, \, U'_{\partial}:= (H_{\partial})^(\gamma_{M \times [0,1]})$$ to be the pull-backs.[^4] Observe that the manifold $U'_{\partial}$ is an $h$-cobordism between ${\partial}\bar W'$ and ${\partial}W'$, since $U_{\partial}$ was an $h$-cobordism between ${\partial}\bar W$ and ${\partial}W$. Further denote $${\partial}\bar \psi' {\colon\!}{\partial}\bar W' {\rightarrow}M \times S^{2k-1}, \quad {\partial}\psi' {\colon\!}{\partial}W' {\rightarrow}M \times S^{2k-1}, \quad H'_{\partial}{\colon\!}U'_{\partial}{\rightarrow}M \times S^{2k-1} \times [0,1]$$ the maps covering ${\partial}\bar \psi$, ${\partial}\psi$, and $H_{\partial}$ respectively. The $S^{k-1}$-bundle projections are denoted by $$q_{\bar W} {\colon\!}{\partial}\bar W' {\rightarrow}{\partial}\bar W, \quad q_{W} {\colon\!}{\partial}W' {\rightarrow}{\partial}W \quad \text{and} \quad q_{U} {\colon\!}U'_{\partial}{\rightarrow}U_{\partial}.$$ By Theorem \[thm:mafs-vs-mcns\] the map $p {\colon\!}{\partial}\bar W \to N$ is a MAF over $N$ with homotopy fibre $S^k$. Obviously $q_{\bar W}$ is also a MAF. As the composition of MAFs is a MAF, the map $$\label{defn:p-prime} p' {\colon\!}{\partial}\bar W' {\xrightarrow}{q_{\bar W}} {\partial}\bar W {\xrightarrow}{p} N$$ is a MAF with homotopy fibre $S^{2k-1}$. Again applying Theorem \[thm:mafs-vs-mcns\] it follows that ${\textup{cyl}}(p')$, the mapping cylinder of $p'$, is a MCN of $N$ with $N$ locally flatly embedded. Denote $$\label{def:N-prime} N' := {\textup{cyl}}(p') \cup U_{\partial}'.$$ The topological manifold $N'$, with boundary ${\partial}W'$, will be the domain of the Cappell-Weinberger map. The $S^{k-1}$-branched cover refers to the projection map ${\textup{cyl}}(p') {\rightarrow}{\textup{cyl}}(p)$ which can be viewed as such, the branching subset being $N \subset {\textup{cyl}}(p)$. Next we define a homotopy equivalence $h' : N' \to M \times D^{2k}$ whose restriction to the boundary $h'\|_{{\partial}W'} : {\partial}W' {\rightarrow}M \times S^{2k-1}$ is a homeomorphism. We regard $M \times D^{2k}$ as the union $(M \times D^{2k}_1)\cup_{M \times S^{2k-1} \times \{ 1 \}} (M \times S^{2k-1} \times [1,2]) $ where $D^{2k}_1$ has radius $1$ and we re-parametrise $H'_{\partial}{\colon\!}U'_{\partial}{\rightarrow}M \times S^{2k-1} \times [1,2]$. Now define $$\label{def:h-prime} h' := {\textup{cyl}}({\partial}\bar \psi',h) \cup H'_{\partial}: N' \to M \times D^{2k}.$$ We remark again that $h'$ is not a map over $h$, since the map ${\partial}\psi$, and hence the maps ${\partial}\psi'$, and $H'_{\partial}$, are not maps over $h$. \[def:CW-map\] The Cappell-Weinberger map is the map $$CW^{k/2} : \mathcal{S}(M) \to \mathcal{S}_{\partial}(M \times D^{2k}), ~~~[h] \longmapsto [h'].$$ The above construction is in fact quite subtle. Note that the map $\bar \psi \cup H_{\partial}$ is $h$-cobordant rel ${\partial}$ to a homeomorphism, hence representing the trivial element in ${\mathcal{S}}_{\partial}(M \times D^{k+1})$. On the other hand, this argument cannot be used for $h'$. The point is that $N'$ cannot be identified with the pull back of $M \times D^{2k} {\rightarrow}M \times D^{k+1}$ along $\bar \psi$. If we wanted to have such an identification we would need $\bar \psi$ to be transverse to $M = M \times \{0\} \in M \times D^{k+1}$. And $\bar \psi$ is in general not transverse to this submanifold (despite the equality $(\bar \psi)^{-1} (M) = N \subset \bar W$). If this were the case, then we would immediately obtain that $h$ is normally cobordant to a homeomorphism, which is not the case in general. The remainder of this subsection is devoted to proving that the structure invariant $[h']$ is independent of all the choices made during its construction from the structure invariant $[h]$. In turn these are: 1. the choice of $h : N \to M$ to represent $[h] \in \mathcal{S}(M)$, 2. the choice of the homotopy equivalence $\omega {\colon\!}(\bar W, {\partial}\bar W) \to (N \times D^{k+1}, N \times S^k)$ representing an element in ${\mathcal{S}}(N \times D^{k+1})$, 3. the choice of the $h$-cobordism $U$, between $\bar W$ and some manifold $W$, with the homotopy equivalence $H {\colon\!}U {\rightarrow}M \times D^{k+1} \times [0,1]$ such that $H\|_{\bar W} = \bar \psi$ and $H\|_{W} = \psi$ which is some homeomorphism. Let $(h,\omega,H)'$ be the structure on $M \times D^{2k}$ produced from a choice of $h$, $\omega$ and $H$. If two homotopy equivalences $h_i {\colon\!}N_i {\rightarrow}M$, $i=0,1$, represent the same element in ${\mathcal{S}}(M)$, then there is an $h$-cobordism $(Z; N_0,N_1)$ and a homotopy equivalence $h_Z {\colon\!}Z {\rightarrow}M \times [0,1]$ with restrictions $h_Z\|_{N_i} = h_i$. The constructions of Lemma \[lem:hutt-1\] can be applied in the relative setting. This means that $h_Z$ can be precomposed with a homotopy equivalence $\omega_Z {\colon\!}\bar X {\rightarrow}Z \times D^{k+1}$ from some $h$-cobordism $(\bar X; \bar W_0, \bar W_1)$, which is also a MCN of $(Z; N_0,N_1)$ and we obtain a homotopy equivalence $$\label{map:bar-Psi} \bar \Psi {\colon\!}\bar X {\rightarrow}M \times D^{k+1} \times [0,1]$$ which restricts to $\bar \Psi\|_{\bar W_i} = \bar \psi_i$ for some homotopy equivalences $\bar \psi_i$ which are $h$-cobordant to homeomorphisms $\psi_i$. The choice of $\omega$ produces a similar outcome. For $i=0,1$, two homotopy equivalences $\omega_{ij} : \bar W_{ij} \to N_i \times D^{k+1}$, for $j=0,1$, as in Lemma \[lem:hutt-1\] represent the same element of ${\mathcal{S}}(N_i \times D^{k+1})$, so again there is an $h$-cobordism $(\bar X_i, \bar W_{i0}, \bar W_{i1})$, which is also a MCN of $(N_i \times [0,1],N_i \times \{0\},N_i \times \{1\})$ and also a homotopy equivalence $\omega_{N_i \times D^{k+1}} {\colon\!}\bar X_i {\rightarrow}N_i \times [0,1] \times D^{k+1}$. Composing with $h_i \times {\textup{id}}{\colon\!}N_i \times [0,1] \times D^{k+1} {\rightarrow}M \times [0,1] \times D^{k+1}$ we obtain $\bar \Psi_i$ with analogous properties as the map $\bar \Psi$ in (\[map:bar-Psi\]). We can glue $\bar X$, $\bar X_0$ and $\bar X_1$ along their common boundary components. There is a corresponding homotopy equivalence to $M \times D^{k+1} \times [0,1]$. This has precisely the same properties as $\bar \Psi$ in (\[map:bar-Psi\]), so we may assume from now on that $\bar \Psi$ represents the difference between choices ($h_0$, $\omega_0$) and ($h_1$, $\omega_1$). Now we come to the choice of $H$. Consider a pair of $h$-cobordisms $U_i$, $i = 0,1$, each from $\bar W_i$ to $W_i$, with homotopy equivalences $H_i$ between $\bar \psi_i$ and homeomorphisms $\psi_i$. We can glue these two $h$-cobordisms onto $\bar X$ along $\bar W_0$ and $\bar W_1$, take a product of the result with $[0,1]$ and rearrange the boundary to obtain a compact manifold $Y$ with boundary. The manifold $Y$ can be seen as an $h$-cobordism either between $h$-cobordisms $U_0$ and $U_1$ or between the $h$-cobordism $\bar X$ and some $h$-cobordism $(X; W_0, W_1)$. We also obtain a homotopy equivalence $$G : (Y;\bar X, X) \to (M \times D^{k+1} \times [0,1] \times [0,1], [0,1] \times \{0\}, [0,1] \times \{1\})$$ restricting to $\bar \Psi$ on the $h$-cobordism $(\bar X; \bar W_0, \bar W_1)$ and to a homotopy equivalence $\Psi : X \to M \times D^{k+1} \times [0,1]$ on $(X; W_0, W_1)$ such that $\Psi\|_{W_i} = \psi_i$, which are homeomorphisms. Further $G\|_{U_i} = H_i$. But we may also view $Y$ as an $h$-cobordism between $U_0 \cup X$ and $U_1$ (just thicken $W_1$ to $W_1 \times [0,1]$ and rearrange the boundary again). The point is that under this change of viewpoint the homotopy equivalence $G$ is a homeomorphism on $W_1 \times [0,1]$. Recall the constructions between Lemma \[lem:hutt-1\] and Definition \[def:CW-map\] which from the homotopy equivalence $H$ of (\[map:H\]) produce the rel ${\partial}$ structure $h'$ on $M \times D^{2k}$ of (\[def:h-prime\]). Using Corollary \[cor:relative-mafs-vs-mcns\] in place of Theorem \[thm:mafs-vs-mcns\] we may now perform the precisely analogous constructions with the homotopy equivalence $G$ to obtain a rel ${\partial}D^{2k} \times [0,1]$ structure on $M \times D^{2k} \times [0,1]$. This structure restricts to $(h_0, \omega_0, H_0 \cup G\|_{X})'$ and $(h_1, \omega_1, H_1)'$ on the respective ends and hence gives an $h$-cobordism rel ${\partial}D^{2k} \times [0,1]$ between these two rel ${\partial}D^{2k}$ structures on $M \times D^{2k}$. (Here the role of $U$ in $(\ref{map:H})$ is played by Y, that of $\bar W$ by $\bar X$, that of $W$ by $W_1 \times [0,1]$ and that of $U_{\partial}$ by the part of the boundary of $Y$ which constitutes the $h$-cobordism between $(U_0)_{\partial}\cup X_{\partial}$ and $(U_1)_{\partial}$ where $X_{\partial}$ is the part of the boundary of $X$ which is the $h$-cobordism between ${\partial}W_0$ and ${\partial}W_1$. Keep in mind that $(\bar X; \bar W_0, \bar W_1)$ is a MCN of $(Z; N_0, N_1)$.) Thus we have proved \[choice1lem\] With $G$ as above $[(h_0, \omega_0, H_0 \cup G\|_{X})'] = [(h_1, \omega_1, H_1)'] \in {\mathcal{S}}_{\partial}(M \times D^{2k})$. It remains to show $[(h_0,\omega_0,H_0)'] = [(h_0, \omega_0, H_0 \cup G\|_{X})']$ which we prove in the following paragraphs. Consider now $G\|_{X} {\colon\!}X {\rightarrow}M \times D^{k+1} \times [0,1]$ which is an h-cobordism between the homeomorphisms $\psi_0$ and $\psi_1$. As such $G\|_{X}$ defines an element of $\mathcal{S}_{{\partial}\{0, 1\}}(M \times D^{k+1} \times [0, 1])$ where the subscript ${\partial}\{0, 1 \}$ indicates that all structure invariants are defined relative to $M \times D^{k+1} \times \{ 0, 1\}$. Using the $S^{k-1}$-branched cover viewpoint we will next show that from $G\|_{X}$ we can obtain a structure from $M \times D^{2k} \times [0, 1]$ relative to $M \times D^{2k} \times \{ 0, 1\}$ which relates $ (h_0,\omega_0,H_0)'$ and $(h_0, \omega_0, H_0 \cup G\|_{X})'$. The Hopf fibration $S^{2k-1} \to S^k$ is given by a free $S^{k-1}$-action on $S^{2k-1}$. If we take the cone of this action we obtain an $S^{k-1}$ action on $D^{2k}$, free except at the centre point, which exhibits $D^{2k}$ as a branched $S^{k-1}$-fibration over $D^{k+1} = D^{2k}/S^{k-1}$. Taking the product with $M$ we have $M \times D^{2k} \to M \times D^{k+1}$ which is a branched $S^{k-1}$-fibration with branch set $M \times \{ 0 \} \subset M \times D^{k+1}$. Now let $f : X \to M \times D^{k+1} \times [0, 1]$ represent $[f] \in \mathcal{S}_{{\partial}\{ 0, 1\}}(M \times D^{k+1} \times [0, 1])$. If we make $f$ transverse to $M \times \{ 0 \} \times [0, 1]$, then we may pull back the branched $S^{k-1}$-fibration $\Gamma : M \times D^{2k} \times [0, 1] \to M \times D^{k+1} \times [0, 1]$ along $f$. [^5] The outcome, $f^(\Gamma)$, is a branched $S^{k-1}$-fibration over $X$ which defines a structure on $M \times D^{2k} \times [0, 1]$ which is relative to $M \times D^{2k} \times \{0, 1\}$. Using transversality along the $h$-cobordisms which define the equivalence relation in $\mathcal{S}_{{\partial}\{0, 1\}}(M \times D^{k+1} \times [0, 1])$ we obtain a well defined map $$\Gamma^ : \mathcal{S}_{{\partial}\{ 0, 1\}}(M \times D^{k+1} \times [0, 1]) \to \mathcal{S}_{{\partial}\{ 0, 1\}}(M \times D^{2k} \times [0, 1]),~~~[f] \mapsto [f^(\Gamma)].$$ We have an obvious map $$R: \mathcal{S}_{{\partial}\{0, 1\}}(M \times D^{2k} \times [0, 1]) \to \mathcal{S}_{{\partial}\{ 0, 1\}}(M \times S^{2k-1} \times [0, 1])$$ and an obvious action $${\rm col} : \mathcal{S}_{{\partial}\{0, 1\}}(M \times S^{2k-1} \times [0,1]) \times \mathcal{S}_{\partial}(M \times D^{2k})\to \mathcal{S}_{\partial}(M \times D^{2k})$$ given, respectively, by restricting to the boundary and by adding a collar. It is straight forward to verify that there is an identity of structures invariants $$\label{eqn:adding-collar} [(h_0,\omega_0, H_0 \cup G\|_{X})'] = {\rm col}((R(\Gamma^([G\|_{X}])), [(h_0, \omega_0, H_0)']) \in \mathcal{S}_{\partial}(M \times D^{2k}).$$ The following two general Lemmas then complete our proof that $CW^{k/2}([h]) = [h']$ is well-defined. \[lem:adding-collar-is-trivial\] Let $M$ be a closed $n$-dimensional manifold and let $l \geq 3$ �be such that $n+l \geq 5$. Then the action $${\rm col} \circ (R \times {\rm Id}) : \mathcal{S}_{{\partial}\{0, 1\}}(M \times D^l \times [0, 1]) \times \mathcal{S}_{\partial}(M \times D^l) \to \mathcal{S}_{\partial}(M \times D^l)$$ is trivial. Let $G : X {\rightarrow}M \times D^l \times [0, 1]$ be a structure representing $[G] \in \mathcal{S}_{{\partial}\{0, 1\}}(M \times D^l \times [0, 1])$. Let $F : N' {\rightarrow}M \times D^l$ represent $[F] \in \mathcal{S}_{\partial}(M \times D^l)$. Consider extending $F$ by ${\rm col}(R(G))$, which can be conveniently denoted as $R(G) \cup F$. We now have two structures, $F$ and $R(G) \cup F$ on $M \times D^l$ and we need to show that they represent the same element of $\mathcal{S}_{\partial}(M \times D^l)$: that is, we need to find an $h$-cobordism between them. One way to think about $G$ is as of a homotopy between two homeomorphisms $G_0$ and $G_1$, where $G_i {\colon\!}X_i {\rightarrow}M \times D^l \times \{i\}$ are the appropriate restrictions. But we can also think of it as of an rel boundary $h$-cobordism between $G_1 \cup R(G)$ and $G_0$. This shows that $G_1 \cup R(G)$, when thought of as an element of the structure set of $M \times D^l$ is trivial, since $G_0$ is a homeomorphism. Now both structures $F$ and $R(G) \cup F$ on $M \times D^l$ can be extended by a homeomorphism to structures on $M \times S^l$, namely $G_1 \cup R(G) \cup F$ and $G_0 \cup F$. But we can now glue $G$ considered as an $h$-cobordism as above with the trivial $h$-cobordism $F \times {\textup{id}}$ to obtain an $s$-cobordism between the two structures $G_1 \cup R(G) \cup F$ and $G_0 \cup F$ on $M \times S^l$. This means that they represent the same element in the structure set of $M \times S^l$. By Lemma \[extension-by-homeo-is-injective\] below the structures $R(G) \cup F$ and $F$ represent the same element of $\mathcal{S}_{\partial}(M \times D^l)$. For the statement of the next lemma, recall that if $P$ if a closed manifold with a decomposition $P = Q \cup C$, where $Q$ and $C$ are codimension $0$ submanifolds. Then there is a well defined a map $E: {\mathcal{S}}_{\partial}(Q) {\rightarrow}{\mathcal{S}}(Y)$ given by extension with a homeomorphism. \[extension-by-homeo-is-injective\] Let $M$ be a closed $n$-dimensional manifold and let $l \geq 3$ with $n+l \geq 5$. Then the extension by a homeomorphism map $${\mathcal{S}}_{\partial}(M \times D^l) {\rightarrow}{\mathcal{S}}(M \times S^l)$$ is injective. Consider the decomposition $$S^l = D_1^l \cup_{S^{l-1}} D_2^l.$$ Taking the product with $M$ we obtain a pushout decomposition of $M \times S^l$ which is a $(\pi-\pi)$-decomposition in the sense of [@Wall(1999) Chapter 4]. Consider the structure set $$\label{split-structure-set} {\mathcal{S}}_\pitchfork (M \times S^l) \subset {\mathcal{S}}(M \times S^l)$$ consisting of homotopy equivalences $h {\colon\!}M {\rightarrow}M \times S^l$ which are split along $M \times S^{l-1}$, which means that they are transverse to $M \times S^{l-1} \subset M \times S^l$ and that the restrictions $h_i {\colon\!}M_i = h^{-1} (M \times D_i^l) {\rightarrow}M \times D_i^l$ are homotopy equivalences of pairs for $i = 1,2$. By [@Wall(1999) Chapter 11] the inclusion (\[split-structure-set\]) is a bijection. We have the following sequence of maps $$\label{exact-sequence-for-splittings} {\mathcal{S}}_{\partial}(M \times D_1^l) {\rightarrow}{\mathcal{S}}_\pitchfork (M \times S^l) {\rightarrow}{\mathcal{S}}(M \times D_2^l)$$ where the second map is given by restriction and the first map is the extension by a homeomorphism. The second map is obviously surjective since we can “double” any structure $h {\colon\!}(M,{\partial}) {\rightarrow}(M \times D^l,{\partial})$ to a structure $h \cup h {\colon\!}M \cup M {\rightarrow}M \times S^l$. The composition of the two maps is clearly trivial. Now the sequence (\[exact-sequence-for-splittings\]) can be seen as the $\pi_0$-part of the long exact sequence of a fibration. Recall from [@Wall(1999) Chapter 17A] that there is a space ${\widetilde{{\mathcal{S}}}}(X)$ such that $\pi_k ( {\widetilde{{\mathcal{S}}}}(X) ) \cong {\mathcal{S}}_{\partial}(X \times D^k)$ for any closed topological manifold $X$ of dimension $n \geq 5$. There are also obvious generalizations when $X$ is a compact manifold with boundary. At this level we also have the sequence $$\label{exact-sequence-for-splittings-spacified} {\widetilde{{\mathcal{S}}}}_{\partial}(M \times D_1^l) {\rightarrow}{\widetilde{{\mathcal{S}}}}_\pitchfork (M \times S^l) {\rightarrow}{\widetilde{{\mathcal{S}}}}(M \times D_2^l)$$ which is in fact a fibration sequence and (\[exact-sequence-for-splittings\]) is the $\pi_0$-part of the corresponding long exact sequence of the homotopy groups. The argument about the surjectivity of the second map in (\[exact-sequence-for-splittings\]) generalises to higher homotopy groups and hence the long exact sequence splits into short exact sequences and hence the inclusion of the fiber induces an injective map on the homotopy groups, in particular on $\pi_0$. Equation (\[eqn:adding-collar\]) and Lemma \[lem:adding-collar-is-trivial\] yield: \[choice2cor\] There is an equality $[(h_0, \omega_0, H_0)'] = [(h_1, \omega_0, H_0 \cup G\|_{X})']$. Lemma \[choice1lem\] and Corollary \[choice2cor\] show that the map $CW^{k/2}$ is well defined. The discussion above includes, in particular, a proof of [@Hutt(1998) Lemma 1.1]: for a fixed choice of $\omega$, a choice of homotopy $H$ does not effect the equivalence class of $(h,\omega,H)' : N' \to M \times D^{2k}$ in $\mathcal{S}_{\partial}(M \times D^{2k})$. The proof of this in the appendix of [@Hutt(1998)] appears to us to be incorrect. Extension by homeomorphism {#subsec:ext-by-homeo} -------------------------- Recall that $M$ is a closed topological manifold, that ${\mathbb{F}}= {\mathbb{C}}$ or ${\mathbb{H}}$ has real dimension $k = 2$ or $4$ and consider the obvious decomposition $$M \times {\mathbb{F}}P^2 = (M \times D^{2k}) \cup (M \times {\mathbb{F}}P^{2\bullet})$$ where ${\mathbb{F}}P^{2\bullet} := {\mathbb{F}}P^2 - D^{2k}$. Associated to this decomposition extension by homeomorphism gives the map of structure sets $$E : \mathcal{S}_{\partial}(M \times D^{2k}) \to \mathcal{S}(M \times {\mathbb{F}}P^2).$$ For the structures $h' {\colon\!}N' \to M \times D^{2k}$ defined in (\[def:h-prime\]) above, we can realize the map $E$ as follows. Observe that ${\mathbb{F}}P^{2\bullet}$ is homeomorphic to the mapping cylinder (disk bundle) of the Hopf map $\gamma {\colon\!}S^{2k-1} {\rightarrow}S^k$ and recall that the restriction of the map $h'$ to ${\partial}N' = {\partial}W'$ is given by the $S^{k-1}$-bundle map ${\partial}\psi' {\colon\!}{\partial}W' {\rightarrow}M \times S^{2k-1}$ over ${\partial}\psi {\colon\!}{\partial}W {\rightarrow}M \times S^{k}$. Thus we can simply extend ${\partial}\psi'$ to the associated $D^k$-bundle map. This amounts to extending to the associated mapping cylinders. To be explicit, recall that $q_W {\colon\!}{\partial}W' {\rightarrow}{\partial}W$ is the projection in the source and define $$\label{defn:N-hat} \widehat N {\colon\!}= N' \cup_{{\partial}W'} {\textup{cyl}}(q_W) = {\textup{cyl}}(p') \cup U'_{\partial}\cup_{{\partial}W'} {\textup{cyl}}(q_W).$$ If follows that a representative of $E[h']$ given by $$\label{defn:h-hat} \widehat h := h' \cup {\textup{cyl}}({\partial}\psi',{\partial}\psi) {\colon\!}\widehat N \to M \times {\mathbb{F}}P^2$$ where ${\textup{cyl}}({\partial}\psi',{\partial}\psi)$ is the mapping cylinder ($D^{k}$-bundle) map associated to the square ${\partial}\psi \circ q = \gamma \circ {\partial}\psi'$. As with the map $h'$ we note that the map $\widehat h : \widehat N \to M \times {\mathbb{F}}P^2$ is not a map over $h : N \to M$ since ${\partial}\psi$ is not a map over $h$. On the other hand, we can replace the structure $\widehat h$ with a structure that is over $h$. In fact one has the following h-decorated version of a key lemma of Hutt. [Cf. [@Hutt(1998) Lemma 1.4].]{} \[lem:phi\] There is a closed manifold $\bar N$ with a map $\bar p {\colon\!}\bar N {\rightarrow}N$ and a homotopy equivalence of closed manifolds $\bar h : \bar N \to M \times {\mathbb{F}}P^2$ covering the map $h$ which represents the same element as $\widehat h$ in ${\mathcal{S}}(M \times {\mathbb{F}}P^2)$. Indeed there is a homotopy equivalence $\varphi : \bar N \to N \times {\mathbb{F}}P^2$ over $N$ such that the following diagram commutes. $$\begin{diagram} \node{\bar N} \arrow{se,t}{\bar h} \arrow[2]{s,l}{\varphi} \\ \node[2]{M \times {\mathbb{F}}P^2} \\ \node{N \times {\mathbb{F}}P^2} \arrow{ne,b}{h \times {\rm Id}} \end{diagram}$$ Recall the projection map $q_{\bar W} {\colon\!}{\partial}\bar W' {\rightarrow}{\partial}\bar W$ and also recall that the mapping cylinder ${\textup{cyl}}(p')$ has boundary ${\partial}\bar W'$. The manifold $\bar N$ is defined as $$\label{defn:N-bar} \bar N := {\textup{cyl}}(p') \cup {\textup{cyl}}(q_{\bar W})$$ and we have the obvious projection map $\bar p {\colon\!}\bar N {\rightarrow}N$. Recall the homotopy equivalence $\omega : \bar W {\rightarrow}N \times D^{k+1}$ over the identity on $N$ from Lemma \[lem:hutt-1\]. Define $\varphi : \bar N \simeq N \times {\mathbb{F}}P^2$ by $$\label{defn:phi} \varphi = {\textup{cyl}}({\partial}\omega',{\textup{id}}) \cup {\textup{cyl}}({\partial}\omega',{\partial}\omega),$$ and $\bar h = (h \times {\rm Id}) \circ \varphi : \bar N \simeq M \times {\mathbb{F}}P^2$. We need to show that $\bar h$ is equivalent to $\widehat h$ and we achieve this with a sort of Alexander trick. We first find an $h$-cobordism $Z$ between the manifolds $\widehat N$ and $\bar N$. To this end think of $\bar N$ as $$\bar N = {\textup{cyl}}(p') \cup ({\partial}\bar W \times [0,1]) \cup {\textup{cyl}}(q_{\bar W}).$$ Now the manifold $Z_3 := {\textup{cyl}}(q_{U_{\partial}})$ yields an $h$-cobordism between ${\textup{cyl}}(q_W)$ and ${\textup{cyl}}(q_{\bar W})$. The product $U'_{\partial}\times [0,1]$ can be viewed as an $h$-cobordism $Z_2$ between $U'_{\partial}$ and ${\partial}\bar W' \times [0,1]$. Let $Z_1$ be the trivial product $h$-cobordism over ${\textup{cyl}}(p')$. Gluing all these together gives the desired global $h$-cobordism. It remains to produce a homotopy equivalence from $Z$ to $M \times {\mathbb{F}}P^2 \times [0,1]$ which restricts to $\widehat h$ and $\bar h$ on the two ends. On $Z_1$ we just take the product of ${\textup{cyl}}({\partial}\bar \psi',{\partial}\bar \psi)$ with the identity. On $Z_2$ we take the product of $H'_{\partial}$ with the identity, but this is modified according to the way we think of $Z_2$. On $Z_3$ we can take the map ${\textup{cyl}}(H'_{\partial},H_{\partial})$. All these maps agree on the boundary and provide the required homotopy equivalence. Notice that the only role played by $h$ in the construction of the map $\varphi$ is to define $\chi \in [N, {\textup{G}}/{\textup{TOP}}]$. Specifically $\varphi$ is determined by ${\partial}\omega$ and ${\partial}\omega$ is determined by $\chi \in [N ;{\textup{G}}/{\textup{TOP}}]$. Moreover the class $\chi$ is all one needs to build the manifolds $N'$, $\widehat N$ and $\bar N$. This observation will be useful in the last Section \[sec:completion\] where coboundaries for $\widehat N$ and $\bar N$ will be constructed. The algebraic theory of surgery {#sec:alg-sur} =============================== We give a brief review of the algebraic theory of surgery. In particular we review how algebraic surgery equips ${\mathcal{S}}(M)$ with an abelian group structure by identifying it with the algebraic structure set ${\mathbb{S}}_{n+1} (M)$. In more detail, the aim of this section is to define, for a closed $n$-dimensional topological manifold, the abelian algebraic structure group ${\mathbb{S}}_{n+1} (M)$ and the map $s {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (M)$ which turns out to be a bijection if $n \geq 5$. Hence one can equip ${\mathcal{S}}(M)$ with an abelian group structure via this bijection. We will also discuss a generalization when $M$ has a boundary. Furthermore we will discuss the functoriality of ${\mathbb{S}}_{n+1} (M)$ in $M$ and we will present a condition which implies that an element in ${\mathbb{S}}_{n+1} (M)$ is zero. All these results will be used in subsequent sections. The principal references are [@Ranicki(1992)], [@Ranicki-I-(1980)], [@Ranicki-II-(1980)], and [@Ranicki-Weiss(2008)]. The abelian group ${\mathbb{S}}_{n+1} (M)$ is defined as a quadratic $L$-group of a certain algebraic bordism category. An [algebraic bordism category]{} $\Lambda = ({\mathbb{A}},{\mathbb{B}},{\mathbb{C}})$ consists of an [additive category with chain duality]{} $({\mathbb{A}},(T,e))$ and two full subcategories ${\mathbb{C}}$, ${\mathbb{B}}\subseteq {\mathbb{B}}({\mathbb{A}})$ of the category of bounded chain complexes in ${\mathbb{A}}$ satisfying certain mild assumptions. The [chain duality]{} $(T,e)$ consists of a contravariant functor $T {\colon\!}{\mathbb{B}}({\mathbb{A}}) {\rightarrow}{\mathbb{B}}({\mathbb{A}})$ and a natural transformation $e {\colon\!}T^2 {\rightarrow}{\textup{id}}$ satisfying certain conditions. It allows one to define the tensor product $C \otimes_{\mathbb{A}}D$ of chain complexes $C,D \in {\mathbb{B}}({\mathbb{A}})$ such that the tensor product $C \otimes_{\mathbb{A}}C$ becomes a chain complex of ${\mathbb{Z}}[{\mathbb{Z}}_2]$-modules. One defines an $n$-dimensional [symmetric]{} structure on $C \in {\mathbb{B}}({\mathbb{A}})$ to be an $n$-cycle $\phi$ in the chain complex $W^{\%} (C) = {\textup{Hom}}_{{\mathbb{Z}}[{\mathbb{Z}}_2]}(W,(C \otimes_{\mathbb{A}}C))$, where $W$ is the standard ${\mathbb{Z}}[{\mathbb{Z}}_2]$-resolution of ${\mathbb{Z}}$. An $n$-dimensional [quadratic]{} structure on $C \in {\mathbb{B}}({\mathbb{A}})$ is defined as an $n$-cycle $\psi$ in the chain complex $W_{\%} (C) = W \otimes_{{\mathbb{Z}}[{\mathbb{Z}}_2]} (C \otimes_{\mathbb{A}}C)$. An $n$-dimensional symmetric structure consists of a collection of chains $\phi = \{ \phi_s \in (C \otimes_{\mathbb{A}}C)_{n+s} \; \| \; s \geq 0 \}$ satisfying certain compatibility connections. An $n$-dimensional quadratic structure consists of a collection of chains $\psi = \{ \psi_s \in (C \otimes_{\mathbb{A}}C)_{n-s} \; \| \; s \geq 0 \}$ satisfying certain compatibility connections. There is also a symmetrization map $(1+T) {\colon\!}W_{\%} (C) {\rightarrow}W^{\%} (C)$. The pair $(C,\phi)$ is called a [symmetric algebraic complex]{}, it is ${\mathbb{C}}$-Poincaré if the mapping cone of $\phi_0$, ${\mathcal{C}}(\phi_0 {\colon\!}\Sigma^n TC {\rightarrow}C)$, lies in ${\mathbb{C}}$. The pair $(C,\psi)$ is called a [quadratic algebraic complex]{}, it is ${\mathbb{C}}$-Poincaré if $(C,(1+T)\psi)$ is ${\mathbb{C}}$-Poincaré. In the above formula recall that $$\label{formula:tensor-over-AA} (C \otimes_{\mathbb{A}}C)_n := {\textup{Hom}}_{\mathbb{A}}(TC,C)_n = {\textup{Hom}}_{\mathbb{A}}(\Sigma^n TC,,C)_0.$$ All these notions are defined in [@Ranicki(1992) Chapters 1,3]. $L$-groups {#l-groups .unnumbered} ---------- There is a well-defined notion of a cobordism of $n$-dimensional quadratic algebraic complexes. The quadratic $L$-groups $L_n (\Lambda)$ are the cobordism group of $n$-dimensional algebraic complexes in $\Lambda$, that means elements are represented by those complexes which are in ${\mathbb{B}}\subset {\mathbb{B}}({\mathbb{A}})$ which are ${\mathbb{C}}$-Poincaré. If ${\mathbb{B}}$ and ${\mathbb{C}}$ are not explicitly stated, we use the convention that ${\mathbb{B}}= {\mathbb{B}}({\mathbb{A}})$ and ${\mathbb{C}}= 0$. See [@Ranicki(1992) Definitions 1.8, 3.4] Example ${\mathbb{A}}[\pi_1 (M)]$ {#example-mathbbapi_1-m .unnumbered} --------------------------------- For any ring with involution $R$, for example for ${\mathbb{Z}}[\pi_1 (M)]$, the category of finitely generated free based $R$-modules ${\mathbb{A}}[R]$ has a chain duality given by $T(M) = {\textup{Hom}}_R (M,R)$. Examples \[expl:sym-con\], \[expl:quad-con\] below explain how to obtain symmetric and quadratic algebraic complexes over the category ${\mathbb{Z}}[\pi_1 (M)]$. The quadratic $L$-groups $L_n ({\mathbb{Z}}[\pi_1 (M)])$ defined as cobordism groups of quadratic algebraic Poincaré complexes agree with the usual Wall surgery $L$-groups defined using quadratic forms or formations [@Ranicki-II-(1980)]. Example ${\mathbb{A}}_\ast (K)$ {#example-mathbba_ast-k .unnumbered} ------------------------------- Let $K$ be a simplicial complex, or more generally a $\Delta$-set, and let ${\mathbb{A}}$ be an additive category with chain duality. The category ${\mathbb{A}}_\ast (K)$ has as its objects the so-called $K$-based objects of ${\mathbb{A}}$, that means objects of ${\mathbb{A}}$ which come as direct sums $$M = \sum_{\sigma \in K} M(\sigma).$$ Morphisms are given by $${\textup{Mor}}_{{\mathbb{A}}_\ast (K)} (M,N) = \{ f = \sum_{\sigma, \tau \in K} f(\tau,\sigma) {\colon\!}M(\sigma) {\rightarrow}N(\tau) \; \| \; f(\tau,\sigma) = 0 \; \textup{unless} \; \sigma \leq \tau \}.$$ The definition of the duality is stated in [@Ranicki(1992) Proposition 5.1], on the objects $M \in {\mathbb{A}}$ it is given by $$(TM)_r (\sigma) = T(M(\sigma))_{r+\|\sigma\|} \quad \textup{if} \quad \sigma \leq \tau, \|\sigma\| = \|\tau\|-1.$$ This formula will not be used in the present paper. What is more important for us is the observation that an $n$-dimensional quadratic algebraic complex $(C,\psi)$ in ${\mathbb{A}}_\ast (K)$ includes in particular for each $\sigma \in K$ a chain complex $C(\sigma)$ and a duality map $\psi_0 (\sigma) {\colon\!}\Sigma^n TC (\sigma) {\rightarrow}C(\sigma)$ (recall (\[formula:tensor-over-AA\])). But it contains more information, there are relations between these data for various simplices and of course the components $\psi_s$ for $s > 0$. See [@Ranicki(1992) Definition 4.1, Proposition 5.1]. Also Examples \[expl:sym-con\], \[expl:frag-quad-con\] below explain how such complexes come from geometry. Functoriality {#functoriality .unnumbered} ------------- Let $\pi {\colon\!}K {\rightarrow}L$ be a $\Delta$-set map. Then we have an additive functor $$\label{functor-add-cat} \pi_\ast {\colon\!}{\mathbb{A}}_\ast (K) {\rightarrow}{\mathbb{A}}_\ast (L) \quad (\pi_\ast M)(\tau) = \sum_{\sigma \in K, \pi(\sigma)=\tau} M(\sigma) \quad \textup{for} \; \tau \in L$$ which induces a functor on the chain complexes which also ‘commutes’ with the chain duality in a suitable sense so that one obtains a map of the $L$-groups $$\pi_\ast {\colon\!}L_{n+1} ({\mathbb{A}}_\ast (K)) {\rightarrow}L_{n+1} ({\mathbb{A}}_\ast (L)).$$ See [@Ranicki(1992) Proposition 5.6, Example 5.8]. Assembly {#assembly .unnumbered} -------- Slightly different functoriality is provided by the [assembly]{} functor $A {\colon\!}{\mathbb{Z}}_\ast (K) {\rightarrow}{\mathbb{Z}}[\pi_1 (K)]$ defined by $$M \mapsto \sum_{\widetilde \sigma \in \widetilde K} = M(p(\widetilde \sigma))$$ which also induces a map of the $L$-groups $$\pi_\ast {\colon\!}L_n ({\mathbb{Z}}_\ast (K)) \cong H_n (K,{\mathbf{L}}_\bullet) {\rightarrow}L_n ({\mathbb{Z}}[\pi_1 (K)]).$$ Here the isomorphism in the source is a calculation, the symbol ${\mathbf{L}}_\bullet$ denotes the quadratic $\langle 0 \rangle$-connective $L$-theory spectrum. See [@Ranicki(1992) Chapter 9] and [@Ranicki(1992) Chapter 13] for the spectrum ${\mathbf{L}}_\bullet$. Algebraic bordism categories {#algebraic-bordism-categories .unnumbered} ---------------------------- So far we have only presented examples of additive categories with chain duality. In order to obtain an algebraic bordism category we need to specify interesting subcategories of ${\mathbb{B}}({\mathbb{A}})$. We will only be interested in the case ${\mathbb{A}}= {\mathbb{A}}_\ast (K)$ for which we will use three such subcategories, denoted ${\mathbb{D}}\subset {\mathbb{C}}\subset {\mathbb{B}}$. Here are the definitions: $$\begin{aligned} \label{alg-bordism-cat} \begin{split} {\mathbb{B}}& := {\mathbb{B}}({\mathbb{A}}_\ast (K)), \\ {\mathbb{C}}& := \{ C \in {\mathbb{B}}\; \| \; A(C) \simeq \ast \}, \\ {\mathbb{D}}&: = \{ C \in {\mathbb{B}}\; \| \; C(\sigma) \simeq \ast \; \; \forall \sigma \in K \}. \end{split}\end{aligned}$$ This gives us three possibilities to construct interesting algebraic bordism categories. Using a suitable notion of a functor between algebraic bordism categories we obtain a sequence $$\label{loc-seq-of-alg-bor-cat} \Lambda'' = ({\mathbb{A}},{\mathbb{C}},{\mathbb{D}}) {\longrightarrow}\Lambda' = ({\mathbb{A}},{\mathbb{B}},{\mathbb{D}}) {\longrightarrow}\Lambda = ({\mathbb{A}},{\mathbb{B}},{\mathbb{C}})$$ which induces a long exact sequence of groups $$\label{loc-seq-alg-sur} \cdots {\rightarrow}L_{n+1} (\Lambda) {\rightarrow}L_n (\Lambda'') {\rightarrow}L_n (\Lambda') {\rightarrow}L_n (\Lambda) {\rightarrow}L_{n-1} (\Lambda'') {\rightarrow}\cdots .$$ This material is discussed in [@Ranicki(1992) Chapter 3]. Connective versions {#connective-versions .unnumbered} ------------------- In order to obtain groups which relate well with geometry we need to use a connective version of the above theory. Let $q \in {\mathbb{Z}}$ and let $\Lambda = ({\mathbb{A}},{\mathbb{B}},{\mathbb{C}})$ be an algebraic bordism category. Define the subcategory ${\mathbb{B}}\langle q \rangle \subset {\mathbb{B}}$ to be the subcategory of chain complexes in ${\mathbb{B}}$ which are homotopy equivalent to $q$-connected chain complexes. Further define ${\mathbb{C}}\langle q \rangle = {\mathbb{B}}\langle q \rangle \cap {\mathbb{C}}$. Then $\Lambda \langle q \rangle = ({\mathbb{A}},{\mathbb{B}}\langle q \rangle,{\mathbb{C}}\langle q \rangle)$ is a new algebraic bordism category. More details are given in [@Ranicki(1992) Chapter 15]. [[@Ranicki(1992) Chapter 17]]{} \[def:alg-str-set\] Let $K$ be a $\Delta$-set, $n \in {\mathbb{N}}$ and let $\Lambda''$ be the algebraic bordism category given by (\[loc-seq-of-alg-bor-cat\]). Define $${\mathbb{S}}_{n+1} (K) := L_n (\Lambda''_\ast \langle 1 \rangle).$$ Algebraic surgery exact sequence {#algebraic-surgery-exact-sequence .unnumbered} -------------------------------- Putting together the previous statements and definitions the long exact sequence (\[loc-seq-alg-sur\]) becomes the algebraic surgery exact sequence $$\label{alg-sur-seq-concrete} \cdots {\rightarrow}L_{n+1} ({\mathbb{Z}}[\pi_1 (K)]) {\rightarrow}{\mathbb{S}}_{n+1} (K) {\rightarrow}H_n (K,{\mathbf{L}}_\bullet \langle 1 \rangle) {\rightarrow}L_n ({\mathbb{Z}}[\pi_1 (K)]) {\rightarrow}\cdots$$ discussed in detail in [@Ranicki(1992) Chapters 14, 15]. We will mostly work directly with the group ${\mathbb{S}}_{n+1} (K)$, but of course many of the properties of this group follow from the existence of the sequence (\[alg-sur-seq-concrete\]). For example recall that the assignment $K \mapsto {\mathbb{S}}_{n+1}(K)$ becomes a covariant functor from $\Delta$-sets to abelian groups via the functoriality described in (\[functor-add-cat\]). So for $\pi {\colon\!}K {\rightarrow}L$ a $\Delta$-set map we obtain the map $$\pi_\ast {\colon\!}{\mathbb{S}}_{n+1} (K) {\rightarrow}{\mathbb{S}}_{n+1} (L).$$ In fact we obtain a map of exact sequences for $K$ and $L$ and it follows that the functor ${\mathbb{S}}_{n+1} (K)$ is a homotopy functor. In particular this means that if $\pi$ is a homotopy equivalence of $\Delta$-sets, then $\pi_\ast$ is an isomorphism of abelian groups, since the other two terms clearly are homotopy functors. Next we review how the above theory relates to topology. We begin with some remarks about topological manifolds in the above setting. When $M$ is a closed $n$-dimensional topological manifold we can apply Definition \[def:alg-str-set\] only if $M$ is triangulated. In that case it is possible to define a map $s {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (M)$, whose construction we recall below and which can be shown to be a bijection, [@Ranicki(1992) Chapter 18]. If $M$ is not triangulated we can choose a homotopy equivalence $r {\colon\!}M {\rightarrow}K$ to a finite $\Delta$-set. Such an $r$ will determine a map $s(r) {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (K)$ which can be shown to be a bijection in the same way. In both cases the bijections $s$ and $s(r)$ can be used to give ${\mathcal{S}}(M)$ the structure of an abelian group. It also turns out that this group structure is independent of both the choice of the triangulation and of the homotopy equivalence $r$. Therefore, following Ranicki, we will abuse notation and write ${\mathbb{S}}_{n+1} (M)$ for ${\mathbb{S}}_{n+1} (K)$ for any choice of a $\Delta$-set $K$ homotopy equivalent to $M$ and $s {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (M)$ for $s(r)$ given by any choice of a homotopy equivalence $r {\colon\!}M {\rightarrow}K$. For the record we make Let $M$ be an $n$-dimensional topological manifold and let $r {\colon\!}M {\rightarrow}K$ be a homotopy equivalence to a finite $\Delta$-set. We write $${\mathbb{S}}_{n+1} (M) = {\mathbb{S}}_{n+1} (K).$$ Now we explain the map $s {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (M)$. First we need some preparation. In the following examples we explain the topological situations which give rise to algebraic complexes, both symmetric and quadratic. \[expl:sym-con\] Let $K$ be a finite $\Delta$-set with barycentric subdivision $K'$. Consider the simplicial chain complex $C := \Delta_\ast (K')$ as an object in ${\mathbb{B}}({\mathbb{Z}})$. Any $n$-cycle $[K]$ in $C_n$ defines via the symmetric construction a symmetric structure $\phi$ on $C$ over ${\mathbb{Z}}$, whose component $\phi_0 {\colon\!}\Sigma^n TC = C^{n-\ast} {\rightarrow}C_\ast$ corresponds to the cap product with $[K]$. If $K$ is a closed oriented $n$-dimensional topological manifold and $[K]$ is its fundamental class, then the resulting symmetric Poincaré complex is denoted $\sigma^\ast (K)$ and is called the [symmetric signature]{} of $K$. See [@Ranicki-II-(1980)] The chain complex $C$ can also be thought of as an object in ${\mathbb{B}}({\mathbb{Z}}_\ast (K))$ with $C(\sigma) = \Delta_\ast (D(\sigma),\partial D(\sigma))$, the simplicial chain complex of the dual cell relative to its boundary. Then we have $\Sigma^n TC (\sigma) \cong \Delta^{n-\|\sigma\|-\ast} (D(\sigma))$. Again by [@Ranicki(1992)] for each $n$-cycle $[K] \in C_n$ there is a refined symmetric construction. Thus we obtain an algebraic symmetric structure $\phi$ over ${\mathbb{Z}}_\ast (K)$, which in particular contains for each $\sigma$ duality maps $\phi_0 (\sigma) {\colon\!}\Sigma^n TC (\sigma) {\rightarrow}C(\sigma)$ which are cap products with certain $(n-\|\sigma\|)$-dimensional classes $[K](\sigma)$. See [@Ranicki(1992) Example 5.5]. Now let $M$ be a topological manifold with a reference homotopy equivalence $r {\colon\!}M {\rightarrow}K$ to a $\Delta$-set, which is transverse to the dual cells of $K'$. Consider the dissection $$\label{dissection-of-M} M = \bigcup_{\sigma \in K} \Big( M(\sigma) = r^{-1} (D(\sigma)). \Big)$$ The chain complex $$\label{dissection-of-chian-cplx} C = \Sigma_{\sigma \in K} \Big( C(\sigma) = C (M(\sigma),{\partial}M(\sigma)) \Big)$$ where $C (M(\sigma),{\partial}M(\sigma))$ is the singular chain complex of the pair $(M(\sigma),{\partial}M(\sigma))$, yields an object in ${\mathbb{B}}({\mathbb{Z}}_\ast (K))$. As an object in ${\mathbb{B}}({\mathbb{Z}})$ it is weakly homotopy equivalent to $C_\ast (M)$, the singular chain complex of $M$. When considered as an object in ${\mathbb{B}}({\mathbb{Z}}_\ast (K))$, then, similarly to above, there is for each $n$-cycle $[M] \in C_n$ a refined symmetric construction, so that we obtain an algebraic symmetric structure $\phi$ over ${\mathbb{Z}}_\ast (K)$, which in particular contains for each $\sigma$ duality maps $\phi_0 (\sigma) {\colon\!}\Sigma^n TC (\sigma) {\rightarrow}C(\sigma)$ which are cap products with certain $(n-\|\sigma\|)$-dimensional classes $[M](\sigma)$. For more details see [@Ranicki(1992) Example 6.2]. \[expl:quad-con\] Let $(f,b) \colon N \rightarrow M$ be a degree one normal map of $n$-dimensional closed manifolds. Denote by $K(f)$ the algebraic mapping cone of the Umkehr map of chain complexes $$f^! \colon C_\ast\widetilde{M} \simeq C^{n - \ast} \widetilde{M} {\xrightarrow}{f^{n -\ast}} C^{n - \ast} \widetilde{N} \simeq C_\ast \widetilde{N}.$$ Then $C_\widetilde M$ comes with a structure of an $n$-dimensional symmetric algebraic Poincaré complex over ${\mathbb{Z}}[\pi_1 (M)]$. This projects to a structure of an $n$-dimensional symmetric algebraic Poincaré complex on $K(f)$. In [@Ranicki-II-(1980)] this is refined to an $n$-dimensional quadratic algebraic Poincaré complex on $(K(f),\psi(f))$. \[expl:frag-quad-con\] Let $(f,b) \colon N \rightarrow M$ be a degree one normal map of closed $n$-dimensional manifolds and let $r\colon M \rightarrow K$ be a map to a $\Delta$-set $K$ such that both $rf$ and $r$ are transverse to the dual cells of $K$. There are $K$-dissections $N \cong \cup N(\sigma)$ and $M \cong \cup M(\sigma)$, so that $C_N$ and $C_M$ can be regarded as objects in ${\mathbb{B}}({\mathbb{Z}}_\ast(K))$. There are preferred structures of $n$–dimensional symmetric algebraic complexes on $C_N$ and $C_M$, as objects of ${\mathbb{B}}({\mathbb{Z}}_\ast(K))$ coming from the fundamental classes. By analogy with Example \[expl:quad-con\], there is an algebraic Umkehr map $$f^!{\colon\!}C_M {\longrightarrow}C_N$$ in ${\mathbb{B}}({\mathbb{Z}}_\ast(K))$ with mapping cone $K(f)$, say. The resulting structure of and $n$–dimensional symmetric algebraic complex on $K(f)$, as an object of ${\mathbb{B}}({\mathbb{Z}}_\ast(K))$, has a preferred refinement to a quadratic structure $\psi(f)$. The chain complex $K(f)(\sigma)$ for a $\sigma \in K$ can be identified with the mapping cone of an algebraic Umkehr map $$C_(M(\sigma),\partial M(\sigma)) {\longrightarrow}C_(N(\sigma),\partial N(\sigma)).$$ See [@Ranicki(1992) Examples 9.13, 9.14] for details. Under assembly, this construction coincides with that in Example \[expl:quad-con\]. \[expl:frag-quad-con-over-htpy-equivalence\] Note that $(K(f),\psi(f))$ is ${\mathbb{D}}$-Poincaré. When in addition $f$ is a homotopy equivalence, then $K(f)$ is contractible after assembly. Furthermore the required connectivity assumptions are fulfilled so that the pair $(K(f),\psi(f))$ represents an element in ${\mathbb{S}}_{n+1} (M)$. [[@Ranicki(1992) Proposition 18.3]]{} The map $$s {\colon\!}{\mathcal{S}}(M) {\rightarrow}{\mathbb{S}}_{n+1} (M), ~~~~~~[f: N \to M] \mapsto [K(f), \psi(f)]$$ is defined by the construction described in Examples \[expl:frag-quad-con\], \[expl:frag-quad-con-over-htpy-equivalence\]. [[@Ranicki(1992) Proposition 18.3]]{} In case we deal with a manifold with boundary $(Y,\partial Y)$, the constructions in Examples \[expl:frag-quad-con\], \[expl:frag-quad-con-over-htpy-equivalence\] yield a map from ${\mathcal{S}}_\partial (Y)$ to ${\mathbb{S}}_{n+1} (Y)$. When $Y = M \times D^k$, then thanks to the homotopy invariance of ${\mathbb{S}}_{\ast} (-)$ we obtain a map $$s {\colon\!}{\mathcal{S}}_\partial (M \times D^k) {\rightarrow}{\mathcal{S}}_{n+1+k} (M).$$ [[@Ranicki(1992) Theorem 18.5]]{} \[thm:identifcation-geo-alg-str-sets\] For a closed manifold $M$ with dimension $n \geq 5$ we have $$s {\colon\!}{\mathcal{S}}(M) {\xrightarrow}{\equiv} {\mathbb{S}}_{n+1} (M), \quad \quad s {\colon\!}{\mathcal{S}}_\partial (M \times D^k) {\xrightarrow}{\cong} {\mathcal{S}}_{n+1+k} (M),$$ where $\equiv$ means a bijection and $\cong$ means an isomorphism of abelian groups. \[functoriality\] Let $h {\colon\!}P {\rightarrow}N$ be a homotopy equivalence of closed $n$-dimensional manifolds representing an element in ${\mathcal{S}}(N)$. Given a homotopy equivalence $r : N \to K$ to a finite $\Delta$-set $K$ we have described $s ([h]) \in {\mathbb{S}}_{n+1} (N)$ in Examples \[expl:frag-quad-con\], \[expl:frag-quad-con-over-htpy-equivalence\]. Let $s {\colon\!}M {\rightarrow}L$ be a homotopy equivalence from another closed $n$-dimensional manifold to a finite $\Delta$-set and let $f {\colon\!}N {\rightarrow}M$ be a map covering via the reference maps a $\Delta$-set map $\pi {\colon\!}K {\rightarrow}L$. Consider $\pi_\ast (s([h])) \in {\mathbb{S}}_{n+1} (M)$. From the description of the functoriality in (\[functoriality\]) we see that for each $\tau \in L$, $\pi_\ast (s([h]))(\tau )$ has the underlying chain complex the algebraic mapping cone of the map $$\big(\bigcup_{\pi(\sigma) = \tau} P(\sigma),\partial \big) {\longrightarrow}\big( \bigcup_{\pi(\sigma) = \tau} N(\sigma),\partial\big).$$ See [@Ranicki(1992) Example 5.8]. \[functoriality-projection\] We will need a special case of the above example when $\pi$ is the projection map $\pi_1 {\colon\!}K \otimes L {\rightarrow}K$. Here $K \otimes L$ is the geometric product of $\Delta$-sets, see [@Rourke-Sanderson(1971)] or [@Ranicki(1992) Chapter 11]. A $p$-simplex of $K \otimes L$ is a triple $$(\sigma,\tau,\lambda) \quad \textup{where} \quad \sigma \in K^{(m)}, \tau \in L^{(n)}, \lambda \in (\Delta^m \otimes \Delta^n)^{(p)}$$ with $\Delta^m \otimes \Delta^n$ the product of ordered simplicial complexes. There is a homeomorphism $\|K \otimes L\| = \|K\| \times \|L\|$ and there is a projection map $\pi_1 {\colon\!}K \otimes L {\rightarrow}L$ which is a $\Delta$-set map (the explicit formula is easy but a little complicated and we do not need it). We have $$\bigcup_{\tau,\lambda} D(\sigma,\tau,\lambda) = D(\sigma) \times L.$$ Let $M$ and $N$ be two closed topological manifolds with reference homotopy equivalences to $\Delta$-sets $r {\colon\!}M {\rightarrow}K$ and $r' {\colon\!}N {\rightarrow}L$ transverse to the dual cells. Then the product map $r \times r' {\colon\!}M \times N {\rightarrow}\|K\| \times \|L\|$ is transverse to the dual cells of the geometric product of the $\Delta$-sets $K \otimes L$. Let $h {\colon\!}P {\rightarrow}M \times N$ be a simple homotopy equivalence representing an element in ${\mathcal{S}}(M \times N)$ which is transverse to the dissection of $M \times N$ induced by $r \times r'$. We have $s([h]) \in {\mathbb{S}}_{m+n+1} (M \times N)$ and this element is represented by an algebraic Poincaré complex over ${\mathbb{Z}}_\ast (K \otimes L)$ whose value at $(\sigma,\tau,\lambda) \in K \otimes L$ has its underlying chain complex the mapping cone of the Umkehr map of the degree one normal map $$\label{eqn:h-before-partial-assembly} h(\sigma,\tau,\lambda) {\colon\!}P (\sigma,\tau,\lambda) = h^{-1} (M \times N) (\sigma,\tau,\lambda) {\rightarrow}(M \times N) (\sigma,\tau,\lambda).$$ Consider the projection map $p {\colon\!}M \times N {\rightarrow}M$. We have $p_\ast (s ([h])) \in {\mathbb{S}}_{m+n+1} (M)$. It follows from the above discussion that this element is represented by an algebraic Poincaré complex in ${\mathbb{A}}_\ast (K)$ whose value at $\sigma \in K$ has its underlying chain complex the mapping cone of the Umkehr map of the degree one normal map $$\label{eqn:partial-assembly} h(\sigma) {\colon\!}P (\sigma) = h^{-1} (M(\sigma) \times N) {\rightarrow}M (\sigma) \times N.$$ It may happen that such $h$ represents a non-trivial element in ${\mathcal{S}}(M \times N)$ and hence $s([h])$ is a non-zero element in ${\mathbb{S}}_{m+n+1} (M \times N)$ and at the same time for each $\sigma \in K$ the map \[eqn:partial-assembly\] is a simple homotopy equivalence. Then the underlying chain complex for each $\sigma$ is contractible and hence the projection $p_\ast(s([h])) = 0 \in {\mathbb{S}}_{m+n+1} (M)$ by the Proposition \[prop:criterion-zero\] below. Such a situation will indeed occur in the next section. \[prop:criterion-zero\] Let $(C,\psi)$ represent an element in ${\mathbb{S}}_{n+1} (M)$. Suppose in addition that $C(\sigma) \simeq \ast$ for each $\sigma \in K$. Then $$[(C,\psi)] = 0 \in {\mathbb{S}}_{n+1} (M).$$ The homotopy equivalences for each $\sigma$ assemble to a null-bordism of chain complexes in the algebraic bordism category $\Lambda'' \langle 1 \rangle$ with $\Lambda'' = ({\mathbb{Z}}_\ast (K),{\mathbb{C}},{\mathbb{D}})$ as in (\[alg-bordism-cat\]) which defines ${\mathbb{S}}_{n+1} (M)$. One of the important and useful features of the algebraic theory of surgery is a particularly easy description of periodicity. Indeed, in case one works with the $0$-connective version of the algebraic structure set, denoted by ${\bar {\mathbb{S}}}_{n+1} (M)$, one obtains the $4$-periodicity given by the so-called skew-double-suspension: $$\label{eqn:4-periodocity-S-bar} {\bar S}^2 : {\bar {\mathbb{S}}}_{n+1} (M) {\rightarrow}{\bar {\mathbb{S}}}_{n+5} (M).$$ If one works with the $1$-connective version and for a positive integer $k$ sets $S^{2k} := (S^2)^k$ one obtains in general an exact sequence $$\label{eqn:4-periodocity-S} 0 {\rightarrow}{\mathbb{S}}_{n+1} (M) {\xrightarrow}{S^{2k}} {\mathbb{S}}_{n+2k+1} (M) {\rightarrow}H_n (M,L_0 ({\mathbb{Z}})) {\rightarrow}\cdots$$ where in fact $H_n (M,L_0 ({\mathbb{Z}})) \cong H_n (M,{\mathbb{Z}})$: see [@Ranicki(1992) Remark 25.4]. This near-periodicity can also be defined using products in $L$-theory. Recall the symmetric signature $\sigma^\ast (M)$ of an $n$-dimensional Poincaré complex $M$, a symmetric algebraic Poincaré complex over ${\mathbb{Z}}$. The products in algebraic surgery [@Ranicki(1992) Appendix B] give for an $n$-dimensional quadratic algebraic ${\mathbb{C}}$-Poincaré complex $(C,\psi)$ representing an element in ${\mathbb{S}}_{n+1} (M)$ a new $(n+4)$-dimensional quadratic algebraic ${\mathbb{C}}$-Poincaré complex $(C,\psi) \otimes \sigma^\ast ({\mathbb{C}}P^2)$ representing an element in ${\mathbb{S}}_{n+5} (M)$. This produces a map which coincides with the double skew-suspension. In geometry this map corresponds to taking a product with the identity on ${\mathbb{C}}P^2$ and projecting algebraically. More generally one has the following identity of injective homomorphisms $$\label{eqn:S=sigma} (\otimes{\sigma^\ast ({\mathbb{F}}P^2)} = S^{k} ) {\colon\!}{\mathbb{S}}_{n+1} (M) {\rightarrow}{\mathbb{S}}_{n+2k+1} (M)$$ where $S^{k}$ is $2k$-skew-suspension map and $\otimes{ \sigma^\ast ({\mathbb{F}}P^2)}$ is the homomorphism defined by taking the product with the symmetric signature of ${\mathbb{C}}P^2$ or ${\mathbb{H}}P^2$ for $k = 2$ or $4$ respectively. Siebenmann periodicity {#sec:sieb-per} ====================== Recall that $M$ is a closed topological manifold of dimension $n \geq 5$ and Theorem \[thm:A\] which states that the Cappell-Weinberger map $CW^2 : {\mathcal{S}}(M) \to {\mathcal{S}}_{\partial}(M \times D^8)$ is an injective homomorphism with cokernel a subgroup of ${\mathbb{Z}}$. Theorem \[thm:A\] is a direct consequence of the Proposition \[prop:sieb\] below. The exactness part follows from the exactness of $S^{4}$ in (\[eqn:4-periodocity-S\]) and the identity $ \otimes \sigma^\ast({\mathbb{H}}P^2) = S^{4}$ of (\[eqn:S=sigma\]). \[prop:sieb\] For ${\mathbb{F}}= {\mathbb{H}}$, hence $k = 4$, the following diagram commutes. $$\xymatrix{ & {\mathcal{S}}(M) \ar[rr]^{CW^{k/2}} \ar[d]_{s} & & {\mathcal{S}}_\partial (M \times D^{2k}) \ar[d]^{s} & & \\ 0 \ar[r] & {\mathbb{S}}_{n+1} (M) \ar[rr]^{\otimes \sigma^\ast ({\mathbb{F}}P^2)} & & {\mathbb{S}}_{n+2k+1} (M) \ar[r] & H_n(M; L_0({\mathbb{Z}})) }$$ Recall that besides the map $CW^{k/2}$ from ${\mathcal{S}}(M)$ to ${\mathcal{S}}_\partial (M \times D^{2k})$ we have also discussed the extension by a homeomorphism map $E$ which brings us further to ${\mathcal{S}}( M \times {\mathbb{F}}P^2)$. This map will be helpful in the proof, in fact the situation can be described by the following diagram: $$\xymatrix{ \mathcal{S}^{}(M) \ar[rrr]^{CW^{k/2}} \ar@{.>}[dr]^{\times {\mathbb{F}}P^{2}} \ar[ddd]_{s} & & & \mathcal{S}^{}_\partial (M \times D^{2k}) \ar[dll]_{E} \ar[ddd]^{s} \\ & \mathcal{S}^{}(M \times {\mathbb{F}}P^2) \ar[dr]^{s_{M \times {\mathbb{F}}P^2}} \\ & & {\mathbb{S}}_{n+5}(M \times {\mathbb{F}}P^2) \ar[dr]^{p_\ast}\\ {\mathbb{S}}_{n+1}(M) \ar[rrr]_{\otimes \sigma^({\mathbb{F}}P^2)} & & & {\mathbb{S}}_{n+2k+1}(M) }$$ The discussion at the end of the last section shows that the lower triangle commutes. The triangle on the right commutes as well. We warn the reader that we do not claim that the upper triangle commutes, in fact it does not (that’s why the arrow is dotted). Nevertheless we will show that the outer square commutes. For this we first recall Lemma \[lem:phi\] which says that $E (CW^2 ([h])) \simeq [(h \times {\textup{id}}) \circ \varphi]$ where $\varphi {\colon\!}\bar N {\rightarrow}N \times {\mathbb{F}}P^2$ is a certain homotopy equivalence over the identity of $N$. The proof of the proposition boils down to the following \[lem:phi-alg-sur\] For ${\mathbb{F}}= {\mathbb{H}}$, hence $k = 4$, there is an equality $$p_\ast \big( s ([\varphi]) \big) = 0 \in {\mathbb{S}}_{n+2k+1} (N)$$ where $p_\ast {\colon\!}{\mathbb{S}}_{n+2k+1} (N \times {\mathbb{F}}P^2) {\rightarrow}{\mathbb{S}}_{n+2k+1} (N)$ is the homomorphism induced by the projection $p {\colon\!}N \times {\mathbb{F}}P^2 {\rightarrow}N$. We finish the proof of the Proposition \[prop:sieb\] and then prove Lemma \[lem:phi-alg-sur\]. We have the following equalities: $$\begin{aligned} \begin{split} s (CW^{k/2} (h)) & = {\textup{pr}}_1 s (E \circ CW^{k/2} ([h])) \\ & = {\textup{pr}}_1 s ([(h \times {\textup{id}}) \circ \varphi]) \\ & = {\textup{pr}}_1 ((h \times {\textup{id}})_\ast s ([\varphi]) + s ([h \times {\textup{id}}])) \\ & = h_\ast {\textup{pr}}_1 s ([\varphi]) + s ([h]) \otimes \sigma^\ast ({\mathbb{F}}P^2) \\ & = s ([h]) \otimes \sigma^\ast ({\mathbb{F}}P^2). \end{split}\end{aligned}$$ The first equality follows from the definitions and the functoriality of ${\mathbb{S}}_?(-)$, the second from Lemma \[lem:phi\], the third from the composition formula of [@Ranicki(2009)], the fourth again from the functoriality of ${\mathbb{S}}_?(-)$ and the fifth from Lemma \[lem:phi-alg-sur\]. Recall the steps that lead from a homotopy equivalence $h : N \to M$ to the homotopy equivalence $\varphi {\colon\!}\bar N {\rightarrow}N \times {\mathbb{F}}P^2$ defined by (\[defn:phi\]): 1. Start with a map $\chi {\colon\!}N {\rightarrow}{\textup{G}}/{\textup{TOP}}$ (which was chosen so that $(h \times {\textup{id}})^\ast\chi = -[h \times {\textup{id}}] \in [M, {\textup{G}}/{\textup{TOP}}]$). 2. Construct the homotopy equivalence of pairs $\omega {\colon\!}(\bar W,{\partial}) {\rightarrow}(N \times D^{k+1},{\partial})$ over the identity from the MCN $p {\colon\!}(\bar W,{\partial}) {\rightarrow}N$, (in Proposition \[prop:constructing-neighborhoods\]). 3. Consider the restriction, ${\partial}\omega {\colon\!}{\partial}\bar W {\rightarrow}N \times S^{k}$ which is a homotopy equivalence over the identity. 4. Construct the homotopy equivalence ${\partial}\omega' {\colon\!}{\partial}\bar W' {\rightarrow}N \times S^{2k-1}$ over the identity as the pullback of $\gamma_N {\colon\!}N \times S^{2k-1} {\rightarrow}N \times S^k$ along ${\partial}\omega$ (recall (\[defn:del-W-prime\])). This yields projection maps $q_{\bar W} {\colon\!}{\partial}\bar W' {\rightarrow}{\partial}\bar W$ and $p' {\colon\!}{\partial}\bar W' {\rightarrow}N$. 5. Define $\bar N = {\textup{cyl}}(p') \cup {\textup{cyl}}(q_{\bar W})$, (Formula (\[defn:N-bar\])). 6. Define $\varphi = {\textup{cyl}}({\partial}\omega',{\textup{id}}) \cup {\textup{cyl}}({\partial}\omega',{\partial}\omega)$, (Formula (\[defn:phi\])). Next recall $r {\colon\!}M {\rightarrow}K$ a homotopy equivalence from $M$ to a $\Delta$-set $K$ which is transverse to the dual cells of $K$ so that we have a dissection $$M = \cup_{\sigma \in K} M(\sigma)$$ with $M(\sigma) = r^{-1} (D(\sigma,K))$ a submanifold with boundary of dimension $(n - \|\sigma\|)$. Further assume that $h {\colon\!}N {\rightarrow}M$ is transverse to $M (\sigma)$ for each $\sigma$ so that $N (\sigma) = h^{-1} (M(\sigma))$ is a submanifold with boundary of dimension $(n-\|\sigma\|)$ and $h (\sigma) {\colon\!}N(\sigma) {\rightarrow}M(\sigma)$ is a degree one normal map. We obtain a dissection $$\label{N-dissection} N = \cup_{\sigma \in K} N(\sigma).$$ Geometry. We show that by a small homotopy it is possible to change $\chi {\colon\!}N {\rightarrow}{\textup{G}}/{\textup{TOP}}$ so that $\bar N$ possess a dissection indexed by simplices $\sigma \in K$, the map $p$ respects the dissections of $\bar N$ and $N$: $$\label{dissection-of-p-hat} \bar p = \cup \bar p(\sigma) {\colon\!}\bar N = \cup \bar N (\sigma) {\rightarrow}N = \cup N(\sigma)$$ and the homotopy equivalence $\varphi {\colon\!}\bar N {\rightarrow}N \times {\mathbb{F}}P^2$ also respects these dissections. Furthermore for each $\sigma \in K$ $$\label{dissection-of-phi} \varphi (\sigma) {\colon\!}\bar N (\sigma) {\rightarrow}N (\sigma) \times {\mathbb{F}}P^2$$ is a homotopy equivalence. To this end modify the map $\chi$ by a small homotopy so that when restricted to the collar of each manifold with boundary $(N(\sigma),{\partial}N(\sigma))$ it is the product map with the identity in the collar direction. Hence we have $$\chi = \cup \chi (\sigma) {\colon\!}N = \cup N(\sigma) {\rightarrow}{\textup{G}}/{\textup{TOP}}.$$ Now simply perform the steps (1) to (6) of the construction of $\varphi$ from $\chi$ to construct $\varphi (\sigma)$ from $\chi (\sigma)$. For this to be possible one should proceed inductively, starting with simplices $\sigma$ of the top-dimension, since then $N(\sigma)$ has the smallest dimension $0$. Over such $N(\sigma)$ the steps (1) to (6) are trivial. To make the inductive step, follow steps (1) to (6), but in the relative setting. Here, Proposition \[prop:constructing-neighborhoods-relative\] should be used in step (2) instead of Proposition \[prop:constructing-neighborhoods\]. Note that as $k = 4$ the dimension restrictions of the proposition are fulfilled, since the dimension of $N(\sigma)$ is $\geq 1$. The steps (3) to (6) have straightforward generalizations to the relative case. The manifold $\bar N$ is the union of all the manifolds $\widehat N(\sigma)$ just constructed and the projection map $\bar p {\colon\!}\bar N {\rightarrow}N$ is the union of the corresponding projections maps $\bar p(\sigma)$. Similarly the homotopy equivalence $\varphi$ is the union of all the homotopy equivalences $\varphi(\sigma)$. Algebraic surgery. The homotopy equivalence $\varphi$ represents an element in the structure set ${\mathcal{S}}(N \times {\mathbb{F}}P^2)$. The map $s {\colon\!}{\mathcal{S}}(N \times {\mathbb{F}}P^2) {\rightarrow}{\mathbb{S}}_{n+2k+1} (N \times {\mathbb{F}}P^2)$ was described in Section \[sec:alg-sur\]. To use it we need to choose a $\Delta$-set homotopy equivalent to $N \times {\mathbb{F}}P^2$. Since ${\mathbb{F}}P^2$ is a triangulable manifold we can choose a triangulation and denote the underlying $\Delta$-set by $L$, the reference map will be denoted $r' {\colon\!}{\mathbb{F}}P^2 {\rightarrow}L$. Then we can pick as our choice the geometric product $K \otimes L$, whose geometric realization we identify with the product $\|K\| \times \|L\|$, and the reference map $\bar r {\colon\!}= (h \circ r) \times r' {\colon\!}N \times {\mathbb{F}}P^2 {\rightarrow}\|K\| \times \|L\|$ is automatically transverse to the dual cells of $K \otimes L$ which we consider as the underlying space of the geometric product of $\Delta$-sets described in Section \[sec:alg-sur\]. Note that each dual cell of $K \otimes L$ is a subspace of $D (\sigma) \times L$ for suitable $\sigma \in K$. Hence also $$(N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda) \subset N (\sigma) \times {\mathbb{F}}P^2$$ for each $\tau, \lambda$. In fact $$\bigcup_{\tau,\lambda} (N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda) \subset N (\sigma) \times {\mathbb{F}}P^2.$$ To determine a representative of $s([\varphi])$ in ${\mathbb{S}}_{n+2k+1} (N \times {\mathbb{F}}P^2)$, where we work over the category ${\mathbb{Z}}_\ast (K \otimes L)$, the map $\varphi$ needs to be made transverse to the submanifolds $(N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda)$ for each $(\sigma,\tau,\lambda) \in K \otimes L$. This can be done by a small homotopy which does not spoil the property that $\varphi$ respects the dissections of $\bar N$ and $N \times {\mathbb{F}}P^2$ over $K$ and that each $\varphi (\sigma)$ is a homotopy equivalence. To achieve this we can again proceed inductively starting from the simplices $\sigma \in K$ of the top dimension. We change each $\varphi (\sigma)$ by a small homotopy to make it transverse to $(N \times {\mathbb{F}}P^2)(\sigma,\tau,\lambda)$ for all choices of $\tau$ and $\lambda$, which if course does not spoil the fact that it is a homotopy equivalence. Hence the new $\varphi$ is the union of the new $\varphi (\sigma)$ and hence is transverse to $(N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda)$ for each $(\sigma,\tau,\lambda) \in K \otimes L$. Now we find ourselves in the situation described in Example \[functoriality-projection\]. We have the homotopy equivalence $\varphi {\colon\!}\bar N {\rightarrow}N \times {\mathbb{F}}P^2$ whose image $s([\varphi]) \in {\mathbb{S}}_{n+2k+1} (N \times {\mathbb{F}}P^2)$ is represented by a quadratic chain complex over the category ${\mathbb{Z}}_\ast (K \otimes L)$ whose value at each $(\sigma,\tau,\lambda) \in K \otimes L$ has its underlying chain complex the mapping cone of the Umkehr map of the degree one normal map $$\label{eqn:phi-before-partial-assembly} \varphi (\sigma,\tau,\lambda) {\colon\!}\bar N (\sigma,\tau,\lambda) = \varphi^{-1} (N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda) {\rightarrow}(N \times {\mathbb{F}}P^2) (\sigma,\tau,\lambda).$$ This may very well be a representative of a non-zero element in ${\mathbb{S}}(N \times {\mathbb{F}}P^2)$. But we are really interested in the projection $p_\ast \big( s ([\varphi]) \big) \in {\mathbb{S}}_{n+2k+1} (N)$. By Example \[functoriality-projection\] this is represented by a quadratic chain complex over the category ${\mathbb{Z}}_\ast (K)$ whose value at each $\sigma \in K$ has its underlying chain complex the mapping cone of the Umkehr map of the degree one normal map $$\label{eqn:phi-partial-assembly} \varphi (\sigma) {\colon\!}\bar N(\sigma) {\rightarrow}N (\sigma) \times {\mathbb{F}}P^2.$$ But $\varphi (\sigma)$ is a homotopy equivalence for each $\sigma$ and so the resulting chain complex over each $\sigma$ is contractible. Thus by Proposition \[prop:criterion-zero\] $p_(s([\varphi])) = 0$ as required. The above proof shows why we chose $k = 4$. If $k = 2$, then the dimension restrictions of Proposition \[prop:constructing-neighborhoods-relative\] are not satisfied.[^6] The bordism groups $\Omega^{{\textup{STOP}}}_{2d-1} (G/{\textup{TOP}}\times BG$) {#sec:bordism-groups} ================================================================================ Let $X$ be a space and let $\Omega_n^{{\textup{STOP}}}(X)$ denote the $n$th oriented topological bordism group of $X$. Recall that $G$ is a finite group and $BG$ is its classifying space. The purpose of this section is to prove the following \[lem:bord\] For all $d > 1$, $\Omega_{2d-1}^{{\textup{STOP}}}({\textup{G}}/{\textup{TOP}}\times BG) {\otimes}{\mathbb{Q}}= 0$. The functor $X \to \Omega_^{{\textup{STOP}}}(X) {\otimes}{\mathbb{Q}}$ is a generalised homology theory with coefficients $\Omega_^{{\textup{STOP}}} {\otimes}{\mathbb{Q}}$. By Theorem \[thm:bord\] for the trivial group $\Omega_{2d-1}^{{\textup{STOP}}}({\rm pt}) {\otimes}{\mathbb{Q}}= 0$. Applying the Atiyah-Hirzebruch spectral sequence to compute $\Omega_^{{\textup{STOP}}}(X)$ $$\bigoplus_{p+q = }H_p(X; \Omega_q^{{\textup{STOP}}} \otimes {\mathbb{Q}}) \Longrightarrow \Omega_{}^{{\textup{STOP}}}(X) \otimes {\mathbb{Q}},$$ we see that if $H_{2d-1}(X; {\mathbb{Q}}) \cong 0$ for all $d$ then $\Omega_{2d-1}^{{\textup{STOP}}}(X) {\otimes}{\mathbb{Q}}= 0$ for all $d$. Now if $Y$ a connected space such that $\pi_{2d-1}(Y) = 0$ for all $d$ then $H_{2d-1}(Y; {\mathbb{Q}}) = 0$ for all $d$. This follows by induction over the Postnikov tower of $Y$: let $n$ be even, let $P_n(Y)$ denote the $n$-th Postnikov of $Y$ and let $$K(\pi_{n+2}, n+2) \to P_{n+2}(Y) \to P_n(Y)$$ be the fibration for the $(n+2)$nd Postnikov stage of $Y$. By induction we have that $H_{2d+1}(P_n(Y); {\mathbb{Q}}) \cong 0$ for all $d$ and the same statement holds for the Eilenberg-MacLane space $K(\pi_{n+2}, n+2)$. From the Leray-Serre spectral sequence: $$\bigoplus_{p+q = } H_p(P_n(Y); H_q(K(\pi_{n+2}, n+2); {\mathbb{Q}})) \Longrightarrow H_{}(P_{n+2}(Y); {\mathbb{Q}})$$ we see that $H_{2d-1}(P_{2n+2}(Y); {\mathbb{Q}}) = 0$ for all $d$. By [@Sullivan(1971); @Kirby-Siebenmann(1977)] $\pi_{2d-1}({\textup{G}}/{\textup{TOP}}) = 0$ for all $d$ and so $H_{2d-1}({\textup{G}}/{\textup{TOP}}; {\mathbb{Q}}) = 0$ for all $k$. Now as $G$ is a finite group $H_(BG; {\mathbb{Q}}) = 0$ for all $> 0$ and so by the Kunneth Theorem we see that $H_{2d-1}({\textup{G}}/{\textup{TOP}}\times BG; {\mathbb{Q}}) = 0$ for all $d$. Thus we conclude that $\Omega_{2d-1}^{{\textup{STOP}}}({\textup{G}}/{\textup{TOP}}\times BG) {\otimes}{\mathbb{Q}}= 0$ for all $d$. Completion of the proof of Theorem \[thm:main\] {#sec:completion} =============================================== In this section we prove Theorem \[thm:B\] which completes the proof of Theorem \[thm:main\]. Recall the definition of the maps ${\widetilde{\rho}}$, ${\widetilde{\rho}}_\partial$ and the $CW^{k/2}$-map and that $k = 2$ or $4$. In addition recall the map defined by extension by a homeomorphism $E {\colon\!}{\mathcal{S}}_\partial (M \times D^{2k}) {\rightarrow}{\mathcal{S}}(M \times {\mathbb{F}}P^2)$. Let $h {\colon\!}N {\rightarrow}M$ represent an element in ${\mathcal{S}}(M)$, let $h' {\colon\!}N' {\rightarrow}M \times D^{2k}$ represent $CW^{k/2} ([h])$. Recall from Lemma \[lem:phi\] that $E([h'])$ can be represented by two homotopy equivalences, namely either by $\widehat h {\colon\!}\widehat N {\rightarrow}M \times {\mathbb{F}}P^2$ (see (\[defn:h-hat\])) or by $\bar h {\colon\!}\bar N {\rightarrow}M \times {\mathbb{F}}P^2$ (see just below (\[defn:phi\])). \[lem-1\] There are identities 1. $\rho (\bar N) = {\widetilde{\rho}}_\partial ([h']) + \rho (M)$, 2. $\rho (\bar N) = \rho (N)$. Combining Lemma \[lem-1\] (1) and (2) we have: $${\widetilde{\rho}}_\partial ([h']) = \rho (N) - \rho (M) = {\widetilde{\rho}}([h]).$$ \(1) The argument is similar to the proof of Proposition \[rho-add-for-bdry\]. From the rel. boundary structure $h' : N' \to M \times D^{k}$ we form the closed manifold $M(h') := N' \cup_{h'}(-M \times D^{2k})$ and by Definition \[defn:reduced-rho-del\], ${\widetilde{\rho}}_{\partial}([h']) = \rho(M(h'))$. Recall the operation $\#_M$ in \[defn:ctd-sum-along-M\] and observe that $$M \times {\mathbb{F}}P^2 = (M \times {\mathbb{F}}P^{2\bullet}) \cup (M \times D^{2k}).$$ If follows that we can form the closed manifold $$\label{last-ctd-sum} M(h') \#_M (M \times {\mathbb{F}}P^2)$$ just as in Definition \[defn:ctd-sum-along-M\]. By the same reasoning as in the proof of Proposition \[rho-add-for-bdry\] we obtain that the $\rho$-invariant of the manifold in (\[last-ctd-sum\]) is the sum of $\rho$-invariants $\rho (M(h')) + \rho (M \times {\mathbb{F}}P^2)$. But from the construction of $\widehat N$ in (\[defn:N-hat\]) we see that there is a homeomorphism $$\widehat N \cong M(h') \#_M (M \times {\mathbb{F}}P^2 ).$$ The statement now follows since $\rho(M \times {\mathbb{F}}P^2) = \rho(M)$ and $\rho (\widehat N) = \rho (\bar N)$, which we have from the $h$-cobordism invariance of the $\rho$-invariant. \(2) We are given $[h : N \to M] \in \mathcal{S}(M)$ and $[ \bar h {\colon\!}\bar N \to M \times {\mathbb{F}}P^2] \in \mathcal{S}(M \times {\mathbb{F}}P^2)$ which represents $E \circ CW^{k/2}([h])$ and we wish to prove that $\rho(\bar N) = \rho(N)$ where the reference map for $\bar N$ is $\lambda(N) \circ \bar p$ and $\bar p : \bar N \to N$ is the map constructed in Section \[subsec:ext-by-homeo\]. Recall from Definition \[defn:rho-1\] the definition of the $\rho$-invariant of a $(2d-1)$-dimensional manifold $N$ equipped with a map $\lambda (N) {\colon\!}N {\rightarrow}BG$ inducing $\lambda (N)_\ast {\colon\!}\pi_1 (N) {\rightarrow}G$, with $G$ a finite group. Because $\Omega^{{\textup{STOP}}}_{2d-1} (BG) \otimes {\mathbb{Q}}= 0$ there is a coboundary for $\sqcup_{i=1}^r N$ over $\lambda (N)$ for some $r \geq 1$. That is, there is a manifold $P$ with boundary ${\partial}P = \sqcup_{i=1}^r N$, and with a map $\lambda (P) {\colon\!}P {\rightarrow}BG$ extending $\sqcup^r \lambda (N)$. The formula is $$\label{rho-of-N} \rho(N) := (1/r){\textup{G-sign}}(P).$$ To show the desired statement it is enough to find a coboundary, say $\bar P$, for $\sqcup_{i=1}^r \bar N$ over $\lambda (\bar N) = (\lambda (N) \circ \bar p) {\colon\!}\bar N {\rightarrow}N {\rightarrow}BG$, such that $\bar P \simeq P \times {\mathbb{F}}P^2$. Then by the multiplicativity of $G$-signature we would obtain $$\label{rho-is-mult} \rho(\bar N) = (1/r){\textup{G-sign}}(\bar P) = (1/r){\textup{G-sign}}(P) \cdot {\textup{sign}}({\mathbb{F}}P^2) = \rho(N).$$ Recall from Section \[sec:cw-map\] that the closed manifold $\bar N$ along with a homotopy equivalence $\varphi {\colon\!}\bar N {\rightarrow}N \times {\mathbb{F}}P^2$ was constructed from a map $\chi (N) {\colon\!}N {\rightarrow}{\textup{G}}/{\textup{TOP}}$. Now, by Lemma \[lem:bord\] we have $\Omega^{{\textup{STOP}}}_{2d-1} (BG \times {\textup{G}}/{\textup{TOP}}) \otimes {\mathbb{Q}}= 0$. This implies that there exists a manifold $P$ with boundary ${\partial}P = \sqcup_{i=1}^r N$ and a map $$\kappa (P) {\colon\!}P {\rightarrow}BG \times {\textup{G}}/{\textup{TOP}}$$ such that $${\textup{pr}}_{{\textup{G}}/{\textup{TOP}}} \circ (\kappa (P)\|_{{\partial}P}) = \sqcup_{i=1}^r \chi (N) {\colon\!}(\sqcup_{i=1}^r N) {\rightarrow}{\textup{G}}/{\textup{TOP}}$$ and $${\textup{pr}}_{BG} \circ (\kappa (P)\|_{{\partial}P}) = \sqcup_{i=1}^r \lambda (N) {\colon\!}(\sqcup_{i=1}^r N ) {\rightarrow}BG.$$ Here ${\textup{pr}}_{{\textup{G}}/{\textup{TOP}}}$ and ${\textup{pr}}_{BG}$ are the obvious projections. We have used the same letter $P$ as above because such a $P$ can be used as a coboundary of $\sqcup_{i=1}^r N$ in \[rho-of-N\]. The improvement is that now $P$ comes equipped with the map $\chi (P) := {\textup{pr}}_{{\textup{G}}/{\textup{TOP}}} \circ \kappa (P) {\colon\!}P {\rightarrow}{\textup{G}}/{\textup{TOP}}$. Recall the recipe for constructing $\bar N$ from $\chi (N) {\colon\!}N {\rightarrow}{\textup{G}}/{\textup{TOP}}$ repeated in the proof of Proposition \[lem:phi-alg-sur\] as steps (1) to (6). In that proof a generalization of steps (1) to (6) was used when one starts with a map from a manifold with boundary to ${\textup{G}}/{\textup{TOP}}$. Using this generalised procedure we construct a manifold with boundary $\bar P$ with a homotopy equivalence $\varphi (P) {\colon\!}\bar P {\rightarrow}P \times {\mathbb{F}}P^2$. The boundary is ${\partial}\bar P = \sqcup_{i=1}^r \bar N$ since the map $\chi (P) {\colon\!}P {\rightarrow}{\textup{G}}/{\textup{TOP}}$, restricts to $\sqcup_{i=1}^r \chi (N) {\colon\!}\sqcup_{i=1}^r N {\rightarrow}{\textup{G}}/{\textup{TOP}}$ on ${\partial}P = \sqcup_{i=1}^r N$. Furthermore if $\bar p (P) : \bar P \to P$ denotes the analogue of $\bar p : \bar N \to N$ obtain from the generalised procedure, then we have the map $$\lambda (\bar P) = (\lambda (P) \circ \bar p (P)) {\colon\!}\bar P {\rightarrow}P {\rightarrow}BG,$$ which restricts to $\sqcup_{i=1}^r \lambda (\bar N)$ on the boundary. If follows that $\bar P$ is the desired coboundary of $\sqcup_{i=1}^r \bar N$ over $\lambda (\bar N)$ which may be used in \[rho-is-mult\]. [^1]: Our results work equally well for the simple structure set, see Remark \[rem:decorations\]. [^2]: In fact Siebenmann mistakenly claimed that $P^j$ is a bijection. In general ${\rm Im}(P^j)$ is a subgroup with ${\mathcal{S}}(M \times D^{4j})/{\rm Im}(P^j)$ isomorphic to a subgroup of $H_0(M; {\mathbb{Z}})$. Therefore to be precise we speak of near-periodicity. A correct statement of near periodicity appeared in [@Nicas(1982)]. [^3]: In fact we do not even have a preferred map from $W$ to $N$. We could use the $h$-cobordism $U$ to obtain some map, but still the map $\psi$ would not be a map over $h$ in general. [^4]: The notation should be understood as $({\partial}W)'$, not ${\partial}(W')$, in fact there is no manifold $W'$. [^5]: Note that $f$ is by definition a homeomorphism on ${\partial}\{0, 1 \}$, hence transverse to anything, hence does not have to be changed on ${\partial}\{0, 1 \}$. [^6]: It is possible that the dimension restrictions in the relevant proposition can be relaxed. This would require careful analysis of all the tools used in the proofs. This might be an interesting problem but lies beyond the scope of this paper.	{ "pile_set_name": "ArXiv" }

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.